diff --git a/.devops/cublas.Dockerfile b/.devops/cublas.Dockerfile
index 60a8edfbb26..e7e51273a0f 100644
--- a/.devops/cublas.Dockerfile
+++ b/.devops/cublas.Dockerfile
@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential git cmake libsdl2-dev
+    apt-get install -y build-essential git cmake libsdl2-dev wget
 
 WORKDIR /app
 
@@ -23,6 +23,6 @@ ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 # Enable cuBLAS
 ENV GGML_CUDA=1
 
-RUN make
+RUN make base.en
 
 ENTRYPOINT ["/app/main"]
diff --git a/.devops/main-cuda.Dockerfile b/.devops/main-cuda.Dockerfile
index 0b7a988697a..a806fa09bde 100644
--- a/.devops/main-cuda.Dockerfile
+++ b/.devops/main-cuda.Dockerfile
@@ -17,7 +17,7 @@ ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
 ENV GGML_CUDA=1
 
 RUN apt-get update && \
-    apt-get install -y build-essential libsdl2-dev \
+    apt-get install -y build-essential libsdl2-dev wget cmake \
     && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
 # Ref: https://stackoverflow.com/a/53464012
@@ -25,7 +25,7 @@ ENV CUDA_MAIN_VERSION=12.3
 ENV LD_LIBRARY_PATH /usr/local/cuda-${CUDA_MAIN_VERSION}/compat:$LD_LIBRARY_PATH
 
 COPY .. .
-RUN make
+RUN make base.en
 
 FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
 ENV CUDA_MAIN_VERSION=12.3
@@ -33,7 +33,7 @@ ENV LD_LIBRARY_PATH /usr/local/cuda-${CUDA_MAIN_VERSION}/compat:$LD_LIBRARY_PATH
 WORKDIR /app
 
 RUN apt-get update && \
-  apt-get install -y curl ffmpeg \
+  apt-get install -y curl ffmpeg wget cmake \
   && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
 COPY --from=build /app /app
diff --git a/.devops/main.Dockerfile b/.devops/main.Dockerfile
index b8170e29561..066d4c83796 100644
--- a/.devops/main.Dockerfile
+++ b/.devops/main.Dockerfile
@@ -2,17 +2,17 @@ FROM ubuntu:22.04 AS build
 WORKDIR /app
 
 RUN apt-get update && \
-  apt-get install -y build-essential \
+  apt-get install -y build-essential wget cmake \
   && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
 COPY .. .
-RUN make
+RUN make base.en
 
 FROM ubuntu:22.04 AS runtime
 WORKDIR /app
 
 RUN apt-get update && \
-  apt-get install -y curl ffmpeg libsdl2-dev \
+  apt-get install -y curl ffmpeg libsdl2-dev wget cmake \
   && rm -rf /var/lib/apt/lists/* /var/cache/apt/archives/*
 
 COPY --from=build /app /app
diff --git a/.github/workflows/bindings-ruby.yml b/.github/workflows/bindings-ruby.yml
index d1d3c341bd2..0b427429da9 100644
--- a/.github/workflows/bindings-ruby.yml
+++ b/.github/workflows/bindings-ruby.yml
@@ -3,61 +3,41 @@ on:
   push:
     paths:
       - bindings/ruby/**
-      - src/whisper.cpp
-      - include/whisper.h
-      - ggml/src/ggml.c
-      - ggml/src/ggml-impl.h
-      - ggml/src/ggml-aarch64.h
-      - ggml/src/ggml-aarch64.c
-      - ggml/src/ggml-alloc.c
-      - ggml/src/ggml-backend-impl.h
-      - ggml/src/ggml-backend.cpp
-      - ggml/src/ggml-common.h
-      - ggml/src/ggml-quants.h
-      - ggml/src/ggml-quants.c
-      - ggml/src/ggml-cpu-impl.h
-      - ggml/src/ggml-metal.m
-      - ggml/src/ggml-metal.metal
-      - ggml/src/ggml-blas.cpp
-      - ggml/include/ggml.h
-      - ggml/include/ggml-alloc.h
-      - ggml/include/ggml-backend.h
-      - ggml/include/ggml-cuda.h
-      - ggml/include/ggml-kompute.h
-      - ggml/include/ggml-metal.h
-      - ggml/include/ggml-sycl.h
-      - ggml/include/ggml-vulkan.h
-      - ggml/include/ggml-blas.h
+      - src/**/*.c
+      - src/**/*.cpp
+      - src/**/*.h
+      - src/**/*.m
+      - src/**/*.metal
+      - include/**/*.c
+      - include/**/*.cpp
+      - include/**/*.h
+      - include/**/*.m
+      - include/**/*.metal
+      - ggml/**/*.c
+      - ggml/**/*.cpp
+      - ggml/**/*.h
+      - ggml/**/*.m
+      - ggml/**/*.metal
       - scripts/get-flags.mk
       - examples/dr_wav.h
   pull_request:
     paths:
       - bindings/ruby/**
-      - src/whisper.cpp
-      - include/whisper.h
-      - ggml/src/ggml.c
-      - ggml/src/ggml-impl.h
-      - ggml/src/ggml-aarch64.h
-      - ggml/src/ggml-aarch64.c
-      - ggml/src/ggml-alloc.c
-      - ggml/src/ggml-backend-impl.h
-      - ggml/src/ggml-backend.cpp
-      - ggml/src/ggml-common.h
-      - ggml/src/ggml-quants.h
-      - ggml/src/ggml-quants.c
-      - ggml/src/ggml-cpu-impl.h
-      - ggml/src/ggml-metal.m
-      - ggml/src/ggml-metal.metal
-      - ggml/src/ggml-blas.cpp
-      - ggml/include/ggml.h
-      - ggml/include/ggml-alloc.h
-      - ggml/include/ggml-backend.h
-      - ggml/include/ggml-cuda.h
-      - ggml/include/ggml-kompute.h
-      - ggml/include/ggml-metal.h
-      - ggml/include/ggml-sycl.h
-      - ggml/include/ggml-vulkan.h
-      - ggml/include/ggml-blas.h
+      - src/**/*.c
+      - src/**/*.cpp
+      - src/**/*.h
+      - src/**/*.m
+      - src/**/*.metal
+      - include/**/*.c
+      - include/**/*.cpp
+      - include/**/*.h
+      - include/**/*.m
+      - include/**/*.metal
+      - ggml/**/*.c
+      - ggml/**/*.cpp
+      - ggml/**/*.h
+      - ggml/**/*.m
+      - ggml/**/*.metal
       - scripts/get-flags.mk
       - examples/dr_wav.h
 
@@ -70,6 +50,6 @@ jobs:
     steps:
       - uses: ruby/setup-ruby@v1
         with:
-          ruby-version: '3.0'
+          ruby-version: '3.1'
       - uses: actions/checkout@v4
       - run: rake test
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 8f741b97091..f48fc7cabee 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -28,9 +28,9 @@ jobs:
             -w /workspace ${{ env.ubuntu_image }} /bin/sh -c '
             set -e
             apt update
-            apt install -y build-essential libsdl2-dev
-            make
-            make stream'
+            apt install -y build-essential libsdl2-dev cmake
+            cmake -B build
+            cmake --build build --config Release -j $(nproc)'
 
   macOS-latest:
     runs-on: macOS-latest
@@ -42,30 +42,30 @@ jobs:
       - name: Dependencies
         run: |
           brew update
-          brew install sdl2
+          brew install sdl2 cmake
 
       - name: Build
         run: |
-          make
-          make stream
-
-  freeBSD-latest:
-    runs-on: macos-12
+          cmake -B build
+          cmake --build build --config Release
 
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-
-      - name: Build
-        uses: cross-platform-actions/action@v0.24.0
-        with:
-          operating_system: freebsd
-          version: '13.3'
-          run: |
-            sudo pkg update
-            sudo pkg install -y gmake sdl2
-            gmake
-            gmake stream
+#  freeBSD-latest:
+#    runs-on: macos-12
+#
+#    steps:
+#      - name: Clone
+#        uses: actions/checkout@v4
+#
+#      - name: Build
+#        uses: cross-platform-actions/action@v0.24.0
+#        with:
+#          operating_system: freebsd
+#          version: '13.3'
+#          run: |
+#            sudo pkg update
+#            sudo pkg install -y gmake sdl2 cmake
+#            cmake -B build
+#            cmake --build build --config Release
 
   ubuntu-latest-gcc:
     runs-on: ubuntu-latest
@@ -280,25 +280,10 @@ jobs:
             mingw-w64-${{matrix.env}}-SDL2
             mingw-w64-${{matrix.env}}-openblas
 
-      - name: Build using make
-        shell: msys2 {0}
-        run: |
-            make -j $(nproc)
-
-      - name: Clean after building using make
-        shell: msys2 {0}
-        run: |
-            make clean
-
-      - name: Build using make w/ OpenBLAS
-        shell: msys2 {0}
-        run: |
-            make GGML_OPENBLAS=1 -j $(nproc)
-
       - name: Build using CMake
         shell: msys2 {0}
         run: |
-            cmake -B build
+            cmake -B build -DWHISPER_SDL2=ON
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
       - name: Clean after building using CMake
@@ -445,71 +430,72 @@ jobs:
           name: whisper-blas-bin-${{ matrix.arch }}
           path: build/bin/${{ matrix.build }}
 
-  windows-cublas:
-    runs-on: windows-2019
-
-    strategy:
-      matrix:
-        build: [Release]
-        arch: [x64]
-        cublas: [ON]
-        sdl2: [ON]
-        cuda-toolkit: [12.2.0, 11.8.0]
-        include:
-          - arch: x64
-            s2arc: x64
-          - sdl2: ON
-            s2ver: 2.28.5
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v4
-
-      - name: Add msbuild to PATH
-        uses: microsoft/setup-msbuild@v2
-
-      - name: Install CUDA Toolkit
-        id: cuda-toolkit
-        uses: Jimver/cuda-toolkit@v0.2.15
-        with:
-          cuda: '${{ matrix.cuda-toolkit }}'
-
-      - name: Fetch SDL2 and set SDL2_DIR
-        if: matrix.sdl2 == 'ON'
-        run: |
-          C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ matrix.s2ver }}/SDL2-devel-${{ matrix.s2ver }}-VC.zip
-          7z x sdl2.zip
-          echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
-
-      - name: Configure
-        run: >
-          cmake -S . -B ./build -A ${{ matrix.arch }}
-          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
-          -DGGML_CUDA=${{ matrix.cublas }}
-          -DWHISPER_SDL2=${{ matrix.sdl2 }}
-
-      - name: Build ${{ matrix.cuda-toolkit }}
-        run: |
-          cd ./build
-          cmake --build . --config ${{ matrix.build }}
-
-      - name: Copy CUDA DLLs
-        run: >
-          Copy-Item -PassThru
-          -Path "${{ steps.cuda-toolkit.outputs.CUDA_PATH }}/bin/*.dll"
-          -Include cudart64_*,cublas64_*,cublasLt64_*
-          -Destination build/bin/${{ matrix.build }}
-
-      - name: Copy SDL2.dll
-        if: matrix.sdl2 == 'ON'
-        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
-
-      - name: Upload binaries
-        if: matrix.sdl2 == 'ON'
-        uses: actions/upload-artifact@v4
-        with:
-          name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}
-          path: build/bin/${{ matrix.build }}
+# TODO: fix and re-enable
+#  windows-cublas:
+#    runs-on: windows-2019
+#
+#    strategy:
+#      matrix:
+#        build: [Release]
+#        arch: [x64]
+#        cublas: [ON]
+#        sdl2: [ON]
+#        cuda-toolkit: [12.2.0, 11.8.0]
+#        include:
+#          - arch: x64
+#            s2arc: x64
+#          - sdl2: ON
+#            s2ver: 2.28.5
+#
+#    steps:
+#      - name: Clone
+#        uses: actions/checkout@v4
+#
+#      - name: Add msbuild to PATH
+#        uses: microsoft/setup-msbuild@v2
+#
+#      - name: Install CUDA Toolkit
+#        id: cuda-toolkit
+#        uses: Jimver/cuda-toolkit@v0.2.15
+#        with:
+#          cuda: '${{ matrix.cuda-toolkit }}'
+#
+#      - name: Fetch SDL2 and set SDL2_DIR
+#        if: matrix.sdl2 == 'ON'
+#        run: |
+#          C:/msys64/usr/bin/wget.exe -qO sdl2.zip https://github.com/libsdl-org/SDL/releases/download/release-${{ matrix.s2ver }}/SDL2-devel-${{ matrix.s2ver }}-VC.zip
+#          7z x sdl2.zip
+#          echo "SDL2_DIR=$env:GITHUB_WORKSPACE/SDL2-${{ matrix.s2ver }}/cmake" >> $env:GITHUB_ENV
+#
+#      - name: Configure
+#        run: >
+#          cmake -S . -B ./build -A ${{ matrix.arch }}
+#          -DCMAKE_BUILD_TYPE=${{ matrix.build }}
+#          -DGGML_CUDA=${{ matrix.cublas }}
+#          -DWHISPER_SDL2=${{ matrix.sdl2 }}
+#
+#      - name: Build ${{ matrix.cuda-toolkit }}
+#        run: |
+#          cd ./build
+#          cmake --build . --config ${{ matrix.build }}
+#
+#      - name: Copy CUDA DLLs
+#        run: >
+#          Copy-Item -PassThru
+#          -Path "${{ steps.cuda-toolkit.outputs.CUDA_PATH }}/bin/*.dll"
+#          -Include cudart64_*,cublas64_*,cublasLt64_*
+#          -Destination build/bin/${{ matrix.build }}
+#
+#      - name: Copy SDL2.dll
+#        if: matrix.sdl2 == 'ON'
+#        run: copy "$env:SDL2_DIR/../lib/${{ matrix.s2arc }}/SDL2.dll" build/bin/${{ matrix.build }}
+#
+#      - name: Upload binaries
+#        if: matrix.sdl2 == 'ON'
+#        uses: actions/upload-artifact@v4
+#        with:
+#          name: whisper-cublas-${{ matrix.cuda-toolkit }}-bin-${{ matrix.arch }}
+#          path: build/bin/${{ matrix.build }}
 
   emscripten:
     runs-on: ubuntu-latest
@@ -533,7 +519,7 @@ jobs:
           emcmake cmake . -DCMAKE_BUILD_TYPE=${{ matrix.build }}
           make
 
-  ios:
+  ios-xcode-build:
     runs-on: macos-latest
 
     strategy:
@@ -541,7 +527,7 @@ jobs:
         build: [Release]
 
     steps:
-      - name: Clone
+      - name: Checkout code
         uses: actions/checkout@v4
 
       - name: Configure
@@ -549,11 +535,34 @@ jobs:
           cp models/for-tests-ggml-base.en.bin models/ggml-base.en.bin
           mkdir models/ggml-base.en-encoder.mlmodelc
 
-      - name: Build objc example
-        run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphonesimulator build
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -G Xcode .. \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DWHISPER_BUILD_EXAMPLES=OFF \
+            -DWHISPER_BUILD_TESTS=OFF \
+            -DWHISPER_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=iOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+          sudo cmake --install . --config Release
+
+      - name: xcodebuild for swift package
+        id: xcodebuild
+        run: |
+          xcodebuild -scheme whisper-Package -destination 'generic/platform=iOS'
+
+#- name: Build objc example
+#  run: xcodebuild -project examples/whisper.objc/whisper.objc.xcodeproj -scheme whisper.objc -configuration ${{ matrix.build }} -sdk iphoneos build
 
       - name: Build swiftui example
-        run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphonesimulator build
+        run: xcodebuild -project examples/whisper.swiftui/whisper.swiftui.xcodeproj -scheme WhisperCppDemo -configuration ${{ matrix.build }} -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
 
   android:
     runs-on: ubuntu-latest
@@ -664,5 +673,6 @@ jobs:
       - name: Test quantize
         run: |
           ./models/download-ggml-model.sh tiny.en
-          make quantize
-          ./quantize models/ggml-tiny.en.bin models/ggml-tiny.en-q4_0.bin q4_0
+          cmake -B build
+          cmake --build build --config Release
+          ./build/bin/quantize models/ggml-tiny.en.bin models/ggml-tiny.en-q4_0.bin q4_0
diff --git a/.gitignore b/.gitignore
index ebb73586796..c1e584dba32 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,5 +1,6 @@
 *.o
 *.a
+*.d
 .cache/
 .coreml/
 .test/
@@ -19,6 +20,9 @@ build-*/
 .swiftpm
 *.metallib
 
+ggml-metal-embed.metal
+ggml-metal-embed.metal.tmp
+
 /main
 /stream
 /command
diff --git a/CMakeLists.txt b/CMakeLists.txt
index c53252bac77..ff6500a84d9 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,6 +1,6 @@
 cmake_minimum_required(VERSION 3.5) # for add_link_options and implicit target directories.
 project("whisper.cpp" C CXX)
-project("whisper.cpp" VERSION 1.7.2)
+project("whisper.cpp" VERSION 1.7.3)
 include(CheckIncludeFileCXX)
 
 set(SOVERSION 1)
diff --git a/Makefile b/Makefile
index a3835dd1653..81fd392029d 100644
--- a/Makefile
+++ b/Makefile
@@ -1,1134 +1,12 @@
-# Define the default target now so that it is always the first target
-BUILD_TARGETS = \
-	main \
-	bench \
-	quantize \
-	server
-
-# Binaries only useful for tests
-TEST_TARGETS = \
-	tests/test-c.o
-
-# Deprecation aliases
-ifdef WHISPER_CUBLAS
-$(error WHISPER_CUBLAS is removed. Use GGML_CUDA instead.)
-endif
-
-ifdef WHISPER_CUDA
-GGML_CUDA := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef WHISPER_KOMPUTE
-GGML_KOMPUTE := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef WHISPER_METAL
-GGML_METAL := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef WHISPER_OPENMP
-GGML_OPENMP := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef WHISPER_RPC
-GGML_RPC := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef WHISPER_SYCL
-GGML_SYCL := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef WHISPER_SYCL_F16
-GGML_SYCL_F16 := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef WHISPER_OPENBLAS
-GGML_OPENBLAS := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef WHISPER_OPENBLAS64
-GGML_OPENBLAS64 := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef WHISPER_BLIS
-GGML_BLIS := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef WHISPER_NO_WHISPERFILE
-GGML_NO_WHISPERFILE := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef WHISPER_NO_ACCELERATE
-GGML_NO_ACCELERATE := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef WHISPER_NO_OPENMP
-GGML_NO_OPENMP := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifdef WHISPER_NO_METAL
-GGML_NO_METAL := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifndef UNAME_S
-UNAME_S := $(shell uname -s)
-endif
-
-ifndef UNAME_P
-UNAME_P := $(shell uname -p)
-endif
-
-ifndef UNAME_M
-UNAME_M := $(shell uname -m)
-endif
-
-# In GNU make default CXX is g++ instead of c++.  Let's fix that so that users
-# of non-gcc compilers don't have to provide g++ alias or wrapper.
-DEFCC  := cc
-DEFCXX := c++
-ifeq ($(origin CC),default)
-CC  := $(DEFCC)
-endif
-ifeq ($(origin CXX),default)
-CXX := $(DEFCXX)
-endif
-
-# Mac OS + Arm can report x86_64
-# ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
-ifeq ($(UNAME_S),Darwin)
-	ifndef GGML_NO_METAL
-		GGML_METAL := 1
-	endif
-
-	GGML_NO_OPENMP := 1
-
-	ifneq ($(UNAME_P),arm)
-		SYSCTL_M := $(shell sysctl -n hw.optional.arm64 2>/dev/null)
-		ifeq ($(SYSCTL_M),1)
-			# UNAME_P := arm
-			# UNAME_M := arm64
-			warn := $(warning Your arch is announced as x86_64, but it seems to actually be ARM64. Not fixing that can lead to bad performance. For more info see: https://github.com/ggerganov/whisper.cpp/issues/66\#issuecomment-1282546789)
-		endif
-	endif
-endif
-
-ifdef GGML_METAL
-	GGML_METAL_EMBED_LIBRARY := 1
-endif
-
-ifdef GGML_RPC
-	BUILD_TARGETS += rpc-server
-endif
-
-ifdef GGML_VULKAN
-	BUILD_TARGETS += vulkan-shaders-gen
-endif
-
-ifeq ($(shell sdl2-config --cflags --libs 2>/dev/null),)
-else
-	BUILD_TARGETS += \
-		command \
-		stream \
-		lsp \
-		talk-llama
-	# talk (TODO: disalbed)
-endif
-
-default: $(BUILD_TARGETS)
-
-test: $(TEST_TARGETS)
-	@failures=0; \
-	for test_target in $(TEST_TARGETS); do \
-		echo "Running test $$test_target..."; \
-		./$$test_target; \
-		if [ $$? -ne 0 ]; then \
-			printf 'Test %s FAILED!\n\n' $$test_target; \
-			failures=$$(( failures + 1 )); \
-		else \
-			printf 'Test %s passed.\n\n' $$test_target; \
-		fi; \
-	done; \
-	failures=$$(( failures + $$? )); \
-	if [ $$failures -gt 0 ]; then \
-		printf '\n%s tests failed.\n' $$failures; \
-		exit 1; \
-	fi
-	@echo 'All tests passed.'
-
-all: $(BUILD_TARGETS) $(TEST_TARGETS)
-
-ifdef RISCV_CROSS_COMPILE
-	CC	:= riscv64-unknown-linux-gnu-gcc
-	CXX	:= riscv64-unknown-linux-gnu-g++
-endif
-
-#
-# Compile flags
-#
-
-# keep standard at C11 and C++11
-MK_CPPFLAGS  = -Iggml/include -Iggml/src -Iinclude -Isrc -Iexamples
-MK_CFLAGS    = -std=c11   -fPIC
-MK_CXXFLAGS  = -std=c++11 -fPIC
-MK_NVCCFLAGS = -std=c++11
-
-ifndef WHISPER_NO_CCACHE
-CCACHE := $(shell which ccache)
-ifdef CCACHE
-export CCACHE_SLOPPINESS = time_macros
-$(info I ccache found, compilation results will be cached. Disable with WHISPER_NO_CCACHE.)
-CC    := $(CCACHE) $(CC)
-CXX   := $(CCACHE) $(CXX)
-else
-$(info I ccache not found. Consider installing it for faster compilation.)
-endif # CCACHE
-endif # WHISPER_NO_CCACHE
-
-# clock_gettime came in POSIX.1b (1993)
-# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
-# posix_memalign came in POSIX.1-2001 / SUSv3
-# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
-MK_CPPFLAGS += -D_XOPEN_SOURCE=600
-
-# Somehow in OpenBSD whenever POSIX conformance is specified
-# some string functions rely on locale_t availability,
-# which was introduced in POSIX.1-2008, forcing us to go higher
-ifeq ($(UNAME_S),OpenBSD)
-	MK_CPPFLAGS += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
-endif
-
-# Data types, macros and functions related to controlling CPU affinity and
-# some memory allocation are available on Linux through GNU extensions in libc
-ifeq ($(UNAME_S),Linux)
-	MK_CPPFLAGS += -D_GNU_SOURCE
-endif
-
-# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
-# and on macOS its availability depends on enabling Darwin extensions
-# similarly on DragonFly, enabling BSD extensions is necessary
-ifeq ($(UNAME_S),Darwin)
-	MK_CPPFLAGS += -D_DARWIN_C_SOURCE
-endif
-ifeq ($(UNAME_S),DragonFly)
-	MK_CPPFLAGS += -D__BSD_VISIBLE
-endif
-
-# alloca is a non-standard interface that is not visible on BSDs when
-# POSIX conformance is specified, but not all of them provide a clean way
-# to enable it in such cases
-ifeq ($(UNAME_S),FreeBSD)
-	MK_CPPFLAGS += -D__BSD_VISIBLE
-endif
-ifeq ($(UNAME_S),NetBSD)
-	MK_CPPFLAGS += -D_NETBSD_SOURCE
-endif
-ifeq ($(UNAME_S),OpenBSD)
-	MK_CPPFLAGS += -D_BSD_SOURCE
-endif
-
-ifdef GGML_SCHED_MAX_COPIES
-	MK_CPPFLAGS += -DGGML_SCHED_MAX_COPIES=$(GGML_SCHED_MAX_COPIES)
-endif
-
-ifdef WHISPER_DEBUG
-	MK_CFLAGS    += -O0 -g
-	MK_CXXFLAGS  += -O0 -g
-	MK_LDFLAGS   += -g
-	MK_NVCCFLAGS += -O0 -g
-
-	ifeq ($(UNAME_S),Linux)
-		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
-	endif
-else
-	MK_CPPFLAGS   += -DNDEBUG
-	MK_CFLAGS     += -O3
-	MK_CXXFLAGS   += -O3
-	MK_NVCCFLAGS  += -O3
-endif
-
-ifdef WHISPER_SANITIZE_THREAD
-	MK_CFLAGS   += -fsanitize=thread -g
-	MK_CXXFLAGS += -fsanitize=thread -g
-	MK_LDFLAGS  += -fsanitize=thread -g
-endif
-
-ifdef WHISPER_SANITIZE_ADDRESS
-	MK_CFLAGS   += -fsanitize=address -fno-omit-frame-pointer -g
-	MK_CXXFLAGS += -fsanitize=address -fno-omit-frame-pointer -g
-	MK_LDFLAGS  += -fsanitize=address -fno-omit-frame-pointer -g
-endif
-
-ifdef WHISPER_SANITIZE_UNDEFINED
-	MK_CFLAGS   += -fsanitize=undefined -g
-	MK_CXXFLAGS += -fsanitize=undefined -g
-	MK_LDFLAGS  += -fsanitize=undefined -g
-endif
-
-ifdef WHISPER_SERVER_VERBOSE
-	MK_CPPFLAGS += -DSERVER_VERBOSE=$(WHISPER_SERVER_VERBOSE)
-endif
-
-ifdef WHISPER_SERVER_SSL
-	MK_CPPFLAGS += -DCPPHTTPLIB_OPENSSL_SUPPORT
-	MK_LDFLAGS  += -lssl -lcrypto
-endif
-
-ifdef WHISPER_DISABLE_LOGS
-	MK_CPPFLAGS += -DLOG_DISABLE_LOGS
-endif # WHISPER_DISABLE_LOGS
-
-# warnings
-WARN_FLAGS = \
-	-Wall \
-	-Wextra \
-	-Wpedantic \
-	-Wcast-qual \
-	-Wno-unused-function
-
-MK_CFLAGS += \
-	$(WARN_FLAGS) \
-	-Wshadow \
-	-Wstrict-prototypes \
-	-Wpointer-arith \
-	-Wmissing-prototypes \
-	-Werror=implicit-int \
-	-Werror=implicit-function-declaration
-
-MK_CXXFLAGS += \
-	$(WARN_FLAGS) \
-	-Wmissing-declarations \
-	-Wmissing-noreturn
-
-ifeq ($(WHISPER_FATAL_WARNINGS),1)
-	MK_CFLAGS   += -Werror
-	MK_CXXFLAGS += -Werror
-endif
-
-# this version of Apple ld64 is buggy
-ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
-	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
-endif
-
-# OS specific
-# TODO: support Windows
-ifneq '' '$(filter $(UNAME_S),Linux Darwin FreeBSD NetBSD OpenBSD Haiku)'
-	MK_CFLAGS   += -pthread
-	MK_CXXFLAGS += -pthread
-endif
-
-# detect Windows
-ifneq ($(findstring _NT,$(UNAME_S)),)
-	_WIN32 := 1
-endif
-
-# library name prefix
-ifneq ($(_WIN32),1)
-	LIB_PRE := lib
-endif
-
-# Dynamic Shared Object extension
-ifneq ($(_WIN32),1)
-	DSO_EXT := .so
-else
-	DSO_EXT := .dll
-endif
-
-# Windows Sockets 2 (Winsock) for network-capable apps
-ifeq ($(_WIN32),1)
-	LWINSOCK2 := -lws2_32
-endif
-
-ifdef WHISPER_GPROF
-	MK_CFLAGS   += -pg
-	MK_CXXFLAGS += -pg
-endif
-
-# Architecture specific
-# TODO: probably these flags need to be tweaked on some architectures
-#       feel free to update the Makefile for your architecture and send a pull request or issue
-
-ifndef RISCV
-
-ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
-	# Use all CPU extensions that are available:
-	MK_CFLAGS     += -march=native -mtune=native
-	HOST_CXXFLAGS += -march=native -mtune=native
-
-	# Usage AVX-only
-	#MK_CFLAGS   += -mfma -mf16c -mavx
-	#MK_CXXFLAGS += -mfma -mf16c -mavx
-
-	# Usage SSSE3-only (Not is SSE3!)
-	#MK_CFLAGS   += -mssse3
-	#MK_CXXFLAGS += -mssse3
-endif
-
-ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))'
-	# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
-	# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
-	# https://github.com/ggerganov/llama.cpp/issues/2922
-	MK_CFLAGS   += -Xassembler -muse-unaligned-vector-move
-	MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move
-
-	# Target Windows 8 for PrefetchVirtualMemory
-	MK_CPPFLAGS += -D_WIN32_WINNT=0x602
-endif
-
-ifneq ($(filter aarch64%,$(UNAME_M)),)
-	# Apple M1, M2, etc.
-	# Raspberry Pi 3, 4, Zero 2 (64-bit)
-	# Nvidia Jetson
-	MK_CFLAGS   += -mcpu=native
-	MK_CXXFLAGS += -mcpu=native
-	JETSON_RELEASE_INFO = $(shell jetson_release)
-	ifdef JETSON_RELEASE_INFO
-		ifneq ($(filter TX2%,$(JETSON_RELEASE_INFO)),)
-			JETSON_EOL_MODULE_DETECT = 1
-			CC = aarch64-unknown-linux-gnu-gcc
-			cxx = aarch64-unknown-linux-gnu-g++
-		endif
-	endif
-endif
-
-ifneq ($(filter armv6%,$(UNAME_M)),)
-	# Raspberry Pi 1, Zero
-	MK_CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
-	MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
-endif
-
-ifneq ($(filter armv7%,$(UNAME_M)),)
-	# Raspberry Pi 2
-	MK_CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
-	MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
-endif
-
-ifneq ($(filter armv8%,$(UNAME_M)),)
-	# Raspberry Pi 3, 4, Zero 2 (32-bit)
-	MK_CFLAGS   += -mfp16-format=ieee -mno-unaligned-access
-	MK_CXXFLAGS += -mfp16-format=ieee -mno-unaligned-access
-endif
-
-ifneq ($(filter ppc64%,$(UNAME_M)),)
-	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
-	ifneq (,$(findstring POWER9,$(POWER9_M)))
-		MK_CFLAGS   += -mcpu=power9
-		MK_CXXFLAGS += -mcpu=power9
-	endif
-endif
-
-ifneq ($(filter ppc64le%,$(UNAME_M)),)
-	MK_CFLAGS   += -mcpu=powerpc64le
-	MK_CXXFLAGS += -mcpu=powerpc64le
-	CUDA_POWER_ARCH = 1
-endif
-
-ifneq ($(filter loongarch64%,$(UNAME_M)),)
-	MK_CFLAGS   += -mlasx
-	MK_CXXFLAGS += -mlasx
-endif
-
-else
-	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
-	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
-endif
-
-ifndef GGML_NO_ACCELERATE
-	# Mac OS - include Accelerate framework.
-	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
-	ifeq ($(UNAME_S),Darwin)
-		MK_CPPFLAGS += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS
-		MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
-		MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
-		MK_LDFLAGS  += -framework Accelerate
-		OBJ_GGML    += ggml/src/ggml-blas.o
-	endif
-endif # GGML_NO_ACCELERATE
-
-ifndef GGML_NO_OPENMP
-	MK_CPPFLAGS += -DGGML_USE_OPENMP
-	MK_CFLAGS   += -fopenmp
-	MK_CXXFLAGS += -fopenmp
-endif # GGML_NO_OPENMP
-
-ifdef GGML_OPENBLAS
-	MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas)
-	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas)
-	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
-	OBJ_GGML    += ggml/src/ggml-blas.o
-endif # GGML_OPENBLAS
-
-ifdef GGML_OPENBLAS64
-	MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas64)
-	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas64)
-	MK_LDFLAGS  += $(shell pkg-config --libs openblas64)
-	OBJ_GGML    += ggml/src/ggml-blas.o
-endif # GGML_OPENBLAS64
-
-ifdef GGML_BLIS
-	MK_CPPFLAGS += -DGGML_USE_BLAS -I/usr/local/include/blis -I/usr/include/blis
-	MK_LDFLAGS  += -lblis -L/usr/local/lib
-	OBJ_GGML    += ggml/src/ggml-blas.o
-endif # GGML_BLIS
-
-ifdef GGML_RPC
-	MK_CPPFLAGS += -DGGML_USE_RPC
-	OBJ_GGML    += ggml/src/ggml-rpc.o
-endif # GGML_RPC
-
-OBJ_CUDA_TMPL      = $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-wmma*.cu))
-OBJ_CUDA_TMPL     += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/mmq*.cu))
-
-ifdef GGML_CUDA_FA_ALL_QUANTS
-	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*.cu))
-else
-	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu))
-	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu))
-	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*f16-f16.cu))
-endif # GGML_CUDA_FA_ALL_QUANTS
-
-ifdef GGML_CUDA
-	ifneq ('', '$(wildcard /opt/cuda)')
-		CUDA_PATH ?= /opt/cuda
-	else
-		CUDA_PATH ?= /usr/local/cuda
-	endif
-
-	#MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
-	#MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcufft -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
-	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
-	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
-	MK_NVCCFLAGS += -use_fast_math
-
-	OBJ_GGML += ggml/src/ggml-cuda.o
-	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
-	OBJ_GGML += $(OBJ_CUDA_TMPL)
-ifdef WHISPER_FATAL_WARNINGS
-	MK_NVCCFLAGS += -Werror all-warnings
-endif # WHISPER_FATAL_WARNINGS
-
-ifndef JETSON_EOL_MODULE_DETECT
-	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
-endif # JETSON_EOL_MODULE_DETECT
-
-ifdef WHISPER_DEBUG
-	MK_NVCCFLAGS += -lineinfo
-endif # WHISPER_DEBUG
-
-ifdef GGML_CUDA_DEBUG
-	MK_NVCCFLAGS += --device-debug
-endif # GGML_CUDA_DEBUG
-
-ifdef GGML_CUDA_NVCC
-	NVCC = $(CCACHE) $(GGML_CUDA_NVCC)
-else
-	NVCC = $(CCACHE) nvcc
-endif #GGML_CUDA_NVCC
-
-ifdef CUDA_DOCKER_ARCH
-	MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
-else ifndef CUDA_POWER_ARCH
-	MK_NVCCFLAGS += -arch=native
-endif # CUDA_DOCKER_ARCH
-
-ifdef GGML_CUDA_FORCE_DMMV
-	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
-endif # GGML_CUDA_FORCE_DMMV
-
-ifdef GGML_CUDA_FORCE_MMQ
-	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
-endif # GGML_CUDA_FORCE_MMQ
-
-ifdef GGML_CUDA_DMMV_X
-	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
-else
-	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
-endif # GGML_CUDA_DMMV_X
-
-ifdef GGML_CUDA_MMV_Y
-	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
-else ifdef GGML_CUDA_DMMV_Y
-	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_DMMV_Y) # for backwards compatibility
-else
-	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
-endif # GGML_CUDA_MMV_Y
-
-ifdef GGML_CUDA_F16
-	MK_NVCCFLAGS += -DGGML_CUDA_F16
-endif # GGML_CUDA_F16
-
-ifdef GGML_CUDA_DMMV_F16
-	MK_NVCCFLAGS += -DGGML_CUDA_F16
-endif # GGML_CUDA_DMMV_F16
-
-ifdef GGML_CUDA_KQUANTS_ITER
-	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
-else
-	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
-endif
-
-ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
-	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE)
-else
-	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
-endif # GGML_CUDA_PEER_MAX_BATCH_SIZE
-
-ifdef GGML_CUDA_NO_PEER_COPY
-	MK_NVCCFLAGS += -DGGML_CUDA_NO_PEER_COPY
-endif # GGML_CUDA_NO_PEER_COPY
-
-ifdef GGML_CUDA_CCBIN
-	MK_NVCCFLAGS += -ccbin $(GGML_CUDA_CCBIN)
-endif # GGML_CUDA_CCBIN
-
-ifdef GGML_CUDA_FA_ALL_QUANTS
-	MK_NVCCFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
-endif # GGML_CUDA_FA_ALL_QUANTS
-
-ifdef JETSON_EOL_MODULE_DETECT
-define NVCC_COMPILE
-	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
-endef # NVCC_COMPILE
-else
-define NVCC_COMPILE
-	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
-endef # NVCC_COMPILE
-endif # JETSON_EOL_MODULE_DETECT
-
-ggml/src/ggml-cuda/%.o: \
-	ggml/src/ggml-cuda/%.cu \
-	ggml/include/ggml.h \
-	ggml/src/ggml-common.h \
-	ggml/src/ggml-cuda/common.cuh
-	$(NVCC_COMPILE)
-
-ggml/src/ggml-cuda.o: \
-	ggml/src/ggml-cuda.cu \
-	ggml/include/ggml.h \
-	ggml/include/ggml-backend.h \
-	ggml/include/ggml-cuda.h \
-	ggml/src/ggml-backend-impl.h \
-	ggml/src/ggml-common.h \
-	$(wildcard ggml/src/ggml-cuda/*.cuh)
-	$(NVCC_COMPILE)
-endif # GGML_CUDA
-
-ifdef GGML_VULKAN
-	MK_CPPFLAGS += -DGGML_USE_VULKAN
-	MK_LDFLAGS  += $(shell pkg-config --libs vulkan)
-	OBJ_GGML    += ggml/src/ggml-vulkan.o ggml/src/ggml-vulkan-shaders.o
-
-ifdef GGML_VULKAN_CHECK_RESULTS
-	MK_CPPFLAGS  += -DGGML_VULKAN_CHECK_RESULTS
-endif
-
-ifdef GGML_VULKAN_DEBUG
-	MK_CPPFLAGS  += -DGGML_VULKAN_DEBUG
-endif
-
-ifdef GGML_VULKAN_MEMORY_DEBUG
-	MK_CPPFLAGS  += -DGGML_VULKAN_MEMORY_DEBUG
-endif
-
-ifdef GGML_VULKAN_PERF
-	MK_CPPFLAGS  += -DGGML_VULKAN_PERF
-endif
-
-ifdef GGML_VULKAN_VALIDATE
-	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
-endif
-
-ifdef GGML_VULKAN_RUN_TESTS
-	MK_CPPFLAGS  += -DGGML_VULKAN_RUN_TESTS
-endif
-
-GLSLC_CMD  = glslc
-_ggml_vk_genshaders_cmd = $(shell pwd)/vulkan-shaders-gen
-_ggml_vk_header = ggml/src/ggml-vulkan-shaders.hpp
-_ggml_vk_source = ggml/src/ggml-vulkan-shaders.cpp
-_ggml_vk_input_dir = ggml/src/vulkan-shaders
-_ggml_vk_shader_deps = $(echo $(_ggml_vk_input_dir)/*.comp)
-
-ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source)
-	$(CXX) $(CXXFLAGS) $(shell pkg-config --cflags vulkan) -c $< -o $@
-
-$(_ggml_vk_header): $(_ggml_vk_source)
-
-$(_ggml_vk_source): $(_ggml_vk_shader_deps) vulkan-shaders-gen
-	$(_ggml_vk_genshaders_cmd) \
-		--glslc      $(GLSLC_CMD) \
-		--input-dir  $(_ggml_vk_input_dir) \
-		--target-hpp $(_ggml_vk_header) \
-		--target-cpp $(_ggml_vk_source)
-
-vulkan-shaders-gen: ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
-	$(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
-
-endif # GGML_VULKAN
-
-ifdef GGML_HIPBLAS
-	ifeq ($(wildcard /opt/rocm),)
-		ROCM_PATH      ?= /usr
-		AMDGPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
-	else
-		ROCM_PATH	?= /opt/rocm
-		AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
-	endif
-
-	GGML_CUDA_DMMV_X       ?= 32
-	GGML_CUDA_MMV_Y        ?= 1
-	GGML_CUDA_KQUANTS_ITER ?= 2
-
-	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
-
-ifdef GGML_HIP_UMA
-	MK_CPPFLAGS += -DGGML_HIP_UMA
-endif # GGML_HIP_UMA
-
-	MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
-	MK_LDFLAGS += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64
-	MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas
-
-	HIPCC ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc
-
-	HIPFLAGS += $(addprefix --offload-arch=,$(AMDGPU_TARGETS))
-	HIPFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
-	HIPFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
-	HIPFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
-
-ifdef GGML_CUDA_FORCE_DMMV
-	HIPFLAGS += -DGGML_CUDA_FORCE_DMMV
-endif # GGML_CUDA_FORCE_DMMV
-
-ifdef GGML_CUDA_NO_PEER_COPY
-	HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY
-endif # GGML_CUDA_NO_PEER_COPY
-
-	OBJ_GGML += ggml/src/ggml-cuda.o
-	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
-	OBJ_GGML += $(OBJ_CUDA_TMPL)
-
-ggml/src/ggml-cuda.o: \
-	ggml/src/ggml-cuda.cu \
-	ggml/include/ggml.h \
-	ggml/include/ggml-backend.h \
-	ggml/include/ggml-cuda.h \
-	ggml/src/ggml-backend-impl.h \
-	ggml/src/ggml-common.h \
-	$(wildcard ggml/src/ggml-cuda/*.cuh)
-	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
-
-ggml/src/ggml-cuda/%.o: \
-	ggml/src/ggml-cuda/%.cu \
-	ggml/include/ggml.h \
-	ggml/src/ggml-common.h \
-	ggml/src/ggml-cuda/common.cuh
-	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
-endif # GGML_HIPBLAS
-
-ifdef GGML_METAL
-	MK_CPPFLAGS += -DGGML_USE_METAL
-	MK_LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
-	OBJ_GGML	+= ggml/src/ggml-metal.o
-ifdef GGML_METAL_NDEBUG
-	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
-endif
-
-ifdef GGML_METAL_EMBED_LIBRARY
-	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
-	OBJ_GGML    += ggml/src/ggml-metal-embed.o
-endif
-endif # GGML_METAL
-
-ifdef WHISPER_COREML
-	MK_CXXFLAGS += -DWHISPER_USE_COREML
-	LDFLAGS     += -framework Foundation -framework CoreML
-
-ifdef WHISPER_COREML_ALLOW_FALLBACK
-	MK_CXXFLAGS += -DWHISPER_COREML_ALLOW_FALLBACK
-endif
-endif
-
-# ===
-
-ifdef GGML_METAL
-ggml/src/ggml-metal.o: \
-	ggml/src/ggml-metal.m \
-	ggml/include/ggml-metal.h \
-	ggml/include/ggml.h
-	$(CC) $(CFLAGS) -c $< -o $@
-
-ifdef GGML_METAL_EMBED_LIBRARY
-ggml/src/ggml-metal-embed.o: \
-	ggml/src/ggml-metal.metal \
-	ggml/src/ggml-common.h
-	@echo "Embedding Metal library"
-	@sed -e '/#include "ggml-common.h"/r ggml/src/ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml/src/ggml-metal.metal > ggml/src/ggml-metal-embed.metal
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f ${TEMP_ASSEMBLY}
-endif
-endif # GGML_METAL
-
-ifdef WHISPER_COREML
-src/coreml/whisper-encoder.o: src/coreml/whisper-encoder.mm src/coreml/whisper-encoder.h
-	$(CXX) -O3 -I . -fobjc-arc -c src/coreml/whisper-encoder.mm -o src/coreml/whisper-encoder.o
-
-src/coreml/whisper-encoder-impl.o: src/coreml/whisper-encoder-impl.m src/coreml/whisper-encoder-impl.h
-	$(CXX) -O3 -I . -fobjc-arc -c src/coreml/whisper-encoder-impl.m -o src/coreml/whisper-encoder-impl.o
-
-OBJ_WHISPER += src/coreml/whisper-encoder.o src/coreml/whisper-encoder-impl.o
-endif
-
-OBJ_GGML += \
-	ggml/src/ggml.o \
-	ggml/src/ggml-cpu.o \
-	ggml/src/ggml-alloc.o \
-	ggml/src/ggml-backend.o \
-	ggml/src/ggml-quants.o \
-	ggml/src/ggml-aarch64.o
-
-OBJ_WHISPER += \
-	src/whisper.o
-
-OBJ_COMMON += \
-	examples/common.o \
-	examples/common-ggml.o \
-	examples/grammar-parser.o
-
-OBJ_SDL += \
-	examples/common-sdl.o
-
-OBJ_ALL = $(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL)
-
-LIB_GGML   = $(LIB_PRE)ggml$(DSO_EXT)
-LIB_GGML_S = $(LIB_PRE)ggml.a
-
-LIB_WHISPER   = $(LIB_PRE)whisper$(DSO_EXT)
-LIB_WHISPER_S = $(LIB_PRE)whisper.a
-
-LIB_COMMON   = $(LIB_PRE)common$(DSO_EXT)
-LIB_COMMON_S = $(LIB_PRE)common.a
-
-LIB_COMMON_SDL   = $(LIB_PRE)common-sdl$(DSO_EXT)
-LIB_COMMON_SDL_S = $(LIB_PRE)common-sdl.a
-
-LIB_ALL   = $(LIB_GGML)   $(LIB_WHISPER)   $(LIB_COMMON)   $(LIB_COMMON_SDL)
-LIB_ALL_S = $(LIB_GGML_S) $(LIB_WHISPER_S) $(LIB_COMMON_S) $(LIB_COMMON_SDL_S)
-
-GF_CC := $(CC)
-include scripts/get-flags.mk
-
-# combine build flags with cmdline overrides
-override CPPFLAGS  := $(MK_CPPFLAGS) $(CPPFLAGS)
-override CFLAGS    := $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
-BASE_CXXFLAGS      := $(MK_CXXFLAGS) $(CXXFLAGS)
-override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS) $(CPPFLAGS)
-override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
-override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
-
-# identify CUDA host compiler
-ifdef GGML_CUDA
-GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
-include scripts/get-flags.mk
-CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
-endif
-
-ifdef WHISPER_CURL
-override CXXFLAGS := $(CXXFLAGS) -DWHISPER_USE_CURL
-override LDFLAGS  := $(LDFLAGS) -lcurl
-endif
-
-#
-# Print build information
-#
-
-$(info I whisper.cpp build info: )
-$(info I UNAME_S:   $(UNAME_S))
-$(info I UNAME_P:   $(UNAME_P))
-$(info I UNAME_M:   $(UNAME_M))
-$(info I CFLAGS:    $(CFLAGS))
-$(info I CXXFLAGS:  $(CXXFLAGS))
-$(info I NVCCFLAGS: $(NVCCFLAGS))
-$(info I LDFLAGS:   $(LDFLAGS))
-$(info I CC:        $(shell $(CC)   --version | head -n 1))
-$(info I CXX:       $(shell $(CXX)  --version | head -n 1))
-ifdef GGML_CUDA
-$(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
-CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
-ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
-
-ifndef CUDA_DOCKER_ARCH
-ifndef CUDA_POWER_ARCH
-$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via environment variable CUDA_DOCKER_ARCH, e.g. by running "export CUDA_DOCKER_ARCH=compute_XX" on Unix-like systems, where XX is the minimum compute capability that the code needs to run on. A list with compute capabilities can be found here: https://developer.nvidia.com/cuda-gpus )
-endif # CUDA_POWER_ARCH
-endif # CUDA_DOCKER_ARCH
-
-endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
-endif # GGML_CUDA
-$(info )
-
-ifdef DEPRECATE_WARNING
-$(info !!! DEPRECATION WARNING !!!)
-$(info The following WHISPER_ options are deprecated and will be removed in the future. Use the GGML_ prefix instead)
-$(info   - WHISPER_CUDA)
-$(info   - WHISPER_METAL)
-$(info   - WHISPER_OPENMP)
-$(info   - WHISPER_RPC)
-$(info   - WHISPER_SYCL)
-$(info   - WHISPER_SYCL_F16)
-$(info   - WHISPER_OPENBLAS)
-$(info   - WHISPER_OPENBLAS64)
-$(info   - WHISPER_BLIS)
-$(info   - WHISPER_NO_LLAMAFILE)
-$(info   - WHISPER_NO_ACCELERATE)
-$(info   - WHISPER_NO_OPENMP)
-$(info   - WHISPER_NO_METAL)
-$(info )
-endif
-
-#
-# Build libraries
-#
-
-# ggml
-
-ggml/src/ggml.o: \
-	ggml/src/ggml.c \
-	ggml/include/ggml.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
-
-ggml/src/ggml-cpu.o: \
-	ggml/src/ggml-cpu.c \
-	ggml/include/ggml.h \
-	ggml/src/ggml-common.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
-
-ggml/src/ggml-alloc.o: \
-	ggml/src/ggml-alloc.c \
-	ggml/include/ggml.h \
-	ggml/include/ggml-alloc.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
-
-ggml/src/ggml-backend.o: \
-	ggml/src/ggml-backend.cpp \
-	ggml/include/ggml.h \
-	ggml/include/ggml-backend.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-ggml/src/ggml-quants.o: \
-	ggml/src/ggml-quants.c \
-	ggml/include/ggml.h \
-	ggml/src/ggml-quants.h \
-	ggml/src/ggml-common.h
-	$(CC) $(CFLAGS)    -c $< -o $@
-
-ggml/src/ggml-aarch64.o: \
-	ggml/src/ggml-aarch64.c \
-	ggml/include/ggml.h \
-	ggml/src/ggml-aarch64.h \
-	ggml/src/ggml-common.h
-	$(CC) $(CFLAGS)    -c $< -o $@
-
-ggml/src/ggml-blas.o: \
-	ggml/src/ggml-blas.cpp \
-	ggml/include/ggml-blas.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-ifdef GGML_LLAMAFILE
-ggml/src/sgemm.o: \
-	ggml/src/sgemm.cpp \
-	ggml/src/sgemm.h \
-	ggml/include/ggml.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-endif # GGML_LLAMAFILE
-
-ifdef GGML_RPC
-ggml/src/ggml-rpc.o: \
-	ggml/src/ggml-rpc.cpp \
-	ggml/include/ggml-rpc.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-endif # GGML_RPC
-
-$(LIB_GGML): \
-	$(OBJ_GGML)
-	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
-
-$(LIB_GGML_S): \
-	$(OBJ_GGML)
-	ar rcs $(LIB_GGML_S) $^
-
-# whisper
-
-src/whisper.o: \
-	src/whisper.cpp \
-	include/whisper.h \
-	ggml/include/ggml.h \
-	ggml/include/ggml-alloc.h \
-	ggml/include/ggml-backend.h \
-	ggml/include/ggml-cuda.h \
-	ggml/include/ggml-metal.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-$(LIB_WHISPER): \
-	$(OBJ_WHISPER) \
-	$(LIB_GGML)
-	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
-
-$(LIB_WHISPER_S): \
-	$(OBJ_WHISPER) \
-	$(OBJ_GGML)
-	ar rcs $(LIB_WHISPER_S) $^
-
-# common
-
-examples/common.o: \
-	examples/common.cpp \
-	examples/common.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-examples/common-ggml.o: \
-	examples/common-ggml.cpp \
-	examples/common-ggml.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-$(LIB_COMMON): \
-	$(OBJ_COMMON)
-	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
-
-$(LIB_COMMON_S): \
-	$(OBJ_COMMON)
-	ar rcs $(LIB_COMMON_S) $^
-
-# common-sdl
-
-CFLAGS_SDL=$(shell sdl2-config --cflags)
-LDFLAGS_SDL=$(shell sdl2-config --libs)
-
-examples/common-sdl.o: \
-	examples/common-sdl.cpp \
-	examples/common-sdl.h
-	$(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $@
-
-$(LIB_COMMON_SDL): \
-	$(OBJ_SDL)
-	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) $(LDFLAGS_SDL)
-
-$(LIB_COMMON_SDL_S): \
-	$(OBJ_SDL)
-	ar rcs $(LIB_COMMON_SDL_S) $^
-
-clean:
-	rm -vrf *.dot $(BUILD_TARGETS) $(TEST_TARGETS)
-	rm -rvf src/*.o
-	rm -rvf src/coreml/*.o
-	rm -rvf tests/*.o
-	rm -rvf examples/*.o
-	rm -rvf *.a
-	rm -rvf *.dll
-	rm -rvf *.so
-	rm -rvf *.dot
-	rm -rvf ggml/*.a
-	rm -rvf ggml/*.dll
-	rm -rvf ggml/*.so
-	rm -vrf ggml/src/*.o
-	rm -vrf ggml/src/ggml-metal-embed.metal
-	rm -vrf ggml/src/ggml-cuda/*.o
-	rm -vrf ggml/src/ggml-cuda/template-instances/*.o
-	rm -rvf $(BUILD_TARGETS)
-	rm -rvf $(TEST_TARGETS)
-	find examples -type f -name "*.o" -delete
-
-#
-# Examples
-#
-
-# $< is the first prerequisite, i.e. the source file.
-# Explicitly compile this to an object file so that it can be cached with ccache.
-# The source file is then filtered out from $^ (the list of all prerequisites) and the object file is added instead.
-
-# Helper function that replaces .c, .cpp, and .cu file endings with .o:
-GET_OBJ_FILE = $(patsubst %.c,%.o,$(patsubst %.cpp,%.o,$(patsubst %.cu,%.o,$(1))))
-
-main: examples/main/main.cpp \
-	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-bench: examples/bench/bench.cpp \
-	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-quantize: examples/quantize/quantize.cpp \
-	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-server: examples/server/server.cpp \
-	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
-
-command: examples/command/command.cpp \
-	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL)
-	$(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL)
-
-stream: examples/stream/stream.cpp \
-	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL)
-	$(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL)
-
-lsp: examples/lsp/lsp.cpp \
-	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL)
-	$(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL)
-
-# TODO: disabled until update
-#       https://github.com/ggerganov/whisper.cpp/issues/1818
-#talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp \
-#	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL)
-#	$(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<)
-#	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL)
-
-talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp examples/talk-llama/llama-vocab.cpp examples/talk-llama/llama-grammar.cpp examples/talk-llama/llama-sampling.cpp examples/talk-llama/unicode.cpp examples/talk-llama/unicode-data.cpp \
-	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL)
-	$(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL)
-
-#
-# Tests
-#
-
-tests: $(TEST_TARGETS)
-
-tests/test-c.o: tests/test-c.c include/whisper.h
-	$(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@
-
 #
 # Audio samples
 #
 
+.PHONY: build
+build:
+	cmake -B build
+	cmake --build build --config Release
+
 # download a few audio samples into folder "./samples":
 .PHONY: samples
 samples:
@@ -1172,8 +50,10 @@ samples:
 .PHONY: large-v3
 .PHONY: large-v3-turbo
 
-tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large-v3 large-v3-turbo: main
+tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large-v3 large-v3-turbo:
 	bash ./models/download-ggml-model.sh $@
+	cmake -B build
+	cmake --build build --config Release
 	@echo ""
 	@echo "==============================================="
 	@echo "Running $@ on all samples in ./samples ..."
@@ -1184,6 +64,6 @@ tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 larg
 		echo "[+] Running $@ on $$f ... (run 'ffplay $$f' to listen)" ; \
 	    echo "----------------------------------------------" ; \
 		echo "" ; \
-		./main -m models/ggml-$@.bin -f $$f ; \
+		./build/bin/main -m models/ggml-$@.bin -f $$f ; \
 		echo "" ; \
 	done
diff --git a/Package.swift b/Package.swift
index 9d224940a37..303ed0d78fb 100644
--- a/Package.swift
+++ b/Package.swift
@@ -14,49 +14,6 @@ let package = Package(
         .library(name: "whisper", targets: ["whisper"]),
     ],
     targets: [
-        .target(
-            name: "whisper",
-            path: ".",
-            exclude: [
-               "build",
-               "bindings",
-               "cmake",
-               "examples",
-               "scripts",
-               "models",
-               "samples",
-               "tests",
-               "CMakeLists.txt",
-               "Makefile",
-               "ggml/src/ggml-metal-embed.metal"
-            ],
-            sources: [
-                "ggml/src/ggml.c",
-                "src/whisper.cpp",
-                "ggml/src/ggml-aarch64.c",
-                "ggml/src/ggml-alloc.c",
-                "ggml/src/ggml-backend.cpp",
-                "ggml/src/ggml-cpu.c",
-                "ggml/src/ggml-quants.c",
-                "ggml/src/ggml-metal.m"
-            ],
-            resources: [.process("ggml/src/ggml-metal.metal")],
-            publicHeadersPath: "spm-headers",
-            cSettings: [
-                .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_ACCELERATE"),
-                .unsafeFlags(["-fno-objc-arc"]),
-                .define("GGML_USE_METAL")
-                // NOTE: NEW_LAPACK will required iOS version 16.4+
-                // We should consider add this in the future when we drop support for iOS 14
-                // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
-                // .define("ACCELERATE_NEW_LAPACK"),
-                // .define("ACCELERATE_LAPACK_ILP64")
-            ],
-            linkerSettings: [
-                .linkedFramework("Accelerate")
-            ]
-        )
-    ],
-    cxxLanguageStandard: .cxx11
+        .systemLibrary(name: "whisper", pkgConfig: "whisper"),
+    ]
 )
diff --git a/README.md b/README.md
index c02f66b8163..faa166d04b2 100644
--- a/README.md
+++ b/README.md
@@ -7,7 +7,7 @@
 [![Conan Center](https://shields.io/conan/v/whisper-cpp)](https://conan.io/center/whisper-cpp)
 [![npm](https://img.shields.io/npm/v/whisper.cpp.svg)](https://www.npmjs.com/package/whisper.cpp/)
 
-Stable: [v1.7.2](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.7.2) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
+Stable: [v1.7.3](https://github.com/ggerganov/whisper.cpp/releases/tag/v1.7.3) / [Roadmap | F.A.Q.](https://github.com/ggerganov/whisper.cpp/discussions/126)
 
 High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisper) automatic speech recognition (ASR) model:
 
@@ -89,10 +89,11 @@ Now build the [main](examples/main) example and transcribe an audio file like th
 
 ```bash
 # build the main example
-make -j
+cmake -B build
+cmake --build build --config Release
 
 # transcribe an audio file
-./main -f samples/jfk.wav
+./build/bin/main -f samples/jfk.wav
 ```
 
 ---
@@ -265,11 +266,12 @@ Here are the steps for creating and using a quantized model:
 
 ```bash
 # quantize a model with Q5_0 method
-make -j quantize
-./quantize models/ggml-base.en.bin models/ggml-base.en-q5_0.bin q5_0
+cmake -B build
+cmake --build build --config Release
+./build/bin/quantize models/ggml-base.en.bin models/ggml-base.en-q5_0.bin q5_0
 
 # run the examples as usual, specifying the quantized model file
-./main -m models/ggml-base.en-q5_0.bin ./samples/gb0.wav
+./build/bin/main -m models/ggml-base.en-q5_0.bin ./samples/gb0.wav
 ```
 
 ## Core ML support
@@ -303,10 +305,6 @@ speed-up - more than x3 faster compared with CPU-only execution. Here are the in
 - Build `whisper.cpp` with Core ML support:
 
   ```bash
-  # using Makefile
-  make clean
-  WHISPER_COREML=1 make -j
-
   # using CMake
   cmake -B build -DWHISPER_COREML=1
   cmake --build build -j --config Release
@@ -426,8 +424,8 @@ First, make sure you have installed `cuda`: https://developer.nvidia.com/cuda-do
 Now build `whisper.cpp` with CUDA support:
 
 ```
-make clean
-GGML_CUDA=1 make -j
+cmake -B build -DGGML_CUDA=1
+cmake --build build -j --config Release
 ```
 
 ## Vulkan GPU support
@@ -436,8 +434,8 @@ First, make sure your graphics card driver provides support for Vulkan API.
 
 Now build `whisper.cpp` with Vulkan support:
 ```
-make clean
-make GGML_VULKAN=1 -j
+cmake -B build -DGGML_VULKAN=1
+cmake --build build -j --config Release
 ```
 
 ## BLAS CPU support via OpenBLAS
@@ -448,28 +446,13 @@ First, make sure you have installed `openblas`: https://www.openblas.net/
 Now build `whisper.cpp` with OpenBLAS support:
 
 ```
-make clean
-GGML_OPENBLAS=1 make -j
-```
-
-## BLAS CPU support via Intel MKL
-
-Encoder processing can be accelerated on the CPU via the BLAS compatible interface of Intel's Math Kernel Library.
-First, make sure you have installed Intel's MKL runtime and development packages: https://www.intel.com/content/www/us/en/developer/tools/oneapi/onemkl-download.html
-
-Now build `whisper.cpp` with Intel MKL BLAS support:
-
-```
-source /opt/intel/oneapi/setvars.sh
-mkdir build
-cd build
-cmake -DWHISPER_MKL=ON ..
-WHISPER_MKL=1 make -j
+cmake -B build -DGGML_BLAS=1
+cmake --build build -j --config Release
 ```
 
 ## Ascend NPU support
 
-Ascend NPU provides inference acceleration via [`CANN`](https://www.hiascend.com/en/software/cann) and AI cores. 
+Ascend NPU provides inference acceleration via [`CANN`](https://www.hiascend.com/en/software/cann) and AI cores.
 
 First, check if your Ascend NPU device is supported:
 
@@ -483,10 +466,8 @@ Then, make sure you have installed [`CANN toolkit`](https://www.hiascend.com/en/
 Now build `whisper.cpp` with CANN support:
 
 ```
-mkdir build
-cd build
-cmake .. -D GGML_CANN=on
-make -j
+cmake -B build -DGGML_CANN=1
+cmake --build build -j --config Release
 ```
 
 Run the inference examples as usual, for example:
@@ -636,8 +617,9 @@ The [stream](examples/stream) tool samples the audio every half a second and run
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
 ```bash
-make stream -j
-./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
+cmake -B build
+cmake --build build --config Release
+./build/bin/stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
 
 https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a80f-28ba83be7d09.mp4
diff --git a/Sources/whisper/module.modulemap b/Sources/whisper/module.modulemap
new file mode 100644
index 00000000000..a266c0f2476
--- /dev/null
+++ b/Sources/whisper/module.modulemap
@@ -0,0 +1,5 @@
+module whisper [system] {
+    header "whisper.h"
+    link "whisper"
+    export *
+}
diff --git a/Sources/whisper/whisper.h b/Sources/whisper/whisper.h
new file mode 100644
index 00000000000..e503c8a38ad
--- /dev/null
+++ b/Sources/whisper/whisper.h
@@ -0,0 +1,4 @@
+#pragma once
+
+#include <whisper.h>
+
diff --git a/bindings/java/README.md b/bindings/java/README.md
index 5255612e398..568965493ed 100644
--- a/bindings/java/README.md
+++ b/bindings/java/README.md
@@ -67,5 +67,5 @@ copy /y ..\..\build\bin\Release\whisper.dll build\generated\resources\main\win32
 
 ## License
 
-The license for the Go bindings is the same as the license for the rest of the whisper.cpp project, which is the MIT License. See the `LICENSE` file for more details.
+The license for the Java bindings is the same as the license for the rest of the whisper.cpp project, which is the MIT License. See the `LICENSE` file for more details.
 
diff --git a/bindings/javascript/package.json b/bindings/javascript/package.json
index a66ba6b6bd9..9eca4c1d733 100644
--- a/bindings/javascript/package.json
+++ b/bindings/javascript/package.json
@@ -1,6 +1,6 @@
 {
   "name": "whisper.cpp",
-  "version": "1.7.2",
+  "version": "1.7.3",
   "description": "Whisper speech recognition",
   "main": "whisper.js",
   "scripts": {
diff --git a/bindings/ruby/.gitignore b/bindings/ruby/.gitignore
index e04a90a9c69..6e3b3be0e24 100644
--- a/bindings/ruby/.gitignore
+++ b/bindings/ruby/.gitignore
@@ -1,3 +1,5 @@
 LICENSE
 pkg/
-lib/whisper.*
+lib/whisper.so
+lib/whisper.bundle
+lib/whisper.dll
diff --git a/bindings/ruby/README.md b/bindings/ruby/README.md
index 05e2279f951..03a8b9e165c 100644
--- a/bindings/ruby/README.md
+++ b/bindings/ruby/README.md
@@ -22,7 +22,7 @@ Usage
 ```ruby
 require "whisper"
 
-whisper = Whisper::Context.new("path/to/model.bin")
+whisper = Whisper::Context.new("base")
 
 params = Whisper::Params.new
 params.language = "en"
@@ -41,21 +41,66 @@ end
 
 ### Preparing model ###
 
-Use script to download model file(s):
+Some models are prepared up-front:
 
-```bash
-git clone https://github.com/ggerganov/whisper.cpp.git
-cd whisper.cpp
-sh ./models/download-ggml-model.sh base.en
+```ruby
+base_en = Whisper::Model.pre_converted_models["base.en"]
+whisper = Whisper::Context.new(base_en)
+```
+
+At first time you use a model, it is downloaded automatically. After that, downloaded cached file is used. To clear cache, call `#clear_cache`:
+
+```ruby
+Whisper::Model.pre_converted_models["base"].clear_cache
+```
+
+You also can use shorthand for pre-converted models:
+
+```ruby
+whisper = Whisper::Context.new("base.en")
+```
+
+You can see the list of prepared model names by `Whisper::Model.preconverted_models.keys`:
+
+```ruby
+puts Whisper::Model.preconverted_model_names
+# tiny
+# tiny.en
+# tiny-q5_1
+# tiny.en-q5_1
+# tiny-q8_0
+# base
+# base.en
+# base-q5_1
+# base.en-q5_1
+# base-q8_0
+#   :
+#   :
+```
+
+You can also use local model files you prepared:
+
+```ruby
+whisper = Whisper::Context.new("path/to/your/model.bin")
+```
+
+Or, you can download model files:
+
+```ruby
+model_uri = Whisper::Model::URI.new("http://example.net/uri/of/your/model.bin")
+whisper = Whisper::Context.new(model_uri)
 ```
 
-There are some types of models. See [models][] page for details.
+See [models][] page for details.
 
 ### Preparing audio file ###
 
 Currently, whisper.cpp accepts only 16-bit WAV files.
 
-### API ###
+API
+---
+
+### Segments ###
 
 Once `Whisper::Context#transcribe` called, you can retrieve segments by `#each_segment`:
 
@@ -85,13 +130,6 @@ end
 You can also add hook to params called on new segment:
 
 ```ruby
-def format_time(time_ms)
-  sec, decimal_part = time_ms.divmod(1000)
-  min, sec = sec.divmod(60)
-  hour, min = min.divmod(60)
-  "%02d:%02d:%02d.%03d" % [hour, min, sec, decimal_part]
-end
-
 # Add hook before calling #transcribe
 params.on_new_segment do |segment|
   line = "[%{st} --> %{ed}] %{text}" % {
@@ -107,10 +145,12 @@ whisper.transcribe("path/to/audio.wav", params)
 
 ```
 
+### Models ###
+
 You can see model information:
 
 ```ruby
-whisper = Whisper::Context.new("path/to/model.bin")
+whisper = Whisper::Context.new("base")
 model = whisper.model
 
 model.n_vocab # => 51864
@@ -128,6 +168,8 @@ model.type # => "base"
 
 ```
 
+### Logging ###
+
 You can set log callback:
 
 ```ruby
@@ -157,9 +199,29 @@ Using this feature, you are also able to suppress log:
 Whisper.log_set ->(level, buffer, user_data) {
   # do nothing
 }, nil
-Whisper::Context.new(MODEL)
+Whisper::Context.new("base")
 ```
 
+### Low-level API to transcribe ###
+
+You can also call `Whisper::Context#full` and `#full_parallel` with a Ruby array as samples. Although `#transcribe` with audio file path is recommended because it extracts PCM samples in C++ and is fast, `#full` and `#full_parallel` give you flexibility.
+
+```ruby
+require "whisper"
+require "wavefile"
+
+reader = WaveFile::Reader.new("path/to/audio.wav", WaveFile::Format.new(:mono, :float, 16000))
+samples = reader.enum_for(:each_buffer).map(&:samples).flatten
+
+whisper = Whisper::Context.new("base")
+whisper.full(Whisper::Params.new, samples)
+whisper.each_segment do |segment|
+  puts segment.text
+end
+```
+
+The second argument `samples` may be an array, an object with `length` method, or a MemoryView. If you can prepare audio data as C array and export it as a MemoryView, whispercpp accepts and works with it with zero copy.
+
 License
 -------
 
diff --git a/bindings/ruby/Rakefile b/bindings/ruby/Rakefile
index d6fc49c8c09..3a7809b7505 100644
--- a/bindings/ruby/Rakefile
+++ b/bindings/ruby/Rakefile
@@ -1,39 +1,30 @@
 require 'rake/clean'
 require "bundler/gem_tasks"
-require "pathname"
-require "yaml"
 require "rake/testtask"
+require_relative "extsources"
 
-extsources = YAML.load_file("extsources.yaml")
 SOURCES = FileList[]
-extsources.each do |src|
+
+EXTSOURCES.each do |src|
   basename = src.pathmap("%f")
-  dest = basename == "LICENSE" ? basename : basename.pathmap("ext/%f")
+  dest = basename == "LICENSE" ? basename : src.pathmap("%{../..,ext}p")
+  dir = dest.pathmap("%d")
   file src
-  file dest => src do |t|
+  directory dir
+  file dest => [src, dir] do |t|
     cp t.source, t.name
   end
   SOURCES.include dest
 end
+
 CLEAN.include SOURCES
-CLEAN.include FileList[
-                "ext/*.o",
-                "ext/*.metal",
-                "ext/whisper.{so,bundle,dll}",
-                "ext/depend"
-              ]
+CLEAN.include FileList["ext/*.o", "ext/*.metal", "ext/whisper.{so,bundle,dll}"]
 
-task build: FileList[
-       "ext/Makefile",
-       "ext/ruby_whisper.h",
-       "ext/ruby_whisper.cpp",
-       "whispercpp.gemspec",
-     ]
+task build: ["ext/Makefile", "ext/ruby_whisper.h", "ext/ruby_whisper.cpp", "whispercpp.gemspec"]
 
 directory "pkg"
 CLOBBER.include "pkg"
 
-TEST_MODEL = "../../models/ggml-base.en.bin"
 LIB_NAME = "whisper".ext(RbConfig::CONFIG["DLEXT"])
 SO_FILE = File.join("ext", LIB_NAME)
 LIB_FILE = File.join("lib", LIB_NAME)
@@ -49,20 +40,25 @@ file SO_FILE => "ext/Makefile" do |t|
     sh "make"
   end
 end
-CLEAN.include LIB_FILE
+CLEAN.include SO_FILE
 
 directory "lib"
 file LIB_FILE => [SO_FILE, "lib"] do |t|
   copy t.source, t.name
 end
+CLEAN.include LIB_FILE
 
 Rake::TestTask.new do |t|
   t.test_files = FileList["tests/test_*.rb"]
 end
-task test: [TEST_MODEL, LIB_FILE]
 
-file TEST_MODEL do
-  Dir.chdir "../.." do
-    sh "./models/download-ggml-model.sh base.en"
+TEST_MEMORY_VIEW = "tests/jfk_reader/jfk_reader.#{RbConfig::CONFIG['DLEXT']}"
+file TEST_MEMORY_VIEW => "tests/jfk_reader/jfk_reader.c" do |t|
+  Dir.chdir "tests/jfk_reader" do
+    ruby "extconf.rb"
+    sh "make"
   end
 end
+CLEAN.include "tests/jfk_reader/jfk_reader.{o,#{RbConfig::CONFIG['DLEXT']}}"
+
+task test: [LIB_FILE, TEST_MEMORY_VIEW]
diff --git a/bindings/ruby/ext/.gitignore b/bindings/ruby/ext/.gitignore
index c9f31967840..e96a8584c94 100644
--- a/bindings/ruby/ext/.gitignore
+++ b/bindings/ruby/ext/.gitignore
@@ -1,35 +1,13 @@
 Makefile
-ggml.c
-ggml.h
-ggml-alloc.c
-ggml-alloc.h
-ggml-aarch64.c
-ggml-aarch64.h
-ggml-backend.cpp
-ggml-backend-impl.h
-ggml-backend.c
-ggml-backend.h
-ggml-common.h
-ggml-cpu-impl.h
-ggml-metal.m
-ggml-metal.metal
-ggml-metal-embed.metal
-ggml-blas.cpp
-ggml-cuda.h
-ggml-impl.h
-ggml-kompute.h
-ggml-metal.h
-ggml-opencl.h
-ggml-quants.c
-ggml-quants.h
-ggml-sycl.h
-ggml-vulkan.h
-ggml-blas.h
-get-flags.mk
-whisper.cpp
-whisper.h
-dr_wav.h
-depend
-whisper.bundle
 whisper.so
+whisper.bundle
 whisper.dll
+scripts/get-flags.mk
+*.o
+*.c
+*.cpp
+*.h
+*.m
+*.metal
+!ruby_whisper.cpp
+!ruby_whisper.h
diff --git a/bindings/ruby/ext/cpu.mk b/bindings/ruby/ext/cpu.mk
new file mode 100644
index 00000000000..a6aa237d7c0
--- /dev/null
+++ b/bindings/ruby/ext/cpu.mk
@@ -0,0 +1,9 @@
+ggml/src/ggml-cpu/ggml-cpu-cpp.o: \
+	ggml/src/ggml-cpu/ggml-cpu.cpp \
+	ggml/include/ggml-backend.h \
+	ggml/include/ggml.h \
+	ggml/include/ggml-alloc.h \
+	ggml/src/ggml-backend-impl.h \
+	ggml/include/ggml-cpu.h \
+	ggml/src/ggml-impl.h
+	$(CXX) $(CXXFLAGS)   -c $< -o $@
diff --git a/bindings/ruby/ext/extconf.rb b/bindings/ruby/ext/extconf.rb
index dd4db09db55..6ffac109e3f 100644
--- a/bindings/ruby/ext/extconf.rb
+++ b/bindings/ruby/ext/extconf.rb
@@ -1,7 +1,7 @@
 require 'mkmf'
 
 # need to use c++ compiler flags
-$CXXFLAGS << ' -std=c++11'
+$CXXFLAGS << ' -std=c++17'
 
 $LDFLAGS << ' -lstdc++'
 
@@ -35,10 +35,10 @@
   $GGML_METAL_EMBED_LIBRARY = true
 end
 
-$MK_CPPFLAGS = ''
+$MK_CPPFLAGS = '-Iggml/include -Iggml/src -Iggml/src/ggml-cpu -Iinclude -Isrc -Iexamples'
 $MK_CFLAGS   = '-std=c11   -fPIC'
-$MK_CXXFLAGS = '-std=c++11 -fPIC'
-$MK_NVCCFLAGS = '-std=c++11'
+$MK_CXXFLAGS = '-std=c++17 -fPIC'
+$MK_NVCCFLAGS = '-std=c++17'
 $MK_LDFLAGS = ''
 
 $OBJ_GGML = []
@@ -111,11 +111,6 @@
     $MK_CFLAGS     << ' -march=native -mtune=native'
     $HOST_CXXFLAGS << ' -march=native -mtune=native'
   end
-
-  if $UNAME_M.match? /aarch64.*/
-    $MK_CFLAGS   << ' -mcpu=native'
-    $MK_CXXFLAGS << ' -mcpu=native'
-  end
 else
   $MK_CFLAGS   << ' -march=rv64gcv -mabi=lp64d'
   $MK_CXXFLAGS << ' -march=rv64gcv -mabi=lp64d'
@@ -123,11 +118,11 @@
 
 unless ENV['GGML_NO_ACCELERATE']
   if $UNAME_S == 'Darwin'
-    $MK_CPPFLAGS << ' -DGGML_USE_ACCELERATE -DGGML_USE_BLAS'
+    $MK_CPPFLAGS << ' -DGGML_USE_ACCELERATE -DGGML_USE_BLAS -DGGML_BLAS_USE_ACCELERATE'
     $MK_CPPFLAGS << ' -DACCELERATE_NEW_LAPACK'
     $MK_CPPFLAGS << ' -DACCELERATE_LAPACK_ILP64'
     $MK_LDFLAGS  << ' -framework Accelerate'
-    $OBJ_GGML    << 'ggml-blas.o'
+    $OBJ_GGML    << 'ggml/src/ggml-blas/ggml-blas.o'
   end
 end
 
@@ -135,20 +130,20 @@
   $MK_CPPFLAGS << " -DGGML_USE_BLAS #{`pkg-config --cflags-only-I openblas`.chomp}"
   $MK_CFLAGS   << " #{`pkg-config --cflags-only-other openblas)`.chomp}"
   $MK_LDFLAGS  << " #{`pkg-config --libs openblas`}"
-  $OBJ_GGML    << 'ggml-blas.o'
+  $OBJ_GGML    << 'ggml/src/ggml-blas/ggml-blas.o'
 end
 
 if ENV['GGML_OPENBLAS64']
   $MK_CPPFLAGS << " -DGGML_USE_BLAS #{`pkg-config --cflags-only-I openblas64`.chomp}"
   $MK_CFLAGS   << " #{`pkg-config --cflags-only-other openblas64)`.chomp}"
   $MK_LDFLAGS  << " #{`pkg-config --libs openblas64`}"
-  $OBJ_GGML    << 'ggml-blas.o'
+  $OBJ_GGML    << 'ggml/src/ggml-blas/ggml-blas.o'
 end
 
 if $GGML_METAL
   $MK_CPPFLAGS << ' -DGGML_USE_METAL'
   $MK_LDFLAGS  << ' -framework Foundation -framework Metal -framework MetalKit'
-  $OBJ_GGML    << 'ggml-metal.o'
+  $OBJ_GGML    << 'ggml/src/ggml-metal/ggml-metal.o'
 
   if ENV['GGML_METAL_NDEBUG']
     $MK_CPPFLAGS << ' -DGGML_METAL_NDEBUG'
@@ -156,20 +151,27 @@
 
   if $GGML_METAL_EMBED_LIBRARY
     $MK_CPPFLAGS << ' -DGGML_METAL_EMBED_LIBRARY'
-    $OBJ_GGML    << 'ggml-metal-embed.o'
+    $OBJ_GGML    << 'ggml/src/ggml-metal/ggml-metal-embed.o'
   end
 end
 
 $OBJ_GGML <<
-  'ggml.o' <<
-  'ggml-cpu.o' <<
-  'ggml-alloc.o' <<
-  'ggml-backend.o' <<
-  'ggml-quants.o' <<
-  'ggml-aarch64.o'
+  'ggml/src/ggml.o' <<
+  'ggml/src/ggml-alloc.o' <<
+  'ggml/src/ggml-backend.o' <<
+  'ggml/src/ggml-backend-reg.o' <<
+  'ggml/src/ggml-opt.o' <<
+  'ggml/src/ggml-quants.o' <<
+  'ggml/src/ggml-threading.o' <<
+  'ggml/src/ggml-cpu/ggml-cpu.o' <<
+  'ggml/src/ggml-cpu/ggml-cpu-cpp.o' <<
+  'ggml/src/ggml-cpu/ggml-cpu-aarch64.o' <<
+  'ggml/src/ggml-cpu/ggml-cpu-hbm.o' <<
+  'ggml/src/ggml-cpu/ggml-cpu-quants.o' <<
+  'ggml/src/ggml-cpu/ggml-cpu-traits.o'
 
 $OBJ_WHISPER <<
-  'whisper.o'
+  'src/whisper.o'
 
 $objs = $OBJ_GGML + $OBJ_WHISPER + $OBJ_COMMON + $OBJ_SDL
 $objs << "ruby_whisper.o"
@@ -184,9 +186,12 @@
 create_makefile('whisper')
 
 File.open 'Makefile', 'a' do |file|
-  file.puts 'include get-flags.mk'
+  file.puts 'include scripts/get-flags.mk'
+  file.puts 'include cpu.mk'
 
   if $GGML_METAL
+    file.puts 'include metal.mk'
+
     if $GGML_METAL_EMBED_LIBRARY
       file.puts 'include metal-embed.mk'
     end
diff --git a/bindings/ruby/ext/metal-embed.mk b/bindings/ruby/ext/metal-embed.mk
index 478b8fd8a8e..cad86f87747 100644
--- a/bindings/ruby/ext/metal-embed.mk
+++ b/bindings/ruby/ext/metal-embed.mk
@@ -1,14 +1,17 @@
-ggml-metal-embed.o: \
-	ggml-metal.metal \
-	ggml-common.h
+ggml/src/ggml-metal/ggml-metal-embed.o: \
+	ggml/src/ggml-metal/ggml-metal.metal \
+	ggml/src/ggml-metal/ggml-metal-impl.h \
+	ggml/src/ggml-common.h
 	@echo "Embedding Metal library"
-	@sed -e '/#include "ggml-common.h"/r ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml-metal.metal > ggml-metal-embed.metal
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"ggml-metal-embed.metal\""          >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f ${TEMP_ASSEMBLY}
+	@sed -e '/__embed_ggml-common.h__/r      ggml/src/ggml-common.h'                -e '/__embed_ggml-common.h__/d'      < ggml/src/ggml-metal/ggml-metal.metal           > ggml/src/ggml-metal/ggml-metal-embed.metal.tmp
+	@sed -e '/#include "ggml-metal-impl.h"/r ggml/src/ggml-metal/ggml-metal-impl.h' -e '/#include "ggml-metal-impl.h"/d' < ggml/src/ggml-metal/ggml-metal-embed.metal.tmp > ggml/src/ggml-metal/ggml-metal-embed.metal
+	$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
+	@echo ".section __DATA, __ggml_metallib"                       >  $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_start"                            >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_start:"                                  >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".incbin \"ggml/src/ggml-metal/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo ".globl _ggml_metallib_end"                              >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	@echo "_ggml_metallib_end:"                                    >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
+	$(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@
+	@rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s
+	@rmdir ${TEMP_ASSEMBLY}
diff --git a/bindings/ruby/ext/metal.mk b/bindings/ruby/ext/metal.mk
new file mode 100644
index 00000000000..2b53cdefd7a
--- /dev/null
+++ b/bindings/ruby/ext/metal.mk
@@ -0,0 +1,6 @@
+ggml/src/ggml-metal/ggml-metal.o: \
+	ggml/src/ggml-metal/ggml-metal.m \
+	ggml/src/ggml-metal/ggml-metal-impl.h \
+	ggml/include/ggml-metal.h \
+	ggml/include/ggml.h
+	$(CC) $(CFLAGS) -c $< -o $@
diff --git a/bindings/ruby/ext/ruby_whisper.cpp b/bindings/ruby/ext/ruby_whisper.cpp
index 3f528ee4790..26e9def4b03 100644
--- a/bindings/ruby/ext/ruby_whisper.cpp
+++ b/bindings/ruby/ext/ruby_whisper.cpp
@@ -1,4 +1,5 @@
 #include <ruby.h>
+#include <ruby/memory_view.h>
 #include "ruby_whisper.h"
 #define DR_WAV_IMPLEMENTATION
 #include "dr_wav.h"
@@ -35,11 +36,20 @@ extern "C" {
 VALUE mWhisper;
 VALUE cContext;
 VALUE cParams;
+VALUE eError;
+
+VALUE cSegment;
+VALUE cModel;
 
 static ID id_to_s;
 static ID id_call;
 static ID id___method__;
 static ID id_to_enum;
+static ID id_length;
+static ID id_next;
+static ID id_new;
+static ID id_to_path;
+static ID id_pre_converted_models;
 
 static bool is_log_callback_finalized = false;
 
@@ -100,13 +110,13 @@ static VALUE ruby_whisper_s_finalize_log_callback(VALUE self, VALUE id) {
  *   log_set ->(level, buffer, user_data) { ... }, user_data -> nil
  */
 static VALUE ruby_whisper_s_log_set(VALUE self, VALUE log_callback, VALUE user_data) {
-  VALUE old_callback = rb_iv_get(self, "@log_callback");
+  VALUE old_callback = rb_iv_get(self, "log_callback");
   if (!NIL_P(old_callback)) {
     rb_undefine_finalizer(old_callback);
   }
 
-  rb_iv_set(self, "@log_callback", log_callback);
-  rb_iv_set(self, "@user_data", user_data);
+  rb_iv_set(self, "log_callback", log_callback);
+  rb_iv_set(self, "user_data", user_data);
 
   VALUE finalize_log_callback = rb_funcall(mWhisper, rb_intern("method"), 1, rb_str_new2("finalize_log_callback"));
   rb_define_finalizer(log_callback, finalize_log_callback);
@@ -115,8 +125,8 @@ static VALUE ruby_whisper_s_log_set(VALUE self, VALUE log_callback, VALUE user_d
     if (is_log_callback_finalized) {
       return;
     }
-    VALUE log_callback = rb_iv_get(mWhisper, "@log_callback");
-    VALUE udata = rb_iv_get(mWhisper, "@user_data");
+    VALUE log_callback = rb_iv_get(mWhisper, "log_callback");
+    VALUE udata = rb_iv_get(mWhisper, "user_data");
     rb_funcall(log_callback, id_call, 3, INT2NUM(level), rb_str_new2(buffer), udata);
   }, nullptr);
 
@@ -181,6 +191,7 @@ static VALUE ruby_whisper_params_allocate(VALUE klass) {
   ruby_whisper_params *rwp;
   rwp = ALLOC(ruby_whisper_params);
   rwp->params = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+  rwp->diarize = false;
   rwp->new_segment_callback_container = rb_whisper_callback_container_allocate();
   rwp->progress_callback_container = rb_whisper_callback_container_allocate();
   rwp->abort_callback_container = rb_whisper_callback_container_allocate();
@@ -189,7 +200,9 @@ static VALUE ruby_whisper_params_allocate(VALUE klass) {
 
 /*
  * call-seq:
+ *   new("base.en") -> Whisper::Context
  *   new("path/to/model.bin") -> Whisper::Context
+ *   new(Whisper::Model::URI.new("https://example.net/uri/of/model.bin")) -> Whisper::Context
  */
 static VALUE ruby_whisper_initialize(int argc, VALUE *argv, VALUE self) {
   ruby_whisper *rw;
@@ -199,6 +212,14 @@ static VALUE ruby_whisper_initialize(int argc, VALUE *argv, VALUE self) {
   rb_scan_args(argc, argv, "01", &whisper_model_file_path);
   Data_Get_Struct(self, ruby_whisper, rw);
 
+  VALUE pre_converted_models = rb_funcall(cModel, id_pre_converted_models, 0);
+  VALUE pre_converted_model = rb_hash_aref(pre_converted_models, whisper_model_file_path);
+  if (!NIL_P(pre_converted_model)) {
+    whisper_model_file_path = pre_converted_model;
+  }
+  if (rb_respond_to(whisper_model_file_path, id_to_path)) {
+    whisper_model_file_path = rb_funcall(whisper_model_file_path, id_to_path, 0);
+  }
   if (!rb_respond_to(whisper_model_file_path, id_to_s)) {
     rb_raise(rb_eRuntimeError, "Expected file path to model to initialize Whisper::Context");
   }
@@ -544,6 +565,168 @@ VALUE ruby_whisper_model_type(VALUE self) {
   return rb_str_new2(whisper_model_type_readable(rw->context));
 }
 
+/*
+ * Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
+ * Not thread safe for same context
+ * Uses the specified decoding strategy to obtain the text.
+ *
+ * call-seq:
+ *   full(params, samples, n_samples) -> nil
+ *   full(params, samples) -> nil
+ *
+ * The second argument +samples+ must be an array of samples, respond to :length, or be a MemoryView of an array of float. It must be 32 bit float PCM audio data.
+ */
+VALUE ruby_whisper_full(int argc, VALUE *argv, VALUE self) {
+  if (argc < 2 || argc > 3) {
+    rb_raise(rb_eArgError, "wrong number of arguments (given %d, expected 2..3)", argc);
+  }
+
+  ruby_whisper *rw;
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  VALUE params = argv[0];
+  Data_Get_Struct(params, ruby_whisper_params, rwp);
+  VALUE samples = argv[1];
+  int n_samples;
+  rb_memory_view_t view;
+  const bool memory_view_available_p = rb_memory_view_available_p(samples);
+  if (argc == 3) {
+    n_samples = NUM2INT(argv[2]);
+    if (TYPE(samples) == T_ARRAY) {
+      if (RARRAY_LEN(samples) < n_samples) {
+        rb_raise(rb_eArgError, "samples length %ld is less than n_samples %d", RARRAY_LEN(samples), n_samples);
+      }
+    }
+    // Should check when samples.respond_to?(:length)?
+  } else {
+    if (TYPE(samples) == T_ARRAY) {
+      n_samples = RARRAY_LEN(samples);
+    } else if (memory_view_available_p) {
+      if (!rb_memory_view_get(samples, &view, RUBY_MEMORY_VIEW_SIMPLE)) {
+        view.obj = Qnil;
+        rb_raise(rb_eArgError, "unable to get a memory view");
+      }
+      n_samples = view.byte_size / view.item_size;
+    } else if (rb_respond_to(samples, id_length)) {
+      n_samples = NUM2INT(rb_funcall(samples, id_length, 0));
+    } else {
+      rb_raise(rb_eArgError, "samples must respond to :length or be a MemoryView of an array of flaot when n_samples is not given");
+    }
+  }
+  float * c_samples = (float *)malloc(n_samples * sizeof(float));
+  if (memory_view_available_p)  {
+    c_samples = (float *)view.data;
+  } else {
+    if (TYPE(samples) == T_ARRAY) {
+      for (int i = 0; i < n_samples; i++) {
+        c_samples[i] = RFLOAT_VALUE(rb_ary_entry(samples, i));
+      }
+    } else {
+      // TODO: use rb_block_call
+      VALUE iter = rb_funcall(samples, id_to_enum, 1, rb_str_new2("each"));
+      for (int i = 0; i < n_samples; i++) {
+        // TODO: check if iter is exhausted and raise ArgumentError appropriately
+        VALUE sample = rb_funcall(iter, id_next, 0);
+        c_samples[i] = RFLOAT_VALUE(sample);
+      }
+    }
+  }
+  const int result = whisper_full(rw->context, rwp->params, c_samples, n_samples);
+  if (0 == result) {
+    return Qnil;
+  } else {
+    rb_exc_raise(rb_funcall(eError, id_new, 1, result));
+  }
+}
+
+/*
+ * Split the input audio in chunks and process each chunk separately using whisper_full_with_state()
+ * Result is stored in the default state of the context
+ * Not thread safe if executed in parallel on the same context.
+ * It seems this approach can offer some speedup in some cases.
+ * However, the transcription accuracy can be worse at the beginning and end of each chunk.
+ *
+ * call-seq:
+ *   full_parallel(params, samples) -> nil
+ *   full_parallel(params, samples, n_samples) -> nil
+ *   full_parallel(params, samples, n_samples, n_processors) -> nil
+ *   full_parallel(params, samples, nil, n_processors) -> nil
+ */
+static VALUE ruby_whisper_full_parallel(int argc, VALUE *argv,VALUE self) {
+  if (argc < 2 || argc > 4) {
+    rb_raise(rb_eArgError, "wrong number of arguments (given %d, expected 2..3)", argc);
+  }
+
+  ruby_whisper *rw;
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper, rw);
+  VALUE params = argv[0];
+  Data_Get_Struct(params, ruby_whisper_params, rwp);
+  VALUE samples = argv[1];
+  int n_samples;
+  int n_processors;
+  rb_memory_view_t view;
+  const bool memory_view_available_p = rb_memory_view_available_p(samples);
+  switch (argc) {
+  case 2:
+    n_processors = 1;
+    break;
+  case 3:
+    n_processors = 1;
+    break;
+  case 4:
+    n_processors = NUM2INT(argv[3]);
+    break;
+  }
+  if (argc >= 3 && !NIL_P(argv[2])) {
+    n_samples = NUM2INT(argv[2]);
+    if (TYPE(samples) == T_ARRAY) {
+      if (RARRAY_LEN(samples) < n_samples) {
+        rb_raise(rb_eArgError, "samples length %ld is less than n_samples %d", RARRAY_LEN(samples), n_samples);
+      }
+    }
+    // Should check when samples.respond_to?(:length)?
+  } else if (memory_view_available_p) {
+    if (!rb_memory_view_get(samples, &view, RUBY_MEMORY_VIEW_SIMPLE)) {
+      view.obj = Qnil;
+      rb_raise(rb_eArgError, "unable to get a memory view");
+    }
+    n_samples = view.byte_size / view.item_size;
+  } else {
+    if (TYPE(samples) == T_ARRAY) {
+      n_samples = RARRAY_LEN(samples);
+    } else if (rb_respond_to(samples, id_length)) {
+      n_samples = NUM2INT(rb_funcall(samples, id_length, 0));
+    } else {
+      rb_raise(rb_eArgError, "samples must respond to :length or be a MemoryView of an array of flaot when n_samples is not given");
+    }
+  }
+  float * c_samples = (float *)malloc(n_samples * sizeof(float));
+  if (memory_view_available_p) {
+    c_samples = (float *)view.data;
+  } else {
+    if (TYPE(samples) == T_ARRAY) {
+      for (int i = 0; i < n_samples; i++) {
+        c_samples[i] = RFLOAT_VALUE(rb_ary_entry(samples, i));
+      }
+    } else {
+      // FIXME: use rb_block_call
+      VALUE iter = rb_funcall(samples, id_to_enum, 1, rb_str_new2("each"));
+      for (int i = 0; i < n_samples; i++) {
+        // TODO: check if iter is exhausted and raise ArgumentError
+        VALUE sample = rb_funcall(iter, id_next, 0);
+        c_samples[i] = RFLOAT_VALUE(sample);
+      }
+    }
+  }
+  const int result = whisper_full_parallel(rw->context, rwp->params, c_samples, n_samples, n_processors);
+  if (0 == result) {
+    return Qnil;
+  } else {
+    rb_exc_raise(rb_funcall(eError, id_new, 1, result));
+  }
+}
+
 /*
  * Number of segments.
  *
@@ -1078,6 +1261,25 @@ static VALUE ruby_whisper_params_set_logprob_thold(VALUE self, VALUE value) {
   rwp->params.logprob_thold = RFLOAT_VALUE(value);
   return value;
 }
+/*
+ * call-seq:
+ *   no_speech_thold -> Float
+ */
+static VALUE ruby_whisper_params_get_no_speech_thold(VALUE self) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  return DBL2NUM(rwp->params.no_speech_thold);
+}
+/*
+ * call-seq:
+ *   no_speech_thold = threshold -> threshold
+ */
+static VALUE ruby_whisper_params_set_no_speech_thold(VALUE self, VALUE value) {
+  ruby_whisper_params *rwp;
+  Data_Get_Struct(self, ruby_whisper_params, rwp);
+  rwp->params.no_speech_thold = RFLOAT_VALUE(value);
+  return value;
+}
 /*
  * Sets new segment callback, called for every newly generated text segment.
  *
@@ -1174,9 +1376,6 @@ typedef struct {
   VALUE context;
 } ruby_whisper_model;
 
-VALUE cSegment;
-VALUE cModel;
-
 static void rb_whisper_segment_mark(ruby_whisper_segment *rws) {
   rb_gc_mark(rws->context);
 }
@@ -1518,15 +1717,61 @@ static VALUE ruby_whisper_c_model_type(VALUE self) {
   return rb_str_new2(whisper_model_type_readable(rw->context));
 }
 
+static VALUE ruby_whisper_error_initialize(VALUE self, VALUE code) {
+  const int c_code = NUM2INT(code);
+  char *raw_message;
+  switch (c_code) {
+  case -2:
+    raw_message = "failed to compute log mel spectrogram";
+    break;
+  case -3:
+    raw_message = "failed to auto-detect language";
+    break;
+  case -4:
+    raw_message = "too many decoders requested";
+    break;
+  case -5:
+    raw_message = "audio_ctx is larger than the maximum allowed";
+    break;
+  case -6:
+    raw_message = "failed to encode";
+    break;
+  case -7:
+    raw_message = "whisper_kv_cache_init() failed for self-attention cache";
+    break;
+  case -8:
+    raw_message = "failed to decode";
+    break;
+  case -9:
+    raw_message = "failed to decode";
+    break;
+  default:
+    raw_message = "unknown error";
+    break;
+  }
+  const VALUE message = rb_str_new2(raw_message);
+  rb_call_super(1, &message);
+  rb_iv_set(self, "@code", code);
+
+  return self;
+}
+
+
 void Init_whisper() {
   id_to_s = rb_intern("to_s");
   id_call = rb_intern("call");
   id___method__ = rb_intern("__method__");
   id_to_enum = rb_intern("to_enum");
+  id_length = rb_intern("length");
+  id_next = rb_intern("next");
+  id_new = rb_intern("new");
+  id_to_path = rb_intern("to_path");
+  id_pre_converted_models = rb_intern("pre_converted_models");
 
   mWhisper = rb_define_module("Whisper");
   cContext = rb_define_class_under(mWhisper, "Context", rb_cObject);
   cParams  = rb_define_class_under(mWhisper, "Params", rb_cObject);
+  eError   = rb_define_class_under(mWhisper, "Error", rb_eStandardError);
 
   rb_define_const(mWhisper, "LOG_LEVEL_NONE", INT2NUM(GGML_LOG_LEVEL_NONE));
   rb_define_const(mWhisper, "LOG_LEVEL_INFO", INT2NUM(GGML_LOG_LEVEL_INFO));
@@ -1564,6 +1809,8 @@ void Init_whisper() {
   rb_define_method(cContext, "full_get_segment_t1", ruby_whisper_full_get_segment_t1, 1);
   rb_define_method(cContext, "full_get_segment_speaker_turn_next", ruby_whisper_full_get_segment_speaker_turn_next, 1);
   rb_define_method(cContext, "full_get_segment_text", ruby_whisper_full_get_segment_text, 1);
+  rb_define_method(cContext, "full", ruby_whisper_full, -1);
+  rb_define_method(cContext, "full_parallel", ruby_whisper_full_parallel, -1);
 
   rb_define_alloc_func(cParams, ruby_whisper_params_allocate);
 
@@ -1615,6 +1862,8 @@ void Init_whisper() {
   rb_define_method(cParams, "entropy_thold=", ruby_whisper_params_set_entropy_thold, 1);
   rb_define_method(cParams, "logprob_thold", ruby_whisper_params_get_logprob_thold, 0);
   rb_define_method(cParams, "logprob_thold=", ruby_whisper_params_set_logprob_thold, 1);
+  rb_define_method(cParams, "no_speech_thold", ruby_whisper_params_get_no_speech_thold, 0);
+  rb_define_method(cParams, "no_speech_thold=", ruby_whisper_params_set_no_speech_thold, 1);
 
   rb_define_method(cParams, "new_segment_callback=", ruby_whisper_params_set_new_segment_callback, 1);
   rb_define_method(cParams, "new_segment_callback_user_data=", ruby_whisper_params_set_new_segment_callback_user_data, 1);
@@ -1623,6 +1872,9 @@ void Init_whisper() {
   rb_define_method(cParams, "abort_callback=", ruby_whisper_params_set_abort_callback, 1);
   rb_define_method(cParams, "abort_callback_user_data=", ruby_whisper_params_set_abort_callback_user_data, 1);
 
+  rb_define_attr(eError, "code", true, false);
+  rb_define_method(eError, "initialize", ruby_whisper_error_initialize, 1);
+
   // High leve
   cSegment  = rb_define_class_under(mWhisper, "Segment", rb_cObject);
 
diff --git a/bindings/ruby/extsources.rb b/bindings/ruby/extsources.rb
new file mode 100644
index 00000000000..1dc900d41b7
--- /dev/null
+++ b/bindings/ruby/extsources.rb
@@ -0,0 +1,6 @@
+require "yaml"
+
+sources = `git ls-files -z ../..`.split("\x0")
+paths = YAML.load_file("../../.github/workflows/bindings-ruby.yml")[true]["push"]["paths"]
+paths.delete "bindings/ruby/**"
+EXTSOURCES = (Dir.glob(paths, base: "../..").collect {|path| "../../#{path}"} << "../../LICENSE") & sources
diff --git a/bindings/ruby/extsources.yaml b/bindings/ruby/extsources.yaml
deleted file mode 100644
index 10c2a563826..00000000000
--- a/bindings/ruby/extsources.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
----
-- ../../src/whisper.cpp
-- ../../include/whisper.h
-- ../../ggml/src/ggml.c
-- ../../ggml/src/ggml-cpu.c
-- ../../ggml/src/ggml-impl.h
-- ../../ggml/src/ggml-aarch64.h
-- ../../ggml/src/ggml-aarch64.c
-- ../../ggml/src/ggml-alloc.c
-- ../../ggml/src/ggml-backend-impl.h
-- ../../ggml/src/ggml-backend.cpp
-- ../../ggml/src/ggml-common.h
-- ../../ggml/src/ggml-quants.h
-- ../../ggml/src/ggml-quants.c
-- ../../ggml/src/ggml-cpu-impl.h
-- ../../ggml/src/ggml-metal.m
-- ../../ggml/src/ggml-metal.metal
-- ../../ggml/src/ggml-blas.cpp
-- ../../ggml/include/ggml.h
-- ../../ggml/include/ggml-alloc.h
-- ../../ggml/include/ggml-backend.h
-- ../../ggml/include/ggml-cpu.h
-- ../../ggml/include/ggml-cuda.h
-- ../../ggml/include/ggml-kompute.h
-- ../../ggml/include/ggml-metal.h
-- ../../ggml/include/ggml-sycl.h
-- ../../ggml/include/ggml-vulkan.h
-- ../../ggml/include/ggml-blas.h
-- ../../scripts/get-flags.mk
-- ../../examples/dr_wav.h
-- ../../LICENSE
diff --git a/bindings/ruby/lib/whisper.rb b/bindings/ruby/lib/whisper.rb
new file mode 100644
index 00000000000..3a0b844e15c
--- /dev/null
+++ b/bindings/ruby/lib/whisper.rb
@@ -0,0 +1,2 @@
+require "whisper.so"
+require "whisper/model/uri"
diff --git a/bindings/ruby/lib/whisper/model/uri.rb b/bindings/ruby/lib/whisper/model/uri.rb
new file mode 100644
index 00000000000..5ca77ed4bb2
--- /dev/null
+++ b/bindings/ruby/lib/whisper/model/uri.rb
@@ -0,0 +1,157 @@
+require "whisper.so"
+require "uri"
+require "net/http"
+require "time"
+require "pathname"
+require "io/console/size"
+
+class Whisper::Model
+  class URI
+    def initialize(uri)
+      @uri = URI(uri)
+    end
+
+    def to_path
+      cache
+      cache_path.to_path
+    end
+
+    def clear_cache
+      path = cache_path
+      path.delete if path.exist?
+    end
+
+    private
+
+    def cache_path
+      base_cache_dir/@uri.host/@uri.path[1..]
+    end
+
+    def base_cache_dir
+      base = case RUBY_PLATFORM
+             when /mswin|mingw/
+               ENV.key?("LOCALAPPDATA") ? Pathname(ENV["LOCALAPPDATA"]) : Pathname(Dir.home)/"AppData/Local"
+             when /darwin/
+               Pathname(Dir.home)/"Library/Caches"
+             else
+               ENV.key?("XDG_CACHE_HOME") ? ENV["XDG_CACHE_HOME"] : Pathname(Dir.home)/".cache"
+             end
+      base/"whisper.cpp"
+    end
+
+    def cache
+      path = cache_path
+      headers = {}
+      headers["if-modified-since"] = path.mtime.httpdate if path.exist?
+      request @uri, headers
+      path
+    end
+
+    def request(uri, headers)
+      Net::HTTP.start uri.host, uri.port, use_ssl: uri.scheme == "https" do |http|
+        request = Net::HTTP::Get.new(uri, headers)
+        http.request request do |response|
+          case response
+          when Net::HTTPNotModified
+            # noop
+          when Net::HTTPOK
+            download response
+          when Net::HTTPRedirection
+            request URI(response["location"]), headers
+          else
+            return if headers.key?("if-modified-since") # Use cache file
+
+            raise "#{response.code} #{response.message}\n#{response.body}"
+          end
+        end
+      end
+    end
+
+    def download(response)
+      path = cache_path
+      path.dirname.mkpath unless path.dirname.exist?
+      downloading_path = Pathname("#{path}.downloading")
+      size = response.content_length
+      downloading_path.open "wb" do |file|
+        downloaded = 0
+        response.read_body do |chunk|
+          file << chunk
+          downloaded += chunk.bytesize
+          show_progress downloaded, size
+        end
+      end
+      downloading_path.rename path
+    end
+
+    def show_progress(current, size)
+      return unless $stderr.tty?
+      return unless size
+
+      unless @prev
+        @prev = Time.now
+        $stderr.puts "Downloading #{@uri}"
+      end
+
+      now = Time.now
+      return if now - @prev < 1 && current < size
+
+      progress_width = 20
+      progress = current.to_f / size
+      arrow_length = progress * progress_width
+      arrow = "=" * (arrow_length - 1) + ">" + " " * (progress_width - arrow_length)
+      line = "[#{arrow}] (#{format_bytesize(current)} / #{format_bytesize(size)})"
+      padding = ' ' * ($stderr.winsize[1] - line.size)
+      $stderr.print "\r#{line}#{padding}"
+      $stderr.puts if current >= size
+      @prev = now
+    end
+
+    def format_bytesize(bytesize)
+      return "0.0 B" if bytesize.zero?
+
+      units = %w[B KiB MiB GiB TiB]
+      exp = (Math.log(bytesize) / Math.log(1024)).to_i
+      format("%.1f %s", bytesize.to_f / 1024 ** exp, units[exp])
+    end
+  end
+
+  @pre_converted_models = {}
+  %w[
+    tiny
+    tiny.en
+    tiny-q5_1
+    tiny.en-q5_1
+    tiny-q8_0
+    base
+    base.en
+    base-q5_1
+    base.en-q5_1
+    base-q8_0
+    small
+    small.en
+    small.en-tdrz
+    small-q5_1
+    small.en-q5_1
+    small-q8_0
+    medium
+    medium.en
+    medium-q5_0
+    medium.en-q5_0
+    medium-q8_0
+    large-v1
+    large-v2
+    large-v2-q5_0
+    large-v2-q8_0
+    large-v3
+    large-v3-q5_0
+    large-v3-turbo
+    large-v3-turbo-q5_0
+    large-v3-turbo-q8_0
+  ].each do |name|
+    @pre_converted_models[name] = URI.new("https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-#{name}.bin")
+  end
+
+  class << self
+    attr_reader :pre_converted_models
+  end
+end
diff --git a/bindings/ruby/tests/helper.rb b/bindings/ruby/tests/helper.rb
index 4172ebc517f..da52f268731 100644
--- a/bindings/ruby/tests/helper.rb
+++ b/bindings/ruby/tests/helper.rb
@@ -1,7 +1,7 @@
 require "test/unit"
 require "whisper"
+require_relative "jfk_reader/jfk_reader"
 
 class TestBase < Test::Unit::TestCase
-  MODEL = File.join(__dir__, "..", "..", "..", "models", "ggml-base.en.bin")
   AUDIO = File.join(__dir__, "..", "..", "..", "samples", "jfk.wav")
 end
diff --git a/bindings/ruby/tests/jfk_reader/.gitignore b/bindings/ruby/tests/jfk_reader/.gitignore
new file mode 100644
index 00000000000..656da8d939c
--- /dev/null
+++ b/bindings/ruby/tests/jfk_reader/.gitignore
@@ -0,0 +1,5 @@
+Makefile
+jfk_reader.o
+jfk_reader.so
+jfk_reader.bundle
+jfk_reader.dll
diff --git a/bindings/ruby/tests/jfk_reader/extconf.rb b/bindings/ruby/tests/jfk_reader/extconf.rb
new file mode 100644
index 00000000000..0d842d03349
--- /dev/null
+++ b/bindings/ruby/tests/jfk_reader/extconf.rb
@@ -0,0 +1,3 @@
+require "mkmf"
+
+create_makefile("jfk_reader")
diff --git a/bindings/ruby/tests/jfk_reader/jfk_reader.c b/bindings/ruby/tests/jfk_reader/jfk_reader.c
new file mode 100644
index 00000000000..6657176e767
--- /dev/null
+++ b/bindings/ruby/tests/jfk_reader/jfk_reader.c
@@ -0,0 +1,68 @@
+#include <ruby.h>
+#include <ruby/memory_view.h>
+#include <ruby/encoding.h>
+
+static VALUE
+jfk_reader_initialize(VALUE self, VALUE audio_path)
+{
+  rb_iv_set(self, "audio_path", audio_path);
+  return Qnil;
+}
+
+static bool
+jfk_reader_get_memory_view(const VALUE obj, rb_memory_view_t *view, int flags)
+{
+  VALUE audio_path = rb_iv_get(obj, "audio_path");
+  const char *audio_path_str = StringValueCStr(audio_path);
+  const int n_samples = 176000;
+  float *data = (float *)malloc(n_samples * sizeof(float));
+  short *samples = (short *)malloc(n_samples * sizeof(short));
+  FILE *file = fopen(audio_path_str, "rb");
+
+  fseek(file, 78, SEEK_SET);
+  fread(samples, sizeof(short), n_samples, file);
+  fclose(file);
+  for (int i = 0; i < n_samples; i++) {
+    data[i] = samples[i]/32768.0;
+  }
+
+  view->obj = obj;
+  view->data = (void *)data;
+  view->byte_size = sizeof(float) * n_samples;
+  view->readonly = true;
+  view->format = "f";
+  view->item_size = sizeof(float);
+  view->item_desc.components = NULL;
+  view->item_desc.length = 0;
+  view->ndim = 1;
+  view->shape = NULL;
+  view->sub_offsets = NULL;
+  view->private_data = NULL;
+
+  return true;
+}
+
+static bool
+jfk_reader_release_memory_view(const VALUE obj, rb_memory_view_t *view)
+{
+  return true;
+}
+
+static bool
+jfk_reader_memory_view_available_p(const VALUE obj)
+{
+  return true;
+}
+
+static const rb_memory_view_entry_t jfk_reader_view_entry = {
+  jfk_reader_get_memory_view,
+  jfk_reader_release_memory_view,
+  jfk_reader_memory_view_available_p
+};
+
+void Init_jfk_reader(void)
+{
+  VALUE cJFKReader = rb_define_class("JFKReader", rb_cObject);
+  rb_memory_view_register(cJFKReader, &jfk_reader_view_entry);
+  rb_define_method(cJFKReader, "initialize", jfk_reader_initialize, 1);
+}
diff --git a/bindings/ruby/tests/test_callback.rb b/bindings/ruby/tests/test_callback.rb
index 1234d31db7a..8e9becf7ef7 100644
--- a/bindings/ruby/tests/test_callback.rb
+++ b/bindings/ruby/tests/test_callback.rb
@@ -1,14 +1,11 @@
-require "test/unit"
-require "whisper"
-
-class TestCallback < Test::Unit::TestCase
-  TOPDIR = File.expand_path(File.join(File.dirname(__FILE__), '..'))
+require_relative "helper"
 
+class TestCallback < TestBase
   def setup
     GC.start
     @params = Whisper::Params.new
-    @whisper = Whisper::Context.new(File.join(TOPDIR, '..', '..', 'models', 'ggml-base.en.bin'))
-    @audio = File.join(TOPDIR, '..', '..', 'samples', 'jfk.wav')
+    @whisper = Whisper::Context.new("base.en")
+    @audio = File.join(AUDIO)
   end
 
   def test_new_segment_callback
diff --git a/bindings/ruby/tests/test_error.rb b/bindings/ruby/tests/test_error.rb
new file mode 100644
index 00000000000..cb92c324fd1
--- /dev/null
+++ b/bindings/ruby/tests/test_error.rb
@@ -0,0 +1,20 @@
+require_relative "helper"
+
+class TestError < TestBase
+  def test_error
+    error = Whisper::Error.new(-2)
+    assert_equal "failed to compute log mel spectrogram", error.message
+    assert_equal -2, error.code
+  end
+
+  def test_unknown_error
+    error = Whisper::Error.new(-20)
+    assert_equal "unknown error", error.message
+  end
+
+  def test_non_int_code
+    assert_raise TypeError do
+      error = Whisper::Error.new("non int")
+    end
+  end
+end
diff --git a/bindings/ruby/tests/test_model.rb b/bindings/ruby/tests/test_model.rb
index 2310522a644..1362fc469bf 100644
--- a/bindings/ruby/tests/test_model.rb
+++ b/bindings/ruby/tests/test_model.rb
@@ -1,13 +1,14 @@
 require_relative "helper"
+require "pathname"
 
 class TestModel < TestBase
   def test_model
-    whisper = Whisper::Context.new(MODEL)
+    whisper = Whisper::Context.new("base.en")
     assert_instance_of Whisper::Model, whisper.model
   end
 
   def test_attributes
-    whisper = Whisper::Context.new(MODEL)
+    whisper = Whisper::Context.new("base.en")
     model = whisper.model
 
     assert_equal 51864, model.n_vocab
@@ -25,7 +26,7 @@ def test_attributes
   end
 
   def test_gc
-    model = Whisper::Context.new(MODEL).model
+    model = Whisper::Context.new("base.en").model
     GC.start
 
     assert_equal 51864, model.n_vocab
@@ -41,4 +42,30 @@ def test_gc
     assert_equal 1, model.ftype
     assert_equal "base", model.type
   end
+
+  def test_pathname
+    path = Pathname(Whisper::Model.pre_converted_models["base.en"].to_path)
+    whisper = Whisper::Context.new(path)
+    model = whisper.model
+
+    assert_equal 51864, model.n_vocab
+    assert_equal 1500, model.n_audio_ctx
+    assert_equal 512, model.n_audio_state
+    assert_equal 8, model.n_audio_head
+    assert_equal 6, model.n_audio_layer
+    assert_equal 448, model.n_text_ctx
+    assert_equal 512, model.n_text_state
+    assert_equal 8, model.n_text_head
+    assert_equal 6, model.n_text_layer
+    assert_equal 80, model.n_mels
+    assert_equal 1, model.ftype
+    assert_equal "base", model.type
+  end
+
+  def test_auto_download
+    path = Whisper::Model.pre_converted_models["base.en"].to_path
+
+    assert_path_exist path
+    assert_equal 147964211, File.size(path)
+  end
 end
diff --git a/bindings/ruby/tests/test_params.rb b/bindings/ruby/tests/test_params.rb
index bf73fd6b29b..d2667ef0d3e 100644
--- a/bindings/ruby/tests/test_params.rb
+++ b/bindings/ruby/tests/test_params.rb
@@ -151,4 +151,10 @@ def test_logprob_thold
     @params.logprob_thold = -0.5
     assert_in_delta -0.5, @params.logprob_thold
   end
+
+  def test_no_speech_thold
+    assert_in_delta 0.6, @params.no_speech_thold
+    @params.no_speech_thold = 0.2
+    assert_in_delta 0.2, @params.no_speech_thold
+  end
 end
diff --git a/bindings/ruby/tests/test_segment.rb b/bindings/ruby/tests/test_segment.rb
index 8129ae5db42..559bcea71e6 100644
--- a/bindings/ruby/tests/test_segment.rb
+++ b/bindings/ruby/tests/test_segment.rb
@@ -5,7 +5,7 @@ class << self
     attr_reader :whisper
 
     def startup
-      @whisper = Whisper::Context.new(TestBase::MODEL)
+      @whisper = Whisper::Context.new("base.en")
       params = Whisper::Params.new
       params.print_timestamps = false
       @whisper.transcribe(TestBase::AUDIO, params)
diff --git a/bindings/ruby/tests/test_whisper.rb b/bindings/ruby/tests/test_whisper.rb
index e37e24c64c4..115569edb3d 100644
--- a/bindings/ruby/tests/test_whisper.rb
+++ b/bindings/ruby/tests/test_whisper.rb
@@ -1,5 +1,6 @@
 require_relative "helper"
 require "stringio"
+require "etc"
 
 # Exists to detect memory-related bug
 Whisper.log_set ->(level, buffer, user_data) {}, nil
@@ -10,7 +11,7 @@ def setup
   end
 
   def test_whisper
-    @whisper = Whisper::Context.new(MODEL)
+    @whisper = Whisper::Context.new("base.en")
     params  = Whisper::Params.new
     params.print_timestamps = false
 
@@ -24,7 +25,7 @@ class << self
       attr_reader :whisper
 
       def startup
-        @whisper = Whisper::Context.new(TestBase::MODEL)
+        @whisper = Whisper::Context.new("base.en")
         params = Whisper::Params.new
         params.print_timestamps = false
         @whisper.transcribe(TestBase::AUDIO, params)
@@ -103,11 +104,11 @@ def test_log_set
       logs << [level, buffer, udata]
     }
     Whisper.log_set log_callback, user_data
-    Whisper::Context.new(MODEL)
+    Whisper::Context.new("base.en")
 
     assert logs.length > 30
     logs.each do |log|
-      assert_equal Whisper::LOG_LEVEL_INFO, log[0]
+      assert_include [Whisper::LOG_LEVEL_DEBUG, Whisper::LOG_LEVEL_INFO, Whisper::LOG_LEVEL_WARN], log[0]
       assert_same user_data, log[2]
     end
   end
@@ -119,9 +120,107 @@ def test_log_suppress
     }, nil
     dev = StringIO.new("")
     $stderr = dev
-    Whisper::Context.new(MODEL)
+    Whisper::Context.new("base.en")
     assert_empty dev.string
   ensure
     $stderr = stderr
   end
+
+  sub_test_case "full" do
+    def setup
+      super
+      @whisper = Whisper::Context.new("base.en")
+      @samples = File.read(AUDIO, nil, 78).unpack("s<*").collect {|i| i.to_f / 2**15}
+    end
+
+    def test_full
+      @whisper.full(@params, @samples, @samples.length)
+
+      assert_equal 1, @whisper.full_n_segments
+      assert_match /ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text
+    end
+
+    def test_full_without_length
+      @whisper.full(@params, @samples)
+
+      assert_equal 1, @whisper.full_n_segments
+      assert_match /ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text
+    end
+
+    def test_full_enumerator
+      samples = @samples.each
+      @whisper.full(@params, samples, @samples.length)
+
+      assert_equal 1, @whisper.full_n_segments
+      assert_match /ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text
+    end
+
+    def test_full_enumerator_without_length
+      samples = @samples.each
+      assert_raise ArgumentError do
+        @whisper.full(@params, samples)
+      end
+    end
+
+    def test_full_enumerator_with_too_large_length
+      samples = @samples.each.take(10).to_enum
+      assert_raise StopIteration do
+        @whisper.full(@params, samples, 11)
+      end
+    end
+
+    def test_full_with_memory_view
+      samples = JFKReader.new(AUDIO)
+      @whisper.full(@params, samples)
+
+      assert_equal 1, @whisper.full_n_segments
+      assert_match /ask not what your country can do for you, ask what you can do for your country/, @whisper.each_segment.first.text
+    end
+
+    def test_full_parallel
+      @whisper.full_parallel(@params, @samples, @samples.length, Etc.nprocessors)
+
+      assert_equal Etc.nprocessors, @whisper.full_n_segments
+      text = @whisper.each_segment.collect(&:text).join
+      assert_match /ask what you can do/i, text
+      assert_match /for your country/i, text
+    end
+
+    def test_full_parallel_with_memory_view
+      samples = JFKReader.new(AUDIO)
+      @whisper.full_parallel(@params, samples, nil, Etc.nprocessors)
+
+      assert_equal Etc.nprocessors, @whisper.full_n_segments
+      text = @whisper.each_segment.collect(&:text).join
+      assert_match /ask what you can do/i, text
+      assert_match /for your country/i, text
+    end
+
+    def test_full_parallel_without_length_and_n_processors
+      @whisper.full_parallel(@params, @samples)
+
+      assert_equal 1, @whisper.full_n_segments
+      text = @whisper.each_segment.collect(&:text).join
+      assert_match /ask what you can do/i, text
+      assert_match /for your country/i, text
+    end
+
+    def test_full_parallel_without_length
+      @whisper.full_parallel(@params, @samples, nil, Etc.nprocessors)
+
+      assert_equal Etc.nprocessors, @whisper.full_n_segments
+      text = @whisper.each_segment.collect(&:text).join
+      assert_match /ask what you can do/i, text
+      assert_match /for your country/i, text
+    end
+
+    def test_full_parallel_without_n_processors
+      @whisper.full_parallel(@params, @samples, @samples.length)
+
+      assert_equal 1, @whisper.full_n_segments
+      text = @whisper.each_segment.collect(&:text).join
+      assert_match /ask what you can do/i, text
+      assert_match /for your country/i, text
+    end
+  end
 end
diff --git a/bindings/ruby/whispercpp.gemspec b/bindings/ruby/whispercpp.gemspec
index 251d03faa06..d8f5c0d81ff 100644
--- a/bindings/ruby/whispercpp.gemspec
+++ b/bindings/ruby/whispercpp.gemspec
@@ -1,4 +1,4 @@
-require "yaml"
+require_relative "extsources"
 
 Gem::Specification.new do |s|
   s.name    = "whispercpp"
@@ -10,24 +10,24 @@ Gem::Specification.new do |s|
   s.extra_rdoc_files = ['LICENSE', 'README.md']
   
   s.files = `git ls-files . -z`.split("\x0") +
-              YAML.load_file("extsources.yaml").collect {|file|
+              EXTSOURCES.collect {|file|
                 basename = File.basename(file)
                 if s.extra_rdoc_files.include?(basename)
                   basename
                 else
-                  File.join("ext", basename)
+                  file.sub("../..", "ext")
                 end
               }
 
   s.summary = %q{Ruby whisper.cpp bindings}
-  s.test_files = ["tests/test_whisper.rb"]
+  s.test_files = s.files.select {|file| file.start_with? "tests/"}
   
   s.extensions << 'ext/extconf.rb'
-  
+  s.required_ruby_version = '>= 3.1.0'
 
   #### Documentation and testing.
   s.homepage = 'https://github.com/ggerganov/whisper.cpp'
-  s.rdoc_options = ['--main', '../../README.md']
+  s.rdoc_options = ['--main', 'README.md']
 
   
     s.platform = Gem::Platform::RUBY
diff --git a/cmake/whisper.pc.in b/cmake/whisper.pc.in
index 50ab4f8285a..00ec7912014 100644
--- a/cmake/whisper.pc.in
+++ b/cmake/whisper.pc.in
@@ -1,10 +1,10 @@
 prefix=@CMAKE_INSTALL_PREFIX@
 exec_prefix=${prefix}
-libdir=@CMAKE_INSTALL_FULL_LIBDIR@
+libdir=${exec_prefix}/lib
 includedir=${prefix}/include
 
 Name: whisper
 Description: Port of OpenAI's Whisper model in C/C++
 Version: @PROJECT_VERSION@
-Libs: -L${libdir} -lwhisper
+Libs: -L${libdir} -lggml  -lggml-base -lwhisper
 Cflags: -I${includedir}
diff --git a/examples/common-ggml.cpp b/examples/common-ggml.cpp
index 760cd1f44be..d59b77fce6b 100644
--- a/examples/common-ggml.cpp
+++ b/examples/common-ggml.cpp
@@ -72,9 +72,6 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_MOSTLY_IQ4_XS:
         case GGML_FTYPE_MOSTLY_IQ1_M:
         case GGML_FTYPE_MOSTLY_BF16:
-        case GGML_FTYPE_MOSTLY_Q4_0_4_4:
-        case GGML_FTYPE_MOSTLY_Q4_0_4_8:
-        case GGML_FTYPE_MOSTLY_Q4_0_8_8:
                 {
                     fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                     return false;
@@ -212,9 +209,6 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_IQ4_XS:
                 case GGML_TYPE_IQ1_M:
                 case GGML_TYPE_BF16:
-                case GGML_TYPE_Q4_0_4_4:
-                case GGML_TYPE_Q4_0_4_8:
-                case GGML_TYPE_Q4_0_8_8:
                 case GGML_TYPE_TQ1_0:
                 case GGML_TYPE_TQ2_0:
                 case GGML_TYPE_COUNT:
diff --git a/examples/common-sdl.cpp b/examples/common-sdl.cpp
index 5329ec739b6..b61f8cff5fd 100644
--- a/examples/common-sdl.cpp
+++ b/examples/common-sdl.cpp
@@ -1,5 +1,7 @@
 #include "common-sdl.h"
 
+#include <cstdio>
+
 audio_async::audio_async(int len_ms) {
     m_len_ms = len_ms;
 
diff --git a/examples/stream/README.md b/examples/stream/README.md
index de5e5a7810d..a8bdf71d0ca 100644
--- a/examples/stream/README.md
+++ b/examples/stream/README.md
@@ -5,7 +5,7 @@ The `stream` tool samples the audio every half a second and runs the transcripti
 More info is available in [issue #10](https://github.com/ggerganov/whisper.cpp/issues/10).
 
 ```bash
-./stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
+./build/bin/stream -m ./models/ggml-base.en.bin -t 8 --step 500 --length 5000
 ```
 
 https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a80f-28ba83be7d09.mp4
@@ -15,7 +15,7 @@ https://user-images.githubusercontent.com/1991296/194935793-76afede7-cfa8-48d8-a
 Setting the `--step` argument to `0` enables the sliding window mode:
 
 ```bash
- ./stream -m ./models/ggml-small.en.bin -t 6 --step 0 --length 30000 -vth 0.6
+ ./build/bin/stream -m ./models/ggml-base.en.bin -t 6 --step 0 --length 30000 -vth 0.6
 ```
 
 In this mode, the tool will transcribe only after some speech activity is detected. A very
@@ -40,21 +40,10 @@ sudo dnf install SDL2 SDL2-devel
 # Install SDL2 on Mac OS
 brew install sdl2
 
-make stream
-```
-
-Ensure you are at the root of the repo when running `make stream`. Not within the `examples/stream` dir
-as the libraries needed like `common-sdl.h` are located within `examples`. Attempting to compile within
-`examples/steam` means your compiler cannot find them and it gives an error it cannot find the file.
+cmake -B build -DWHISPER_SDL2=ON
+cmake --build build --config Release
 
-```bash
-whisper.cpp/examples/stream$ make stream
-g++     stream.cpp   -o stream
-stream.cpp:6:10: fatal error: common/sdl.h: No such file or directory
-    6 | #include "common/sdl.h"
-      |          ^~~~~~~~~~~~~~
-compilation terminated.
-make: *** [<builtin>: stream] Error 1
+./build/bin/stream
 ```
 
 ## Web version
diff --git a/examples/talk-llama/llama-sampling.cpp b/examples/talk-llama/llama-sampling.cpp
index fd8ca8a9edf..bebff77cfa0 100644
--- a/examples/talk-llama/llama-sampling.cpp
+++ b/examples/talk-llama/llama-sampling.cpp
@@ -1396,19 +1396,15 @@ struct llama_sampler * llama_sampler_init_grammar_impl(const struct llama_vocab
 // penalties
 
 struct llama_sampler_penalties {
-    const int32_t     n_vocab;
-    const llama_token special_eos_id;
-    const llama_token linefeed_id;
-
     const int32_t penalty_last_n;
     const float   penalty_repeat;
     const float   penalty_freq;
     const float   penalty_present;
 
-    const bool    penalize_nl;
-    const bool    ignore_eos;
-
     ring_buffer<llama_token> prev;
+
+    // a frequency map to count token occurrences
+    std::unordered_map<llama_token, int> token_count;
 };
 
 static const char * llama_sampler_penalties_name(const struct llama_sampler * /*smpl*/) {
@@ -1421,76 +1417,50 @@ static void llama_sampler_penalties_accept(struct llama_sampler * smpl, llama_to
         return;
     }
 
-    ctx->prev.push_back(token);
-}
-
-static void llama_sampler_penalties_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
-    auto * ctx = (llama_sampler_penalties *) smpl->ctx;
+    ctx->token_count[token]++;
 
-    if (ctx->ignore_eos) {
-        assert(ctx->special_eos_id >= 0);
+    // if the ring buffer is full, remove the oldest token
+    if (ctx->prev.size() >= (size_t) ctx->penalty_last_n) {
+        const auto old = ctx->prev.front();
 
-        // optimistically check if the candidates are not yet sorted/shuffled/truncated
-        if (cur_p->size > (size_t) ctx->special_eos_id && cur_p->data[ctx->special_eos_id].id == ctx->special_eos_id) {
-            cur_p->data[ctx->special_eos_id].logit = -INFINITY;
-        } else {
-            // else, search for the special EOS token
-            for (size_t i = 0; i < cur_p->size; ++i) {
-                if (cur_p->data[i].id == ctx->special_eos_id) {
-                    cur_p->data[i].logit = -INFINITY;
-                    break;
-                }
-            }
+        ctx->token_count[old]--;
+        if (ctx->token_count[old] == 0) {
+            ctx->token_count.erase(old);
         }
     }
 
-    if ((ctx->penalty_last_n == 0) ||
-        (ctx->penalty_repeat == 1.0f && ctx->penalty_freq == 0.0f && ctx->penalty_present == 0.0f)) {
-        return;
-    }
-
-    bool nl_found = false;
-    size_t nl_idx = 0;
-    float nl_logit = -INFINITY;
-    if (!ctx->penalize_nl) {
-        assert(ctx->linefeed_id >= 0);
+    ctx->prev.push_back(token);
 
-        // optimistically check if the candidates are not yet sorted/shuffled/truncated
-        if (cur_p->size > (size_t) ctx->linefeed_id && cur_p->data[ctx->linefeed_id].id == ctx->linefeed_id) {
-            nl_found = true;
-            nl_idx = ctx->linefeed_id;
-            nl_logit = cur_p->data[ctx->linefeed_id].logit;
-        } else {
-            // else, search for the linefeed token
-            for (size_t i = 0; i < cur_p->size; ++i) {
-                if (cur_p->data[i].id == ctx->linefeed_id) {
-                    nl_found = true;
-                    nl_idx = i;
-                    nl_logit = cur_p->data[i].logit;
-                    break;
-                }
-            }
-        }
+#if 0
+    // sanity check
+    std::unordered_map<llama_token, int> tmp;
+    for (int i = 0; i < std::min<int>(ctx->penalty_last_n, ctx->prev.size()); ++i) {
+        tmp[ctx->prev.rat(i)]++;
     }
 
-    // Create a frequency map to count occurrences of each token in last_tokens
-    // TODO: optimize this by maintaining the token count in the sampler context
-    using llama_token_cnt = std::unordered_map<llama_token, int>;
-    llama_token_cnt token_count;
+    assert(ctx->token_count == tmp);
+#endif
+}
+
+static void llama_sampler_penalties_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (llama_sampler_penalties *) smpl->ctx;
 
-    for (int i = 0; i < std::min<int>(ctx->penalty_last_n, ctx->prev.size()); ++i) {
-        token_count[ctx->prev.rat(i)]++;
+    if ((ctx->penalty_last_n == 0) ||
+        (ctx->penalty_repeat == 1.0f && ctx->penalty_freq == 0.0f && ctx->penalty_present == 0.0f)) {
+        return;
     }
 
     // Apply frequency and presence penalties to the cur_p
     for (size_t i = 0; i < cur_p->size; ++i) {
-        const auto token_iter = token_count.find(cur_p->data[i].id);
-        if (token_iter == token_count.end()) {
+        const auto token_iter = ctx->token_count.find(cur_p->data[i].id);
+        if (token_iter == ctx->token_count.end()) {
             continue;
         }
 
         const int count = token_iter->second;
 
+        assert(count > 0 && count <= ctx->penalty_last_n);
+
         // The academic publication that described this technique actually just only divided, but that would cause tokens with negative logits to become more likely, which is obviously wrong.
         // This is common fix for this problem, which is to multiply by the penalty instead of dividing.
         if (cur_p->data[i].logit <= 0) {
@@ -1503,30 +1473,21 @@ static void llama_sampler_penalties_apply(struct llama_sampler * smpl, llama_tok
     }
 
     cur_p->sorted = false;
-
-    if (!ctx->penalize_nl && nl_found) {
-        // restore the logit of the newline token if it was penalized
-        cur_p->data[nl_idx].logit = nl_logit;
-    }
 }
 
 static void llama_sampler_penalties_reset(struct llama_sampler * smpl) {
     auto * ctx = (llama_sampler_penalties *) smpl->ctx;
     ctx->prev.clear();
+    ctx->token_count.clear();
 }
 
 static struct llama_sampler * llama_sampler_penalties_clone(const struct llama_sampler * smpl) {
     const auto * ctx = (const llama_sampler_penalties *) smpl->ctx;
     auto * result = llama_sampler_init_penalties(
-            ctx->n_vocab,
-            ctx->special_eos_id,
-            ctx->linefeed_id,
             ctx->penalty_last_n,
             ctx->penalty_repeat,
             ctx->penalty_freq,
-            ctx->penalty_present,
-            ctx->penalize_nl,
-            ctx->ignore_eos);
+            ctx->penalty_present);
 
     // copy the state
     {
@@ -1552,38 +1513,21 @@ static struct llama_sampler_i llama_sampler_penalties_i = {
 };
 
 struct llama_sampler * llama_sampler_init_penalties(
-        int32_t n_vocab,
-        llama_token special_eos_id,
-        llama_token linefeed_id,
         int32_t penalty_last_n,
         float penalty_repeat,
         float penalty_freq,
-        float penalty_present,
-        bool penalize_nl,
-        bool ignore_eos) {
-    if (linefeed_id == LLAMA_TOKEN_NULL) {
-        penalize_nl = true;
-    }
-
-    if (special_eos_id == LLAMA_TOKEN_NULL) {
-        ignore_eos = false;
-    }
-
+        float penalty_present) {
     penalty_last_n = std::max(penalty_last_n, 0);
 
     return new llama_sampler {
         /* .iface = */ &llama_sampler_penalties_i,
         /* .ctx   = */ new llama_sampler_penalties {
-            /* .n_vocab         = */ n_vocab,
-            /* .special_eos_id  = */ special_eos_id,
-            /* .linefeed_id     = */ linefeed_id,
             /* .penalty_last_n  = */ penalty_last_n,
             /* .penalty_repeat  = */ penalty_repeat,
             /* .penalty_freq    = */ penalty_freq,
             /* .penalty_present = */ penalty_present,
-            /* .penalize_nl     = */ penalize_nl,
-            /* .ignore_eos      = */ ignore_eos,
             /* .prev            = */ ring_buffer<llama_token>(penalty_last_n),
+            /* .token_count     = */ {},
         },
     };
 }
@@ -1611,7 +1555,8 @@ static void get_overlapping_token_sequences(const llama_vocab & vocab, const std
         if (word.find(str) != std::string::npos) {
             token_sequences.emplace(token_id, std::vector<llama_token>());
         } else {
-            size_t word_len = word.size(), str_len = str.size();
+            size_t word_len = word.size();
+            size_t str_len = str.size();
             size_t pos = -1;
             while ((pos = word.find(str[0], pos + 1)) != std::string::npos) {
                 bool match = true;
diff --git a/examples/talk-llama/llama-vocab.cpp b/examples/talk-llama/llama-vocab.cpp
index d1dc96276c2..e38e5985323 100644
--- a/examples/talk-llama/llama-vocab.cpp
+++ b/examples/talk-llama/llama-vocab.cpp
@@ -418,6 +418,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
             case LLAMA_VOCAB_PRE_TYPE_SMOLLM:
             case LLAMA_VOCAB_PRE_TYPE_CODESHELL:
             case LLAMA_VOCAB_PRE_TYPE_EXAONE:
+            case LLAMA_VOCAB_PRE_TYPE_MINERVA:
                 regex_exprs = {
                     "\\p{N}",
                     "'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)",
@@ -737,7 +738,7 @@ struct llm_tokenizer_wpm_session {
         std::vector<std::string> words(1, "");
 
         for (const uint32_t cpt : cpts_nfd) {
-            const auto flags = unicode_cpt_flags(cpt);
+            const auto flags = unicode_cpt_flags_from_cpt(cpt);
 
             if (flags.is_whitespace) {
                 if (words.back().size()) {  // finish previous word if any
diff --git a/examples/talk-llama/llama.cpp b/examples/talk-llama/llama.cpp
index 97eee26a577..1cc8a93323b 100644
--- a/examples/talk-llama/llama.cpp
+++ b/examples/talk-llama/llama.cpp
@@ -163,6 +163,7 @@ enum llm_arch {
     LLM_ARCH_QWEN,
     LLM_ARCH_QWEN2,
     LLM_ARCH_QWEN2MOE,
+    LLM_ARCH_QWEN2VL,
     LLM_ARCH_PHI2,
     LLM_ARCH_PHI3,
     LLM_ARCH_PLAMO,
@@ -179,9 +180,11 @@ enum llm_arch {
     LLM_ARCH_COMMAND_R,
     LLM_ARCH_DBRX,
     LLM_ARCH_OLMO,
+    LLM_ARCH_OLMO2,
     LLM_ARCH_OLMOE,
     LLM_ARCH_OPENELM,
     LLM_ARCH_ARCTIC,
+    LLM_ARCH_DEEPSEEK,
     LLM_ARCH_DEEPSEEK2,
     LLM_ARCH_CHATGLM,
     LLM_ARCH_BITNET,
@@ -216,6 +219,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_QWEN,            "qwen"         },
     { LLM_ARCH_QWEN2,           "qwen2"        },
     { LLM_ARCH_QWEN2MOE,        "qwen2moe"     },
+    { LLM_ARCH_QWEN2VL,         "qwen2vl"      },
     { LLM_ARCH_PHI2,            "phi2"         },
     { LLM_ARCH_PHI3,            "phi3"         },
     { LLM_ARCH_PLAMO,           "plamo"        },
@@ -232,9 +236,11 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_COMMAND_R,       "command-r"    },
     { LLM_ARCH_DBRX,            "dbrx"         },
     { LLM_ARCH_OLMO,            "olmo"         },
+    { LLM_ARCH_OLMO2,           "olmo2"        },
     { LLM_ARCH_OLMOE,           "olmoe"        },
     { LLM_ARCH_OPENELM,         "openelm"      },
     { LLM_ARCH_ARCTIC,          "arctic"       },
+    { LLM_ARCH_DEEPSEEK,        "deepseek"     },
     { LLM_ARCH_DEEPSEEK2,       "deepseek2"    },
     { LLM_ARCH_CHATGLM,         "chatglm"      },
     { LLM_ARCH_BITNET,          "bitnet"       },
@@ -306,6 +312,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_SCALE,
 
     LLM_KV_ROPE_DIMENSION_COUNT,
+    LLM_KV_ROPE_DIMENSION_SECTIONS,
     LLM_KV_ROPE_FREQ_BASE,
     LLM_KV_ROPE_SCALE_LINEAR,
     LLM_KV_ROPE_SCALING_TYPE,
@@ -422,6 +429,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SCALE,                  "%s.attention.scale"                  },
 
     { LLM_KV_ROPE_DIMENSION_COUNT,             "%s.rope.dimension_count"                 },
+    { LLM_KV_ROPE_DIMENSION_SECTIONS,          "%s.rope.dimension_sections"              },
     { LLM_KV_ROPE_FREQ_BASE,                   "%s.rope.freq_base"                       },
     { LLM_KV_ROPE_SCALE_LINEAR,                "%s.rope.scale_linear"                    },
     { LLM_KV_ROPE_SCALING_TYPE,                "%s.rope.scaling.type"                    },
@@ -896,6 +904,23 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_QWEN2VL,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_QWEN2MOE,
         {
@@ -1034,6 +1059,8 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
             { LLM_TENSOR_OUTPUT,          "output" },
             { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ROPE_FACTORS_LONG,  "rope_factors_long" },
+            { LLM_TENSOR_ROPE_FACTORS_SHORT, "rope_factors_short" },
             { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
             { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
             { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
@@ -1207,6 +1234,25 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
         },
     },
+    {
+        LLM_ARCH_OLMO2,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
+            { LLM_TENSOR_ATTN_Q_NORM,     "blk.%d.attn_q_norm" },
+            { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+        },
+    },
     {
         LLM_ARCH_OLMOE,
         {
@@ -1265,6 +1311,33 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_DEEPSEEK,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,         "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,        "output_norm" },
+            { LLM_TENSOR_OUTPUT,             "output" },
+            { LLM_TENSOR_ROPE_FREQS,         "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,          "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,             "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,             "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,             "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,           "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,      "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_GATE_INP,       "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,           "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,           "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,           "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,             "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXPS,      "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,      "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
+            { LLM_TENSOR_FFN_GATE_INP_SHEXP, "blk.%d.ffn_gate_inp_shexp" },
+            { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
+            { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
+            { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
+        },
+    },
     {
         LLM_ARCH_DEEPSEEK2,
         {
@@ -1528,6 +1601,71 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
     },
 };
 
+enum llm_chat_template {
+    LLM_CHAT_TEMPLATE_CHATML,
+    LLM_CHAT_TEMPLATE_LLAMA_2,
+    LLM_CHAT_TEMPLATE_LLAMA_2_SYS,
+    LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS,
+    LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP,
+    LLM_CHAT_TEMPLATE_MISTRAL_V1,
+    LLM_CHAT_TEMPLATE_MISTRAL_V3,
+    LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN,
+    LLM_CHAT_TEMPLATE_MISTRAL_V7,
+    LLM_CHAT_TEMPLATE_PHI_3,
+    LLM_CHAT_TEMPLATE_FALCON_3,
+    LLM_CHAT_TEMPLATE_ZEPHYR,
+    LLM_CHAT_TEMPLATE_MONARCH,
+    LLM_CHAT_TEMPLATE_GEMMA,
+    LLM_CHAT_TEMPLATE_ORION,
+    LLM_CHAT_TEMPLATE_OPENCHAT,
+    LLM_CHAT_TEMPLATE_VICUNA,
+    LLM_CHAT_TEMPLATE_VICUNA_ORCA,
+    LLM_CHAT_TEMPLATE_DEEPSEEK,
+    LLM_CHAT_TEMPLATE_DEEPSEEK_2,
+    LLM_CHAT_TEMPLATE_COMMAND_R,
+    LLM_CHAT_TEMPLATE_LLAMA_3,
+    LLM_CHAT_TEMPLATE_CHATGML_3,
+    LLM_CHAT_TEMPLATE_CHATGML_4,
+    LLM_CHAT_TEMPLATE_MINICPM,
+    LLM_CHAT_TEMPLATE_EXAONE_3,
+    LLM_CHAT_TEMPLATE_RWKV_WORLD,
+    LLM_CHAT_TEMPLATE_GRANITE,
+    LLM_CHAT_TEMPLATE_GIGACHAT,
+    LLM_CHAT_TEMPLATE_UNKNOWN,
+};
+
+static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
+    { "chatml",            LLM_CHAT_TEMPLATE_CHATML            },
+    { "llama2",            LLM_CHAT_TEMPLATE_LLAMA_2           },
+    { "llama2-sys",        LLM_CHAT_TEMPLATE_LLAMA_2_SYS       },
+    { "llama2-sys-bos",    LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS   },
+    { "llama2-sys-strip",  LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP },
+    { "mistral-v1",        LLM_CHAT_TEMPLATE_MISTRAL_V1        },
+    { "mistral-v3",        LLM_CHAT_TEMPLATE_MISTRAL_V3        },
+    { "mistral-v3-tekken", LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN },
+    { "mistral-v7",        LLM_CHAT_TEMPLATE_MISTRAL_V7        },
+    { "phi3",              LLM_CHAT_TEMPLATE_PHI_3             },
+    { "falcon3",           LLM_CHAT_TEMPLATE_FALCON_3          },
+    { "zephyr",            LLM_CHAT_TEMPLATE_ZEPHYR            },
+    { "monarch",           LLM_CHAT_TEMPLATE_MONARCH           },
+    { "gemma",             LLM_CHAT_TEMPLATE_GEMMA             },
+    { "orion",             LLM_CHAT_TEMPLATE_ORION             },
+    { "openchat",          LLM_CHAT_TEMPLATE_OPENCHAT          },
+    { "vicuna",            LLM_CHAT_TEMPLATE_VICUNA            },
+    { "vicuna-orca",       LLM_CHAT_TEMPLATE_VICUNA_ORCA       },
+    { "deepseek",          LLM_CHAT_TEMPLATE_DEEPSEEK          },
+    { "deepseek2",         LLM_CHAT_TEMPLATE_DEEPSEEK_2        },
+    { "command-r",         LLM_CHAT_TEMPLATE_COMMAND_R         },
+    { "llama3",            LLM_CHAT_TEMPLATE_LLAMA_3           },
+    { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGML_3         },
+    { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGML_4         },
+    { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
+    { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
+    { "rwkv-world",        LLM_CHAT_TEMPLATE_RWKV_WORLD        },
+    { "granite",           LLM_CHAT_TEMPLATE_GRANITE           },
+    { "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },
+};
+
 static llm_arch llm_arch_from_string(const std::string & name) {
     for (const auto & kv : LLM_ARCH_NAMES) { // NOLINT
         if (kv.second == name) {
@@ -1601,9 +1739,10 @@ struct LLM_TN {
 //
 
 static const std::map<llama_rope_scaling_type, const char *> LLAMA_ROPE_SCALING_TYPES = {
-    { LLAMA_ROPE_SCALING_TYPE_NONE,   "none"   },
-    { LLAMA_ROPE_SCALING_TYPE_LINEAR, "linear" },
-    { LLAMA_ROPE_SCALING_TYPE_YARN,   "yarn"   },
+    { LLAMA_ROPE_SCALING_TYPE_NONE,       "none"       },
+    { LLAMA_ROPE_SCALING_TYPE_LINEAR,     "linear"     },
+    { LLAMA_ROPE_SCALING_TYPE_YARN,       "yarn"       },
+    { LLAMA_ROPE_SCALING_TYPE_LONGROPE,   "longrope"   },
 };
 
 static llama_rope_scaling_type llama_rope_scaling_type_from_string(const std::string & name) {
@@ -1709,7 +1848,7 @@ struct llama_file {
         DWORD bufLen = FormatMessageA(FORMAT_MESSAGE_ALLOCATE_BUFFER | FORMAT_MESSAGE_FROM_SYSTEM | FORMAT_MESSAGE_IGNORE_INSERTS,
                                     NULL, error_code, MAKELANGID(LANG_NEUTRAL, SUBLANG_DEFAULT), (LPSTR)&lpMsgBuf, 0, NULL);
         if (!bufLen) {
-            ret = format("Win32 error code: %s", error_code);
+            ret = format("Win32 error code: %lx", error_code);
         } else {
             ret = lpMsgBuf;
             LocalFree(lpMsgBuf);
@@ -2047,7 +2186,7 @@ struct llama_mmap {
             HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
 
             // may fail on pre-Windows 8 systems
-            pPrefetchVirtualMemory = reinterpret_cast<decltype(pPrefetchVirtualMemory)> (GetProcAddress(hKernel32, "PrefetchVirtualMemory"));
+            pPrefetchVirtualMemory = (decltype(pPrefetchVirtualMemory))(void *) GetProcAddress(hKernel32, "PrefetchVirtualMemory");
 
             if (pPrefetchVirtualMemory) {
                 // advise the kernel to preload the mapped memory
@@ -2320,6 +2459,7 @@ enum e_model {
     MODEL_16B,
     MODEL_20B,
     MODEL_30B,
+    MODEL_32B,
     MODEL_34B,
     MODEL_35B,
     MODEL_40B,
@@ -2388,11 +2528,12 @@ struct llama_hparams {
     uint32_t time_decay_extra_dim = 0;
     uint32_t wkv_head_size = 0;
 
-    float    rope_attn_factor = 1.0f;
-    float    rope_freq_base_train;
-    float    rope_freq_scale_train;
-    uint32_t n_ctx_orig_yarn;
-    float    rope_yarn_log_mul;
+    float     rope_attn_factor = 1.0f;
+    float     rope_freq_base_train;
+    float     rope_freq_scale_train;
+    uint32_t  n_ctx_orig_yarn;
+    float     rope_yarn_log_mul;
+    int       rope_sections[4];
 
     // for State Space Models
     uint32_t ssm_d_conv  = 0;
@@ -2449,6 +2590,9 @@ struct llama_hparams {
 
         if (this->rope_finetuned  != other.rope_finetuned)  return true;
         if (this->n_ctx_orig_yarn != other.n_ctx_orig_yarn) return true;
+        if (std::equal(std::begin(this->rope_sections),
+                       std::end(this->rope_sections),
+                       std::begin(other.rope_sections)))    return true;
 
         if (this->ssm_d_conv  != other.ssm_d_conv)  return true;
         if (this->ssm_d_inner != other.ssm_d_inner) return true;
@@ -2907,9 +3051,15 @@ struct llama_model {
     // for quantize-stats only
     std::vector<std::pair<std::string, struct ggml_tensor *>> tensors_by_name;
 
-    int64_t t_load_us = 0;
+    int64_t t_load_us  = 0;
     int64_t t_start_us = 0;
 
+    // total number of parameters in the model
+    uint64_t n_elements = 0;
+
+    // total size of all the tensors in the model in bytes
+    size_t  n_bytes     = 0;
+
     // keep track of loaded lora adapters
     std::set<struct llama_lora_adapter *> lora_adapters;
 
@@ -3286,6 +3436,11 @@ struct llama_context {
     // whether we are computing encoder output or decoder output
     bool is_encoding = false;
 
+    // TODO: find a better way to accommodate mutli-dimension position encoding methods
+    // number of position id each token get, 1 for each token in most cases.
+    // when using m-rope, it will be 3 position ids per token to representing 3 dimension coordinate.
+    int n_pos_per_token = 1;
+
     // output of the encoder part of the encoder-decoder models
     std::vector<float> embd_enc;
     std::vector<std::set<llama_seq_id>> seq_ids_enc;
@@ -3454,21 +3609,13 @@ static bool llama_kv_cache_init(
         const uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i) + hparams.n_embd_k_s();
         const uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i) + hparams.n_embd_v_s();
 
-        const llama_model::buft_list_t * buft_list;
+        ggml_backend_buffer_type_t buft;
         if (offload) {
-            buft_list = model.dev_layer.at(i).buft_list;
+            auto * dev = model.dev_layer.at(i).dev;
+            buft = ggml_backend_dev_buffer_type(dev);
         } else {
-            buft_list = &model.cpu_buft_list;
+            buft = ggml_backend_cpu_buffer_type();
         }
-        ggml_backend_buffer_type_t buft = select_buft(*buft_list,
-            [&](ggml_context * ctx) {
-                ggml_tensor * k = ggml_new_tensor_1d(ctx, type_k, n_embd_k_gqa*kv_size);
-                if (hparams.rope_type == LLAMA_ROPE_TYPE_NONE) {
-                    return k;
-                }
-                ggml_tensor * p = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 1);
-                return ggml_rope(ctx, k, p, hparams.n_rot, hparams.rope_type);
-            });
         ggml_context * ctx = ctx_for_buft(buft);
 
         if (!ctx) {
@@ -4275,8 +4422,8 @@ struct llama_model_loader {
     int n_tensors = 0;
     int n_created = 0;
 
-    int64_t n_elements = 0;
-    size_t  n_bytes    = 0;
+    uint64_t n_elements = 0;
+    size_t  n_bytes     = 0;
 
     bool use_mmap = false;
     bool check_tensors;
@@ -4494,9 +4641,6 @@ struct llama_model_loader {
                 case GGML_TYPE_IQ4_NL:  ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL;  break;
                 case GGML_TYPE_IQ4_XS:  ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS;  break;
                 case GGML_TYPE_IQ3_S:   ftype = LLAMA_FTYPE_MOSTLY_IQ3_S;   break;
-                case GGML_TYPE_Q4_0_4_4: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_4; break;
-                case GGML_TYPE_Q4_0_4_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_4_8; break;
-                case GGML_TYPE_Q4_0_8_8: ftype = LLAMA_FTYPE_MOSTLY_Q4_0_8_8; break;
                 default:
                     {
                         LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@@ -4847,7 +4991,9 @@ struct llama_model_loader {
             mappings.reserve(files.size());
             mmaps_used.reserve(files.size());
             for (const auto & file : files) {
-                std::unique_ptr<llama_mmap> mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, ggml_is_numa()));
+                auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
+                auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa");
+                std::unique_ptr<llama_mmap> mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, is_numa_fn()));
                 mmaps_used.emplace_back(mapping->size, 0);
                 if (mlock_mmaps) {
                     std::unique_ptr<llama_mlock> mlock_mmap(new llama_mlock());
@@ -5258,9 +5404,6 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_IQ4_XS:   return "IQ4_XS - 4.25 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ3_S:    return "IQ3_S - 3.4375 bpw";
         case LLAMA_FTYPE_MOSTLY_IQ3_M:    return "IQ3_S mix - 3.66 bpw";
-        case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: return "Q4_0_4_4";
-        case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: return "Q4_0_4_8";
-        case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: return "Q4_0_8_8";
 
         default: return "unknown, may not work";
     }
@@ -5309,6 +5452,7 @@ static const char * llama_model_type_name(e_model type) {
         case MODEL_16B:           return "16B";
         case MODEL_20B:           return "20B";
         case MODEL_30B:           return "30B";
+        case MODEL_32B:           return "32B";
         case MODEL_34B:           return "34B";
         case MODEL_35B:           return "35B";
         case MODEL_40B:           return "40B";
@@ -5344,6 +5488,11 @@ static const char * llama_model_vocab_type_name(enum llama_vocab_type type){
     }
 }
 
+static void llm_load_stats(llama_model_loader & ml, llama_model & model) {
+    model.n_elements = ml.n_elements;
+    model.n_bytes = ml.n_bytes;
+}
+
 static void llm_load_arch(llama_model_loader & ml, llama_model & model) {
     model.arch = ml.get_arch();
     if (model.arch == LLM_ARCH_UNKNOWN) {
@@ -5491,8 +5640,12 @@ static void llm_load_hparams(
         case LLM_ARCH_MINICPM:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale);
+                ml.get_key(LLM_KV_RESIDUAL_SCALE, hparams.f_residual_scale);
+                ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale);
 
                 switch (hparams.n_layer) {
+                    case 52: model.type = e_model::MODEL_1B; break;
                     case 40: model.type = e_model::MODEL_2B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
@@ -5657,6 +5810,13 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_QWEN2VL:
+            {
+                std::array<int, 4> section_dims;
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, section_dims, 4, true);
+                std::copy(section_dims.begin(), section_dims.begin() + 4, std::begin(hparams.rope_sections));
+            }
+            // fall through
         case LLM_ARCH_QWEN2:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -5664,7 +5824,10 @@ static void llm_load_hparams(
                     case 24: model.type = hparams.n_embd == 1024 ? e_model::MODEL_0_5B : e_model::MODEL_1B; break;
                     case 28: model.type = hparams.n_embd == 1536 ? e_model::MODEL_1_5B : e_model::MODEL_7B; break;
                     case 32: model.type = e_model::MODEL_7B; break;
+                    case 36: model.type = e_model::MODEL_3B; break;
                     case 40: model.type = hparams.n_head() == 20 ? e_model::MODEL_4B : e_model::MODEL_13B; break;
+                    case 48: model.type = e_model::MODEL_14B; break;
+                    case 64: model.type = e_model::MODEL_32B; break;
                     case 80: model.type = e_model::MODEL_70B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
@@ -5874,6 +6037,17 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_OLMO2:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+                switch (hparams.n_layer) {
+                    case 16: model.type = e_model::MODEL_1B; break;
+                    case 32: model.type = e_model::MODEL_7B; break;
+                    case 40: model.type = e_model::MODEL_13B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_OLMOE:
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -5953,6 +6127,19 @@ static void llm_load_hparams(
                     model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_DEEPSEEK:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+                ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared);
+                ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale);
+
+                switch (hparams.n_layer) {
+                    case 28: model.type = e_model::MODEL_20B; break;
+                    default: model.type = e_model::MODEL_UNKNOWN;
+                }
+            } break;
         case LLM_ARCH_DEEPSEEK2:
             {
                 bool is_lite = (hparams.n_layer == 27);
@@ -6288,6 +6475,11 @@ static void llm_load_vocab(
             } else if (
                     tokenizer_pre == "falcon") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_FALCON;
+            } else if (
+                    tokenizer_pre == "falcon3") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_LLAMA3;
+                vocab.tokenizer_ignore_merges = true;
+                vocab.tokenizer_add_bos = true;
             } else if (
                     tokenizer_pre == "mpt") {
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_MPT;
@@ -6299,6 +6491,7 @@ static void llm_load_vocab(
                     tokenizer_pre == "phi-2"   ||
                     tokenizer_pre == "jina-es" ||
                     tokenizer_pre == "jina-de" ||
+                    tokenizer_pre == "gigachat"   ||
                     tokenizer_pre == "jina-v1-en" ||
                     tokenizer_pre == "jina-v2-es" ||
                     tokenizer_pre == "jina-v2-de" ||
@@ -6369,6 +6562,9 @@ static void llm_load_vocab(
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_CHAMELEON;
                 vocab.tokenizer_add_bos = true;
                 vocab.tokenizer_clean_spaces = false;
+            } else if (
+                tokenizer_pre == "minerva-7b") {
+                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_MINERVA;
             } else {
                 throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
             }
@@ -6947,6 +7143,13 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
 
     LLAMA_LOG_INFO("%s: max token length = %d\n", __func__, vocab.max_token_len);
 
+    if (model.arch == LLM_ARCH_DEEPSEEK) {
+        LLAMA_LOG_INFO("%s: n_layer_dense_lead   = %d\n",     __func__, hparams.n_layer_dense_lead);
+        LLAMA_LOG_INFO("%s: n_ff_exp             = %d\n",     __func__, hparams.n_ff_exp);
+        LLAMA_LOG_INFO("%s: n_expert_shared      = %d\n",     __func__, hparams.n_expert_shared);
+        LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n",   __func__, hparams.expert_weights_scale);
+    }
+
     if (model.arch == LLM_ARCH_DEEPSEEK2) {
         LLAMA_LOG_INFO("%s: n_layer_dense_lead   = %d\n",     __func__, hparams.n_layer_dense_lead);
         LLAMA_LOG_INFO("%s: n_lora_q             = %d\n",     __func__, hparams.n_lora_q);
@@ -6962,7 +7165,7 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
         LLAMA_LOG_INFO("%s: n_ff_shexp       = %d\n",     __func__, hparams.n_ff_shexp);
     }
 
-    if (model.arch == LLM_ARCH_GRANITE || model.arch == LLM_ARCH_GRANITE_MOE) {
+    if (model.arch == LLM_ARCH_MINICPM || model.arch == LLM_ARCH_GRANITE || model.arch == LLM_ARCH_GRANITE_MOE) {
         LLAMA_LOG_INFO("%s: f_embedding_scale = %f\n", __func__, hparams.f_embedding_scale);
         LLAMA_LOG_INFO("%s: f_residual_scale  = %f\n", __func__, hparams.f_residual_scale);
         LLAMA_LOG_INFO("%s: f_attention_scale = %f\n", __func__, hparams.f_attention_scale);
@@ -7146,12 +7349,12 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w
             } break;
         case GGML_OP_ADD:
             {
-                ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, w->ne[0], 512);
+                ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], w->ne[1], w->ne[2], w->ne[3]);
                 op_tensor = ggml_add(ctx, a, w);
             } break;
         case GGML_OP_MUL:
             {
-                ggml_tensor * a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, w->ne[0], 512);
+                ggml_tensor * a = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, w->ne[0], w->ne[1], w->ne[2], w->ne[3]);
                 op_tensor = ggml_mul(ctx, a, w);
             } break;
         case GGML_OP_DIV:
@@ -7254,7 +7457,7 @@ static llama_model::buft_list_t make_cpu_buft_list(llama_model & model) {
     auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
     auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
     auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
-        ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_cpu_get_extra_bufts");
+        ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
     if (ggml_backend_dev_get_extra_bufts_fn) {
         ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
         while (extra_bufts && *extra_bufts) {
@@ -7521,7 +7724,7 @@ static bool llm_load_tensors(
 
             // avoid using a host buffer when using mmap
             auto * buft_dev = ggml_backend_buft_get_device(buft);
-            if (ml.use_mmap && buft == ggml_backend_dev_host_buffer_type(buft_dev)) {
+            if (ml.use_mmap && buft_dev && buft == ggml_backend_dev_host_buffer_type(buft_dev)) {
                 auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
                 buft = ggml_backend_dev_buffer_type(cpu_dev);
             }
@@ -7587,7 +7790,13 @@ static bool llm_load_tensors(
 
                         layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
 
-                        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
+                            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                        }
+                        else {
+                            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, llama_model_loader::TENSOR_NOT_REQUIRED | (i != 0 ? llama_model_loader::TENSOR_DUPLICATED : 0));
+                        }
 
                         if (n_expert == 0) {
                             layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
@@ -8054,6 +8263,7 @@ static bool llm_load_tensors(
                     }
                 } break;
             case LLM_ARCH_QWEN2:
+            case LLM_ARCH_QWEN2VL:
                 {
                     model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -8556,6 +8766,31 @@ static bool llm_load_tensors(
                         layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
                     }
                 } break;
+            case LLM_ARCH_OLMO2:
+                {
+                    model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    model.output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    model.output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = model.layers[i];
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+                        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, 0);
+                        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+                    }
+                } break;
             case LLM_ARCH_OLMOE:
                 {
                     model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -8689,6 +8924,55 @@ static bool llm_load_tensors(
                         layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
                     }
                 } break;
+            case LLM_ARCH_DEEPSEEK:
+                {
+
+                    const int64_t n_ff_exp        = hparams.n_ff_exp;
+                    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+                    model.tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+                    // output
+                    model.output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+                    model.output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+                        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd}, 0);
+                        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, 0);
+                        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+                        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+                        if (i < (int) hparams.n_layer_dense_lead) {
+                            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+                            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+                            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+                        } else {
+                            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+
+                            if (n_expert == 0) {
+                                throw std::runtime_error("n_expert must be > 0");
+                            }
+                            if (n_expert_used == 0) {
+                                throw std::runtime_error("n_expert_used must be > 0");
+                            }
+
+                            // MoE branch
+                            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+                            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+                            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+
+                            // Shared expert branch
+                            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+                            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, 0);
+                            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+                        }
+                    }
+                } break;
             case LLM_ARCH_DEEPSEEK2:
                 {
                     const bool is_lite = (hparams.n_layer == 27);
@@ -9128,6 +9412,10 @@ static bool llm_load_tensors(
 
         // check if it is possible to use buffer_from_host_ptr with this buffer type
         ggml_backend_dev_t dev = ggml_backend_buft_get_device(buft);
+        if (!dev) {
+            // FIXME: workaround for CPU backend buft having a NULL device
+            dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        }
         ggml_backend_dev_props props;
         ggml_backend_dev_get_props(dev, &props);
         bool buffer_from_host_ptr_supported = props.caps.buffer_from_host_ptr;
@@ -9252,6 +9540,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
             throw std::runtime_error("error loading model vocabulary: " + std::string(e.what()));
         }
 
+        llm_load_stats(ml, model);
         llm_load_print_meta(ml, model);
 
         if (model.vocab.type != LLAMA_VOCAB_TYPE_NONE &&
@@ -12413,12 +12702,8 @@ struct llm_build_context {
         return gf;
     }
 
-    struct ggml_cgraph * build_qwen2moe() {
+    struct ggml_cgraph * build_qwen2vl() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
-
-        // mutable variable, needed during the last layer of the computation to skip unused tokens
-        int32_t n_tokens = this->n_tokens;
-
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
         GGML_ASSERT(n_embd_head == hparams.n_rot);
@@ -12429,10 +12714,15 @@ struct llm_build_context {
         inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
 
         // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
+        lctx.inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens * 4);
+        cb(lctx.inp_pos, "inp_pos", -1);
+        ggml_set_input(lctx.inp_pos);
+        struct ggml_tensor * inp_pos = lctx.inp_pos;
 
         // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
         struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+        int sections[4];
+        std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
 
         for (int il = 0; il < n_layer; ++il) {
             struct ggml_tensor * inpSA = inpL;
@@ -12443,7 +12733,7 @@ struct llm_build_context {
                     LLM_NORM_RMS, cb, il);
             cb(cur, "attn_norm", il);
 
-            // self_attention
+            // self-attention
             {
                 // compute Q and K and RoPE them
                 struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
@@ -12461,16 +12751,18 @@ struct llm_build_context {
                 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
                 cb(Vcur, "Vcur", il);
 
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                Qcur = ggml_rope_multi(
+                    ctx0,
+                    ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Qcur, "Qcur", il);
 
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                Kcur = ggml_rope_multi(
+                    ctx0,
+                    ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -12483,7 +12775,6 @@ struct llm_build_context {
             if (il == n_layer - 1) {
                 // skip computing output for unused tokens
                 struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                n_tokens = n_outputs;
                 cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
                 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
             }
@@ -12491,49 +12782,19 @@ struct llm_build_context {
             struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
             cb(ffn_inp, "ffn_inp", il);
 
-            // MoE branch
+            // feed-forward network
             cur = llm_build_norm(ctx0, ffn_inp, hparams,
                     model.layers[il].ffn_norm, NULL,
                     LLM_NORM_RMS, cb, il);
             cb(cur, "ffn_norm", il);
 
-            ggml_tensor * moe_out =
-                    llm_build_moe_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_gate_inp,
-                        model.layers[il].ffn_up_exps,
-                        model.layers[il].ffn_gate_exps,
-                        model.layers[il].ffn_down_exps,
-                        n_expert, n_expert_used,
-                        LLM_FFN_SILU, false,
-                        false, 0.0,
-                        cb, il);
-            cb(cur, "ffn_moe_out", il);
-
-            // FFN shared expert
-            {
-                ggml_tensor * cur_gate_inp = llm_build_lora_mm(lctx, ctx0, model.layers[il].ffn_gate_inp_shexp, cur);
-                cb(cur_gate_inp, "ffn_shexp_gate_inp", il);
-
-                // sigmoid
-                ggml_tensor * cur_gate = ggml_div(ctx0, ggml_silu(ctx0, cur_gate_inp), cur_gate_inp);
-                cb(cur_gate, "ffn_shexp_gate", il);
-
-                ggml_tensor * cur_ffn = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up_shexp,   NULL, NULL,
-                        model.layers[il].ffn_gate_shexp, NULL, NULL,
-                        model.layers[il].ffn_down_shexp, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur_ffn, "ffn_shexp", il);
-
-                ggml_tensor * ffn_shexp_out = ggml_mul(ctx0, cur_ffn, cur_gate);
-                cb(ffn_shexp_out, "ffn_shexp_out", il);
-
-                moe_out = ggml_add(ctx0, moe_out, ffn_shexp_out);
-                cb(moe_out, "ffn_out", il);
-
-                cur = moe_out;
-            }
+            cur = llm_build_ffn(ctx0, lctx, cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
 
             cur = ggml_add(ctx0, cur, ffn_inp);
             cur = lctx.cvec.apply_to(ctx0, cur, il);
@@ -12559,7 +12820,153 @@ struct llm_build_context {
         return gf;
     }
 
-    struct ggml_cgraph * build_phi2() {
+    struct ggml_cgraph * build_qwen2moe() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self_attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // MoE branch
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            ggml_tensor * moe_out =
+                    llm_build_moe_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_gate_inp,
+                        model.layers[il].ffn_up_exps,
+                        model.layers[il].ffn_gate_exps,
+                        model.layers[il].ffn_down_exps,
+                        n_expert, n_expert_used,
+                        LLM_FFN_SILU, false,
+                        false, 0.0,
+                        cb, il);
+            cb(cur, "ffn_moe_out", il);
+
+            // FFN shared expert
+            {
+                ggml_tensor * cur_gate_inp = llm_build_lora_mm(lctx, ctx0, model.layers[il].ffn_gate_inp_shexp, cur);
+                cb(cur_gate_inp, "ffn_shexp_gate_inp", il);
+
+                // sigmoid
+                ggml_tensor * cur_gate = ggml_div(ctx0, ggml_silu(ctx0, cur_gate_inp), cur_gate_inp);
+                cb(cur_gate, "ffn_shexp_gate", il);
+
+                ggml_tensor * cur_ffn = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up_shexp,   NULL, NULL,
+                        model.layers[il].ffn_gate_shexp, NULL, NULL,
+                        model.layers[il].ffn_down_shexp, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur_ffn, "ffn_shexp", il);
+
+                ggml_tensor * ffn_shexp_out = ggml_mul(ctx0, cur_ffn, cur_gate);
+                cb(ffn_shexp_out, "ffn_shexp_out", il);
+
+                moe_out = ggml_add(ctx0, moe_out, ffn_shexp_out);
+                cb(moe_out, "ffn_out", il);
+
+                cur = moe_out;
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
+    struct ggml_cgraph * build_phi2() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -13364,153 +13771,6 @@ struct llm_build_context {
         return gf;
     }
 
-    // ref: https://arxiv.org/abs/2203.03466
-    //      https://github.com/ggerganov/llama.cpp/issues/5276#issuecomment-1925774738
-    // based on the original build_llama() function
-    struct ggml_cgraph * build_minicpm() {
-        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
-
-        const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-        GGML_ASSERT(n_embd_head == hparams.n_rot);
-
-        const int64_t n_embd = hparams.n_embd;
-        //TODO: if the model varies, these parameters need to be read from the model
-        const int64_t n_embd_base = 256;
-        const float scale_embd  = 12.0f;
-        const float scale_depth = 1.4f;
-
-        struct ggml_tensor * cur;
-        struct ggml_tensor * inpL;
-
-        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
-
-        // scale the input embeddings
-        inpL = ggml_scale(ctx0, inpL, scale_embd);
-        cb(inpL, "inp_scaled", -1);
-
-        // inp_pos - contains the positions
-        struct ggml_tensor * inp_pos = build_inp_pos();
-
-        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-
-        for (int il = 0; il < n_layer; ++il) {
-            struct ggml_tensor * inpSA = inpL;
-
-            // norm
-            cur = llm_build_norm(ctx0, inpL, hparams,
-                    model.layers[il].attn_norm, NULL,
-                    LLM_NORM_RMS, cb, il);
-            cb(cur, "attn_norm", il);
-
-            // self-attention
-            {
-                // compute Q and K and RoPE them
-                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
-                cb(Qcur, "Qcur", il);
-                if (model.layers[il].bq) {
-                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
-                    cb(Qcur, "Qcur", il);
-                }
-
-                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
-                cb(Kcur, "Kcur", il);
-                if (model.layers[il].bk) {
-                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
-                    cb(Kcur, "Kcur", il);
-                }
-
-                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
-                cb(Vcur, "Vcur", il);
-                if (model.layers[il].bv) {
-                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
-                    cb(Vcur, "Vcur", il);
-                }
-
-                Qcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Qcur, "Qcur", il);
-
-                Kcur = ggml_rope_ext(
-                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                );
-                cb(Kcur, "Kcur", il);
-
-                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
-                        model.layers[il].wo, model.layers[il].bo,
-                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
-            }
-
-            if (il == n_layer - 1) {
-                // skip computing output for unused tokens
-                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-            }
-
-            // scale_res - scale the hidden states for residual connection
-            const float scale_res = scale_depth/sqrtf(float(n_layer));
-            cur = ggml_scale(ctx0, cur, scale_res);
-            cb(cur, "hidden_scaled", -1);
-
-            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-            cb(ffn_inp, "ffn_inp", il);
-
-            // feed-forward network
-            {
-                cur = llm_build_norm(ctx0, ffn_inp, hparams,
-                        model.layers[il].ffn_norm, NULL,
-                        LLM_NORM_RMS, cb, il);
-                cb(cur, "ffn_norm", il);
-
-                cur = llm_build_ffn(ctx0, lctx, cur,
-                        model.layers[il].ffn_up,   NULL, NULL,
-                        model.layers[il].ffn_gate, NULL, NULL,
-                        model.layers[il].ffn_down, NULL, NULL,
-                        NULL,
-                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
-                cb(cur, "ffn_out", il);
-            }
-
-            // scale the hidden states for residual connection
-            cur = ggml_scale(ctx0, cur, scale_res);
-            cb(cur, "hidden_scaled_ffn", -1);
-
-            cur = ggml_add(ctx0, cur, ffn_inp);
-            cur = lctx.cvec.apply_to(ctx0, cur, il);
-            cb(cur, "l_out", il);
-
-            // input for next layer
-            inpL = cur;
-        }
-
-        cur = inpL;
-
-        cur = llm_build_norm(ctx0, cur, hparams,
-                model.output_norm, NULL,
-                LLM_NORM_RMS, cb, -1);
-        cb(cur, "result_norm", -1);
-
-        // lm_head scaling
-        const float scale_lmhead = float(n_embd_base)/float(n_embd);
-        cur = ggml_scale(ctx0, cur, scale_lmhead);
-        cb(cur, "lmhead_scaling", -1);
-
-        // lm_head
-        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
-        cb(cur, "result_output", -1);
-
-        ggml_build_forward_expand(gf, cur);
-
-        return gf;
-    }
-
     struct ggml_cgraph * build_minicpm3() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
@@ -14416,6 +14676,130 @@ struct llm_build_context {
         return gf;
     }
 
+    struct ggml_cgraph * build_olmo2() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            cur = inpL;
+
+            // self_attention
+            {
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = llm_build_norm(ctx0, Qcur, hparams, model.layers[il].attn_q_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(Qcur, "Qcur_normed", il);
+
+                Kcur = llm_build_norm(ctx0, Kcur, hparams, model.layers[il].attn_k_norm, NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(Kcur, "Kcur_normed", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+
+                Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur_rope", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, nullptr,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur_rope", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+            }
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                    model.layers[il].attn_post_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_post_norm", il);
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // feed-forward network
+            cur = llm_build_ffn(ctx0, lctx, ffn_inp,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+            cb(cur, "ffn_out", il);
+
+            cur = llm_build_norm(ctx0, cur, hparams,
+                model.layers[il].ffn_post_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+            cb(cur, "ffn_post_norm", -1);
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cb(cur, "ffn_out", il);
+
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
     // based on the build_qwen2moe() function, changes:
     //   * removed shared experts
     //   * removed bias
@@ -14943,6 +15327,161 @@ struct llm_build_context {
         return gf;
     }
 
+    struct ggml_cgraph * build_deepseek() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
+
+        // mutable variable, needed during the last layer of the computation to skip unused tokens
+        int32_t n_tokens = this->n_tokens;
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, ubatch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+        const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            // norm
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                // rope freq factors for llama3; may return nullptr for llama2 and other models
+                struct ggml_tensor * rope_factors = build_rope_factors(il);
+
+                // compute Q and K and RoPE them
+                struct ggml_tensor * Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+                if (model.layers[il].bq) {
+                    Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    cb(Qcur, "Qcur", il);
+                }
+
+                struct ggml_tensor * Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+                if (model.layers[il].bk) {
+                    Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    cb(Kcur, "Kcur", il);
+                }
+
+                struct ggml_tensor * Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+                if (model.layers[il].bv) {
+                    Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    cb(Vcur, "Vcur", il);
+                }
+
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur", il);
+
+                cur = llm_build_kv(ctx0, lctx, kv_self, gf,
+                        model.layers[il].wo, model.layers[il].bo,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, kq_scale, cb, il);
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                n_tokens = n_outputs;
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "ffn_norm", il);
+
+            if ((uint32_t) il < hparams.n_layer_dense_lead) {
+                cur = llm_build_ffn(ctx0, lctx, cur,
+                        model.layers[il].ffn_up,   NULL, NULL,
+                        model.layers[il].ffn_gate, NULL, NULL,
+                        model.layers[il].ffn_down, NULL, NULL,
+                        NULL,
+                        LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                cb(cur, "ffn_out", il);
+            } else {
+                // MoE branch
+                ggml_tensor * moe_out =
+                        llm_build_moe_ffn(ctx0, lctx, cur,
+                            model.layers[il].ffn_gate_inp,
+                            model.layers[il].ffn_up_exps,
+                            model.layers[il].ffn_gate_exps,
+                            model.layers[il].ffn_down_exps,
+                            n_expert, n_expert_used,
+                            LLM_FFN_SILU, false,
+                            false, hparams.expert_weights_scale,
+                            cb, il);
+                cb(moe_out, "ffn_moe_out", il);
+
+                // FFN shared expert
+                {
+                    ggml_tensor * ffn_shexp = llm_build_ffn(ctx0, lctx, cur,
+                            model.layers[il].ffn_up_shexp,   NULL, NULL,
+                            model.layers[il].ffn_gate_shexp, NULL, NULL,
+                            model.layers[il].ffn_down_shexp, NULL, NULL,
+                            NULL,
+                            LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+                    cb(ffn_shexp, "ffn_shexp", il);
+
+                    cur = ggml_add(ctx0, moe_out, ffn_shexp);
+                    cb(cur, "ffn_out", il);
+                }
+            }
+
+            cur = ggml_add(ctx0, cur, ffn_inp);
+            cur = lctx.cvec.apply_to(ctx0, cur, il);
+            cb(cur, "l_out", il);
+
+            // input for next layer
+            inpL = cur;
+        }
+
+        cur = inpL;
+
+        cur = llm_build_norm(ctx0, cur, hparams,
+                model.output_norm, NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        // lm_head
+        cur = llm_build_lora_mm(lctx, ctx0, model.output, cur);
+
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
+
     struct ggml_cgraph * build_deepseek2() {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, llama_model_max_nodes(model), false);
 
@@ -16485,6 +17024,7 @@ static struct ggml_cgraph * llama_build_graph(
 
     switch (model.arch) {
         case LLM_ARCH_LLAMA:
+        case LLM_ARCH_MINICPM:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
             {
@@ -16536,6 +17076,11 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_qwen2();
             } break;
+        case LLM_ARCH_QWEN2VL:
+            {
+                lctx.n_pos_per_token = 4;
+                result = llm.build_qwen2vl();
+            } break;
         case LLM_ARCH_QWEN2MOE:
             {
                 result = llm.build_qwen2moe();
@@ -16568,10 +17113,6 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_internlm2();
             } break;
-        case LLM_ARCH_MINICPM:
-            {
-                result = llm.build_minicpm();
-            } break;
         case LLM_ARCH_MINICPM3:
             {
                 result = llm.build_minicpm3();
@@ -16608,6 +17149,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_olmo();
             } break;
+        case LLM_ARCH_OLMO2:
+            {
+                result = llm.build_olmo2();
+            } break;
         case LLM_ARCH_OLMOE:
             {
                 result = llm.build_olmoe();
@@ -16624,6 +17169,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_arctic();
             } break;
+        case LLM_ARCH_DEEPSEEK:
+            {
+                result = llm.build_deepseek();
+            } break;
         case LLM_ARCH_DEEPSEEK2:
             {
                 result = llm.build_deepseek2();
@@ -16754,8 +17303,8 @@ static void llama_set_inputs(llama_context & lctx, const llama_ubatch & ubatch)
 
     if (ubatch.pos && lctx.inp_pos) {
         const int64_t n_tokens = ubatch.n_tokens;
-
-        ggml_backend_tensor_set(lctx.inp_pos, ubatch.pos, 0, n_tokens*ggml_element_size(lctx.inp_pos));
+        auto n_pos = lctx.n_pos_per_token;
+        ggml_backend_tensor_set(lctx.inp_pos, ubatch.pos, 0, n_tokens*n_pos*ggml_element_size(lctx.inp_pos));
     }
 
     if (hparams.causal_attn || cparams.pooling_type == LLAMA_POOLING_TYPE_NONE) {
@@ -17250,8 +17799,9 @@ static enum ggml_status llama_graph_compute(
                     int   n_threads,
         ggml_threadpool * threadpool) {
     if (lctx.backend_cpu != nullptr) {
-        ggml_backend_cpu_set_threadpool(lctx.backend_cpu, threadpool);
-        ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data);
+        auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(lctx.backend_cpu));
+        auto * set_threadpool_fn = (decltype(ggml_backend_cpu_set_threadpool) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_set_threadpool");
+        set_threadpool_fn(lctx.backend_cpu, threadpool);
     }
 
     // set the number of threads for all the backends
@@ -18018,13 +18568,13 @@ static void llama_kv_cache_defrag_internal(struct llama_context & lctx) {
 static void llama_kv_cache_update_internal(struct llama_context & lctx) {
     bool need_reserve = false;
 
-    // apply K-shift if needed
-    if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE && lctx.kv_self.has_shift) {
-        if (lctx.model.arch == LLM_ARCH_DEEPSEEK2) { // not supported due to MLA
-            GGML_ABORT("Deepseek2 does not support K-shift");
+    if (lctx.kv_self.has_shift) {
+        if (!llama_kv_cache_can_shift(&lctx)) {
+            GGML_ABORT("The current context does not support K-shift");
         }
 
-        {
+        // apply K-shift if needed
+        if (lctx.model.hparams.rope_type != LLAMA_ROPE_TYPE_NONE) {
             ggml_backend_sched_reset(lctx.sched.get());
 
             ggml_cgraph * gf = llama_build_graph_k_shift(lctx);
@@ -18239,10 +18789,6 @@ static ggml_type llama_tensor_get_type(quantize_state_internal & qs, ggml_type n
             else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
                 new_type = GGML_TYPE_IQ3_S;
             }
-            else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 ||
-                     new_type == GGML_TYPE_Q4_0_8_8) {
-                new_type = GGML_TYPE_Q4_0;
-            }
             else if (ftype == LLAMA_FTYPE_MOSTLY_TQ1_0 || ftype == LLAMA_FTYPE_MOSTLY_TQ2_0) {
                 new_type = GGML_TYPE_Q4_K;
             }
@@ -18565,9 +19111,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_IQ4_XS:  default_type = GGML_TYPE_IQ4_XS;  break;
         case LLAMA_FTYPE_MOSTLY_IQ3_S:   default_type = GGML_TYPE_IQ3_S;   break;
         case LLAMA_FTYPE_MOSTLY_IQ3_M:   default_type = GGML_TYPE_IQ3_S;   break;
-        case LLAMA_FTYPE_MOSTLY_Q4_0_4_4: default_type = GGML_TYPE_Q4_0_4_4; break;
-        case LLAMA_FTYPE_MOSTLY_Q4_0_4_8: default_type = GGML_TYPE_Q4_0_4_8; break;
-        case LLAMA_FTYPE_MOSTLY_Q4_0_8_8: default_type = GGML_TYPE_Q4_0_8_8; break;
 
         default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
     }
@@ -18597,6 +19140,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     llama_model model;
     llm_load_arch(ml, model);
     llm_load_hparams(ml, model);
+    llm_load_stats(ml, model);
 
     struct quantize_state_internal qs(model, params);
 
@@ -18905,14 +19449,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 f32_data = (float *) f32_conv_buf.data();
             }
 
-            int chunk_size_multiplier = 1;
-            if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8 || new_type == GGML_TYPE_Q4_0_8_8) {
-                if ((new_type == GGML_TYPE_Q4_0_8_8) && (tensor->ne[1] % 8 != 0)) new_type = GGML_TYPE_Q4_0;
-                else if (tensor->ne[1] % 4 != 0) new_type = GGML_TYPE_Q4_0;
-                if (new_type == GGML_TYPE_Q4_0_8_8) chunk_size_multiplier = 8;
-                else if (new_type == GGML_TYPE_Q4_0_4_4 || new_type == GGML_TYPE_Q4_0_4_8) chunk_size_multiplier = 4;
-            }
-
             LLAMA_LOG_INFO("converting to %s .. ", ggml_type_name(new_type));
             fflush(stdout);
 
@@ -18925,8 +19461,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             const int64_t nrows = tensor->ne[1];
 
             static const int64_t min_chunk_size = 32 * 512;
-            const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row)) *
-                                       chunk_size_multiplier;
+            const int64_t chunk_size = (n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row));
 
             const int64_t nelements_matrix = tensor->ne[0] * tensor->ne[1];
             const int64_t nchunk = (nelements_matrix + chunk_size - 1)/chunk_size;
@@ -19167,6 +19702,7 @@ void llama_lora_adapter_free(struct llama_lora_adapter * adapter) {
 //
 struct llama_model_params llama_model_default_params() {
     struct llama_model_params result = {
+        /*.devices                     =*/ nullptr,
         /*.n_gpu_layers                =*/ 0,
         /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
         /*.main_gpu                    =*/ 0,
@@ -19284,7 +19820,11 @@ void llama_backend_init(void) {
 
 void llama_numa_init(enum ggml_numa_strategy numa) {
     if (numa != GGML_NUMA_STRATEGY_DISABLED) {
-        ggml_numa_init(numa);
+        auto * dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        GGML_ASSERT(dev && "CPU backend is not loaded");
+        auto * reg = ggml_backend_dev_backend_reg(dev);
+        auto * numa_init_fn = (decltype(ggml_numa_init) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_numa_init");
+        numa_init_fn(numa);
     }
 }
 
@@ -19375,19 +19915,24 @@ struct llama_model * llama_load_model_from_file(
     }
 
     // create list of devices to use with this model
-    // currently, we use all available devices
-    // TODO: rework API to give user more control over device selection
-    for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
-        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
-        switch (ggml_backend_dev_type(dev)) {
-            case GGML_BACKEND_DEVICE_TYPE_CPU:
-            case GGML_BACKEND_DEVICE_TYPE_ACCEL:
-                // skip CPU backends since they are handled separately
-                break;
+    if (params.devices) {
+        for (ggml_backend_dev_t * dev = params.devices; *dev; ++dev) {
+            model->devices.push_back(*dev);
+        }
+    } else {
+        // use all available devices
+        for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+            ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+            switch (ggml_backend_dev_type(dev)) {
+                case GGML_BACKEND_DEVICE_TYPE_CPU:
+                case GGML_BACKEND_DEVICE_TYPE_ACCEL:
+                    // skip CPU backends since they are handled separately
+                    break;
 
-            case GGML_BACKEND_DEVICE_TYPE_GPU:
-                model->devices.push_back(dev);
-                break;
+                case GGML_BACKEND_DEVICE_TYPE_GPU:
+                    model->devices.push_back(dev);
+                    break;
+            }
         }
     }
 
@@ -19558,9 +20103,6 @@ struct llama_context * llama_new_context_with_model(
                 __func__, n_ctx_per_seq, hparams.n_ctx_train);
     }
 
-    ctx->abort_callback      = params.abort_callback;
-    ctx->abort_callback_data = params.abort_callback_data;
-
     ctx->logits_all = params.logits_all;
 
     // build worst-case graph for encoder if a model contains encoder
@@ -19609,7 +20151,7 @@ struct llama_context * llama_new_context_with_model(
         }
 
         // add CPU backend
-        ctx->backend_cpu = ggml_backend_cpu_init();
+        ctx->backend_cpu = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
         if (ctx->backend_cpu == nullptr) {
             LLAMA_LOG_ERROR("%s: failed to initialize CPU backend\n", __func__);
             llama_free(ctx);
@@ -19629,6 +20171,8 @@ struct llama_context * llama_new_context_with_model(
             }
         }
 
+        llama_set_abort_callback(ctx, params.abort_callback, params.abort_callback_data);
+
         if (!llama_kv_cache_init(ctx->kv_self, ctx, type_k, type_v, kv_size, cparams.offload_kqv)) {
             LLAMA_LOG_ERROR("%s: llama_kv_cache_init() failed for self-attention cache\n", __func__);
             llama_free(ctx);
@@ -19674,7 +20218,8 @@ struct llama_context * llama_new_context_with_model(
             std::vector<ggml_backend_t> backend_ptrs;
             for (auto & backend : ctx->backends) {
                 auto * buft = ggml_backend_get_default_buffer_type(backend.get());
-                if (ggml_backend_is_cpu(backend.get()) && !model->devices.empty()) {
+                auto backend_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get()));
+                if (backend_type == GGML_BACKEND_DEVICE_TYPE_CPU && !model->devices.empty()) {
                     // use the host buffer of the first device CPU for faster transfer of the intermediate state
                     auto * dev = model->devices[0];
                     auto * host_buft = ggml_backend_dev_host_buffer_type(dev);
@@ -19702,7 +20247,8 @@ struct llama_context * llama_new_context_with_model(
             // pipeline parallelism requires support for async compute and events in all devices
             if (pipeline_parallel) {
                 for (auto & backend : ctx->backends) {
-                    if (ggml_backend_is_cpu(backend.get())) {
+                    auto dev_type = ggml_backend_dev_type(ggml_backend_get_device(backend.get()));
+                    if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU) {
                         // ignore CPU backend
                         continue;
                     }
@@ -19858,6 +20404,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_COMMAND_R:
         case LLM_ARCH_OLMO:
         case LLM_ARCH_ARCTIC:
+        case LLM_ARCH_DEEPSEEK:
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_CHATGLM:
         case LLM_ARCH_GRANITE:
@@ -19876,6 +20423,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_QWEN:
         case LLM_ARCH_QWEN2:
         case LLM_ARCH_QWEN2MOE:
+        case LLM_ARCH_OLMO2:
         case LLM_ARCH_OLMOE:
         case LLM_ARCH_PHI2:
         case LLM_ARCH_PHI3:
@@ -19890,6 +20438,9 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_MINICPM3:
             return LLAMA_ROPE_TYPE_NEOX;
 
+        case LLM_ARCH_QWEN2VL:
+            return LLAMA_ROPE_TYPE_MROPE;
+
         // all model arches should be listed explicitly here
         case LLM_ARCH_UNKNOWN:
             GGML_ABORT("unknown architecture");
@@ -19949,19 +20500,11 @@ int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t bu
 }
 
 uint64_t llama_model_size(const struct llama_model * model) {
-    uint64_t size = 0;
-    for (const auto & it : model->tensors_by_name) {
-        size += ggml_nbytes(it.second);
-    }
-    return size;
+    return model->n_bytes;
 }
 
 uint64_t llama_model_n_params(const struct llama_model * model) {
-    uint64_t nparams = 0;
-    for (const auto & it : model->tensors_by_name) {
-        nparams += ggml_nelements(it.second);
-    }
-    return nparams;
+    return model->n_elements;
 }
 
 struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name) {
@@ -20275,6 +20818,10 @@ void llama_kv_cache_update(struct llama_context * ctx) {
     llama_kv_cache_update_internal(*ctx);
 }
 
+bool llama_kv_cache_can_shift(struct llama_context * ctx) {
+    return !ctx->kv_self.recurrent && ctx->model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA
+}
+
 // deprecated
 size_t llama_get_state_size(struct llama_context * ctx) {
     return llama_state_get_size(ctx);
@@ -21259,6 +21806,14 @@ int32_t llama_n_threads_batch(struct llama_context * ctx) {
 void llama_set_abort_callback(struct llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) {
     ctx->abort_callback      = abort_callback;
     ctx->abort_callback_data = abort_callback_data;
+
+    for (auto & backend : ctx->backends) {
+        auto * reg = ggml_backend_dev_backend_reg(ggml_backend_get_device(backend.get()));
+        auto * set_abort_callback_fn = (ggml_backend_set_abort_callback_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_abort_callback");
+        if (set_abort_callback_fn) {
+            set_abort_callback_fn(backend.get(), ctx->abort_callback, ctx->abort_callback_data);
+        }
+    }
 }
 
 void llama_set_embeddings(struct llama_context * ctx, bool embeddings) {
@@ -21454,7 +22009,7 @@ float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i) {
                 throw std::runtime_error(format("negative index out of range [0, %d)", ctx->n_outputs));
             }
         } else if ((size_t) i >= ctx->output_ids.size()) {
-            throw std::runtime_error(format("out of range [0, %lu)", ctx->output_ids.size()));
+            throw std::runtime_error(format("out of range [0, %zu)", ctx->output_ids.size()));
         } else {
             j = ctx->output_ids[i];
         }
@@ -21625,18 +22180,113 @@ int32_t llama_detokenize(
 // chat templates
 //
 
+static llm_chat_template llama_chat_detect_template(const std::string & tmpl) {
+    if (LLM_CHAT_TEMPLATES.find(tmpl) != LLM_CHAT_TEMPLATES.end()) {
+        return LLM_CHAT_TEMPLATES.at(tmpl);
+    }
+    auto tmpl_contains = [&tmpl](const char * haystack) -> bool {
+        return tmpl.find(haystack) != std::string::npos;
+    };
+    if (tmpl_contains("<|im_start|>")) {
+        return LLM_CHAT_TEMPLATE_CHATML;
+    } else if (tmpl.find("mistral") == 0 || tmpl_contains("[INST]")) {
+        if (tmpl_contains("[SYSTEM_PROMPT]")) {
+            return LLM_CHAT_TEMPLATE_MISTRAL_V7;
+        } else if (
+            // catches official 'v1' template
+            tmpl_contains("' [INST] ' + system_message")
+            // catches official 'v3' and 'v3-tekken' templates
+            || tmpl_contains("[AVAILABLE_TOOLS]")
+        ) {
+            // Official mistral 'v1', 'v3' and 'v3-tekken' templates
+            // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/chat_templates.md
+            // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/templates.md
+            if (tmpl_contains(" [INST]")) {
+                return LLM_CHAT_TEMPLATE_MISTRAL_V1;
+            } else if (tmpl_contains("\"[INST]\"")) {
+                return LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN;
+            }
+            return LLM_CHAT_TEMPLATE_MISTRAL_V3;
+        } else {
+            // llama2 template and its variants
+            // [variant] support system message
+            // See: https://huggingface.co/blog/llama2#how-to-prompt-llama-2
+            bool support_system_message = tmpl_contains("<<SYS>>");
+            bool add_bos_inside_history = tmpl_contains("bos_token + '[INST]");
+            bool strip_message = tmpl_contains("content.strip()");
+            if (strip_message) {
+                return LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP;
+            } else if (add_bos_inside_history) {
+                return LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS;
+            } else if (support_system_message) {
+                return LLM_CHAT_TEMPLATE_LLAMA_2_SYS;
+            } else {
+                return LLM_CHAT_TEMPLATE_LLAMA_2;
+            }
+        }
+    } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) {
+        return LLM_CHAT_TEMPLATE_PHI_3;
+    } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
+        return LLM_CHAT_TEMPLATE_FALCON_3;
+    } else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) {
+        return LLM_CHAT_TEMPLATE_ZEPHYR;
+    } else if (tmpl_contains("bos_token + message['role']")) {
+        return LLM_CHAT_TEMPLATE_MONARCH;
+    } else if (tmpl_contains("<start_of_turn>")) {
+        return LLM_CHAT_TEMPLATE_GEMMA;
+    } else if (tmpl_contains("'\\n\\nAssistant: ' + eos_token")) {
+        // OrionStarAI/Orion-14B-Chat
+        return LLM_CHAT_TEMPLATE_ORION;
+    } else if (tmpl_contains("GPT4 Correct ")) {
+        // openchat/openchat-3.5-0106
+        return LLM_CHAT_TEMPLATE_OPENCHAT;
+    } else if (tmpl_contains("USER: ") && tmpl_contains("ASSISTANT: ")) {
+        // eachadea/vicuna-13b-1.1 (and Orca variant)
+        if (tmpl_contains("SYSTEM: ")) {
+            return LLM_CHAT_TEMPLATE_VICUNA_ORCA;
+        }
+        return LLM_CHAT_TEMPLATE_VICUNA;
+    } else if (tmpl_contains("### Instruction:") && tmpl_contains("<|EOT|>")) {
+        // deepseek-ai/deepseek-coder-33b-instruct
+        return LLM_CHAT_TEMPLATE_DEEPSEEK;
+    } else if (tmpl_contains("<|START_OF_TURN_TOKEN|>") && tmpl_contains("<|USER_TOKEN|>")) {
+        // CohereForAI/c4ai-command-r-plus
+        return LLM_CHAT_TEMPLATE_COMMAND_R;
+    } else if (tmpl_contains("<|start_header_id|>") && tmpl_contains("<|end_header_id|>")) {
+        return LLM_CHAT_TEMPLATE_LLAMA_3;
+    } else if (tmpl_contains("[gMASK]sop")) {
+        // chatglm3-6b
+        return LLM_CHAT_TEMPLATE_CHATGML_3;
+    } else if (tmpl_contains("[gMASK]<sop>")) {
+        return LLM_CHAT_TEMPLATE_CHATGML_4;
+    } else if (tmpl_contains(LU8("<用户>"))) {
+        // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
+        return LLM_CHAT_TEMPLATE_MINICPM;
+    } else if (tmpl_contains("'Assistant: ' + message['content'] + eos_token")) {
+        return LLM_CHAT_TEMPLATE_DEEPSEEK_2;
+    } else if (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]")) {
+        // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb
+        // EXAONE-3.0-7.8B-Instruct
+        return LLM_CHAT_TEMPLATE_EXAONE_3;
+    } else if (tmpl_contains("rwkv-world")) {
+        return LLM_CHAT_TEMPLATE_RWKV_WORLD;
+    } else if (tmpl_contains("<|start_of_role|>")) {
+        return LLM_CHAT_TEMPLATE_GRANITE;
+    } else if (tmpl_contains("message['role'] + additional_special_tokens[0] + message['content'] + additional_special_tokens[1]")) {
+        return LLM_CHAT_TEMPLATE_GIGACHAT;
+    }
+    return LLM_CHAT_TEMPLATE_UNKNOWN;
+}
+
 // Simple version of "llama_apply_chat_template" that only works with strings
 // This function uses heuristic checks to determine commonly used template. It is not a jinja parser.
 static int32_t llama_chat_apply_template_internal(
-    const std::string & tmpl,
+    const llm_chat_template tmpl,
     const std::vector<const llama_chat_message *> & chat,
     std::string & dest, bool add_ass) {
     // Taken from the research: https://github.com/ggerganov/llama.cpp/issues/5527
     std::stringstream ss;
-    auto tmpl_contains = [&tmpl](std::string haystack) -> bool {
-        return tmpl.find(haystack) != std::string::npos;
-    };
-    if (tmpl == "chatml" || tmpl_contains("<|im_start|>")) {
+    if (tmpl == LLM_CHAT_TEMPLATE_CHATML) {
         // chatml template
         for (auto message : chat) {
             ss << "<|im_start|>" << message->role << "\n" << message->content << "<|im_end|>\n";
@@ -21644,16 +22294,59 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|im_start|>assistant\n";
         }
-    } else if (tmpl == "llama2" || tmpl == "mistral" || tmpl_contains("[INST]")) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V7) {
+        // Official mistral 'v7' template
+        // See: https://huggingface.co/mistralai/Mistral-Large-Instruct-2411#basic-instruct-template-v7
+        for (auto message : chat) {
+            std::string role(message->role);
+            std::string content(message->content);
+            if (role == "system") {
+                ss << "[SYSTEM_PROMPT] " << content << "[/SYSTEM_PROMPT]";
+            } else if (role == "user") {
+                ss << "[INST] " << content << "[/INST]";
+            }
+            else {
+                ss << " " << content << "</s>";
+            }
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1
+            || tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3
+            || tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN) {
+        // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/chat_templates.md
+        // See: https://github.com/mistralai/cookbook/blob/main/concept-deep-dive/tokenization/templates.md
+        std::string leading_space = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V1 ? " " : "";
+        std::string trailing_space = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3_TEKKEN ? "" : " ";
+        bool trim_assistant_message = tmpl == LLM_CHAT_TEMPLATE_MISTRAL_V3;
+        bool is_inside_turn = false;
+        for (auto message : chat) {
+            if (!is_inside_turn) {
+                ss << leading_space << "[INST]" << trailing_space;
+                is_inside_turn = true;
+            }
+            std::string role(message->role);
+            std::string content(message->content);
+            if (role == "system") {
+                ss << content << "\n\n";
+            } else if (role == "user") {
+                ss << content << leading_space << "[/INST]";
+            } else {
+                ss << trailing_space << (trim_assistant_message ? trim(content) : content) << "</s>";
+                is_inside_turn = false;
+            }
+        }
+    } else if (
+            tmpl == LLM_CHAT_TEMPLATE_LLAMA_2
+            || tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS
+            || tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS
+            || tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP) {
         // llama2 template and its variants
         // [variant] support system message
-        bool support_system_message = tmpl_contains("<<SYS>>") || tmpl == "mistral";
-        // [variant] space before + after response
-        bool space_around_response = tmpl_contains("' ' + eos_token");
+        // See: https://huggingface.co/blog/llama2#how-to-prompt-llama-2
+        bool support_system_message = tmpl != LLM_CHAT_TEMPLATE_LLAMA_2;
         // [variant] add BOS inside history
-        bool add_bos_inside_history = tmpl_contains("bos_token + '[INST]");
+        bool add_bos_inside_history = tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_BOS;
         // [variant] trim spaces from the input message
-        bool strip_message = tmpl_contains("content.strip()");
+        bool strip_message = tmpl == LLM_CHAT_TEMPLATE_LLAMA_2_SYS_STRIP;
         // construct the prompt
         bool is_inside_turn = true; // skip BOS at the beginning
         ss << "[INST] ";
@@ -21674,12 +22367,11 @@ static int32_t llama_chat_apply_template_internal(
             } else if (role == "user") {
                 ss << content << " [/INST]";
             } else {
-                ss << (space_around_response ? " " : "") << content << (space_around_response ? " " : "") << "</s>";
+                ss << content << "</s>";
                 is_inside_turn = false;
             }
         }
-        // llama2 templates seem to not care about "add_generation_prompt"
-    } else if (tmpl == "phi3" || (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>"))) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_PHI_3) {
         // Phi 3
         for (auto message : chat) {
             std::string role(message->role);
@@ -21688,7 +22380,16 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|assistant|>\n";
         }
-    } else if (tmpl == "zephyr" || tmpl_contains("<|user|>")) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_FALCON_3) {
+        // Falcon 3
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>\n" << message->content << "\n";
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_ZEPHYR) {
         // zephyr template
         for (auto message : chat) {
             ss << "<|" << message->role << "|>" << "\n" << message->content << "<|endoftext|>\n";
@@ -21696,7 +22397,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|assistant|>\n";
         }
-    } else if (tmpl == "monarch" || tmpl_contains("bos_token + message['role']")) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_MONARCH) {
         // mlabonne/AlphaMonarch-7B template (the <s> is included inside history)
         for (auto message : chat) {
             std::string bos = (message == chat.front()) ? "" : "<s>"; // skip BOS for first message
@@ -21705,7 +22406,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<s>assistant\n";
         }
-    } else if (tmpl == "gemma" || tmpl == "gemma2" || tmpl_contains("<start_of_turn>")) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GEMMA) {
         // google/gemma-7b-it
         std::string system_prompt = "";
         for (auto message : chat) {
@@ -21727,7 +22428,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<start_of_turn>model\n";
         }
-    } else if (tmpl == "orion" || tmpl_contains("'\\n\\nAssistant: ' + eos_token")) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_ORION) {
         // OrionStarAI/Orion-14B-Chat
         std::string system_prompt = "";
         for (auto message : chat) {
@@ -21747,7 +22448,7 @@ static int32_t llama_chat_apply_template_internal(
                 ss << message->content << "</s>";
             }
         }
-    } else if (tmpl == "openchat" || tmpl_contains("GPT4 Correct ")) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_OPENCHAT) {
         // openchat/openchat-3.5-0106,
         for (auto message : chat) {
             std::string role(message->role);
@@ -21761,13 +22462,13 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "GPT4 Correct Assistant:";
         }
-    } else if (tmpl == "vicuna" || tmpl == "vicuna-orca" || (tmpl_contains("USER: ") && tmpl_contains("ASSISTANT: "))) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_VICUNA || tmpl == LLM_CHAT_TEMPLATE_VICUNA_ORCA) {
         // eachadea/vicuna-13b-1.1 (and Orca variant)
         for (auto message : chat) {
             std::string role(message->role);
             if (role == "system") {
                 // Orca-Vicuna variant uses a system prefix
-                if (tmpl == "vicuna-orca" || tmpl_contains("SYSTEM: ")) {
+                if (tmpl == LLM_CHAT_TEMPLATE_VICUNA_ORCA) {
                     ss << "SYSTEM: " << message->content << "\n";
                 } else {
                     ss << message->content << "\n\n";
@@ -21781,7 +22482,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "ASSISTANT:";
         }
-    } else if (tmpl == "deepseek" || (tmpl_contains("### Instruction:") && tmpl_contains("<|EOT|>"))) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_DEEPSEEK) {
         // deepseek-ai/deepseek-coder-33b-instruct
         for (auto message : chat) {
             std::string role(message->role);
@@ -21796,7 +22497,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "### Response:\n";
         }
-    } else if (tmpl == "command-r" || (tmpl_contains("<|START_OF_TURN_TOKEN|>") && tmpl_contains("<|USER_TOKEN|>"))) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_COMMAND_R) {
         // CohereForAI/c4ai-command-r-plus
         for (auto message : chat) {
             std::string role(message->role);
@@ -21811,7 +22512,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>";
         }
-    } else if (tmpl == "llama3" || (tmpl_contains("<|start_header_id|>") && tmpl_contains("<|end_header_id|>"))) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_LLAMA_3) {
         // Llama 3
         for (auto message : chat) {
             std::string role(message->role);
@@ -21820,7 +22521,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|start_header_id|>assistant<|end_header_id|>\n\n";
         }
-    } else if (tmpl == "chatglm3" || tmpl_contains("[gMASK]sop")) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_3) {
         // chatglm3-6b
         ss << "[gMASK]" << "sop";
         for (auto message : chat) {
@@ -21830,7 +22531,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|assistant|>";
         }
-    } else if (tmpl == "chatglm4" || tmpl_contains("[gMASK]<sop>")) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_4) {
         ss << "[gMASK]" << "<sop>";
         for (auto message : chat) {
             std::string role(message->role);
@@ -21839,7 +22540,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|assistant|>";
         }
-    } else if (tmpl == "minicpm" || tmpl_contains(LU8("<用户>"))) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_MINICPM) {
         // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
         for (auto message : chat) {
             std::string role(message->role);
@@ -21851,7 +22552,7 @@ static int32_t llama_chat_apply_template_internal(
                 ss << trim(message->content);
             }
         }
-    } else if (tmpl == "deepseek2" || tmpl_contains("'Assistant: ' + message['content'] + eos_token")) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_DEEPSEEK_2) {
         // DeepSeek-V2
         for (auto message : chat) {
             std::string role(message->role);
@@ -21866,7 +22567,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "Assistant:";
         }
-    } else if (tmpl == "exaone3" || (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]"))) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_EXAONE_3) {
         // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb
         // EXAONE-3.0-7.8B-Instruct
         for (auto message : chat) {
@@ -21882,7 +22583,7 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "[|assistant|]";
         }
-    } else if (tmpl == "rwkv-world" || tmpl_contains("rwkv-world")) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) {
         // this template requires the model to have "\n\n" as EOT token
         for (auto message : chat) {
             std::string role(message->role);
@@ -21892,7 +22593,7 @@ static int32_t llama_chat_apply_template_internal(
                 ss << message->content << "\n\n";
             }
         }
-    } else if (tmpl == "granite" || tmpl_contains("<|start_of_role|>")) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GRANITE) {
         // IBM Granite template
         for (const auto & message : chat) {
             std::string role(message->role);
@@ -21905,6 +22606,32 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|start_of_role|>assistant<|end_of_role|>\n";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GIGACHAT) {
+        // GigaChat template
+        bool has_system = !chat.empty() && std::string(chat[0]->role) == "system";
+
+        // Handle system message if present
+        if (has_system) {
+            ss << "<s>" << chat[0]->content << "<|message_sep|>";
+        } else {
+            ss << "<s>";
+        }
+
+        // Process remaining messages
+        for (size_t i = has_system ? 1 : 0; i < chat.size(); i++) {
+            std::string role(chat[i]->role);
+            if (role == "user") {
+                ss << "user<|role_sep|>" << chat[i]->content << "<|message_sep|>"
+                << "available functions<|role_sep|>[]<|message_sep|>";
+            } else if (role == "assistant") {
+                ss << "assistant<|role_sep|>" << chat[i]->content << "<|message_sep|>";
+            }
+        }
+
+        // Add generation prompt if needed
+        if (add_ass) {
+            ss << "assistant<|role_sep|>";
+        }
     } else {
         // template not supported
         return -1;
@@ -21944,7 +22671,11 @@ int32_t llama_chat_apply_template(
     }
 
     std::string formatted_chat;
-    int32_t res = llama_chat_apply_template_internal(curr_tmpl, chat_vec, formatted_chat, add_ass);
+    llm_chat_template detected_tmpl = llama_chat_detect_template(curr_tmpl);
+    if (detected_tmpl == LLM_CHAT_TEMPLATE_UNKNOWN) {
+        return -1;
+    }
+    int32_t res = llama_chat_apply_template_internal(detected_tmpl, chat_vec, formatted_chat, add_ass);
     if (res < 0) {
         return res;
     }
@@ -21954,6 +22685,15 @@ int32_t llama_chat_apply_template(
     return res;
 }
 
+int32_t llama_chat_builtin_templates(const char ** output, size_t len) {
+    auto it = LLM_CHAT_TEMPLATES.begin();
+    for (size_t i = 0; i < std::min(len, LLM_CHAT_TEMPLATES.size()); i++) {
+        output[i] = it->first.c_str();
+        std::advance(it, 1);
+    }
+    return (int32_t) LLM_CHAT_TEMPLATES.size();
+}
+
 //
 // sampling
 //
@@ -22000,33 +22740,23 @@ int llama_split_prefix(char * dest, size_t maxlen, const char * split_path, int
 }
 
 const char * llama_print_system_info(void) {
-    ggml_cpu_init(); // some ARM features are detected at runtime
-
     static std::string s;
 
-    s  = "";
-    s += "AVX = "         + std::to_string(ggml_cpu_has_avx())         + " | ";
-    s += "AVX_VNNI = "    + std::to_string(ggml_cpu_has_avx_vnni())    + " | ";
-    s += "AVX2 = "        + std::to_string(ggml_cpu_has_avx2())        + " | ";
-    s += "AVX512 = "      + std::to_string(ggml_cpu_has_avx512())      + " | ";
-    s += "AVX512_VBMI = " + std::to_string(ggml_cpu_has_avx512_vbmi()) + " | ";
-    s += "AVX512_VNNI = " + std::to_string(ggml_cpu_has_avx512_vnni()) + " | ";
-    s += "AVX512_BF16 = " + std::to_string(ggml_cpu_has_avx512_bf16()) + " | ";
-    s += "AMX_INT8 = "    + std::to_string(ggml_cpu_has_amx_int8())    + " | ";
-    s += "FMA = "         + std::to_string(ggml_cpu_has_fma())         + " | ";
-    s += "NEON = "        + std::to_string(ggml_cpu_has_neon())        + " | ";
-    s += "SVE = "         + std::to_string(ggml_cpu_has_sve())         + " | ";
-    s += "ARM_FMA = "     + std::to_string(ggml_cpu_has_arm_fma())     + " | ";
-    s += "F16C = "        + std::to_string(ggml_cpu_has_f16c())        + " | ";
-    s += "FP16_VA = "     + std::to_string(ggml_cpu_has_fp16_va())     + " | ";
-    s += "RISCV_VECT = "  + std::to_string(ggml_cpu_has_riscv_v())     + " | ";
-    s += "WASM_SIMD = "   + std::to_string(ggml_cpu_has_wasm_simd())   + " | ";
-    s += "BLAS = "        + std::to_string(ggml_cpu_has_blas())        + " | ";
-    s += "SSE3 = "        + std::to_string(ggml_cpu_has_sse3())        + " | ";
-    s += "SSSE3 = "       + std::to_string(ggml_cpu_has_ssse3())       + " | ";
-    s += "VSX = "         + std::to_string(ggml_cpu_has_vsx())         + " | ";
-    s += "MATMUL_INT8 = " + std::to_string(ggml_cpu_has_matmul_int8()) + " | ";
-    s += "LLAMAFILE = "   + std::to_string(ggml_cpu_has_llamafile())   + " | ";
+    for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
+        auto * reg = ggml_backend_reg_get(i);
+        auto * get_features_fn = (ggml_backend_get_features_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_get_features");
+        if (get_features_fn) {
+            ggml_backend_feature * features = get_features_fn(reg);
+            s += ggml_backend_reg_name(reg);
+            s += " : ";
+            for (; features->name; features++) {
+                s += features->name;
+                s += " = ";
+                s += features->value;
+                s += " | ";
+            }
+        }
+    }
 
     return s.c_str();
 }
@@ -22067,28 +22797,6 @@ void llama_perf_context_reset(struct llama_context * ctx) {
     ctx->t_p_eval_us = ctx->n_p_eval = 0;
 }
 
-void llama_perf_dump_yaml(FILE * stream, const llama_context * ctx) {
-    fprintf(stream, "\n");
-    fprintf(stream, "###########\n");
-    fprintf(stream, "# Timings #\n");
-    fprintf(stream, "###########\n");
-    fprintf(stream, "\n");
-
-    fprintf(stream, "mst_eval: %.2f  # ms / token during generation\n",
-            1.0e-3 * ctx->t_eval_us / ctx->n_eval);
-    fprintf(stream, "mst_p_eval: %.2f  # ms / token during prompt processing\n",
-            1.0e-3 * ctx->t_p_eval_us / ctx->n_p_eval);
-    fprintf(stream, "n_eval: %d  # number of tokens generated (excluding the first one)\n", ctx->n_eval);
-    fprintf(stream, "n_p_eval: %d  # number of tokens processed in batches at the beginning\n", ctx->n_p_eval);
-    fprintf(stream, "t_eval_us: %" PRId64 "  # total microseconds spent generating tokens\n", ctx->t_eval_us);
-    fprintf(stream, "t_load_us: %" PRId64 "  # total microseconds spent loading the model\n", ctx->t_load_us);
-    fprintf(stream, "t_p_eval_us: %" PRId64 "  # total microseconds spent prompt processing\n", ctx->t_p_eval_us);
-    fprintf(stream, "ts_eval: %.2f  # tokens / second during generation\n",
-            1.0e6 * ctx->n_eval / ctx->t_eval_us);
-    fprintf(stream, "ts_p_eval: %.2f  # tokens / second during prompt processing\n",
-            1.0e6 * ctx->n_p_eval / ctx->t_p_eval_us);
-}
-
 // For internal test use
 const std::vector<std::pair<std::string, struct ggml_tensor *>> & llama_internal_get_tensor_map(
     struct llama_context * ctx
diff --git a/examples/talk-llama/llama.h b/examples/talk-llama/llama.h
index 5e742642eec..efbb27d2152 100644
--- a/examples/talk-llama/llama.h
+++ b/examples/talk-llama/llama.h
@@ -104,12 +104,15 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH   = 24,
         LLAMA_VOCAB_PRE_TYPE_EXAONE         = 25,
         LLAMA_VOCAB_PRE_TYPE_CHAMELEON      = 26,
+        LLAMA_VOCAB_PRE_TYPE_MINERVA        = 27,
     };
 
     enum llama_rope_type {
-        LLAMA_ROPE_TYPE_NONE = -1,
-        LLAMA_ROPE_TYPE_NORM = 0,
-        LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX,
+        LLAMA_ROPE_TYPE_NONE   = -1,
+        LLAMA_ROPE_TYPE_NORM   = 0,
+        LLAMA_ROPE_TYPE_NEOX   = GGML_ROPE_TYPE_NEOX,
+        LLAMA_ROPE_TYPE_MROPE  = GGML_ROPE_TYPE_MROPE,
+        LLAMA_ROPE_TYPE_VISION = GGML_ROPE_TYPE_VISION,
     };
 
     enum llama_token_type { //TODO: remove, required until per token attributes are available from GGUF file
@@ -171,9 +174,9 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_IQ4_XS        = 30, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_IQ1_M         = 31, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_BF16          = 32, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0_4_4      = 33, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0_4_8      = 34, // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0_8_8      = 35, // except 1d tensors
+        //LLAMA_FTYPE_MOSTLY_Q4_0_4_4      = 33, // removed from gguf files, use Q4_0 and runtime repack
+        //LLAMA_FTYPE_MOSTLY_Q4_0_4_8      = 34, // removed from gguf files, use Q4_0 and runtime repack
+        //LLAMA_FTYPE_MOSTLY_Q4_0_8_8      = 35, // removed from gguf files, use Q4_0 and runtime repack
         LLAMA_FTYPE_MOSTLY_TQ1_0         = 36, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_TQ2_0         = 37, // except 1d tensors
 
@@ -185,7 +188,8 @@ extern "C" {
         LLAMA_ROPE_SCALING_TYPE_NONE        = 0,
         LLAMA_ROPE_SCALING_TYPE_LINEAR      = 1,
         LLAMA_ROPE_SCALING_TYPE_YARN        = 2,
-        LLAMA_ROPE_SCALING_TYPE_MAX_VALUE   = LLAMA_ROPE_SCALING_TYPE_YARN,
+        LLAMA_ROPE_SCALING_TYPE_LONGROPE    = 3,
+        LLAMA_ROPE_SCALING_TYPE_MAX_VALUE   = LLAMA_ROPE_SCALING_TYPE_LONGROPE,
     };
 
     enum llama_pooling_type {
@@ -272,6 +276,9 @@ extern "C" {
     };
 
     struct llama_model_params {
+        // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used)
+        ggml_backend_dev_t * devices;
+
         int32_t n_gpu_layers; // number of layers to store in VRAM
         enum llama_split_mode split_mode; // how to split the model across multiple GPUs
 
@@ -451,6 +458,7 @@ extern "C" {
     // Functions to access the model's GGUF metadata scalar values
     // - The functions return the length of the string on success, or -1 on failure
     // - The output string is always null-terminated and cleared on failure
+    // - When retrieving a string, an extra byte must be allocated to account for the null terminator
     // - GGUF array values are not supported by these functions
 
     // Get metadata value as a string by key name
@@ -667,6 +675,9 @@ extern "C" {
     // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
     LLAMA_API void llama_kv_cache_update(struct llama_context * ctx);
 
+    // Check if the context supports KV cache shifting
+    LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx);
+
     //
     // State / sessions
     //
@@ -984,6 +995,9 @@ extern "C" {
                                   char * buf,
                                int32_t   length);
 
+    // Get list of built-in chat templates
+    LLAMA_API int32_t llama_chat_builtin_templates(const char ** output, size_t len);
+
     //
     // Sampling API
     //
@@ -1125,16 +1139,12 @@ extern "C" {
                           const char * grammar_str,
                           const char * grammar_root);
 
+    /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first.
     LLAMA_API struct llama_sampler * llama_sampler_init_penalties(
-                             int32_t   n_vocab,         // llama_n_vocab()
-                         llama_token   special_eos_id,  // llama_token_eos()
-                         llama_token   linefeed_id,     // llama_token_nl()
-                             int32_t   penalty_last_n,  // last n tokens to penalize (0 = disable penalty, -1 = context size)
-                               float   penalty_repeat,  // 1.0 = disabled
-                               float   penalty_freq,    // 0.0 = disabled
-                               float   penalty_present, // 0.0 = disabled
-                                bool   penalize_nl,     // consider newlines as a repeatable token
-                                bool   ignore_eos);     // ignore the end-of-sequence token
+                             int32_t   penalty_last_n,   // last n tokens to penalize (0 = disable penalty, -1 = context size)
+                               float   penalty_repeat,   // 1.0 = disabled
+                               float   penalty_freq,     // 0.0 = disabled
+                               float   penalty_present); // 0.0 = disabled
 
     ///  @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982
     LLAMA_API struct llama_sampler *    llama_sampler_init_dry(
@@ -1244,8 +1254,6 @@ extern "C" {
     LLAMA_API void                           llama_perf_sampler_print(const struct llama_sampler * chain);
     LLAMA_API void                           llama_perf_sampler_reset(      struct llama_sampler * chain);
 
-    LLAMA_API void llama_perf_dump_yaml(FILE * stream, const struct llama_context * ctx);
-
 #ifdef __cplusplus
 }
 #endif
diff --git a/examples/talk-llama/unicode.cpp b/examples/talk-llama/unicode.cpp
index 50b35bbbc91..8ed6b1a51c2 100644
--- a/examples/talk-llama/unicode.cpp
+++ b/examples/talk-llama/unicode.cpp
@@ -71,15 +71,15 @@ uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset) {
     throw std::invalid_argument("failed to convert utf8 to codepoint");
 }
 
-//static std::vector<uint16_t> unicode_cpt_to_utf16(uint32_t cp) {
+//static std::vector<uint16_t> unicode_cpt_to_utf16(uint32_t cpt) {
 //    std::vector<uint16_t> result;
-//    if (/* 0x0000 <= cp && */ cp <= 0xffff) {
-//        result.emplace_back(cp);
+//    if (/* 0x0000 <= cpt && */ cpt <= 0xffff) {
+//        result.emplace_back(cpt);
 //        return result;
 //    }
-//    if (0x10000 <= cp && cp <= 0x10ffff) {
-//        result.emplace_back(0xd800 | ((cp - 0x10000) >> 10));
-//        result.emplace_back(0xdc00 | ((cp - 0x10000) & 0x03ff));
+//    if (0x10000 <= cpt && cpt <= 0x10ffff) {
+//        result.emplace_back(0xd800 | ((cpt - 0x10000) >> 10));
+//        result.emplace_back(0xdc00 | ((cpt - 0x10000) & 0x03ff));
 //        return result;
 //    }
 //    throw std::invalid_argument("failed to convert codepoint to utf16");
@@ -120,8 +120,8 @@ uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset) {
 //    return result;
 //}
 
-static std::vector<codepoint_flags> unicode_cpt_flags_array() {
-    std::vector<codepoint_flags> cpt_flags(MAX_CODEPOINTS, codepoint_flags::UNDEFINED);
+static std::vector<unicode_cpt_flags> unicode_cpt_flags_array() {
+    std::vector<unicode_cpt_flags> cpt_flags(MAX_CODEPOINTS, unicode_cpt_flags::UNDEFINED);
 
     assert (unicode_ranges_flags.begin()[0].first == 0);
     assert (unicode_ranges_flags.begin()[unicode_ranges_flags.size()-1].first == MAX_CODEPOINTS);
@@ -201,7 +201,18 @@ static std::unordered_map<std::string, uint8_t> unicode_utf8_to_byte_map() {
 }
 
 static inline std::wstring unicode_wstring_from_utf8(const std::string & s) {
+#if defined(__clang__)
+    // disable C++17 deprecation warning for std::codecvt_utf8
+#    pragma clang diagnostic push
+#    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#endif
+
     std::wstring_convert<std::codecvt_utf8<wchar_t>> conv;
+
+#if defined(__clang__)
+#    pragma clang diagnostic pop
+#endif
+
     return conv.from_bytes(s);
 }
 
@@ -242,8 +253,8 @@ static std::vector<size_t> unicode_regex_split_custom_gpt2(const std::string & t
             return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
         };
 
-        auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
-            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : codepoint_flags{};
+        auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
         };
 
         size_t _prev_end = offset_ini;
@@ -360,8 +371,8 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
             return (offset_ini <= pos && pos < offset_end) ? cpts[pos] : OUT_OF_RANGE;
         };
 
-        auto _get_flags = [&] (const size_t pos) -> codepoint_flags {
-            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags(cpts[pos]) : codepoint_flags{};
+        auto _get_flags = [&] (const size_t pos) -> unicode_cpt_flags {
+            return (offset_ini <= pos && pos < offset_end) ? unicode_cpt_flags_from_cpt(cpts[pos]) : unicode_cpt_flags{};
         };
 
         size_t _prev_end = offset_ini;
@@ -561,29 +572,29 @@ static std::vector<size_t> unicode_regex_split_custom(const std::string & text,
 // interface
 //
 
-std::string unicode_cpt_to_utf8(uint32_t cp) {
+std::string unicode_cpt_to_utf8(uint32_t cpt) {
     std::string result;
 
-    if (/* 0x00 <= cp && */ cp <= 0x7f) {
-        result.push_back(cp);
+    if (/* 0x00 <= cpt && */ cpt <= 0x7f) {
+        result.push_back(cpt);
         return result;
     }
-    if (0x80 <= cp && cp <= 0x7ff) {
-        result.push_back(0xc0 | ((cp >> 6) & 0x1f));
-        result.push_back(0x80 | (cp & 0x3f));
+    if (0x80 <= cpt && cpt <= 0x7ff) {
+        result.push_back(0xc0 | ((cpt >> 6) & 0x1f));
+        result.push_back(0x80 | (cpt & 0x3f));
         return result;
     }
-    if (0x800 <= cp && cp <= 0xffff) {
-        result.push_back(0xe0 | ((cp >> 12) & 0x0f));
-        result.push_back(0x80 | ((cp >> 6) & 0x3f));
-        result.push_back(0x80 | (cp & 0x3f));
+    if (0x800 <= cpt && cpt <= 0xffff) {
+        result.push_back(0xe0 | ((cpt >> 12) & 0x0f));
+        result.push_back(0x80 | ((cpt >> 6) & 0x3f));
+        result.push_back(0x80 | (cpt & 0x3f));
         return result;
     }
-    if (0x10000 <= cp && cp <= 0x10ffff) {
-        result.push_back(0xf0 | ((cp >> 18) & 0x07));
-        result.push_back(0x80 | ((cp >> 12) & 0x3f));
-        result.push_back(0x80 | ((cp >> 6) & 0x3f));
-        result.push_back(0x80 | (cp & 0x3f));
+    if (0x10000 <= cpt && cpt <= 0x10ffff) {
+        result.push_back(0xf0 | ((cpt >> 18) & 0x07));
+        result.push_back(0x80 | ((cpt >> 12) & 0x3f));
+        result.push_back(0x80 | ((cpt >> 6) & 0x3f));
+        result.push_back(0x80 | (cpt & 0x3f));
         return result;
     }
 
@@ -613,19 +624,19 @@ std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8) {
     return result;
 }
 
-codepoint_flags unicode_cpt_flags(const uint32_t cp) {
-    static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+unicode_cpt_flags unicode_cpt_flags_from_cpt(const uint32_t cpt) {
+    static const unicode_cpt_flags undef(unicode_cpt_flags::UNDEFINED);
     static const auto cpt_flags = unicode_cpt_flags_array();
-    return cp < cpt_flags.size() ? cpt_flags[cp] : undef;
+    return cpt < cpt_flags.size() ? cpt_flags[cpt] : undef;
 }
 
-codepoint_flags unicode_cpt_flags(const std::string & utf8) {
-    static const codepoint_flags undef(codepoint_flags::UNDEFINED);
+unicode_cpt_flags unicode_cpt_flags_from_utf8(const std::string & utf8) {
+    static const unicode_cpt_flags undef(unicode_cpt_flags::UNDEFINED);
     if (utf8.empty()) {
         return undef;  // undefined
     }
     size_t offset = 0;
-    return unicode_cpt_flags(unicode_cpt_from_utf8(utf8, offset));
+    return unicode_cpt_flags_from_cpt(unicode_cpt_from_utf8(utf8, offset));
 }
 
 std::string unicode_byte_to_utf8(uint8_t byte) {
@@ -638,41 +649,41 @@ uint8_t unicode_utf8_to_byte(const std::string & utf8) {
     return map.at(utf8);
 }
 
-uint32_t unicode_tolower(uint32_t cp) {
+uint32_t unicode_tolower(uint32_t cpt) {
     // binary search
-    auto it = std::lower_bound(unicode_map_lowercase.begin(), unicode_map_lowercase.end(), cp,
+    auto it = std::lower_bound(unicode_map_lowercase.begin(), unicode_map_lowercase.end(), cpt,
         [](const std::pair<uint32_t, uint32_t> & pair, uint32_t value) {
             return pair.first < value;
         });
-    if (it != unicode_map_lowercase.end() && it->first == cp) {
+    if (it != unicode_map_lowercase.end() && it->first == cpt) {
         return it->second;
     }
-    return cp;  // Return the original code point if no lowercase mapping is found
+    return cpt;  // Return the original code point if no lowercase mapping is found
 }
 
 std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs) {
     // unicode categories
     static const std::map<std::string, int> k_ucat_enum = {
-        { "\\p{N}", codepoint_flags::NUMBER },
-        { "\\p{L}", codepoint_flags::LETTER },
-        { "\\p{P}", codepoint_flags::PUNCTUATION },
+        { "\\p{N}", unicode_cpt_flags::NUMBER },
+        { "\\p{L}", unicode_cpt_flags::LETTER },
+        { "\\p{P}", unicode_cpt_flags::PUNCTUATION },
     };
 
     static const std::map<int, int> k_ucat_cpt = {
-        { codepoint_flags::NUMBER,        0xD1 },
-        { codepoint_flags::LETTER,        0xD2 },
-        { codepoint_flags::PUNCTUATION,   0xD3 },
+        { unicode_cpt_flags::NUMBER,      0xD1 },
+        { unicode_cpt_flags::LETTER,      0xD2 },
+        { unicode_cpt_flags::PUNCTUATION, 0xD3 },
     };
 
     static const std::map<int, std::string> k_ucat_map = {
-        { codepoint_flags::NUMBER,        "\x30-\x39" }, // 0-9
-        { codepoint_flags::LETTER,        "\x41-\x5A\x61-\x7A" }, // A-Za-z
-        { codepoint_flags::PUNCTUATION,   "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
+        { unicode_cpt_flags::NUMBER,      "\x30-\x39" }, // 0-9
+        { unicode_cpt_flags::LETTER,      "\x41-\x5A\x61-\x7A" }, // A-Za-z
+        { unicode_cpt_flags::PUNCTUATION, "\x21-\x23\x25-\x2A\x2C-\x2F\x3A-\x3B\x3F-\x40\\\x5B-\\\x5D\x5F\\\x7B\\\x7D" }, // !-#%-*,-/:-;?-@\[-\]_\{\}
     };
 
     // compute collapsed codepoints only if needed by at least one regex
     bool need_collapse = false;
-    for (auto & regex_expr : regex_exprs) {
+    for (const auto & regex_expr : regex_exprs) {
         // search for unicode categories
         for (const auto & ucat : k_ucat_enum) {
             if (std::string::npos != regex_expr.find(ucat.first)) {
@@ -698,7 +709,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                 continue;
             }
 
-            const auto flags = unicode_cpt_flags(cpts[i]);
+            const auto flags = unicode_cpt_flags_from_cpt(cpts[i]);
 
             if (flags.is_whitespace) {
                 //NOTE: C++ std::regex \s does not mach 0x85, Rust and Python regex does.
@@ -714,7 +725,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
 
     std::vector<size_t> bpe_offsets = { cpts.size() };
 
-    for (auto & regex_expr : regex_exprs) {
+    for (const auto & regex_expr : regex_exprs) {
         // first, see if we have an efficient custom regex implementation
         auto tmp = unicode_regex_split_custom(text, regex_expr, bpe_offsets);
 
@@ -728,7 +739,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
             // if a unicode category is used in the regex, we use the collapsed text and replace the unicode category
             // with the corresponding collapsed representation
             bool use_collapsed = false;
-            for (auto & ucat : k_ucat_enum) {
+            for (const auto & ucat : k_ucat_enum) {
                 if (std::string::npos != regex_expr.find(ucat.first)) {
                     use_collapsed = true;
                     break;
@@ -794,7 +805,7 @@ std::vector<std::string> unicode_regex_split(const std::string & text, const std
                 // std::wregex \s does not mach non-ASCII whitespaces, using 0x0B as fallback
                 std::wstring wtext(cpts.begin(), cpts.end());
                 for (size_t i = 0; i < wtext.size(); ++i) {
-                    if (wtext[i] > 0x7F && unicode_cpt_flags(wtext[i]).is_whitespace) {
+                    if (wtext[i] > 0x7F && unicode_cpt_flags_from_cpt(wtext[i]).is_whitespace) {
                         wtext[i] = 0x0B;
                     }
                 }
diff --git a/examples/talk-llama/unicode.h b/examples/talk-llama/unicode.h
index 008532a242a..c27098df7d4 100644
--- a/examples/talk-llama/unicode.h
+++ b/examples/talk-llama/unicode.h
@@ -4,9 +4,7 @@
 #include <string>
 #include <vector>
 
-// TODO: prefix all symbols with "llama_"
-
-struct codepoint_flags {
+struct unicode_cpt_flags {
     enum {
         UNDEFINED       = 0x0001,
         NUMBER          = 0x0002,  // regex: \p{N}
@@ -35,7 +33,7 @@ struct codepoint_flags {
     uint16_t is_nfd         : 1;
 
     // decode from uint16
-    inline codepoint_flags(const uint16_t flags=0) {
+    inline unicode_cpt_flags(const uint16_t flags = 0) {
         *reinterpret_cast<uint16_t*>(this) = flags;
     }
 
@@ -50,18 +48,19 @@ struct codepoint_flags {
 
 size_t unicode_len_utf8(char src);
 
-std::string unicode_cpt_to_utf8(uint32_t cp);
-uint32_t unicode_cpt_from_utf8(const std::string & utf8, size_t & offset);
+std::string unicode_cpt_to_utf8  (uint32_t cpt);
+uint32_t    unicode_cpt_from_utf8(const std::string & utf8, size_t & offset);
+
 std::vector<uint32_t> unicode_cpts_from_utf8(const std::string & utf8);
 
 std::vector<uint32_t> unicode_cpts_normalize_nfd(const std::vector<uint32_t> & cpts);
 
-codepoint_flags unicode_cpt_flags(const uint32_t cp);
-codepoint_flags unicode_cpt_flags(const std::string & utf8);
+unicode_cpt_flags unicode_cpt_flags_from_cpt (uint32_t cpt);
+unicode_cpt_flags unicode_cpt_flags_from_utf8(const std::string & utf8);
 
 std::string unicode_byte_to_utf8(uint8_t byte);
-uint8_t unicode_utf8_to_byte(const std::string & utf8);
+uint8_t     unicode_utf8_to_byte(const std::string & utf8);
 
-uint32_t unicode_tolower(uint32_t cp);
+uint32_t unicode_tolower(uint32_t cpt);
 
 std::vector<std::string> unicode_regex_split(const std::string & text, const std::vector<std::string> & regex_exprs);
diff --git a/examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt b/examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt
index 2e44317cb21..9bf4c669560 100644
--- a/examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt
+++ b/examples/whisper.android/lib/src/main/jni/whisper/CMakeLists.txt
@@ -2,11 +2,11 @@ cmake_minimum_required(VERSION 3.10)
 
 project(whisper.cpp)
 
-set(CMAKE_CXX_STANDARD 11)
+set(CMAKE_CXX_STANDARD 17)
 set(WHISPER_LIB_DIR ${CMAKE_SOURCE_DIR}/../../../../../../..)
 
 # Path to external GGML, otherwise uses the copy in whisper.cpp.
-option(GGML_HOME       "whisper: Path to external GGML source" OFF)
+option(GGML_HOME "whisper: Path to external GGML source" OFF)
 
 set(
     SOURCE_FILES
@@ -14,16 +14,24 @@ set(
     ${CMAKE_SOURCE_DIR}/jni.c
     )
 
+# TODO: this needs to be updated to work with the new ggml CMakeLists
+
 if (NOT GGML_HOME)
     set(
         SOURCE_FILES
         ${SOURCE_FILES}
         ${WHISPER_LIB_DIR}/ggml/src/ggml.c
-        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu.c
-        ${WHISPER_LIB_DIR}/ggml/src/ggml-aarch64.c
         ${WHISPER_LIB_DIR}/ggml/src/ggml-alloc.c
         ${WHISPER_LIB_DIR}/ggml/src/ggml-backend.cpp
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-backend-reg.cpp
         ${WHISPER_LIB_DIR}/ggml/src/ggml-quants.c
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-threading.cpp
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu.c
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu.cpp
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-hbm.cpp
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-quants.c
+        ${WHISPER_LIB_DIR}/ggml/src/ggml-cpu/ggml-cpu-traits.cpp
         )
 endif()
 
@@ -81,3 +89,5 @@ include_directories(${WHISPER_LIB_DIR}/src)
 include_directories(${WHISPER_LIB_DIR}/include)
 include_directories(${WHISPER_LIB_DIR}/ggml/include)
 include_directories(${WHISPER_LIB_DIR}/ggml/src)
+include_directories(${WHISPER_LIB_DIR}/ggml/src/ggml-cpu)
+
diff --git a/examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj b/examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj
index b1287f78916..fcf0614179e 100644
--- a/examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj
+++ b/examples/whisper.objc/whisper.objc.xcodeproj/project.pbxproj
@@ -25,6 +25,11 @@
 		18ABE15A2AF556340044A204 /* ggml-backend.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18ABE1572AF556340044A204 /* ggml-backend.cpp */; };
 		18ABE15B2AF556340044A204 /* ggml-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 18ABE1592AF556340044A204 /* ggml-quants.c */; };
 		18E864A92CE73C1E0094B8B3 /* ggml-cpu.c in Sources */ = {isa = PBXBuildFile; fileRef = 18E864A82CE73C1E0094B8B3 /* ggml-cpu.c */; };
+		18F8C0BC2CEDF4DC00CAD607 /* ggml-threading.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0BB2CEDF4DC00CAD607 /* ggml-threading.cpp */; };
+		18F8C0BE2CEDF50700CAD607 /* ggml-cpu.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0BD2CEDF50700CAD607 /* ggml-cpu.cpp */; };
+		18F8C0C42CEDF52700CAD607 /* ggml-cpu-aarch64.c in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.c */; };
+		18F8C0C52CEDF52700CAD607 /* ggml-cpu-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0C32CEDF52700CAD607 /* ggml-cpu-quants.c */; };
+		18F8C0C72CEDF7AB00CAD607 /* ggml-backend-reg.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 18F8C0C62CEDF7AB00CAD607 /* ggml-backend-reg.cpp */; };
 		7FE3424B2A0C3FA20015A058 /* whisper-encoder-impl.m in Sources */ = {isa = PBXBuildFile; fileRef = 7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */; };
 		7FE3424C2A0C3FA20015A058 /* whisper-encoder.mm in Sources */ = {isa = PBXBuildFile; fileRef = 7FE342472A0C3FA20015A058 /* whisper-encoder.mm */; };
 		7FE3424D2A0C3FA20015A058 /* whisper-decoder-impl.m in Sources */ = {isa = PBXBuildFile; fileRef = 7FE3424A2A0C3FA20015A058 /* whisper-decoder-impl.m */; };
@@ -50,8 +55,8 @@
 		18133C7F2C64E342005CEAAC /* ggml-aarch64.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-aarch64.c"; path = "../../../ggml/src/ggml-aarch64.c"; sourceTree = "<group>"; };
 		184447182AB211A2007D6BFE /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../../ggml/src/ggml-alloc.c"; sourceTree = "<group>"; };
 		184447192AB211A2007D6BFE /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../../ggml/include/ggml-alloc.h"; sourceTree = "<group>"; };
-		1844471B2AB21655007D6BFE /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../../ggml/src/ggml-metal.m"; sourceTree = "<group>"; };
-		1844471D2AB2195F007D6BFE /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../../ggml/src/ggml-metal.metal"; sourceTree = "<group>"; };
+		1844471B2AB21655007D6BFE /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../../ggml/src/ggml-metal/ggml-metal.m"; sourceTree = "<group>"; };
+		1844471D2AB2195F007D6BFE /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../../ggml/src/ggml-metal/ggml-metal.metal"; sourceTree = "<group>"; };
 		18627C7629052BDF00BD2A04 /* whisper.objc.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = whisper.objc.app; sourceTree = BUILT_PRODUCTS_DIR; };
 		18627C7929052BDF00BD2A04 /* AppDelegate.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = AppDelegate.h; sourceTree = "<group>"; };
 		18627C7A29052BDF00BD2A04 /* AppDelegate.m */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.objc; path = AppDelegate.m; sourceTree = "<group>"; };
@@ -77,8 +82,17 @@
 		18ABE1572AF556340044A204 /* ggml-backend.cpp */ = {isa = PBXFileReference; explicitFileType = sourcecode.cpp.cpp; fileEncoding = 4; name = "ggml-backend.cpp"; path = "../../../ggml/src/ggml-backend.cpp"; sourceTree = "<group>"; };
 		18ABE1582AF556340044A204 /* ggml-impl.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-impl.h"; path = "../../../ggml/src/ggml-impl.h"; sourceTree = "<group>"; };
 		18ABE1592AF556340044A204 /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../../ggml/src/ggml-quants.c"; sourceTree = "<group>"; };
-		18E864A82CE73C1E0094B8B3 /* ggml-cpu.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu.c"; path = "../../../ggml/src/ggml-cpu.c"; sourceTree = "<group>"; };
+		18E864A82CE73C1E0094B8B3 /* ggml-cpu.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu.c"; path = "../../../ggml/src/ggml-cpu/ggml-cpu.c"; sourceTree = "<group>"; };
 		18E864AA2CE73C580094B8B3 /* ggml-cpu.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu.h"; path = "../../../ggml/include/ggml-cpu.h"; sourceTree = "<group>"; };
+		18F8C0BA2CEDF4DC00CAD607 /* ggml-threading.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-threading.h"; path = "../../../ggml/src/ggml-threading.h"; sourceTree = "<group>"; };
+		18F8C0BB2CEDF4DC00CAD607 /* ggml-threading.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; name = "ggml-threading.cpp"; path = "../../../ggml/src/ggml-threading.cpp"; sourceTree = "<group>"; };
+		18F8C0BD2CEDF50700CAD607 /* ggml-cpu.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; name = "ggml-cpu.cpp"; path = "../../../ggml/src/ggml-cpu/ggml-cpu.cpp"; sourceTree = "<group>"; };
+		18F8C0BF2CEDF52700CAD607 /* ggml-cpu-aarch64.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu-aarch64.h"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-aarch64.h"; sourceTree = "<group>"; };
+		18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu-aarch64.c"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-aarch64.c"; sourceTree = "<group>"; };
+		18F8C0C12CEDF52700CAD607 /* ggml-cpu-impl.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu-impl.h"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-impl.h"; sourceTree = "<group>"; };
+		18F8C0C22CEDF52700CAD607 /* ggml-cpu-quants.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-cpu-quants.h"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-quants.h"; sourceTree = "<group>"; };
+		18F8C0C32CEDF52700CAD607 /* ggml-cpu-quants.c */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.c; name = "ggml-cpu-quants.c"; path = "../../../ggml/src/ggml-cpu/ggml-cpu-quants.c"; sourceTree = "<group>"; };
+		18F8C0C62CEDF7AB00CAD607 /* ggml-backend-reg.cpp */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.cpp.cpp; name = "ggml-backend-reg.cpp"; path = "../../../ggml/src/ggml-backend-reg.cpp"; sourceTree = "<group>"; };
 		7FE342452A0C3FA20015A058 /* whisper-encoder-impl.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; path = "whisper-encoder-impl.m"; sourceTree = "<group>"; };
 		7FE342462A0C3FA20015A058 /* whisper-encoder.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; path = "whisper-encoder.h"; sourceTree = "<group>"; };
 		7FE342472A0C3FA20015A058 /* whisper-encoder.mm */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.objcpp; path = "whisper-encoder.mm"; sourceTree = "<group>"; };
@@ -118,6 +132,15 @@
 		18627C7829052BDF00BD2A04 /* whisper.objc */ = {
 			isa = PBXGroup;
 			children = (
+				18F8C0C62CEDF7AB00CAD607 /* ggml-backend-reg.cpp */,
+				18F8C0BF2CEDF52700CAD607 /* ggml-cpu-aarch64.h */,
+				18F8C0C02CEDF52700CAD607 /* ggml-cpu-aarch64.c */,
+				18F8C0C12CEDF52700CAD607 /* ggml-cpu-impl.h */,
+				18F8C0C22CEDF52700CAD607 /* ggml-cpu-quants.h */,
+				18F8C0C32CEDF52700CAD607 /* ggml-cpu-quants.c */,
+				18F8C0BD2CEDF50700CAD607 /* ggml-cpu.cpp */,
+				18F8C0BA2CEDF4DC00CAD607 /* ggml-threading.h */,
+				18F8C0BB2CEDF4DC00CAD607 /* ggml-threading.cpp */,
 				18E864AA2CE73C580094B8B3 /* ggml-cpu.h */,
 				18E864A82CE73C1E0094B8B3 /* ggml-cpu.c */,
 				18133C7F2C64E342005CEAAC /* ggml-aarch64.c */,
@@ -252,11 +275,16 @@
 				18627C9629052C5800BD2A04 /* ggml.c in Sources */,
 				18627C7B29052BDF00BD2A04 /* AppDelegate.m in Sources */,
 				7FE3424D2A0C3FA20015A058 /* whisper-decoder-impl.m in Sources */,
+				18F8C0C72CEDF7AB00CAD607 /* ggml-backend-reg.cpp in Sources */,
+				18F8C0BE2CEDF50700CAD607 /* ggml-cpu.cpp in Sources */,
 				1844471A2AB211A2007D6BFE /* ggml-alloc.c in Sources */,
+				18F8C0C42CEDF52700CAD607 /* ggml-cpu-aarch64.c in Sources */,
+				18F8C0C52CEDF52700CAD607 /* ggml-cpu-quants.c in Sources */,
 				18E864A92CE73C1E0094B8B3 /* ggml-cpu.c in Sources */,
 				18ABE15A2AF556340044A204 /* ggml-backend.cpp in Sources */,
 				18627C8C29052BE000BD2A04 /* main.m in Sources */,
 				18627C7E29052BDF00BD2A04 /* SceneDelegate.m in Sources */,
+				18F8C0BC2CEDF4DC00CAD607 /* ggml-threading.cpp in Sources */,
 				1844471C2AB21655007D6BFE /* ggml-metal.m in Sources */,
 				7FE3424B2A0C3FA20015A058 /* whisper-encoder-impl.m in Sources */,
 			);
@@ -335,6 +363,7 @@
 				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
 				GCC_WARN_UNUSED_FUNCTION = YES;
 				GCC_WARN_UNUSED_VARIABLE = YES;
+				HEADER_SEARCH_PATHS = ../../../ggml/src/;
 				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
 				MTL_ENABLE_DEBUG_INFO = INCLUDE_SOURCE;
 				MTL_FAST_MATH = YES;
@@ -388,6 +417,7 @@
 				GCC_WARN_UNINITIALIZED_AUTOS = YES_AGGRESSIVE;
 				GCC_WARN_UNUSED_FUNCTION = YES;
 				GCC_WARN_UNUSED_VARIABLE = YES;
+				HEADER_SEARCH_PATHS = ../../../ggml/src/;
 				IPHONEOS_DEPLOYMENT_TARGET = 16.0;
 				MTL_ENABLE_DEBUG_INFO = NO;
 				MTL_FAST_MATH = YES;
@@ -410,6 +440,7 @@
 				DEVELOPMENT_TEAM = P8JZH34X63;
 				GCC_WARN_64_TO_32_BIT_CONVERSION = NO;
 				GENERATE_INFOPLIST_FILE = YES;
+				HEADER_SEARCH_PATHS = ../../../ggml/src/;
 				INFOPLIST_FILE = whisper.objc/Info.plist;
 				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
 				INFOPLIST_KEY_UILaunchStoryboardName = LaunchScreen;
@@ -439,6 +470,7 @@
 				DEVELOPMENT_TEAM = P8JZH34X63;
 				GCC_WARN_64_TO_32_BIT_CONVERSION = NO;
 				GENERATE_INFOPLIST_FILE = YES;
+				HEADER_SEARCH_PATHS = ../../../ggml/src/;
 				INFOPLIST_FILE = whisper.objc/Info.plist;
 				INFOPLIST_KEY_UIApplicationSupportsIndirectInputEvents = YES;
 				INFOPLIST_KEY_UILaunchStoryboardName = LaunchScreen;
diff --git a/examples/whisper.swiftui/whisper.cpp.swift/LibWhisper.swift b/examples/whisper.swiftui/whisper.cpp.swift/LibWhisper.swift
index 54cb3bd41e0..3c769e7af8e 100644
--- a/examples/whisper.swiftui/whisper.cpp.swift/LibWhisper.swift
+++ b/examples/whisper.swiftui/whisper.cpp.swift/LibWhisper.swift
@@ -66,9 +66,7 @@ actor WhisperContext {
 
     private func systemInfo() -> String {
         var info = ""
-        if (ggml_cpu_has_neon() != 0) { info += "NEON " }
-        if (ggml_cpu_has_metal() != 0) { info += "METAL " }
-        if (ggml_cpu_has_blas() != 0) { info += "BLAS " }
+        //if (ggml_cpu_has_neon() != 0) { info += "NEON " }
         return String(info.dropLast())
     }
 
@@ -77,45 +75,45 @@ actor WhisperContext {
         if (whisper_set_mel(context, nil, 0, nMels) != 0) {
             return "error: failed to set mel"
         }
-        
+
         // heat encoder
         if (whisper_encode(context, 0, nThreads) != 0) {
             return "error: failed to encode"
         }
-        
+
         var tokens = [whisper_token](repeating: 0, count: 512)
-        
+
         // prompt heat
         if (whisper_decode(context, &tokens, 256, 0, nThreads) != 0) {
             return "error: failed to decode"
         }
-        
+
         // text-generation heat
         if (whisper_decode(context, &tokens, 1, 256, nThreads) != 0) {
             return "error: failed to decode"
         }
-        
+
         whisper_reset_timings(context)
-        
+
         // actual run
         if (whisper_encode(context, 0, nThreads) != 0) {
             return "error: failed to encode"
         }
-        
+
         // text-generation
         for i in 0..<256 {
             if (whisper_decode(context, &tokens, 1, Int32(i), nThreads) != 0) {
                 return "error: failed to decode"
             }
         }
-        
+
         // batched decoding
         for _ in 0..<64 {
             if (whisper_decode(context, &tokens, 5, 0, nThreads) != 0) {
                 return "error: failed to decode"
             }
         }
-        
+
         // prompt processing
         for _ in 0..<16 {
             if (whisper_decode(context, &tokens, 256, 0, nThreads) != 0) {
diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index 81b7a02f519..3442142adbb 100644
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -32,7 +32,15 @@ else()
     endif()
 endif()
 
+# remove the lib prefix on win32 mingw
+if (WIN32)
+    set(CMAKE_STATIC_LIBRARY_PREFIX "")
+    set(CMAKE_SHARED_LIBRARY_PREFIX "")
+    set(CMAKE_SHARED_MODULE_PREFIX  "")
+endif()
+
 option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
+option(GGML_BACKEND_DL   "ggml: build backends as dynamic libraries (requires BUILD_SHARED_LIBS)" OFF)
 
 #
 # option list
@@ -91,31 +99,38 @@ else()
     set(INS_ENB ON)
 endif()
 
-option(GGML_CPU_HBM     "ggml: use memkind for CPU HBM" OFF)
-
-option(GGML_AVX         "ggml: enable AVX"              ${INS_ENB})
-option(GGML_AVX2        "ggml: enable AVX2"             ${INS_ENB})
-option(GGML_AVX512      "ggml: enable AVX512"           OFF)
-option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI"      OFF)
-option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI"      OFF)
-option(GGML_AVX512_BF16 "ggml: enable AVX512-BF16"      OFF)
-option(GGML_AMX_TILE    "ggml: enable AMX-TILE"         OFF)
-option(GGML_AMX_INT8    "ggml: enable AMX-INT8"         OFF)
-option(GGML_AMX_BF16    "ggml: enable AMX-BF16"         OFF)
-option(GGML_FMA         "ggml: enable FMA"              ${INS_ENB})
+option(GGML_CPU_HBM          "ggml: use memkind for CPU HBM" OFF)
+option(GGML_CPU_AARCH64      "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
+option(GGML_AVX              "ggml: enable AVX"              ${INS_ENB})
+option(GGML_AVX_VNNI         "ggml: enable AVX-VNNI"         OFF)
+option(GGML_AVX2             "ggml: enable AVX2"             ${INS_ENB})
+option(GGML_AVX512           "ggml: enable AVX512F"          OFF)
+option(GGML_AVX512_VBMI      "ggml: enable AVX512-VBMI"      OFF)
+option(GGML_AVX512_VNNI      "ggml: enable AVX512-VNNI"      OFF)
+option(GGML_AVX512_BF16      "ggml: enable AVX512-BF16"      OFF)
 if (NOT MSVC)
-    option(GGML_F16C    "ggml: enable F16C"             ${INS_ENB}) # in MSVC F16C is implied with AVX2/AVX512
+    # in MSVC F16C and FMA is implied with AVX2/AVX512
+    option(GGML_FMA          "ggml: enable FMA"              ${INS_ENB})
+    option(GGML_F16C         "ggml: enable F16C"             ${INS_ENB})
+    # MSVC does not seem to support AMX
+    option(GGML_AMX_TILE     "ggml: enable AMX-TILE"         OFF)
+    option(GGML_AMX_INT8     "ggml: enable AMX-INT8"         OFF)
+    option(GGML_AMX_BF16     "ggml: enable AMX-BF16"         OFF)
 endif()
-option(GGML_LASX        "ggml: enable lasx"             ON)
-option(GGML_LSX         "ggml: enable lsx"              ON)
-option(GGML_SVE         "ggml: enable SVE"              OFF)
+option(GGML_LASX             "ggml: enable lasx"             ON)
+option(GGML_LSX              "ggml: enable lsx"              ON)
+option(GGML_RVV              "ggml: enable rvv"              ON)
+option(GGML_SVE              "ggml: enable SVE"              OFF)
+option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
+
 
 if (WIN32)
-    set(GGML_WIN_VER "0x602" CACHE STRING "ggml: Windows Version")
+    set(GGML_WIN_VER "0x602" CACHE STRING   "ggml: Windows version")
 endif()
 
 # ggml core
 set(GGML_SCHED_MAX_COPIES  "4" CACHE STRING "ggml: max input copies for pipeline parallelism")
+option(GGML_CPU                             "ggml: enable CPU backend"                        ON)
 
 # 3rd party libs / backends
 option(GGML_ACCELERATE                      "ggml: enable Accelerate framework"               ON)
@@ -126,14 +141,9 @@ option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"
 
 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
-option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
-set   (GGML_CUDA_DMMV_X   "32" CACHE STRING "ggml: x stride for dmmv CUDA kernels")
-set   (GGML_CUDA_MMV_Y     "1" CACHE STRING "ggml: y block size for mmv CUDA kernels")
 option(GGML_CUDA_F16                        "ggml: use 16 bit floats for some calculations"   OFF)
-set   (GGML_CUDA_KQUANTS_ITER "2" CACHE STRING
-                                            "ggml: iters./thread per block for Q2_K/Q6_K")
 set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                             "ggml: max. batch size for using peer access")
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
@@ -141,7 +151,7 @@ option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"
 option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
 option(GGML_CUDA_GRAPHS                     "ggml: use CUDA graphs (llama.cpp only)"          ${GGML_CUDA_GRAPHS_DEFAULT})
 
-option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               OFF)
+option(GGML_HIP                             "ggml: use HIP"                                   OFF)
 option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
@@ -162,11 +172,17 @@ set   (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
 set   (GGML_METAL_STD "" CACHE STRING       "ggml: metal standard version (-std flag)")
 option(GGML_OPENMP                          "ggml: use OpenMP"                                ON)
 option(GGML_RPC                             "ggml: use RPC"                                   OFF)
-option(GGML_AMX                             "ggml: use AMX"                                   OFF)
 option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
 set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
                                             "ggml: sycl target device")
+set   (GGML_SYCL_DEVICE_ARCH "" CACHE STRING
+                                            "ggml: sycl device architecture")
+
+option(GGML_OPENCL                          "ggml: use OpenCL"                                OFF)
+option(GGML_OPENCL_PROFILING                "ggml: use OpenCL profiling (increases overhead)" OFF)
+option(GGML_OPENCL_EMBED_KERNELS            "ggml: embed kernels"                             ON)
+option(GGML_OPENCL_USE_ADRENO_KERNELS       "ggml: use optimized kernels for Adreno"          ON)
 
 # extra artifacts
 option(GGML_BUILD_TESTS    "ggml: build tests"    ${GGML_STANDALONE})
@@ -179,11 +195,7 @@ option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE})
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
 
-if (GGML_SYCL)
-    set(CMAKE_CXX_STANDARD 17)
-else()
-    set(CMAKE_CXX_STANDARD 11)
-endif()
+set(CMAKE_CXX_STANDARD 17)
 set(CMAKE_CXX_STANDARD_REQUIRED true)
 
 set(THREADS_PREFER_PTHREAD_FLAG ON)
@@ -226,6 +238,7 @@ set(GGML_PUBLIC_HEADERS
     include/ggml-cann.h
     include/ggml-cuda.h
     include/ggml-kompute.h
+    include/ggml-opt.h
     include/ggml-metal.h
     include/ggml-rpc.h
     include/ggml-sycl.h
@@ -235,15 +248,14 @@ set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
 #if (GGML_METAL)
 #    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
 #endif()
-install(TARGETS ggml PUBLIC_HEADER)
-
-if (BUILD_SHARED_LIBS)
-    install(TARGETS ggml LIBRARY)
-endif()
+install(TARGETS ggml LIBRARY PUBLIC_HEADER)
+install(TARGETS ggml-base LIBRARY)
 
+# FIXME: this should be done in the backend cmake files
 if (GGML_METAL)
+    # FIXME: does this need to be installed with GGML_METAL_EMBED_LIBRARY?
     install(
-        FILES src/ggml-metal.metal
+        FILES src/ggml-metal/ggml-metal.metal
         PERMISSIONS
             OWNER_READ
             OWNER_WRITE
diff --git a/ggml/ggml_vk_generate_shaders.py b/ggml/ggml_vk_generate_shaders.py
deleted file mode 100644
index 38914eedb79..00000000000
--- a/ggml/ggml_vk_generate_shaders.py
+++ /dev/null
@@ -1,220 +0,0 @@
-#!/usr/bin/env python
-
-import logging
-import argparse
-import asyncio
-import os
-from tempfile import gettempdir
-
-logger = logging.getLogger("ggml-vk-generate-shaders")
-
-GLSLC = "glslc"
-
-type_names = [
-    "f32",
-    "f16",
-    "q4_0",
-    "q4_1",
-    "q5_0",
-    "q5_1",
-    "q8_0",
-    "q2_k",
-    "q3_k",
-    "q4_k",
-    "q5_k",
-    "q6_k",
-]
-
-ASYNCIO_CONCURRENCY = 64
-
-input_dir = "vulkan-shaders"
-output_dir = gettempdir()
-
-lock = asyncio.Lock()
-shader_fnames = []
-
-
-async def string_to_spv(name, in_fname, defines, fp16=True):
-    name = f"{name}{'_fp32' if not fp16 else ''}"
-    out_fname = os.path.join(output_dir, f"{name}.spv")
-
-    in_path = os.path.join(input_dir, in_fname)
-
-    cmd = [GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o", out_fname]
-
-    cmd.extend([f"-D{key}={value}" for key, value in defines.items()])
-
-    proc = await asyncio.create_subprocess_exec(*cmd, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE)
-
-    stdout, stderr = await proc.communicate()
-
-    stdout = stdout.decode()
-    error = stderr.decode()
-
-    if proc.returncode:
-        cmd = " ".join(cmd)
-        logger.error(f"cannot compile {name}\n\n{cmd}\n\n{error}")
-        return
-
-    async with lock:
-        shader_fnames.append((name, out_fname))
-
-
-def matmul_shaders(tasks, fp16, matmul_id):
-    if fp16:
-        load_vec = "8"
-        aligned_b_type_f32 = "mat2x4"
-        aligned_b_type_f16 = "f16mat2x4"
-    else:
-        load_vec = "4"
-        aligned_b_type_f32 = "vec4"
-        aligned_b_type_f16 = "f16vec4"
-
-    base_dict = {"FLOAT_TYPE": "float" if not fp16 else "float16_t"}
-    shader_name = "matmul"
-
-    if matmul_id:
-        base_dict["MUL_MAT_ID"] = "1"
-        shader_name = "matmul_id"
-
-    if fp16:
-        base_dict["FLOAT16"] = "1"
-
-    # Shaders with f16 B_TYPE
-    tasks.append(string_to_spv(f"{shader_name}_f32_f16", "mul_mm.comp", base_dict | {"DATA_A_F32": "1", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
-    tasks.append(string_to_spv(f"{shader_name}_f32_f16_aligned", "mul_mm.comp", base_dict | {"DATA_A_F32": "1", "LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f16, "D_TYPE": "float"}, fp16))
-
-    tasks.append(string_to_spv(f"{shader_name}_f16", "mul_mm.comp", base_dict | {"DATA_A_F16": "1", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
-    tasks.append(string_to_spv(f"{shader_name}_f16_aligned", "mul_mm.comp", base_dict | {"DATA_A_F16": "1", "LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f16, "D_TYPE": "float"}, fp16))
-
-    for tname in type_names:
-        data_a_key = f"DATA_A_{tname.upper()}"
-        load_vec_a = load_vec if tname in ("f32", "f16") else "2"
-        tasks.append(string_to_spv(f"{shader_name}_{tname}_f32", "mul_mm.comp", base_dict | {data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
-        tasks.append(string_to_spv(f"{shader_name}_{tname}_f32_aligned", "mul_mm.comp", base_dict | {data_a_key: "2", "LOAD_VEC_A": load_vec_a, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f32, "D_TYPE": "float"}, fp16))
-
-
-async def main():
-    logger.info("ggml_vulkan: Generating and compiling shaders to SPIR-V")
-
-    tasks = []
-
-    for fp16 in (False, True):
-        # MUL_MAT
-        matmul_shaders(tasks, fp16, False)
-        # MUL_MAT_ID
-        matmul_shaders(tasks, fp16, True)
-
-    for tname in type_names:
-        base_dict = {"FLOAT_TYPE": "float"}
-
-        # mul mat vec
-        data_a_key = f"DATA_A_{tname.upper()}"
-        shader = f"mul_mat_vec_{tname}.comp" if tname.endswith("_k") else "mul_mat_vec.comp"
-
-        tasks.append(string_to_spv(f"mul_mat_vec_{tname}_f32_f32", shader, base_dict | {data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}))
-        tasks.append(string_to_spv(f"mul_mat_vec_{tname}_f16_f32", shader, base_dict | {data_a_key: "1", "B_TYPE": "float16_t", "D_TYPE": "float"}))
-
-        tasks.append(string_to_spv(f"mul_mat_vec_id_{tname}_f32", shader, base_dict | {"MUL_MAT_ID": "1", data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}))
-
-        # Dequant shaders
-        if tname != "f16":
-            tasks.append(string_to_spv(f"dequant_{tname}", f"dequant_{tname}.comp", base_dict | {data_a_key: "1", "D_TYPE": "float16_t"}))
-
-        # get_rows
-        if not tname.endswith("_k"):
-            shader = "get_rows.comp" if tname in ("f32", "f16") else "get_rows_quant.comp"
-
-            if tname == "f16":
-                tasks.append(string_to_spv(f"get_rows_{tname}", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float16_t", "OPTIMIZATION_ERROR_WORKAROUND": "1"}))
-            else:
-                tasks.append(string_to_spv(f"get_rows_{tname}", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float16_t"}))
-            tasks.append(string_to_spv(f"get_rows_{tname}_f32", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}))
-
-    # Norms
-    tasks.append(string_to_spv("norm_f32", "norm.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("rms_norm_f32", "rms_norm.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("cpy_f32_f32", "copy.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("cpy_f32_f16", "copy.comp", {"A_TYPE": "float", "D_TYPE": "float16_t"}))
-    tasks.append(string_to_spv("cpy_f16_f16", "copy.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t", "OPTIMIZATION_ERROR_WORKAROUND": "1"}))
-
-    tasks.append(string_to_spv("add_f32", "add.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {}))
-
-    tasks.append(string_to_spv("mul_f32", "mul.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("div_f32", "div.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("scale_f32", "scale.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("sqr_f32", "square.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("clamp_f32", "clamp.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
-
-    tasks.append(string_to_spv("gelu_f32", "gelu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("silu_f32", "silu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("relu_f32", "relu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("soft_max_f32", "soft_max.comp", base_dict | {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("soft_max_f32_f16", "soft_max.comp", base_dict | {"A_TYPE": "float", "B_TYPE": "float16_t", "D_TYPE": "float"}))
-
-    tasks.append(string_to_spv("rope_norm_f32", "rope_norm.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("rope_norm_f16", "rope_norm.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
-
-    tasks.append(string_to_spv("rope_neox_f32", "rope_neox.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
-    tasks.append(string_to_spv("rope_neox_f16", "rope_neox.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
-
-    tasks.append(string_to_spv("argsort_f32", "argsort.comp", {"A_TYPE": "float"}))
-
-    tasks.append(string_to_spv("sum_rows_f32", "sum_rows.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
-
-    # Helper to decorate tasks with semaphore acquisition.
-    async def withSemaphore(sem, task):
-        async with sem:
-            return await task
-
-    # Run tasks concurrently guarded by a concurrency limit.
-    sem = asyncio.Semaphore(ASYNCIO_CONCURRENCY)
-    await asyncio.gather(*(withSemaphore(sem, task) for task in tasks))
-
-    with open("ggml-vulkan-shaders.hpp", "w") as f:
-        f.write("#include <cstdint>\n\n")
-        for name, path in sorted(shader_fnames):
-
-            with open(path, "rb") as spv:
-                counter = 0
-                newline_counter = 0
-                f.write(f"unsigned char {name}_data[] = {{\n")
-                for val in spv.read():
-                    f.write(f"0x{val:02x},")
-                    newline_counter += 1
-                    counter += 1
-                    if newline_counter >= 12:
-                        newline_counter = 0
-                        f.write("\n")
-            f.write("\n};\n")
-            f.write(f"const uint64_t {name}_len = {counter};\n\n")
-            os.remove(path)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(description="GGML Vulkan Shader Generator")
-
-    parser.add_argument("--glslc", help="Path to glslc")
-    parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
-
-    args = parser.parse_args()
-
-    logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
-
-    if args.glslc:
-        GLSLC = args.glslc
-
-    asyncio.run(main())
diff --git a/ggml/include/ggml-amx.h b/ggml/include/ggml-amx.h
deleted file mode 100644
index 22b3f70f43a..00000000000
--- a/ggml/include/ggml-amx.h
+++ /dev/null
@@ -1,25 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-// buffer_type API
-GGML_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
-
-GGML_API bool ggml_backend_is_amx(ggml_backend_t backend);
-
-// backend API
-GGML_API ggml_backend_t ggml_backend_amx_init(void);
-
-GGML_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
-
-GGML_API ggml_backend_reg_t ggml_backend_amx_reg(void);
-
-#ifdef  __cplusplus
-}
-#endif
diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
index 125413d1bfd..7221a083092 100644
--- a/ggml/include/ggml-backend.h
+++ b/ggml/include/ggml-backend.h
@@ -3,6 +3,20 @@
 #include "ggml.h"
 #include "ggml-alloc.h"
 
+#ifdef GGML_BACKEND_SHARED
+#    if defined(_WIN32) && !defined(__MINGW32__)
+#        ifdef GGML_BACKEND_BUILD
+#            define GGML_BACKEND_API __declspec(dllexport) extern
+#        else
+#            define GGML_BACKEND_API __declspec(dllimport) extern
+#        endif
+#    else
+#        define GGML_BACKEND_API __attribute__ ((visibility ("default"))) extern
+#    endif
+#else
+#    define GGML_BACKEND_API extern
+#endif
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
@@ -72,7 +86,7 @@ extern "C" {
     GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
 
-    // "offset" refers to the offset of the tensor data for setting/getting data
+    // "offset" refers to the offset in tensor->data for setting/getting data
     GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_memset(   struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
@@ -176,6 +190,14 @@ extern "C" {
     typedef void                         (*ggml_backend_set_n_threads_t)(ggml_backend_t backend, int n_threads);
     // Get additional buffer types provided by the device (returns a NULL-terminated array)
     typedef ggml_backend_buffer_type_t * (*ggml_backend_dev_get_extra_bufts_t)(ggml_backend_dev_t device);
+    // Set the abort callback for the backend
+    typedef void                         (*ggml_backend_set_abort_callback_t)(ggml_backend_t backend, ggml_abort_callback abort_callback, void * abort_callback_data);
+    // Get a list of feature flags supported by the backend (returns a NULL-terminated array)
+    struct ggml_backend_feature {
+        const char * name;
+        const char * value;
+    };
+    typedef struct ggml_backend_feature * (*ggml_backend_get_features_t)(ggml_backend_reg_t reg);
 
     //
     // Backend registry
@@ -200,6 +222,14 @@ extern "C" {
     // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU) OR ggml_backend_dev_by_type(CPU), NULL)
     GGML_API ggml_backend_t ggml_backend_init_best(void);
 
+    // Load a backend from a dynamic library and register it
+    GGML_API ggml_backend_reg_t ggml_backend_load(const char * path);
+    // Unload a backend if loaded dynamically and unregister it
+    GGML_API void               ggml_backend_unload(ggml_backend_reg_t reg);
+    // Load all known backends from dynamic libraries
+    GGML_API void               ggml_backend_load_all(void);
+    GGML_API void               ggml_backend_load_all_from_path(const char * dir_path);
+
     //
     // Backend scheduler
     //
@@ -228,14 +258,20 @@ extern "C" {
         ggml_backend_sched_reserve(sched, reserve_graph);
 
         // compute
-        graph = build_graph(sched);
-        ggml_backend_sched_graph_compute(sched, graph);
+        graph = build_graph(sched); // the graph and its tensors are single-use in terms of allocation, multi-use in terms of computation
+        for (int i = 0; i < 10; ++i) {
+            ggml_backend_sched_graph_compute(sched, graph); // on the first iteration the graph is allocated automatically
+        }
 
         // if there are graph inputs:
-        ggml_backend_sched_reset(sched);
-        ggml_backend_sched_alloc_graph(sched, graph);
-        ggml_backend_tensor_set(input_tensor, ...);
-        ggml_backend_sched_graph_compute(sched, graph);
+        graph = build_graph(sched); // get a new graph that is not allocated (the metadata for the old graph is freed once ggml_free is called)
+        ggml_backend_sched_reset(sched); // clear the allocation of the previous graph
+        ggml_backend_sched_alloc_graph(sched, graph); // explicitly allocate the new graph but do not execute it
+        ggml_backend_tensor_set(input_tensor, ...); // copy data to the newly allocated graph tensors
+        ggml_backend_sched_graph_compute(sched, graph); // execute the graph
+
+        // as an alternative to the above it is also possible to assign the inputs to a dedicated context and
+        // allocate them statically via ggml_backend_alloc_ctx_tensors
     }
     */
 
@@ -250,7 +286,7 @@ extern "C" {
     //
     typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
 
-    // Initialize a backend scheduler
+    // Initialize a backend scheduler, backends with low index are given priority over backends with high index
     GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
     GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
 
@@ -275,7 +311,9 @@ extern "C" {
     GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
     GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
 
-    // Reset all assignments and allocators - must be called before changing the node backends
+    // Reset all assignments and allocators - must be called before changing the node backends or allocating a new graph.
+    // This in effect deallocates all tensors that were previously allocated and leaves them with dangling pointers.
+    // The correct way to use this API is to discard the deallocated tensors and create new ones.
     GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
 
     // Set a callback to be called for each resulting node during graph compute
diff --git a/ggml/include/ggml-blas.h b/ggml/include/ggml-blas.h
index 25b2e637fb4..87a81b36348 100644
--- a/ggml/include/ggml-blas.h
+++ b/ggml/include/ggml-blas.h
@@ -9,15 +9,15 @@ extern "C" {
 #endif
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_blas_init(void);
+GGML_BACKEND_API ggml_backend_t ggml_backend_blas_init(void);
 
-GGML_API bool ggml_backend_is_blas(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_blas(ggml_backend_t backend);
 
 // number of threads used for conversion to float
 // for openblas and blis, this will also set the number of threads used for blas operations
-GGML_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
+GGML_BACKEND_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
 
-GGML_API ggml_backend_reg_t ggml_backend_blas_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_blas_reg(void);
 
 
 #ifdef  __cplusplus
diff --git a/ggml/include/ggml-cann.h b/ggml/include/ggml-cann.h
index 52897549388..b469e228d06 100644
--- a/ggml/include/ggml-cann.h
+++ b/ggml/include/ggml-cann.h
@@ -34,7 +34,7 @@ extern "C" {
  */
 #define GGML_CANN_MAX_DEVICES 16
 
-GGML_API ggml_backend_reg_t ggml_backend_cann_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cann_reg(void);
 
 /**
  * @brief Initializes the CANN backend for a specified device.
@@ -46,7 +46,7 @@ GGML_API ggml_backend_reg_t ggml_backend_cann_reg(void);
  * @param device The index of the device to initialize.
  * @return A pointer to the initialized backend instance, or nullptr on failure.
  */
-GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_cann_init(int32_t device);
 
 /**
  * @brief Checks if a given backend is a CANN backend.
@@ -57,7 +57,7 @@ GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device);
  * @param backend The backend instance to check.
  * @return True if the backend is a CANN backend, false otherwise.
  */
-GGML_API bool ggml_backend_is_cann(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_cann(ggml_backend_t backend);
 
 /**
  * @brief Retrieves the CANN buffer type for a specified device.
@@ -69,7 +69,7 @@ GGML_API bool ggml_backend_is_cann(ggml_backend_t backend);
  * @return A pointer to the buffer type interface for the specified device, or
  * nullptr if the device index is out of range.
  */
-GGML_API ggml_backend_buffer_type_t
+GGML_BACKEND_API ggml_backend_buffer_type_t
 ggml_backend_cann_buffer_type(int32_t device);
 
 /**
@@ -80,14 +80,14 @@ ggml_backend_cann_buffer_type(int32_t device);
  *
  * @return The number of CANN devices available.
  */
-GGML_API int32_t ggml_backend_cann_get_device_count(void);
+GGML_BACKEND_API int32_t ggml_backend_cann_get_device_count(void);
 
 /**
  * @brief pinned host buffer for use with the CPU backend for faster copies between CPU and NPU.
  *
  * @return A pointer to the host buffer type interface.
  */
-GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
 
 /**
  * @brief Retrieves the description of a specific CANN device.
@@ -99,7 +99,7 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
  * @param description Pointer to a buffer where the description will be written.
  * @param description_size Size of the description buffer.
  */
-GGML_API void ggml_backend_cann_get_device_description(
+GGML_BACKEND_API void ggml_backend_cann_get_device_description(
     int32_t device, char* description, size_t description_size);
 
 /**
@@ -114,7 +114,7 @@ GGML_API void ggml_backend_cann_get_device_description(
  * @param total Pointer to a variable where the total memory size will be
  * stored.
  */
-GGML_API void ggml_backend_cann_get_device_memory(int32_t device,
+GGML_BACKEND_API void ggml_backend_cann_get_device_memory(int32_t device,
                                                   size_t* free,
                                                   size_t* total);
 
diff --git a/ggml/include/ggml-cpu.h b/ggml/include/ggml-cpu.h
index 7f1ee757310..3aa71badb5f 100644
--- a/ggml/include/ggml-cpu.h
+++ b/ggml/include/ggml-cpu.h
@@ -7,29 +7,6 @@
 extern "C" {
 #endif
 
-    // Scheduling priorities
-    enum ggml_sched_priority {
-        GGML_SCHED_PRIO_NORMAL,
-        GGML_SCHED_PRIO_MEDIUM,
-        GGML_SCHED_PRIO_HIGH,
-        GGML_SCHED_PRIO_REALTIME
-    };
-
-    // Threadpool params
-    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
-    struct ggml_threadpool_params {
-        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
-        int                 n_threads;                   // number of threads
-        enum ggml_sched_priority prio;                   // thread priority
-        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
-        bool                strict_cpu;                  // strict cpu placement
-        bool                paused;                      // start in paused state
-    };
-
-    struct ggml_threadpool;     // forward declaration, see ggml.c
-
-    typedef struct ggml_threadpool * ggml_threadpool_t;
-
     // the compute plan that needs to be prepared for ggml_graph_compute()
     // since https://github.com/ggerganov/ggml/issues/287
     struct ggml_cplan {
@@ -54,96 +31,104 @@ extern "C" {
         GGML_NUMA_STRATEGY_COUNT
     };
 
-    GGML_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
-    GGML_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
+    GGML_BACKEND_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
+    GGML_BACKEND_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
 
-    GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
-    GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
+    GGML_BACKEND_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
+    GGML_BACKEND_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
 
-    GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
-    GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
+    GGML_BACKEND_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
+    GGML_BACKEND_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
 
-    GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
-    GGML_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
+    GGML_BACKEND_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
+    GGML_BACKEND_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
 
-    GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
-    GGML_API void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
+    GGML_BACKEND_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_BACKEND_API void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
 
-    GGML_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
-    GGML_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
+    GGML_BACKEND_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
+    GGML_BACKEND_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
 
-    GGML_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
-    GGML_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+    GGML_BACKEND_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_BACKEND_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
 
-    GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
-    GGML_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
-    GGML_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
-    GGML_API struct ggml_threadpool *      ggml_threadpool_new          (struct ggml_threadpool_params  * params);
-    GGML_API void                          ggml_threadpool_free         (struct ggml_threadpool * threadpool);
-    GGML_API int                           ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
-    GGML_API void                          ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
-    GGML_API void                          ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API struct ggml_threadpool *      ggml_threadpool_new           (struct ggml_threadpool_params  * params);
+    GGML_BACKEND_API void                          ggml_threadpool_free          (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API int                           ggml_threadpool_get_n_threads (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API void                          ggml_threadpool_pause         (struct ggml_threadpool * threadpool);
+    GGML_BACKEND_API void                          ggml_threadpool_resume        (struct ggml_threadpool * threadpool);
 
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
-    GGML_API struct ggml_cplan ggml_graph_plan(
+    GGML_BACKEND_API struct ggml_cplan ggml_graph_plan(
                   const struct ggml_cgraph * cgraph,
                                        int   n_threads, /* = GGML_DEFAULT_N_THREADS */
                     struct ggml_threadpool * threadpool /* = NULL */ );
-    GGML_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+    GGML_BACKEND_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
 
     // same as ggml_graph_compute() but the work data is allocated as a part of the context
     // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
-    GGML_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
+    GGML_BACKEND_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
 
-    // TODO: move to backend interface
-    GGML_API int ggml_cpu_has_neon       (void);
-    GGML_API int ggml_cpu_has_sve        (void);
-    GGML_API int ggml_cpu_has_matmul_int8(void);
-    // get the sve vector length in bytes
-    GGML_API int ggml_cpu_get_sve_cnt(void);
+    //
+    // system info
+    //
+
+    // x86
+    GGML_BACKEND_API int ggml_cpu_has_sse3       (void);
+    GGML_BACKEND_API int ggml_cpu_has_ssse3      (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx        (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx_vnni   (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx2       (void);
+    GGML_BACKEND_API int ggml_cpu_has_f16c       (void);
+    GGML_BACKEND_API int ggml_cpu_has_fma        (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512     (void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512_vbmi(void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512_vnni(void);
+    GGML_BACKEND_API int ggml_cpu_has_avx512_bf16(void);
+    GGML_BACKEND_API int ggml_cpu_has_amx_int8   (void);
+    // ARM
+    GGML_BACKEND_API int ggml_cpu_has_neon       (void);
+    GGML_BACKEND_API int ggml_cpu_has_arm_fma    (void);
+    GGML_BACKEND_API int ggml_cpu_has_fp16_va    (void);
+    GGML_BACKEND_API int ggml_cpu_has_dotprod    (void);
+    GGML_BACKEND_API int ggml_cpu_has_matmul_int8(void);
+    GGML_BACKEND_API int ggml_cpu_has_sve        (void);
+    GGML_BACKEND_API int ggml_cpu_get_sve_cnt    (void);  // sve vector length in bytes
+    // other
+    GGML_BACKEND_API int ggml_cpu_has_riscv_v    (void);
+    GGML_BACKEND_API int ggml_cpu_has_vsx        (void);
+    GGML_BACKEND_API int ggml_cpu_has_wasm_simd  (void);
+    GGML_BACKEND_API int ggml_cpu_has_llamafile  (void);
 
     // Internal types and functions exposed for tests and benchmarks
 
-    typedef void (*ggml_from_float_to_mat_t)
-                                     (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
     typedef void (*ggml_vec_dot_t)  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
                                        const void * GGML_RESTRICT y, size_t by, int nrc);
-    typedef void (*ggml_gemv_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
-                                       const void * GGML_RESTRICT y, int nr, int nc);
-    typedef void (*ggml_gemm_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
-                                       const void * GGML_RESTRICT y, int nr, int nc);
 
     struct ggml_type_traits_cpu {
-        ggml_from_float_to_mat_t from_float_to_mat;
+        ggml_from_float_t        from_float;
         ggml_vec_dot_t           vec_dot;
         enum ggml_type           vec_dot_type;
         int64_t                  nrows; // number of rows to process simultaneously
-        int64_t                  ncols; // number of columns to process simultaneously
-        ggml_gemv_t              gemv;
-        ggml_gemm_t              gemm;
     };
 
-    GGML_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
+    GGML_BACKEND_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
 
-    GGML_API void ggml_cpu_init(void);
+    GGML_BACKEND_API void ggml_cpu_init(void);
 
     //
     // CPU backend
     //
 
-    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
+    GGML_BACKEND_API ggml_backend_t ggml_backend_cpu_init(void);
 
-    GGML_API bool ggml_backend_is_cpu                (ggml_backend_t backend);
-    GGML_API void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
-    GGML_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
-    GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
+    GGML_BACKEND_API bool ggml_backend_is_cpu                (ggml_backend_t backend);
+    GGML_BACKEND_API void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
+    GGML_BACKEND_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
+    GGML_BACKEND_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
 
-    GGML_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
-
-#ifdef GGML_USE_CPU_HBM
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
-#endif
+    GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
 #ifdef __cplusplus
 }
diff --git a/ggml/include/ggml-cuda.h b/ggml/include/ggml-cuda.h
index 305d0b636df..22ad2c00963 100644
--- a/ggml/include/ggml-cuda.h
+++ b/ggml/include/ggml-cuda.h
@@ -7,7 +7,7 @@
 extern "C" {
 #endif
 
-#ifdef GGML_USE_HIPBLAS
+#ifdef GGML_USE_HIP
 #define GGML_CUDA_NAME "ROCm"
 #define GGML_CUBLAS_NAME "hipBLAS"
 #elif defined(GGML_USE_MUSA)
@@ -20,27 +20,27 @@ extern "C" {
 #define GGML_CUDA_MAX_DEVICES       16
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_cuda_init(int device);
 
-GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_cuda(ggml_backend_t backend);
 
 // device buffer
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
 
 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
 
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
 
-GGML_API int  ggml_backend_cuda_get_device_count(void);
-GGML_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
-GGML_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
+GGML_BACKEND_API int  ggml_backend_cuda_get_device_count(void);
+GGML_BACKEND_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
+GGML_BACKEND_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
 
-GGML_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
-GGML_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);
+GGML_BACKEND_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
+GGML_BACKEND_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);
 
-GGML_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
 
 #ifdef  __cplusplus
 }
diff --git a/ggml/include/ggml-kompute.h b/ggml/include/ggml-kompute.h
index c0c43521b73..154aa56a742 100644
--- a/ggml/include/ggml-kompute.h
+++ b/ggml/include/ggml-kompute.h
@@ -37,13 +37,13 @@ struct ggml_vk_device ggml_vk_current_device(void);
 // forward declaration
 typedef struct ggml_backend * ggml_backend_t;
 
-GGML_API ggml_backend_t ggml_backend_kompute_init(int device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
 
-GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
 
-GGML_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
 
 #ifdef __cplusplus
 }
diff --git a/ggml/include/ggml-metal.h b/ggml/include/ggml-metal.h
index b8d3f678b71..669c1f84ae6 100644
--- a/ggml/include/ggml-metal.h
+++ b/ggml/include/ggml-metal.h
@@ -39,27 +39,27 @@ extern "C" {
 // user-code should use only these functions
 //
 
-GGML_API ggml_backend_t ggml_backend_metal_init(void);
+GGML_BACKEND_API ggml_backend_t ggml_backend_metal_init(void);
 
-GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_metal(ggml_backend_t backend);
 
 GGML_DEPRECATED(
-        GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
+        GGML_BACKEND_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size),
         "obsoleted by the new device interface - https://github.com/ggerganov/llama.cpp/pull/9713");
 
-GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
+GGML_BACKEND_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
 
 // helper to check if the device supports a specific family
 // ideally, the user code should be doing these checks
 // ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
-GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
+GGML_BACKEND_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
 
 // capture all command buffers committed the next time `ggml_backend_graph_compute` is called
-GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
+GGML_BACKEND_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
 
-GGML_API ggml_backend_reg_t ggml_backend_metal_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_metal_reg(void);
 
 #ifdef __cplusplus
 }
diff --git a/ggml/include/ggml-opencl.h b/ggml/include/ggml-opencl.h
new file mode 100644
index 00000000000..6b61771358f
--- /dev/null
+++ b/ggml/include/ggml-opencl.h
@@ -0,0 +1,26 @@
+#ifndef GGML_OPENCL_H
+#define GGML_OPENCL_H
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+//
+// backend API
+//
+GGML_BACKEND_API ggml_backend_t ggml_backend_opencl_init(void);
+GGML_BACKEND_API bool ggml_backend_is_opencl(ggml_backend_t backend);
+
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_opencl_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_opencl_host_buffer_type(void);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_opencl_reg(void);
+
+#ifdef  __cplusplus
+}
+#endif
+
+#endif // GGML_OPENCL_H
diff --git a/ggml/include/ggml-opt.h b/ggml/include/ggml-opt.h
new file mode 100644
index 00000000000..eb5eab9de67
--- /dev/null
+++ b/ggml/include/ggml-opt.h
@@ -0,0 +1,216 @@
+// This file contains functionality for training models using GGML.
+// It is not strictly needed vs. just vanilla GGML but it provides a more high-level interface for common needs such as datasets.
+// At the bottom of this file especially there are relatively high-level functions that are suitable use or adaptation in user code.
+//
+// Module maintainer: Johannes Gäßler (@JohannesGaessler, johannesg@5d6.de)
+
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#include <stdint.h>
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    struct ggml_opt_dataset;
+    struct ggml_opt_context;
+    struct ggml_opt_result;
+
+    typedef struct ggml_opt_dataset * ggml_opt_dataset_t;
+    typedef struct ggml_opt_context * ggml_opt_context_t;
+    typedef struct ggml_opt_result  * ggml_opt_result_t;
+
+    // ====== Loss ======
+
+    // built-in loss types, i.e. the built-in quantities minimized by the optimizer
+    // custom loss types can be defined via mean or sum which simply reduce the outputs for all datapoints to a single value
+    enum ggml_opt_loss_type {
+        GGML_OPT_LOSS_TYPE_MEAN,
+        GGML_OPT_LOSS_TYPE_SUM,
+        GGML_OPT_LOSS_TYPE_CROSS_ENTROPY,
+        GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR,
+    };
+
+    // ====== Dataset ======
+
+    GGML_API ggml_opt_dataset_t ggml_opt_dataset_init(
+            int64_t ne_datapoint, // number of elements per datapoint
+            int64_t ne_label,     // number of elements per label
+            int64_t ndata,        // total number of datapoints/labels
+            int64_t ndata_shard); // number of datapoints/labels per shard (unit at which the dataset is shuffled/copied)
+    GGML_API void ggml_opt_dataset_free(ggml_opt_dataset_t dataset);
+
+    // get underlying tensors that store the data
+    GGML_API struct ggml_tensor * ggml_opt_dataset_data  (ggml_opt_dataset_t dataset); // shape = [ne_datapoint, ndata]
+    GGML_API struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset); // shape = [nd_label,     ndata]
+
+    // shuffle idata first datapoints from dataset with RNG from opt_ctx, shuffle all datapoints if idata is negative
+    GGML_API void ggml_opt_dataset_shuffle(ggml_opt_context_t opt_ctx, ggml_opt_dataset_t dataset, int64_t idata);
+
+    // get batch at position ibatch from dataset and copy the data to data_batch and labels_batch
+    GGML_API void ggml_opt_dataset_get_batch(
+            ggml_opt_dataset_t   dataset,
+            struct ggml_tensor * data_batch,   // shape = [ne_datapoint, ndata_batch]
+            struct ggml_tensor * labels_batch, // shape = [ne_label,     ndata_batch]
+            int64_t              ibatch);
+
+    // ====== Model / Context ======
+
+    enum ggml_opt_build_type {
+        GGML_OPT_BUILD_TYPE_FORWARD,
+        GGML_OPT_BUILD_TYPE_GRAD,
+        GGML_OPT_BUILD_TYPE_OPT,
+    };
+
+    // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
+    struct ggml_opt_optimizer_params {
+        // AdamW optimizer parameters
+        struct {
+            float alpha; // learning rate
+            float beta1;
+            float beta2;
+            float eps;   // epsilon for numerical stability
+            float wd;    // weight decay for AdamW, use 0.0f to disable
+        } adamw;
+    };
+
+    // callback to calculate optimizer parameters prior to a backward pass
+    // userdata can be used to pass arbitrary data
+    typedef struct ggml_opt_optimizer_params (*ggml_opt_get_optimizer_params)(void * userdata);
+
+    // returns the default optimizer params (constant)
+    // userdata is not used
+    GGML_API struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata);
+
+    // parameters for initializing a new optimization context
+    struct ggml_opt_params {
+        ggml_backend_sched_t backend_sched; // defines which backends are used to construct the compute graphs
+
+        struct ggml_context * ctx_compute; // created in user code, holds non-static tensors
+
+        // the forward graph is defined by inputs and outputs
+        // those tensors and all tensors inbetween are not intended to be reusable between multiple optimization contexts
+        struct ggml_tensor * inputs;
+        struct ggml_tensor * outputs;
+
+        enum ggml_opt_loss_type  loss_type;
+        enum ggml_opt_build_type build_type;
+
+        int32_t opt_period; // after how many gradient accumulation steps an optimizer step should be done
+
+        ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
+        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+    };
+
+    // get parameters for an optimization context with defaults set where possible
+    // parameters for which no sensible defaults exist are supplied as arguments to this function
+    GGML_API ggml_opt_params ggml_opt_default_params(
+            ggml_backend_sched_t      backend_sched,
+            struct ggml_context     * ctx_compute,
+            struct ggml_tensor      * inputs,
+            struct ggml_tensor      * outputs,
+            enum ggml_opt_loss_type   loss_type);
+
+    GGML_API ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params);
+    GGML_API void ggml_opt_free(ggml_opt_context_t opt_ctx);
+
+    // set gradients to zero, initilize loss, and optionally reset the optimizer
+    GGML_API void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer);
+
+    // get underlying tensors that store data
+    GGML_API struct ggml_tensor * ggml_opt_inputs(  ggml_opt_context_t opt_ctx); // forward graph input tensor
+    GGML_API struct ggml_tensor * ggml_opt_outputs( ggml_opt_context_t opt_ctx); // forward graph output tensor
+    GGML_API struct ggml_tensor * ggml_opt_labels(  ggml_opt_context_t opt_ctx); // labels to compare outputs against
+    GGML_API struct ggml_tensor * ggml_opt_loss(    ggml_opt_context_t opt_ctx); // scalar tensor that contains the loss
+    GGML_API struct ggml_tensor * ggml_opt_pred(    ggml_opt_context_t opt_ctx); // predictions made by outputs
+    GGML_API struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx); // number of matching predictions between outputs and labels
+
+    GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
+
+    // ====== Optimization Result ======
+
+    GGML_API ggml_opt_result_t ggml_opt_result_init();
+    GGML_API void ggml_opt_result_free(ggml_opt_result_t result);
+    GGML_API void ggml_opt_result_reset(ggml_opt_result_t result);
+
+    // get data from result, uncertainties are optional and can be ignored by passing NULL
+    GGML_API void ggml_opt_result_ndata(   ggml_opt_result_t result, int64_t * ndata);                  // writes 1 value, number of datapoints
+    GGML_API void ggml_opt_result_loss(    ggml_opt_result_t result, double  * loss,     double * unc); // writes 1 value
+    GGML_API void ggml_opt_result_pred(    ggml_opt_result_t result, int32_t * pred);                   // writes ndata values
+    GGML_API void ggml_opt_result_accuracy(ggml_opt_result_t result, double  * accuracy, double * unc); // writes 1 value
+
+    // ====== Computation ======
+
+    // do forward pass, increment result if not NULL
+    GGML_API void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+
+    // do forward pass, increment result if not NULL, do backward pass
+    GGML_API void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result_t result);
+
+    // ############################################################################
+    // ## The high-level functions start here. They do not depend on any private ##
+    // ## functions or structs and can be copied to and adapted for user code.   ##
+    // ############################################################################
+
+    // ====== Intended Usage ======
+    //
+    // 1. Select the appropriate loss for your problem.
+    // 2. Create a dataset and set the data for the "data" tensor. Also set the "labels" tensor if your loss needs them.
+    //    Setting the shard size to 1 will be fine, it's the granularity with which data is shuffled/loaded (bigger values are faster).
+    // 3. Create a GGML graph for your model with no_alloc == true. Use two separate contexts for the tensors.
+    //    The first context should contain the model parameters and inputs and be allocated statically in user code.
+    //    The second context should contain all other tensors and will be (re)allocated automatically.
+    //    Due to this automated allocation the data of the second context is not defined when accessed in user code.
+    //    Note that the second dimension of the inputs/outputs are interpreted as the number of datapoints in those tensors.
+    // 4. Call ggml_opt_fit. If you need more control you can use ggml_opt_epoch instead.
+
+    // signature for a callback while evaluating opt_ctx on dataset, called after an evaluation
+    typedef void (*ggml_opt_epoch_callback)(
+            bool               train,       // true after training evaluation, false after validation evaluation
+            ggml_opt_context_t opt_ctx,
+            ggml_opt_dataset_t dataset,
+            ggml_opt_result_t  result,      // result associated with the dataset subsection
+            int64_t            ibatch,      // number of batches that have been evaluated so far
+            int64_t            ibatch_max,  // total number of batches in this dataset subsection
+            int64_t            t_start_us); // time at which the evaluation on the dataset subsection was started
+
+    // do training on front of dataset, do evaluation only on back of dataset
+    GGML_API void ggml_opt_epoch(
+            ggml_opt_context_t      opt_ctx,
+            ggml_opt_dataset_t      dataset,
+            ggml_opt_result_t       result_train,   // result to increment during training, ignored if NULL
+            ggml_opt_result_t       result_eval,    // result to increment during evaluation, ignored if NULL
+            int64_t                 idata_split,    // data index at which to split training and evaluation
+            ggml_opt_epoch_callback callback_train,
+            ggml_opt_epoch_callback callback_eval);
+
+    // callback that prints a progress bar on stderr
+    GGML_API void ggml_opt_epoch_callback_progress_bar(
+            bool               train,
+            ggml_opt_context_t opt_ctx,
+            ggml_opt_dataset_t dataset,
+            ggml_opt_result_t  result,
+            int64_t            ibatch,
+            int64_t            ibatch_max,
+            int64_t            t_start_us);
+
+    // fit model defined by inputs and outputs to dataset
+    GGML_API void ggml_opt_fit(
+            ggml_backend_sched_t            backend_sched,  // backend scheduler for constructing the compute graphs
+            ggml_context                  * ctx_compute,    // context with temporarily allocated tensors to calculate the outputs
+            ggml_tensor                   * inputs,         // input tensor with shape [ne_datapoint, ndata_batch]
+            ggml_tensor                   * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
+            ggml_opt_dataset_t              dataset,        // dataset with data and optionally also labels
+            enum ggml_opt_loss_type         loss_type,      // loss to minimize
+            ggml_opt_get_optimizer_params   get_opt_pars,   // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)
+            int64_t                         nepoch,         // how many times the dataset should be iterated over
+            int64_t                         nbatch_logical, // datapoints optimizer step, must be a multiple of ndata_batch in inputs/outputs
+            float                           val_split,      // fraction of the dataset to use for validation, must be in [0.0f, 1.0f)
+            bool                            silent);        // whether or not info prints to stderr should be suppressed
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/ggml/include/ggml-rpc.h b/ggml/include/ggml-rpc.h
index d5796736821..ade6c3b0ef1 100644
--- a/ggml/include/ggml-rpc.h
+++ b/ggml/include/ggml-rpc.h
@@ -10,18 +10,18 @@ extern "C" {
 #define GGML_RPC_MAX_SERVERS       16
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
-GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend);
+GGML_BACKEND_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
+GGML_BACKEND_API bool ggml_backend_is_rpc(ggml_backend_t backend);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
 
-GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
+GGML_BACKEND_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
 
-GGML_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+GGML_BACKEND_API void ggml_backend_rpc_start_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
 
-GGML_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_rpc_reg(void);
 
-GGML_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
+GGML_BACKEND_API ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint);
 
 #ifdef  __cplusplus
 }
diff --git a/ggml/include/ggml-sycl.h b/ggml/include/ggml-sycl.h
index af521f59930..5ce349a880e 100644
--- a/ggml/include/ggml-sycl.h
+++ b/ggml/include/ggml-sycl.h
@@ -17,32 +17,32 @@ extern "C" {
 #endif
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
+GGML_BACKEND_API ggml_backend_t ggml_backend_sycl_init(int device);
 
-GGML_API bool ggml_backend_is_sycl(ggml_backend_t backend);
+GGML_BACKEND_API bool ggml_backend_is_sycl(ggml_backend_t backend);
 
 // devide buffer
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
 
 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
 
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
 
-GGML_API void ggml_backend_sycl_print_sycl_devices(void);
-GGML_API void ggml_backend_sycl_get_gpu_list(int *id_list, int max_len);
-GGML_API void ggml_backend_sycl_get_device_description(int device,
+GGML_BACKEND_API void ggml_backend_sycl_print_sycl_devices(void);
+GGML_BACKEND_API void ggml_backend_sycl_get_gpu_list(int *id_list, int max_len);
+GGML_BACKEND_API void ggml_backend_sycl_get_device_description(int device,
                                                        char *description,
                                                        size_t description_size);
-GGML_API int  ggml_backend_sycl_get_device_count();
-GGML_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
+GGML_BACKEND_API int  ggml_backend_sycl_get_device_count();
+GGML_BACKEND_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
 
 // SYCL doesn't support registering host memory, keep here for reference
-// GGML_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
-// GGML_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
+// GGML_BACKEND_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
+// GGML_BACKEND_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
 
-GGML_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
 
 #ifdef  __cplusplus
 }
diff --git a/ggml/include/ggml-vulkan.h b/ggml/include/ggml-vulkan.h
index c03bbfe5e37..53cdba072c2 100644
--- a/ggml/include/ggml-vulkan.h
+++ b/ggml/include/ggml-vulkan.h
@@ -10,21 +10,21 @@ extern "C" {
 #define GGML_VK_NAME "Vulkan"
 #define GGML_VK_MAX_DEVICES 16
 
-GGML_API void ggml_vk_instance_init(void);
+GGML_BACKEND_API void ggml_vk_instance_init(void);
 
 // backend API
-GGML_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
+GGML_BACKEND_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
 
-GGML_API bool ggml_backend_is_vk(ggml_backend_t backend);
-GGML_API int  ggml_backend_vk_get_device_count(void);
-GGML_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
-GGML_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
+GGML_BACKEND_API bool ggml_backend_is_vk(ggml_backend_t backend);
+GGML_BACKEND_API int  ggml_backend_vk_get_device_count(void);
+GGML_BACKEND_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
+GGML_BACKEND_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
+GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
 
-GGML_API ggml_backend_reg_t ggml_backend_vk_reg(void);
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_vk_reg(void);
 
 #ifdef  __cplusplus
 }
diff --git a/ggml/include/ggml.h b/ggml/include/ggml.h
index 73ede181331..b0c1ac9ce2b 100644
--- a/ggml/include/ggml.h
+++ b/ggml/include/ggml.h
@@ -176,15 +176,15 @@
 #ifdef GGML_SHARED
 #    if defined(_WIN32) && !defined(__MINGW32__)
 #        ifdef GGML_BUILD
-#            define GGML_API __declspec(dllexport)
+#            define GGML_API __declspec(dllexport) extern
 #        else
-#            define GGML_API __declspec(dllimport)
+#            define GGML_API __declspec(dllimport) extern
 #        endif
 #    else
-#        define GGML_API __attribute__ ((visibility ("default")))
+#        define GGML_API __attribute__ ((visibility ("default"))) extern
 #    endif
 #else
-#    define GGML_API
+#    define GGML_API extern
 #endif
 
 // TODO: support for clang
@@ -237,7 +237,9 @@
 #define GGML_EXIT_SUCCESS 0
 #define GGML_EXIT_ABORTED 1
 
-#define GGML_ROPE_TYPE_NEOX 2
+#define GGML_ROPE_TYPE_NEOX   2
+#define GGML_ROPE_TYPE_MROPE  8
+#define GGML_ROPE_TYPE_VISION 24
 
 #define GGUF_MAGIC "GGUF"
 
@@ -384,12 +386,15 @@ extern "C" {
         GGML_TYPE_F64     = 28,
         GGML_TYPE_IQ1_M   = 29,
         GGML_TYPE_BF16    = 30,
-        GGML_TYPE_Q4_0_4_4 = 31,
-        GGML_TYPE_Q4_0_4_8 = 32,
-        GGML_TYPE_Q4_0_8_8 = 33,
+        // GGML_TYPE_Q4_0_4_4 = 31, support has been removed from gguf files
+        // GGML_TYPE_Q4_0_4_8 = 32,
+        // GGML_TYPE_Q4_0_8_8 = 33,
         GGML_TYPE_TQ1_0   = 34,
         GGML_TYPE_TQ2_0   = 35,
-        GGML_TYPE_COUNT,
+        // GGML_TYPE_IQ4_NL_4_4 = 36,
+        // GGML_TYPE_IQ4_NL_4_8 = 37,
+        // GGML_TYPE_IQ4_NL_8_8 = 38,
+        GGML_TYPE_COUNT   = 39,
     };
 
     // precision
@@ -430,9 +435,6 @@ extern "C" {
         GGML_FTYPE_MOSTLY_IQ4_XS  = 22, // except 1d tensors
         GGML_FTYPE_MOSTLY_IQ1_M   = 23, // except 1d tensors
         GGML_FTYPE_MOSTLY_BF16    = 24, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_0_4_4 = 25, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_0_4_8 = 26, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_0_8_8 = 27, // except 1d tensors
     };
 
     // available tensor operations:
@@ -496,6 +498,7 @@ extern "C" {
         GGML_OP_POOL_2D_BACK,
         GGML_OP_UPSCALE, // nearest interpolate
         GGML_OP_PAD,
+        GGML_OP_PAD_REFLECT_1D,
         GGML_OP_ARANGE,
         GGML_OP_TIMESTEP_EMBEDDING,
         GGML_OP_ARGSORT,
@@ -602,7 +605,6 @@ extern "C" {
 
         int32_t flags;
 
-        struct ggml_tensor * grad;
         struct ggml_tensor * src[GGML_MAX_SRC];
 
         // source tensor and offset for views
@@ -615,7 +617,7 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        // char padding[4];
+        char padding[8];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
@@ -1443,6 +1445,22 @@ extern "C" {
             float                 beta_fast,
             float                 beta_slow);
 
+    GGML_API struct ggml_tensor * ggml_rope_multi(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
+            int                   n_dims,
+            int                   sections[4],
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
     // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_rope_ext_inplace(
             struct ggml_context * ctx,
@@ -1490,7 +1508,7 @@ extern "C" {
         "use ggml_rope_ext_inplace instead");
 
     // compute correction dims for YaRN RoPE scaling
-    void ggml_rope_yarn_corr_dims(
+    GGML_API void ggml_rope_yarn_corr_dims(
         int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]);
 
     // rotary position embedding backward, i.e compute dx from dy
@@ -1693,6 +1711,13 @@ extern "C" {
             int                  p2,
             int                  p3);
 
+    // pad each dimension with reflection: [a, b, c, d] -> [b, a, b, c, d, c]
+    GGML_API struct ggml_tensor * ggml_pad_reflect_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   p0,
+            int                   p1);
+
     // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
     // timesteps: [N,]
     // return: [N, dim]
@@ -1985,28 +2010,20 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * grad,
-            float                 alpha,
-            float                 beta1,
-            float                 beta2,
-            float                 eps,
-            float                 wd); // weight decay
+            struct ggml_tensor  * m,
+            struct ggml_tensor  * v,
+            struct ggml_tensor  * adamw_params); // parameters such a the learning rate
 
     //
     // automatic differentiation
     //
 
-    GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
-    GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate);
-
-    GGML_API void ggml_build_opt_adamw(
-            struct ggml_context * ctx,
-            struct ggml_cgraph  * gf,
-            struct ggml_cgraph  * gb,
-            float                 alpha,
-            float                 beta1,
-            float                 beta2,
-            float                 eps,
-            float                 wd); // weight decay
+    GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
+    GGML_API void ggml_build_backward_expand(
+        struct ggml_context * ctx_static,  // context for static gradients (loss + gradient accumulation)
+        struct ggml_context * ctx_compute, // context for gradient computation
+        struct ggml_cgraph  * cgraph,
+        bool                  accumulate); // whether or not gradients should be accumulated, requires static allocation of tensors in ctx_static
 
     // graph allocation in a context
     GGML_API struct ggml_cgraph * ggml_new_graph       (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
@@ -2026,7 +2043,9 @@ extern "C" {
     GGML_API size_t ggml_graph_overhead(void);
     GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
 
-    GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
+    GGML_API struct ggml_tensor * ggml_graph_get_tensor  (const struct ggml_cgraph * cgraph, const char * name);
+    GGML_API struct ggml_tensor * ggml_graph_get_grad    (const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
+    GGML_API struct ggml_tensor * ggml_graph_get_grad_acc(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
 
     GGML_API void                 ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
     GGML_API struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
@@ -2037,198 +2056,15 @@ extern "C" {
     // dump the graph into a file using the dot format
     GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
 
-    // build gradient checkpointing backward graph gb for gf using provided checkpoints
-    // gb_tmp will contain original backward graph with rewritten backward process nodes,
-    // but without the second forward pass nodes.
-    GGML_API void ggml_build_backward_gradient_checkpointing(
-            struct ggml_context   * ctx,
-            struct ggml_cgraph    * gf,
-            struct ggml_cgraph    * gb,
-            struct ggml_cgraph    * gb_tmp,
-            struct ggml_tensor  * * checkpoints,
-            int                     n_checkpoints);
-    //
-    // optimization
-    //
-
-    // optimization methods
-    enum ggml_opt_type {
-        GGML_OPT_TYPE_ADAM,
-        GGML_OPT_TYPE_LBFGS,
-    };
-
-    // linesearch methods
-    enum ggml_linesearch {
-        GGML_LINESEARCH_DEFAULT = 1,
-
-        GGML_LINESEARCH_BACKTRACKING_ARMIJO       = 0,
-        GGML_LINESEARCH_BACKTRACKING_WOLFE        = 1,
-        GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 2,
-    };
-
-    // optimization return values
-    enum ggml_opt_result {
-        GGML_OPT_RESULT_OK = 0,
-        GGML_OPT_RESULT_DID_NOT_CONVERGE,
-        GGML_OPT_RESULT_NO_CONTEXT,
-        GGML_OPT_RESULT_INVALID_WOLFE,
-        GGML_OPT_RESULT_FAIL,
-        GGML_OPT_RESULT_CANCEL,
-
-        GGML_LINESEARCH_FAIL = -128,
-        GGML_LINESEARCH_MINIMUM_STEP,
-        GGML_LINESEARCH_MAXIMUM_STEP,
-        GGML_LINESEARCH_MAXIMUM_ITERATIONS,
-        GGML_LINESEARCH_INVALID_PARAMETERS,
-    };
-
-    typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel);
+    // TODO these functions were sandwiched in the old optimization interface, is there a better place for them?
     typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);
 
     // Set callback for all future logging events.
     // If this is not called, or NULL is supplied, everything is output on stderr.
     GGML_API void ggml_log_set(ggml_log_callback log_callback, void * user_data);
 
-    // optimization parameters
-    //
-    //   see ggml.c (ggml_opt_default_params) for default values
-    //
-    struct ggml_opt_params {
-        enum ggml_opt_type type;
-
-        size_t graph_size;
-
-        int n_threads;
-
-        // delta-based convergence test
-        //
-        //   if past == 0 - disabled
-        //   if past > 0:
-        //     stop if |f(x) - f(x_past)| < delta * max(1, |f(x)|)
-        //
-        int past;
-        float delta;
-
-        // maximum number of iterations without improvement
-        //
-        //   if 0 - disabled
-        //   if > 0:
-        //     assume convergence if no cost improvement in this number of iterations
-        //
-        int max_no_improvement;
-
-        bool print_forward_graph;
-        bool print_backward_graph;
-
-        int n_gradient_accumulation;
-
-        // ADAM parameters
-        struct {
-            int n_iter;
-
-            float sched; // schedule multiplier (fixed, decay or warmup)
-            float decay; // weight decay for AdamW, use 0.0f to disable
-            int   decay_min_ndim; // minimum number of tensor dimension to apply weight decay
-            float alpha; // learning rate
-            float beta1;
-            float beta2;
-            float eps;   // epsilon for numerical stability
-            float eps_f; // epsilon for convergence test
-            float eps_g; // epsilon for convergence test
-            float gclip; // gradient clipping
-        } adam;
-
-        // LBFGS parameters
-        struct {
-            int m; // number of corrections to approximate the inv. Hessian
-            int n_iter;
-            int max_linesearch;
-
-            float eps;      // convergence tolerance
-            float ftol;     // line search tolerance
-            float wolfe;
-            float min_step;
-            float max_step;
-
-            enum ggml_linesearch linesearch;
-        } lbfgs;
-    };
-
-    struct ggml_opt_context {
-        struct ggml_context * ctx;
-        struct ggml_opt_params params;
-
-        int iter;
-        int64_t nx; // number of parameter elements
-
-        bool just_initialized;
-
-        float loss_before;
-        float loss_after;
-
-        struct {
-            struct ggml_tensor * g;  // current gradient
-            struct ggml_tensor * m;  // first moment
-            struct ggml_tensor * v;  // second moment
-            struct ggml_tensor * pf; // past function values
-            float fx_best;
-            float fx_prev;
-            int n_no_improvement;
-        } adam;
-
-        struct {
-            struct ggml_tensor * x;    // current parameters
-            struct ggml_tensor * xp;   // previous parameters
-            struct ggml_tensor * g;    // current gradient
-            struct ggml_tensor * gp;   // previous gradient
-            struct ggml_tensor * d;    // search direction
-            struct ggml_tensor * pf;   // past function values
-            struct ggml_tensor * lmal; // the L-BFGS memory alpha
-            struct ggml_tensor * lmys; // the L-BFGS memory ys
-            struct ggml_tensor * lms;  // the L-BFGS memory s
-            struct ggml_tensor * lmy;  // the L-BFGS memory y
-            float fx_best;
-            float step;
-            int j;
-            int k;
-            int end;
-            int n_no_improvement;
-        } lbfgs;
-    };
-
     GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
 
-    GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
-
-    // optimize the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt(
-            struct ggml_context * ctx,
-            struct ggml_opt_params params,
-            struct ggml_tensor * f);
-
-    // initialize optimizer context
-    GGML_API void ggml_opt_init(
-            struct ggml_context     * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_opt_params    params,
-            int64_t                   nx);
-
-    // continue optimizing the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt_resume(
-            struct ggml_context * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_tensor * f);
-
-    // continue optimizing the function defined by the tensor f
-    GGML_API enum ggml_opt_result ggml_opt_resume_g(
-            struct ggml_context * ctx,
-            struct ggml_opt_context * opt,
-            struct ggml_tensor * f,
-            struct ggml_cgraph * gf,
-            struct ggml_cgraph * gb,
-            ggml_opt_callback callback,
-            void * callback_data);
-
     //
     // quantization
     //
@@ -2384,43 +2220,19 @@ extern "C" {
     GGML_API size_t gguf_get_meta_size(const struct gguf_context * ctx);
     GGML_API void   gguf_get_meta_data(const struct gguf_context * ctx, void * data);
 
-    //
-    // system info
-    //
-
-    GGML_API int ggml_cpu_has_avx        (void);
-    GGML_API int ggml_cpu_has_avx_vnni   (void);
-    GGML_API int ggml_cpu_has_avx2       (void);
-    GGML_API int ggml_cpu_has_avx512     (void);
-    GGML_API int ggml_cpu_has_avx512_vbmi(void);
-    GGML_API int ggml_cpu_has_avx512_vnni(void);
-    GGML_API int ggml_cpu_has_avx512_bf16(void);
-    GGML_API int ggml_cpu_has_amx_int8   (void);
-    GGML_API int ggml_cpu_has_fma        (void);
-    GGML_API int ggml_cpu_has_arm_fma    (void);
-    GGML_API int ggml_cpu_has_metal      (void);
-    GGML_API int ggml_cpu_has_f16c       (void);
-    GGML_API int ggml_cpu_has_fp16_va    (void);
-    GGML_API int ggml_cpu_has_wasm_simd  (void);
-    GGML_API int ggml_cpu_has_blas       (void);
-    GGML_API int ggml_cpu_has_cuda       (void);
-    GGML_API int ggml_cpu_has_vulkan     (void);
-    GGML_API int ggml_cpu_has_kompute    (void);
-    GGML_API int ggml_cpu_has_gpublas    (void);
-    GGML_API int ggml_cpu_has_sse3       (void);
-    GGML_API int ggml_cpu_has_ssse3      (void);
-    GGML_API int ggml_cpu_has_riscv_v    (void);
-    GGML_API int ggml_cpu_has_sycl       (void);
-    GGML_API int ggml_cpu_has_rpc        (void);
-    GGML_API int ggml_cpu_has_vsx        (void);
-    GGML_API int ggml_cpu_has_cann       (void);
-    GGML_API int ggml_cpu_has_llamafile  (void);
-
-#ifdef  __cplusplus
-// restrict not standard in C++
-#define GGML_RESTRICT
+#ifdef __cplusplus
+    // restrict not standard in C++
+#    if defined(__GNUC__)
+#        define GGML_RESTRICT __restrict__
+#    elif defined(__clang__)
+#        define GGML_RESTRICT __restrict
+#    elif defined(_MSC_VER)
+#        define GGML_RESTRICT __restrict
+#    else
+#        define GGML_RESTRICT
+#    endif
 #else
-#define GGML_RESTRICT restrict
+#    define GGML_RESTRICT restrict
 #endif
     typedef void (*ggml_to_float_t)  (const void  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
     typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int64_t k);
@@ -2432,12 +2244,42 @@ extern "C" {
         size_t                   type_size;
         bool                     is_quantized;
         ggml_to_float_t          to_float;
-        ggml_from_float_t        from_float;
         ggml_from_float_t        from_float_ref;
     };
 
     GGML_API const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type);
 
+    // ggml threadpool
+    // TODO: currently, only a few functions are in the base ggml API, while the rest are in the CPU backend
+    // the goal should be to create an API that other backends can use move everything to the ggml base
+
+    // scheduling priorities
+    enum ggml_sched_priority {
+        GGML_SCHED_PRIO_NORMAL,
+        GGML_SCHED_PRIO_MEDIUM,
+        GGML_SCHED_PRIO_HIGH,
+        GGML_SCHED_PRIO_REALTIME
+    };
+
+    // threadpool params
+    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
+    struct ggml_threadpool_params {
+        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
+        int                 n_threads;                   // number of threads
+        enum ggml_sched_priority prio;                   // thread priority
+        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
+        bool                strict_cpu;                  // strict cpu placement
+        bool                paused;                      // start in paused state
+    };
+
+    struct ggml_threadpool;     // forward declaration, see ggml.c
+
+    typedef struct ggml_threadpool * ggml_threadpool_t;
+
+    GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
+    GGML_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
+    GGML_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
+
 #ifdef  __cplusplus
 }
 #endif
diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
index 48df3561860..bf5ee5fc273 100644
--- a/ggml/src/CMakeLists.txt
+++ b/ggml/src/CMakeLists.txt
@@ -1,7 +1,5 @@
 include(CheckCXXCompilerFlag)
 
-unset(GGML_CDEF_PUBLIC)
-
 add_compile_definitions(GGML_SCHED_MAX_COPIES=${GGML_SCHED_MAX_COPIES})
 
 # enable libstdc++ assertions for debug builds
@@ -26,930 +24,7 @@ if (NOT MSVC)
     endif()
 endif()
 
-unset(GGML_EXTRA_LIBS_PRIVATE)
-unset(GGML_EXTRA_LIBS_PUBLIC)
-
-if (APPLE AND GGML_ACCELERATE)
-    find_library(ACCELERATE_FRAMEWORK Accelerate)
-    if (ACCELERATE_FRAMEWORK)
-        message(STATUS "Accelerate framework found")
-
-        add_compile_definitions(GGML_USE_ACCELERATE)
-        add_compile_definitions(ACCELERATE_NEW_LAPACK)
-        add_compile_definitions(ACCELERATE_LAPACK_ILP64)
-
-        list(APPEND GGML_EXTRA_LIBS_PRIVATE ${ACCELERATE_FRAMEWORK})
-    else()
-        message(WARNING "Accelerate framework not found")
-    endif()
-endif()
-
-if (GGML_METAL)
-    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
-    find_library(METAL_FRAMEWORK    Metal      REQUIRED)
-    find_library(METALKIT_FRAMEWORK MetalKit   REQUIRED)
-
-    message(STATUS "Metal framework found")
-    set(GGML_HEADERS_METAL ../include/ggml-metal.h)
-    set(GGML_SOURCES_METAL ggml-metal.m)
-
-    list(APPEND GGML_CDEF_PUBLIC GGML_USE_METAL)
-    if (GGML_METAL_NDEBUG)
-        add_compile_definitions(GGML_METAL_NDEBUG)
-    endif()
-
-    if (GGML_METAL_USE_BF16)
-        add_compile_definitions(GGML_METAL_USE_BF16)
-    endif()
-
-    # copy ggml-common.h and ggml-metal.metal to bin directory
-    configure_file(ggml-common.h    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h    COPYONLY)
-    configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
-
-    if (GGML_METAL_EMBED_LIBRARY)
-        enable_language(ASM)
-
-        add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
-
-        set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/ggml-common.h")
-        set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-
-        file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
-
-        # merge ggml-common.h and ggml-metal.metal into a single file
-        set(METALLIB_EMBED_ASM    "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
-        set(METALLIB_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
-
-        add_custom_command(
-            OUTPUT ${METALLIB_EMBED_ASM}
-            COMMAND echo "Embedding Metal library"
-            COMMAND sed -e '/\#include \"ggml-common.h\"/r ${METALLIB_COMMON}' -e '/\#include \"ggml-common.h\"/d' < ${METALLIB_SOURCE} > ${METALLIB_SOURCE_EMBED}
-            COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
-            COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
-            COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
-            COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
-            COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
-            COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
-            DEPENDS ggml-metal.metal ggml-common.h
-            COMMENT "Generate assembly for embedded Metal library"
-        )
-
-        list(APPEND GGML_SOURCES_METAL ${METALLIB_EMBED_ASM})
-    else()
-        if (GGML_METAL_SHADER_DEBUG)
-            # custom command to do the following:
-            #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-            #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-            #
-            # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-            #       disabling fast math is needed in order to pass tests/test-backend-ops
-            # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
-            # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
-            #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
-            set(XC_FLAGS -fno-fast-math -fno-inline -g)
-        else()
-            set(XC_FLAGS -O3)
-        endif()
-
-        # Append macOS metal versioning flags
-        if (GGML_METAL_MACOSX_VERSION_MIN)
-            message(STATUS "Adding  -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
-            list   (APPEND XC_FLAGS -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN})
-        endif()
-
-        if (GGML_METAL_STD)
-            message(STATUS "Adding  -std=${GGML_METAL_STD} flag to metal compilation")
-            list   (APPEND XC_FLAGS -std=${GGML_METAL_STD})
-        endif()
-
-        add_custom_command(
-            OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
-            COMMAND xcrun -sdk macosx metallib                ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air   -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
-            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
-            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
-            DEPENDS ggml-metal.metal ggml-common.h
-            COMMENT "Compiling Metal kernels"
-            )
-
-        add_custom_target(
-            ggml-metal ALL
-            DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            )
-    endif() # GGML_METAL_EMBED_LIBRARY
-
-    list(APPEND GGML_EXTRA_LIBS_PRIVATE
-        ${FOUNDATION_LIBRARY}
-        ${METAL_FRAMEWORK}
-        ${METALKIT_FRAMEWORK}
-        )
-endif()
-
-if (GGML_MUSA)
-    set(CMAKE_C_COMPILER clang)
-    set(CMAKE_C_EXTENSIONS OFF)
-    set(CMAKE_CXX_COMPILER clang++)
-    set(CMAKE_CXX_EXTENSIONS OFF)
-
-    set(GGML_CUDA ON)
-
-    list(APPEND GGML_CDEF_PUBLIC GGML_USE_MUSA)
-endif()
-
-if (GGML_OPENMP)
-    find_package(OpenMP)
-    if (OpenMP_FOUND)
-        message(STATUS "OpenMP found")
-
-        add_compile_definitions(GGML_USE_OPENMP)
-
-        list(APPEND GGML_EXTRA_LIBS_PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
-
-        if (GGML_MUSA)
-            list(APPEND GGML_EXTRA_INCLUDES     "/usr/lib/llvm-14/lib/clang/14.0.0/include")
-            list(APPEND GGML_EXTRA_LIBS_PRIVATE "/usr/lib/llvm-14/lib/libomp.so")
-        endif()
-    else()
-        message(WARNING "OpenMP not found")
-    endif()
-endif()
-
-if (GGML_BLAS)
-    if (GGML_STATIC)
-        set(BLA_STATIC ON)
-    endif()
-    #if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
-    #    set(BLA_SIZEOF_INTEGER 8)
-    #endif()
-
-    set(BLA_VENDOR ${GGML_BLAS_VENDOR})
-    find_package(BLAS)
-
-    if (BLAS_FOUND)
-        message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
-
-        if (("${BLAS_INCLUDE_DIRS}" STREQUAL "") AND NOT (${GGML_BLAS_VENDOR} MATCHES "Apple"))
-            # BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
-            # see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
-            find_package(PkgConfig REQUIRED)
-            if (${GGML_BLAS_VENDOR} MATCHES "Generic")
-                pkg_check_modules(DepBLAS blas)
-            elseif (${GGML_BLAS_VENDOR} MATCHES "OpenBLAS")
-                # As of openblas v0.3.22, the 64-bit is named openblas64.pc
-                pkg_check_modules(DepBLAS openblas64)
-                if (NOT DepBLAS_FOUND)
-                    pkg_check_modules(DepBLAS openblas)
-                endif()
-            elseif (${GGML_BLAS_VENDOR} MATCHES "FLAME")
-                add_compile_definitions(GGML_BLAS_USE_BLIS)
-                pkg_check_modules(DepBLAS blis)
-            elseif (${GGML_BLAS_VENDOR} MATCHES "ATLAS")
-                pkg_check_modules(DepBLAS blas-atlas)
-            elseif (${GGML_BLAS_VENDOR} MATCHES "FlexiBLAS")
-                pkg_check_modules(DepBLAS flexiblas_api)
-            elseif (${GGML_BLAS_VENDOR} MATCHES "Intel")
-                add_compile_definitions(GGML_BLAS_USE_MKL)
-                # all Intel* libraries share the same include path
-                pkg_check_modules(DepBLAS mkl-sdl)
-            elseif (${GGML_BLAS_VENDOR} MATCHES "NVHPC")
-                # this doesn't provide pkg-config
-                # suggest to assign BLAS_INCLUDE_DIRS on your own
-                if ("${NVHPC_VERSION}" STREQUAL "")
-                    message(WARNING "Better to set NVHPC_VERSION")
-                else()
-                    set(DepBLAS_FOUND ON)
-                    set(DepBLAS_INCLUDE_DIRS "/opt/nvidia/hpc_sdk/${CMAKE_SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}/${NVHPC_VERSION}/math_libs/include")
-                endif()
-            endif()
-            if (DepBLAS_FOUND)
-                set(BLAS_INCLUDE_DIRS ${DepBLAS_INCLUDE_DIRS})
-            else()
-                message(WARNING "BLAS_INCLUDE_DIRS neither been provided nor been automatically"
-                " detected by pkgconfig, trying to find cblas.h from possible paths...")
-                find_path(BLAS_INCLUDE_DIRS
-                    NAMES cblas.h
-                    HINTS
-                        /usr/include
-                        /usr/local/include
-                        /usr/include/openblas
-                        /opt/homebrew/opt/openblas/include
-                        /usr/local/opt/openblas/include
-                        /usr/include/x86_64-linux-gnu/openblas/include
-                )
-            endif()
-        endif()
-
-        message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
-
-        add_compile_options(${BLAS_LINKER_FLAGS})
-
-        list(APPEND GGML_CDEF_PUBLIC GGML_USE_BLAS)
-
-        if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
-            add_compile_definitions(GGML_BLAS_USE_MKL)
-        endif()
-
-        set(GGML_HEADERS_BLAS ../include/ggml-blas.h)
-        set(GGML_SOURCES_BLAS ggml-blas.cpp)
-
-        list(APPEND GGML_EXTRA_LIBS_PRIVATE ${BLAS_LIBRARIES})
-        list(APPEND GGML_EXTRA_INCLUDES     ${BLAS_INCLUDE_DIRS})
-    else()
-        message(WARNING "BLAS not found, please refer to "
-        "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
-        " to set correct GGML_BLAS_VENDOR")
-    endif()
-endif()
-
-if (GGML_LLAMAFILE)
-    message(STATUS "Using llamafile")
-
-    add_compile_definitions(GGML_USE_LLAMAFILE)
-
-    set(GGML_HEADERS_LLAMAFILE llamafile/sgemm.h)
-    set(GGML_SOURCES_LLAMAFILE llamafile/sgemm.cpp)
-endif()
-
-if (GGML_AMX)
-    if (CMAKE_COMPILER_IS_GNUCC AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 11.0)
-    else()
-        set(GGML_AMX OFF)
-        message(WARNING "AMX requires gcc version > 11.0. Turning off GGML_AMX.")
-    endif()
-
-    if (GGML_AMX)
-        message(STATUS "Using AMX")
-
-        list(APPEND GGML_CDEF_PUBLIC GGML_USE_AMX)
-
-        file(GLOB   GGML_HEADERS_AMX "ggml-amx/*.h")
-        list(APPEND GGML_HEADERS_AMX "../include/ggml-amx.h")
-
-        file(GLOB   GGML_SOURCES_AMX "ggml-amx/*.cpp")
-        list(APPEND GGML_SOURCES_AMX "ggml-amx.cpp")
-    endif()
-endif()
-
-if (GGML_CUDA)
-    cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
-
-    if (GGML_MUSA)
-        list(APPEND CMAKE_MODULE_PATH "/usr/local/musa/cmake/")
-        find_package(MUSAToolkit)
-        set(CUDAToolkit_FOUND ${MUSAToolkit_FOUND})
-    else()
-        find_package(CUDAToolkit)
-    endif()
-
-    if (CUDAToolkit_FOUND)
-        message(STATUS "CUDA found")
-
-        if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-            # 52 == lowest CUDA 12 standard
-            # 60 == FP16 CUDA intrinsics
-            # 61 == integer CUDA intrinsics
-            # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
-            if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
-                set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
-            else()
-                set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
-                #set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
-            endif()
-        endif()
-        message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
-
-        if (GGML_MUSA)
-            set(CMAKE_CUDA_COMPILER ${MUSAToolkit_MCC_EXECUTABLE})
-        else()
-            enable_language(CUDA)
-        endif()
-
-        file(GLOB   GGML_HEADERS_CUDA "ggml-cuda/*.cuh")
-        list(APPEND GGML_HEADERS_CUDA "../include/ggml-cuda.h")
-
-        file(GLOB   GGML_SOURCES_CUDA "ggml-cuda/*.cu")
-        list(APPEND GGML_SOURCES_CUDA "ggml-cuda.cu")
-        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-wmma*.cu")
-        list(APPEND GGML_SOURCES_CUDA ${SRCS})
-        file(GLOB   SRCS "ggml-cuda/template-instances/mmq*.cu")
-        list(APPEND GGML_SOURCES_CUDA ${SRCS})
-
-        if (GGML_CUDA_FA_ALL_QUANTS)
-            file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*.cu")
-            list(APPEND GGML_SOURCES_CUDA ${SRCS})
-            add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
-        else()
-            file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
-            list(APPEND GGML_SOURCES_CUDA ${SRCS})
-            file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
-            list(APPEND GGML_SOURCES_CUDA ${SRCS})
-            file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
-            list(APPEND GGML_SOURCES_CUDA ${SRCS})
-        endif()
-
-        list(APPEND GGML_CDEF_PUBLIC GGML_USE_CUDA)
-
-        add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
-        add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
-        add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
-        add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
-
-        if (GGML_CUDA_GRAPHS)
-            add_compile_definitions(GGML_CUDA_USE_GRAPHS)
-        endif()
-
-        if (GGML_CUDA_FORCE_DMMV)
-            add_compile_definitions(GGML_CUDA_FORCE_DMMV)
-        endif()
-
-        if (GGML_CUDA_FORCE_MMQ)
-            add_compile_definitions(GGML_CUDA_FORCE_MMQ)
-        endif()
-
-        if (GGML_CUDA_FORCE_CUBLAS)
-            add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
-        endif()
-
-        if (GGML_CUDA_NO_VMM)
-            add_compile_definitions(GGML_CUDA_NO_VMM)
-        endif()
-
-        if (DEFINED GGML_CUDA_DMMV_Y)
-            add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_DMMV_Y}) # for backwards compatibility
-        endif()
-
-        if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
-            add_compile_definitions(GGML_CUDA_F16)
-        endif()
-
-        if (GGML_CUDA_NO_PEER_COPY)
-            add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
-        endif()
-
-        if (GGML_MUSA)
-            set_source_files_properties(${GGML_SOURCES_CUDA} PROPERTIES LANGUAGE CXX)
-            foreach(SOURCE ${GGML_SOURCES_CUDA})
-                set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS "-x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22")
-            endforeach()
-        endif()
-
-        if (GGML_STATIC)
-            if (WIN32)
-                # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
-                list(APPEND GGML_EXTRA_LIBS_PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
-            else ()
-                if (GGML_MUSA)
-                    list(APPEND GGML_EXTRA_LIBS_PRIVATE MUSA::musart_static MUSA::mublas_static)
-                else()
-                    list(APPEND GGML_EXTRA_LIBS_PRIVATE CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
-                endif()
-            endif()
-        else()
-            if (GGML_MUSA)
-                list(APPEND GGML_EXTRA_LIBS_PRIVATE MUSA::musart MUSA::mublas)
-            else()
-                list(APPEND GGML_EXTRA_LIBS_PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
-            endif()
-        endif()
-
-        if (GGML_CUDA_NO_VMM)
-            # No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
-        else()
-            if (GGML_MUSA)
-                list(APPEND GGML_EXTRA_LIBS_PRIVATE MUSA::musa_driver) # required by muDeviceGetAttribute(), muMemGetAllocationGranularity(...), ...
-            else()
-                list(APPEND GGML_EXTRA_LIBS_PRIVATE CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
-            endif()
-        endif()
-    else()
-        message(WARNING "CUDA not found")
-    endif()
-endif()
-
-if (GGML_HIPBLAS)
-    if (NOT EXISTS $ENV{ROCM_PATH})
-        if (NOT EXISTS /opt/rocm)
-            set(ROCM_PATH /usr)
-        else()
-            set(ROCM_PATH /opt/rocm)
-        endif()
-    else()
-        set(ROCM_PATH $ENV{ROCM_PATH})
-    endif()
-
-    list(APPEND CMAKE_PREFIX_PATH  ${ROCM_PATH})
-    list(APPEND CMAKE_PREFIX_PATH "${ROCM_PATH}/lib64/cmake")
-
-    # CMake on Windows doesn't support the HIP language yet
-    if (WIN32)
-        set(CXX_IS_HIPCC TRUE)
-    else()
-        string(REGEX MATCH "hipcc(\.bat)?$" CXX_IS_HIPCC "${CMAKE_CXX_COMPILER}")
-    endif()
-
-    if (CXX_IS_HIPCC)
-        if (LINUX)
-            if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
-                message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
-            endif()
-
-            message(WARNING "Setting hipcc as the C++ compiler is legacy behavior."
-                    " Prefer setting the HIP compiler directly. See README for details.")
-        endif()
-    else()
-        # Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
-        if (AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
-            set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
-        endif()
-        cmake_minimum_required(VERSION 3.21)
-        enable_language(HIP)
-    endif()
-
-    find_package(hip     REQUIRED)
-    find_package(hipblas REQUIRED)
-    find_package(rocblas REQUIRED)
-
-    message(STATUS "HIP and hipBLAS found")
-
-    file(GLOB   GGML_HEADERS_ROCM "ggml-cuda/*.cuh")
-    list(APPEND GGML_HEADERS_ROCM "../include/ggml-cuda.h")
-
-    file(GLOB   GGML_SOURCES_ROCM "ggml-cuda/*.cu")
-    list(APPEND GGML_SOURCES_ROCM "ggml-cuda.cu")
-    file(GLOB   SRCS "ggml-cuda/template-instances/fattn-wmma*.cu")
-    list(APPEND GGML_SOURCES_ROCM ${SRCS})
-    file(GLOB   SRCS "ggml-cuda/template-instances/mmq*.cu")
-    list(APPEND GGML_SOURCES_ROCM ${SRCS})
-
-    if (GGML_CUDA_FA_ALL_QUANTS)
-        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*.cu")
-        list(APPEND GGML_SOURCES_ROCM ${SRCS})
-        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
-    else()
-        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
-        list(APPEND GGML_SOURCES_ROCM ${SRCS})
-        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
-        list(APPEND GGML_SOURCES_ROCM ${SRCS})
-        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
-        list(APPEND GGML_SOURCES_ROCM ${SRCS})
-    endif()
-
-    list(APPEND GGML_CDEF_PUBLIC GGML_USE_CUDA)
-
-    add_compile_definitions(GGML_USE_HIPBLAS)
-    add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
-    add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
-    add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
-
-    if (GGML_HIP_UMA)
-        add_compile_definitions(GGML_HIP_UMA)
-    endif()
-
-    if (GGML_CUDA_FORCE_DMMV)
-        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
-    endif()
-
-    if (GGML_CUDA_FORCE_MMQ)
-        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
-    endif()
-
-    if (GGML_CUDA_FORCE_CUBLAS)
-        add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
-    endif()
-
-    if (GGML_CUDA_NO_PEER_COPY)
-        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
-    endif()
-
-    if (CXX_IS_HIPCC)
-        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
-        list(APPEND GGML_EXTRA_LIBS_PRIVATE hip::device)
-    else()
-        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE HIP)
-    endif()
-
-    if (GGML_STATIC)
-        message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
-    endif()
-
-    list(APPEND GGML_EXTRA_LIBS_PUBLIC hip::host roc::rocblas roc::hipblas)
-endif()
-
-if (GGML_SYCL)
-    if (NOT GGML_SYCL_TARGET MATCHES "^(INTEL|NVIDIA|AMD)$")
-        message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL, NVIDIA, or AMD")
-    endif()
-
-    check_cxx_compiler_flag("-fsycl" SUPPORTS_SYCL)
-
-    if (DEFINED ENV{ONEAPI_ROOT})
-        message(STATUS "Using oneAPI Release SYCL compiler (icpx).")
-    elseif(SUPPORTS_SYCL)
-        message(WARNING "Using open-source SYCL compiler (clang++). Didn't detect ENV {ONEAPI_ROOT}.
-         If you expected the oneAPI Release compiler, please install oneAPI & source it, like:
-         source /opt/intel/oneapi/setvars.sh")
-    else()
-        message(FATAL_ERROR, "C++ compiler lacks SYCL support.")
-    endif()
-    message(STATUS "SYCL found")
-    #todo: AOT
-
-    list(APPEND GGML_CDEF_PUBLIC GGML_USE_SYCL)
-
-    if (GGML_SYCL_F16)
-        if (GGML_SYCL_TARGET STREQUAL "AMD")
-            message(WARNING "AMD target does not entirely support FP16 in the SYCL backend.")
-        endif()
-        add_compile_definitions(GGML_SYCL_F16)
-    endif()
-
-    if (GGML_CUDA_FORCE_MMQ)
-        add_compile_definitions(GGML_SYCL_FORCE_MMQ)
-    endif()
-
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
-
-    if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
-        add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
-    elseif (GGML_SYCL_TARGET STREQUAL "AMD")
-        # INFO: Allowed Sub_group_sizes are not consistent through all
-        # hip targets. For example, 64 is used for certain models, but the backend
-        # does not support it.
-        # Target archs tested working: gfx1030, gfx1031, (Only tested sub_group_size = 32)
-        add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
-    else()
-        add_compile_definitions(GGML_SYCL_WARP_SIZE=16)
-    endif()
-
-    file(GLOB   GGML_HEADERS_SYCL "ggml-sycl/*.hpp")
-    list(APPEND GGML_HEADERS_SYCL "../include/ggml-sycl.h")
-
-    file(GLOB   GGML_SOURCES_SYCL "ggml-sycl/*.cpp")
-    list(APPEND GGML_SOURCES_SYCL "ggml-sycl.cpp")
-
-    find_package(DNNL)
-    message("-- DNNL found:" ${DNNL_FOUND})
-
-    if (GGML_SYCL_TARGET STREQUAL "INTEL")
-        add_compile_definitions(GGML_SYCL_DNNL=${DNNL_FOUND})
-    else()
-        add_compile_definitions(GGML_SYCL_DNNL=0)
-    endif()
-
-    if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
-        list(APPEND GGML_EXTRA_LIBS_PRIVATE DNNL::dnnl)
-    endif()
-
-    if (WIN32)
-        find_package(IntelSYCL REQUIRED)
-        find_package(MKL REQUIRED)
-        list(APPEND GGML_EXTRA_LIBS_PRIVATE IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
-    else()
-        if (GGML_SYCL_TARGET STREQUAL "INTEL")
-            list(APPEND GGML_EXTRA_LIBS_PRIVATE sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
-        elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA")
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
-            list(APPEND GGML_EXTRA_LIBS_PRIVATE sycl pthread m dl onemkl)
-        elseif (GGML_SYCL_TARGET STREQUAL "AMD")
-            if (GGML_SYCL_HIP_TARGET STREQUAL "")
-                message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_HIP_TARGET has not been set.")
-            endif()
-            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=amdgcn-amd-amdhsa -Xsycl-target-backend --offload-arch=${GGML_SYCL_HIP_TARGET}")
-            list(APPEND GGML_EXTRA_LIBS_PRIVATE sycl pthread m dl onemkl)
-        endif()
-    endif()
-endif()
-
-if (GGML_RPC)
-    message(STATUS "RPC found")
-
-    list(APPEND GGML_CDEF_PUBLIC GGML_USE_RPC)
-
-    if (WIN32)
-        list(APPEND GGML_EXTRA_LIBS_PRIVATE ws2_32)
-    endif()
-
-    set(GGML_HEADERS_RPC ../include/ggml-rpc.h)
-    set(GGML_SOURCES_RPC ggml-rpc.cpp)
-endif()
-
-if (GGML_VULKAN)
-    find_package(Vulkan COMPONENTS glslc REQUIRED)
-
-    if (Vulkan_FOUND)
-        message(STATUS "Vulkan found")
-
-        list(APPEND GGML_CDEF_PUBLIC GGML_USE_VULKAN)
-
-        # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
-        # Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
-        if (MSVC AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
-            add_compile_definitions(_ITERATOR_DEBUG_LEVEL=0)
-        endif()
-
-        if (GGML_VULKAN_CHECK_RESULTS)
-            add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
-        endif()
-
-        if (GGML_VULKAN_DEBUG)
-            add_compile_definitions(GGML_VULKAN_DEBUG)
-        endif()
-
-        if (GGML_VULKAN_MEMORY_DEBUG)
-            add_compile_definitions(GGML_VULKAN_MEMORY_DEBUG)
-        endif()
-
-        if (GGML_VULKAN_SHADER_DEBUG_INFO)
-            add_compile_definitions(GGML_VULKAN_SHADER_DEBUG_INFO)
-        endif()
-
-        if (GGML_VULKAN_PERF)
-            add_compile_definitions(GGML_VULKAN_PERF)
-        endif()
-
-        if (GGML_VULKAN_VALIDATE)
-            add_compile_definitions(GGML_VULKAN_VALIDATE)
-        endif()
-
-        if (GGML_VULKAN_RUN_TESTS)
-            add_compile_definitions(GGML_VULKAN_RUN_TESTS)
-        endif()
-
-        add_subdirectory(vulkan-shaders)
-
-        set (_ggml_vk_genshaders_cmd vulkan-shaders-gen)
-        set (_ggml_vk_header     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp)
-        set (_ggml_vk_source     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp)
-        set (_ggml_vk_input_dir  ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders)
-        set (_ggml_vk_output_dir ${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv)
-
-        file(GLOB _ggml_vk_shader_deps "${_ggml_vk_input_dir}/*.comp")
-
-        add_custom_command(
-            OUTPUT ${_ggml_vk_header}
-                   ${_ggml_vk_source}
-
-            COMMAND ${_ggml_vk_genshaders_cmd}
-                --glslc      ${Vulkan_GLSLC_EXECUTABLE}
-                --input-dir  ${_ggml_vk_input_dir}
-                --output-dir ${_ggml_vk_output_dir}
-                --target-hpp ${_ggml_vk_header}
-                --target-cpp ${_ggml_vk_source}
-                --no-clean
-
-            DEPENDS ${_ggml_vk_shader_deps}
-            COMMENT "Generate vulkan shaders"
-        )
-
-        set(GGML_HEADERS_VULKAN ${CMAKE_CURRENT_SOURCE_DIR}/../include/ggml-vulkan.h ${_ggml_vk_header})
-        set(GGML_SOURCES_VULKAN ggml-vulkan.cpp ${_ggml_vk_source})
-
-        list(APPEND GGML_EXTRA_LIBS_PRIVATE Vulkan::Vulkan)
-        list(APPEND GGML_EXTRA_INCLUDES     ${CMAKE_CURRENT_BINARY_DIR})
-    else()
-        message(WARNING "Vulkan not found")
-    endif()
-endif()
-
-if (GGML_KOMPUTE)
-    add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
-
-    find_package(Vulkan COMPONENTS glslc REQUIRED)
-    find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
-
-    if (NOT glslc_executable)
-        message(FATAL_ERROR "glslc not found")
-    endif()
-
-    function(compile_shader)
-        set(options)
-        set(oneValueArgs)
-        set(multiValueArgs SOURCES)
-        cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-        foreach(source ${compile_shader_SOURCES})
-            get_filename_component(filename ${source} NAME)
-            set(spv_file ${filename}.spv)
-            add_custom_command(
-                OUTPUT ${spv_file}
-                DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
-                COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-                COMMENT "Compiling ${source} to ${spv_file}"
-                )
-
-            get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
-            set(FILE_NAME "shader${RAW_FILE_NAME}")
-            string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
-            string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
-            string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
-            set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
-            message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
-            if(CMAKE_GENERATOR MATCHES "Visual Studio")
-                add_custom_command(
-                    OUTPUT ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    DEPENDS ${spv_file} xxd
-                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
-                    )
-            else()
-                add_custom_command(
-                    OUTPUT ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    DEPENDS ${spv_file} xxd
-                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
-                    )
-            endif()
-        endforeach()
-    endfunction()
-
-    if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
-        message(STATUS "Kompute found")
-        set(KOMPUTE_OPT_LOG_LEVEL Error CACHE STRING "Kompute log level")
-        add_subdirectory(kompute)
-
-        # Compile our shaders
-        compile_shader(SOURCES
-            kompute-shaders/op_scale.comp
-            kompute-shaders/op_scale_8.comp
-            kompute-shaders/op_add.comp
-            kompute-shaders/op_addrow.comp
-            kompute-shaders/op_mul.comp
-            kompute-shaders/op_silu.comp
-            kompute-shaders/op_relu.comp
-            kompute-shaders/op_gelu.comp
-            kompute-shaders/op_softmax.comp
-            kompute-shaders/op_norm.comp
-            kompute-shaders/op_rmsnorm.comp
-            kompute-shaders/op_diagmask.comp
-            kompute-shaders/op_mul_mat_mat_f32.comp
-            kompute-shaders/op_mul_mat_f16.comp
-            kompute-shaders/op_mul_mat_q8_0.comp
-            kompute-shaders/op_mul_mat_q4_0.comp
-            kompute-shaders/op_mul_mat_q4_1.comp
-            kompute-shaders/op_mul_mat_q4_k.comp
-            kompute-shaders/op_mul_mat_q6_k.comp
-            kompute-shaders/op_getrows_f32.comp
-            kompute-shaders/op_getrows_f16.comp
-            kompute-shaders/op_getrows_q4_0.comp
-            kompute-shaders/op_getrows_q4_1.comp
-            kompute-shaders/op_getrows_q6_k.comp
-            kompute-shaders/op_rope_f16.comp
-            kompute-shaders/op_rope_f32.comp
-            kompute-shaders/op_cpy_f16_f16.comp
-            kompute-shaders/op_cpy_f16_f32.comp
-            kompute-shaders/op_cpy_f32_f16.comp
-            kompute-shaders/op_cpy_f32_f32.comp
-        )
-
-        # Create a custom target for our generated shaders
-        add_custom_target(generated_shaders DEPENDS
-            shaderop_scale.h
-            shaderop_scale_8.h
-            shaderop_add.h
-            shaderop_addrow.h
-            shaderop_mul.h
-            shaderop_silu.h
-            shaderop_relu.h
-            shaderop_gelu.h
-            shaderop_softmax.h
-            shaderop_norm.h
-            shaderop_rmsnorm.h
-            shaderop_diagmask.h
-            shaderop_mul_mat_mat_f32.h
-            shaderop_mul_mat_f16.h
-            shaderop_mul_mat_q8_0.h
-            shaderop_mul_mat_q4_0.h
-            shaderop_mul_mat_q4_1.h
-            shaderop_mul_mat_q4_k.h
-            shaderop_mul_mat_q6_k.h
-            shaderop_getrows_f32.h
-            shaderop_getrows_f16.h
-            shaderop_getrows_q4_0.h
-            shaderop_getrows_q4_1.h
-            shaderop_getrows_q6_k.h
-            shaderop_rope_f16.h
-            shaderop_rope_f32.h
-            shaderop_cpy_f16_f16.h
-            shaderop_cpy_f16_f32.h
-            shaderop_cpy_f32_f16.h
-            shaderop_cpy_f32_f32.h
-        )
-
-        # Create a custom command that depends on the generated_shaders
-        add_custom_command(
-            OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
-            COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
-            DEPENDS generated_shaders
-            COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
-        )
-
-        # Add the stamp to the main sources to ensure dependency tracking
-        set(GGML_SOURCES_KOMPUTE ggml-kompute.cpp           ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
-        set(GGML_HEADERS_KOMPUTE ../include/ggml-kompute.h  ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
-
-        list(APPEND GGML_CDEF_PUBLIC GGML_USE_KOMPUTE)
-
-        list(APPEND GGML_EXTRA_LIBS_PRIVATE kompute)
-        list(APPEND GGML_EXTRA_INCLUDES     ${CMAKE_CURRENT_BINARY_DIR})
-    else()
-        message(WARNING "Kompute not found")
-    endif()
-endif()
-
-if (GGML_CPU_HBM)
-    find_library(memkind memkind REQUIRED)
-
-    message(STATUS "Using memkind for CPU HBM")
-
-    add_compile_definitions(GGML_USE_CPU_HBM)
-
-    target_link_libraries(ggml PUBLIC memkind)
-endif()
-
-if (GGML_CANN)
-    if ("cann${CANN_INSTALL_DIR}" STREQUAL "cann" AND DEFINED ENV{ASCEND_TOOLKIT_HOME})
-        set(CANN_INSTALL_DIR $ENV{ASCEND_TOOLKIT_HOME})
-        message(STATUS "CANN: updated CANN_INSTALL_DIR from ASCEND_TOOLKIT_HOME=$ENV{ASCEND_TOOLKIT_HOME}")
-    endif()
-
-    if (CANN_INSTALL_DIR)
-        # Only Support Linux.
-        if (GGML_CANN)
-            if (NOT UNIX)
-                set(GGML_CANN OFF)
-                message(WARNING "CANN: CANN toolkit supports unix but not ${CMAKE_SYSTEM_NAME}. Turning off GGML_CANN")
-            endif()
-        endif()
-
-        # Supported platforms: x86-64, arm64
-        if (GGML_CANN)
-            if (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
-            elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "amd64")
-            else()
-                set(GGML_CANN OFF)
-                message(WARNING "CANN: CANN toolkit supports x86-64 and arm64 but not ${CMAKE_SYSTEM_PROCESSOR}. Turning off GGML_CANN")
-            endif()
-        endif()
-
-        # Set header and libs
-        if(GGML_CANN)
-            set(CANN_INCLUDE_DIRS
-                ${CANN_INSTALL_DIR}/include
-                ${CANN_INSTALL_DIR}/include/aclnn
-                ${CANN_INSTALL_DIR}/acllib/include
-            )
-
-            add_subdirectory(ggml-cann/kernels)
-            list(APPEND CANN_LIBRARIES
-                ascendcl
-                nnopbase
-                opapi
-                acl_op_compiler
-                ascendc_kernels
-            )
-
-            set(GGML_HEADERS_CANN "../include/ggml-cann.h")
-            file(GLOB GGML_SOURCES_CANN "ggml-cann/*.cpp")
-            list(APPEND GGML_SOURCES_CANN "ggml-cann.cpp")
-
-            message(STATUS "CANN: CANN_INCLUDE_DIRS =  ${CANN_INCLUDE_DIRS}")
-            message(STATUS "CANN: CANN_LIBRARIES =  ${CANN_LIBRARIES}")
-
-            list(APPEND GGML_EXTRA_LIBS_PRIVATE ${CANN_LIBRARIES} )
-            list(APPEND GGML_EXTRA_INCLUDES     ${CANN_INCLUDE_DIRS})
-            list(APPEND GGML_EXTRA_LIBDIRS      ${CANN_INSTALL_DIR}/lib64)
-
-            list(APPEND GGML_CDEF_PUBLIC GGML_USE_CANN)
-        endif()
-    else()
-        set(GGML_CANN OFF)
-        message(WARNING "CANN: Can't find CANN_INSTALL_DIR, do you forget to source set_var.sh. Turning off GGML_CANN")
-    endif()
-
-    if(NOT GGML_CANN)
-        message(WARNING "CANN: GGML_CANN is turned OFF, see above for details.")
-    endif()
-endif()
-
-function(get_flags CCID CCVER)
+function(ggml_get_flags CCID CCVER)
     set(C_FLAGS "")
     set(CXX_FLAGS "")
 
@@ -967,11 +42,6 @@ function(get_flags CCID CCVER)
         set(C_FLAGS   -Wdouble-promotion)
         set(CXX_FLAGS -Wno-array-bounds)
 
-        if (NOT GGML_MUSA)
-            if (CCVER VERSION_GREATER_EQUAL 7.1.0)
-                list(APPEND CXX_FLAGS -Wno-format-truncation)
-            endif()
-        endif()
         if (CCVER VERSION_GREATER_EQUAL 8.1.0)
             list(APPEND CXX_FLAGS -Wextra-semi)
         endif()
@@ -1000,7 +70,7 @@ if (GGML_ALL_WARNINGS)
         list(APPEND C_FLAGS   ${WARNING_FLAGS})
         list(APPEND CXX_FLAGS ${WARNING_FLAGS})
 
-        get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
+        ggml_get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
 
         add_compile_options("$<$<COMPILE_LANGUAGE:C>:${C_FLAGS};${GF_C_FLAGS}>"
                             "$<$<COMPILE_LANGUAGE:CXX>:${CXX_FLAGS};${GF_CXX_FLAGS}>")
@@ -1011,54 +81,6 @@ if (GGML_ALL_WARNINGS)
     endif()
 endif()
 
-set(CUDA_CXX_FLAGS "")
-
-if (GGML_CUDA)
-    set(CUDA_FLAGS -use_fast_math)
-
-    if (GGML_FATAL_WARNINGS)
-        list(APPEND CUDA_FLAGS -Werror all-warnings)
-    endif()
-
-    if (GGML_ALL_WARNINGS AND NOT MSVC)
-        set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
-        if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
-            list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
-        endif()
-
-        execute_process(
-            COMMAND ${NVCC_CMD} -Xcompiler --version
-            OUTPUT_VARIABLE CUDA_CCFULLVER
-            ERROR_QUIET
-        )
-
-        if (NOT CUDA_CCFULLVER MATCHES clang)
-            set(CUDA_CCID "GNU")
-            execute_process(
-                COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
-                OUTPUT_VARIABLE CUDA_CCVER
-                ERROR_QUIET
-            )
-        else()
-            if (CUDA_CCFULLVER MATCHES Apple)
-                set(CUDA_CCID "AppleClang")
-            else()
-                set(CUDA_CCID "Clang")
-            endif()
-            string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
-        endif()
-
-        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
-
-        get_flags(${CUDA_CCID} ${CUDA_CCVER})
-        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
-    endif()
-
-    if (NOT MSVC)
-        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
-    endif()
-endif()
-
 if (GGML_LTO)
     include(CheckIPOSupported)
     check_ipo_supported(RESULT result OUTPUT output)
@@ -1116,197 +138,6 @@ if (NOT MSVC)
     endif()
 endif()
 
-set(ARCH_FLAGS "")
-
-if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
-    CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
-    (NOT CMAKE_OSX_ARCHITECTURES      AND
-     NOT CMAKE_GENERATOR_PLATFORM_LWR AND
-         CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
-
-    message(STATUS "ARM detected")
-
-    if (MSVC)
-        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
-        add_compile_definitions(__ARM_NEON)
-        add_compile_definitions(__ARM_FEATURE_FMA)
-
-        set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
-        string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
-
-        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
-        if (GGML_COMPILER_SUPPORT_DOTPROD)
-            add_compile_definitions(__ARM_FEATURE_DOTPROD)
-        endif ()
-
-        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
-
-        if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
-            add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
-        endif ()
-
-        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
-        if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
-            add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
-        endif ()
-
-        set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
-    else()
-        check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
-        if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
-            list(APPEND ARCH_FLAGS -mfp16-format=ieee)
-        endif()
-        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
-            # Raspberry Pi 1, Zero
-            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
-        endif()
-        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
-            if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
-                # Android armeabi-v7a
-                list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
-            else()
-                # Raspberry Pi 2
-                list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
-            endif()
-        endif()
-        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
-            # Android arm64-v8a
-            # Raspberry Pi 3, 4, Zero 2 (32-bit)
-            list(APPEND ARCH_FLAGS -mno-unaligned-access)
-        endif()
-        if (GGML_SVE)
-            list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
-        endif()
-    endif()
-elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
-        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
-         CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$"))
-    message(STATUS "x86 detected")
-    if (MSVC)
-        # instruction set detection for MSVC only
-        if (GGML_NATIVE)
-            # TODO: improve, should not reference files from the parent folder
-            include(../cmake/FindSIMD.cmake)
-        endif ()
-        if (GGML_AVX512)
-            list(APPEND ARCH_FLAGS /arch:AVX512)
-            # MSVC has no compile-time flags enabling specific
-            # AVX512 extensions, neither it defines the
-            # macros corresponding to the extensions.
-            # Do it manually.
-            if (GGML_AVX512_VBMI)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
-            endif()
-            if (GGML_AVX512_VNNI)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
-            endif()
-            if (GGML_AVX512_BF16)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
-            endif()
-            if (GGML_AMX_TILE)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
-            endif()
-            if (GGML_AMX_INT8)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
-            endif()
-            if (GGML_AMX_BF16)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
-                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_BF16__>)
-            endif()
-        elseif (GGML_AVX2)
-            list(APPEND ARCH_FLAGS /arch:AVX2)
-        elseif (GGML_AVX)
-            list(APPEND ARCH_FLAGS /arch:AVX)
-        endif()
-    else()
-        if (GGML_NATIVE)
-            list(APPEND ARCH_FLAGS -march=native)
-        endif()
-        if (GGML_F16C)
-            list(APPEND ARCH_FLAGS -mf16c)
-        endif()
-        if (GGML_FMA)
-            list(APPEND ARCH_FLAGS -mfma)
-        endif()
-        if (GGML_AVX)
-            list(APPEND ARCH_FLAGS -mavx)
-        endif()
-        if (GGML_AVX2)
-            list(APPEND ARCH_FLAGS -mavx2)
-        endif()
-        if (GGML_AVX512)
-            list(APPEND ARCH_FLAGS -mavx512f)
-            list(APPEND ARCH_FLAGS -mavx512dq)
-            list(APPEND ARCH_FLAGS -mavx512bw)
-        endif()
-        if (GGML_AVX512_VBMI)
-            list(APPEND ARCH_FLAGS -mavx512vbmi)
-        endif()
-        if (GGML_AVX512_VNNI)
-            list(APPEND ARCH_FLAGS -mavx512vnni)
-        endif()
-        if (GGML_AVX512_BF16)
-            list(APPEND ARCH_FLAGS -mavx512bf16)
-        endif()
-        if (GGML_AMX_TILE)
-            list(APPEND ARCH_FLAGS -mamx-tile)
-        endif()
-        if (GGML_AMX_INT8)
-            list(APPEND ARCH_FLAGS -mamx-int8)
-        endif()
-        if (GGML_AMX_BF16)
-            list(APPEND ARCH_FLAGS -mamx-bf16)
-        endif()
-    endif()
-elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
-    message(STATUS "PowerPC detected")
-    execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1" OUTPUT_VARIABLE POWER10_M)
-    string(FIND "${POWER10_M}" "POWER10" substring_index)
-    if (NOT DEFINED substring_index OR "${substring_index}" STREQUAL "")
-        set(substring_index -1)
-    endif()
-
-    if (${substring_index} GREATER_EQUAL 0)
-       list(APPEND ARCH_FLAGS -mcpu=power10)
-    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
-       list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
-    else()
-        list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
-        #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
-    endif()
-elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
-    message(STATUS "loongarch64 detected")
-
-    list(APPEND ARCH_FLAGS -march=loongarch64)
-    if (GGML_LASX)
-        list(APPEND ARCH_FLAGS -mlasx)
-    endif()
-    if (GGML_LSX)
-        list(APPEND ARCH_FLAGS -mlsx)
-    endif()
-else()
-    message(STATUS "Unknown architecture")
-endif()
-
-add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
-add_compile_options("$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
-
-if (GGML_CUDA)
-    list(APPEND CUDA_CXX_FLAGS ${ARCH_FLAGS})
-    list(JOIN   CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
-
-    if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
-        list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
-    endif()
-
-    add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
-endif()
-
 if (MINGW)
     # Target Windows 8 for PrefetchVirtualMemory
     add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
@@ -1320,14 +151,14 @@ endif()
 # CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
 # posix_memalign came in POSIX.1-2001 / SUSv3
 # M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
-add_compile_definitions(_XOPEN_SOURCE=600)
 
 # Somehow in OpenBSD whenever POSIX conformance is specified
 # some string functions rely on locale_t availability,
 # which was introduced in POSIX.1-2008, forcing us to go higher
 if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
-    remove_definitions(-D_XOPEN_SOURCE=600)
     add_compile_definitions(_XOPEN_SOURCE=700)
+else()
+    add_compile_definitions(_XOPEN_SOURCE=600)
 endif()
 
 # Data types, macros and functions related to controlling CPU affinity and
@@ -1363,74 +194,144 @@ endif()
 
 if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
-
-    if (BUILD_SHARED_LIBS)
-        # TODO: should not use this
-        set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
-    endif()
 endif()
 
-#
-# libraries
-#
-
 # ggml
 
-add_library(ggml
+if (GGML_BACKEND_DL AND NOT BUILD_SHARED_LIBS)
+    message(FATAL_ERROR "GGML_BACKEND_DL requires BUILD_SHARED_LIBS")
+endif()
+
+add_library(ggml-base
             ../include/ggml.h
-            ../include/ggml-cpu.h
             ../include/ggml-alloc.h
             ../include/ggml-backend.h
             ../include/ggml-cpp.h
+            ../include/ggml-opt.h
             ggml.c
-            ggml-cpu.c
             ggml-alloc.c
             ggml-backend.cpp
+            ggml-opt.cpp
+            ggml-threading.cpp
+            ggml-threading.h
             ggml-quants.c
-            ggml-quants.h
-            ${GGML_SOURCES_CUDA}      ${GGML_HEADERS_CUDA}
-            ${GGML_SOURCES_METAL}     ${GGML_HEADERS_METAL}
-            ${GGML_SOURCES_RPC}       ${GGML_HEADERS_RPC}
-            ${GGML_SOURCES_EXTRA}     ${GGML_HEADERS_EXTRA}
-            ${GGML_SOURCES_SYCL}      ${GGML_HEADERS_SYCL}
-            ${GGML_SOURCES_KOMPUTE}   ${GGML_HEADERS_KOMPUTE}
-            ${GGML_SOURCES_VULKAN}    ${GGML_HEADERS_VULKAN}
-            ${GGML_SOURCES_ROCM}      ${GGML_HEADERS_ROCM}
-            ${GGML_SOURCES_BLAS}      ${GGML_HEADERS_BLAS}
-            ${GGML_SOURCES_LLAMAFILE} ${GGML_HEADERS_LLAMAFILE}
-            ${GGML_SOURCES_AMX}       ${GGML_HEADERS_AMX}
-            ${GGML_SOURCES_CANN}      ${GGML_HEADERS_CANN}
-            ggml-aarch64.c            ggml-aarch64.h
-            )
+            ggml-quants.h)
 
-if (EMSCRIPTEN)
-    set_target_properties(ggml PROPERTIES COMPILE_FLAGS "-msimd128")
+target_include_directories(ggml-base PRIVATE .)
+
+add_library(ggml
+            ggml-backend-reg.cpp)
+
+target_link_libraries(ggml PUBLIC ggml-base)
+
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    target_link_libraries(ggml PRIVATE dl)
 endif()
 
-target_compile_definitions(ggml PUBLIC    ${GGML_CDEF_PUBLIC})
-target_include_directories(ggml PUBLIC    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)
-target_include_directories(ggml PRIVATE . ${GGML_EXTRA_INCLUDES})
-target_link_directories   (ggml PRIVATE   ${GGML_EXTRA_LIBDIRS})
-target_compile_features   (ggml PRIVATE c_std_11) # don't bump
+function(ggml_add_backend_library backend)
+    if (GGML_BACKEND_DL)
+        add_library(${backend} MODULE ${ARGN})
+        # write the shared library to the output directory
+        set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
+        target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
+    else()
+        add_library(${backend} ${ARGN})
+        target_link_libraries(ggml PUBLIC ${backend})
+        install(TARGETS ${backend} LIBRARY)
+    endif()
+
+    target_link_libraries(${backend} PRIVATE ggml-base)
+    target_include_directories(${backend} PRIVATE ..)
+
+    if (${BUILD_SHARED_LIBS})
+        target_compile_definitions(${backend} PRIVATE GGML_BACKEND_BUILD)
+        target_compile_definitions(${backend} PUBLIC  GGML_BACKEND_SHARED)
+    endif()
+endfunction()
+
+function(ggml_add_backend backend)
+    string(TOUPPER "GGML_${backend}" backend_id)
+    if (${backend_id})
+        string(TOLOWER "ggml-${backend}" backend_target)
+        add_subdirectory(${backend_target})
+        message(STATUS "Including ${backend} backend")
+        if (NOT GGML_BACKEND_DL)
+            string(TOUPPER "GGML_USE_${backend}" backend_use)
+            target_compile_definitions(ggml PUBLIC ${backend_use})
+        endif()
+    endif()
+endfunction()
+
+function(ggml_add_cpu_backend_variant tag_name)
+    set(GGML_CPU_TAG_NAME ${tag_name})
+    # other: OPENMP LLAMAFILE CPU_HBM
+    foreach (feat NATIVE
+                  AVX AVX2 AVX_VNNI FMA F16C
+                  AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16
+                  AMX_TILE AMX_INT8 AMX_BF16)
+        set(GGML_${feat} OFF)
+    endforeach()
+
+    foreach (feat ${ARGN})
+        set(GGML_${feat} ON)
+    endforeach()
+
+    ggml_add_cpu_backend_variant_impl(${tag_name})
+endfunction()
 
-list(APPEND GGML_EXTRA_LIBS_PRIVATE Threads::Threads)
+ggml_add_backend(CPU)
+
+if (GGML_CPU_ALL_VARIANTS)
+    if (NOT GGML_BACKEND_DL)
+        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
+    endif()
+    ggml_add_cpu_backend_variant(sandybridge    AVX)
+    ggml_add_cpu_backend_variant(haswell        AVX F16C AVX2 FMA)
+    ggml_add_cpu_backend_variant(skylakex       AVX F16C AVX2 FMA AVX512)
+    ggml_add_cpu_backend_variant(icelake        AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
+    if (NOT MSVC)
+        # MSVC doesn't support AVX-VNNI or AMX
+        ggml_add_cpu_backend_variant(alderlake      AVX F16C AVX2 FMA AVX_VNNI)
+        ggml_add_cpu_backend_variant(sapphirerapids AVX F16C AVX2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
+    endif()
+else ()
+    ggml_add_cpu_backend_variant_impl("")
+endif()
+
+ggml_add_backend(BLAS)
+ggml_add_backend(CANN)
+ggml_add_backend(CUDA)
+ggml_add_backend(HIP)
+ggml_add_backend(Kompute)
+ggml_add_backend(METAL)
+ggml_add_backend(MUSA)
+ggml_add_backend(RPC)
+ggml_add_backend(SYCL)
+ggml_add_backend(Vulkan)
+ggml_add_backend(OpenCL)
+
+foreach (target ggml-base ggml)
+    target_include_directories(${target} PUBLIC    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)
+    target_compile_features   (${target} PRIVATE c_std_11 cxx_std_17) # don't bump
+endforeach()
+
+target_link_libraries(ggml-base PRIVATE Threads::Threads)
 
 find_library(MATH_LIBRARY m)
 if (MATH_LIBRARY)
     if (NOT WIN32 OR NOT DEFINED ENV{ONEAPI_ROOT})
-        list(APPEND GGML_EXTRA_LIBS_PRIVATE m)
+        target_link_libraries(ggml-base PRIVATE m)
     endif()
 endif()
 
 if (CMAKE_SYSTEM_NAME MATCHES "Android")
-    list(APPEND GGML_EXTRA_LIBS_PRIVATE dl) # Must be linked explicitly
+    target_link_libraries(ggml-base PRIVATE dl)
 endif()
 
-list(REMOVE_DUPLICATES GGML_EXTRA_LIBS_PRIVATE)
-list(REMOVE_DUPLICATES GGML_EXTRA_LIBS_PUBLIC)
-target_link_libraries(ggml PRIVATE ${GGML_EXTRA_LIBS_PRIVATE} PUBLIC ${GGML_EXTRA_LIBS_PUBLIC})
-
 if (BUILD_SHARED_LIBS)
-    set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    target_compile_definitions(ggml PRIVATE GGML_SHARED GGML_BUILD)
+    foreach (target ggml-base ggml)
+        set_target_properties(${target} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        target_compile_definitions(${target} PRIVATE GGML_BUILD)
+        target_compile_definitions(${target} PUBLIC  GGML_SHARED)
+    endforeach()
 endif()
diff --git a/ggml/src/ggml-aarch64.h b/ggml/src/ggml-aarch64.h
deleted file mode 100644
index 517babaf169..00000000000
--- a/ggml/src/ggml-aarch64.h
+++ /dev/null
@@ -1,39 +0,0 @@
-// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
-#pragma once
-
-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
-#include "ggml.h"
-
-// GGML internal header
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// Quantization
-void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
-
-// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-// GEMV
-void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-
-// GEMM
-void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
-
-#ifdef __cplusplus
-}
-#endif
-
diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c
index 041de9e3efc..8dc8226ac49 100644
--- a/ggml/src/ggml-alloc.c
+++ b/ggml/src/ggml-alloc.c
@@ -466,18 +466,12 @@ static bool ggml_gallocr_is_own(ggml_gallocr_t galloc, struct ggml_tensor * t) {
     return ggml_gallocr_hash_get(galloc, t)->allocated;
 }
 
-static void ggml_gallocr_set_node_offset(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, size_t offset) {
-    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
-    hn->buffer_id = buffer_id;
-    hn->offset = offset;
-    hn->allocated = true;
-}
-
 static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor * t) {
     return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
 }
 
 static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
+    GGML_ASSERT(buffer_id >= 0);
     struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
 
     if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_is_view(node)) {
@@ -540,7 +534,6 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
         size_t offset = ggml_dyn_tallocr_alloc(alloc, size, node);
         hn->buffer_id = buffer_id;
         hn->offset = offset;
-        return;
     }
 }
 
@@ -816,7 +809,11 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
 }
 
 static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
-    size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
+    size_t node_size = 0;
+    if (!node->data && !node->view_src) {
+        GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API
+        node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
+    }
     return talloc->size_max >= node_size;
 }
 
diff --git a/ggml/src/ggml-amx/CMakeLists.txt b/ggml/src/ggml-amx/CMakeLists.txt
new file mode 100644
index 00000000000..d6676f3f67b
--- /dev/null
+++ b/ggml/src/ggml-amx/CMakeLists.txt
@@ -0,0 +1,107 @@
+if (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
+        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+         CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$") AND
+        CMAKE_COMPILER_IS_GNUCC AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 11.0)
+    message(STATUS "Using AMX")
+
+    file(GLOB   GGML_HEADERS_AMX "*.h")
+    list(APPEND GGML_HEADERS_AMX "../../include/ggml-amx.h")
+
+    file(GLOB   GGML_SOURCES_AMX "*.cpp")
+
+    add_library(ggml-amx
+                ${GGML_HEADERS_AMX}
+                ${GGML_SOURCES_AMX})
+
+    target_link_libraries(ggml-amx PRIVATE ggml-base)
+    target_include_directories(ggml-amx PRIVATE . ..)
+
+    # this is duplicated from the CPU backend, since the AMX backend also depends on the architecture flags
+    # TODO: integrate AMX backend into the CPU backend
+    if (MSVC)
+        # instruction set detection for MSVC only
+        if (GGML_NATIVE)
+            # TODO: improve, should not reference files from the parent folder
+            include(../ggml-cpu/cmake/FindSIMD.cmake)
+        endif ()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS /arch:AVX512)
+            # MSVC has no compile-time flags enabling specific
+            # AVX512 extensions, neither it defines the
+            # macros corresponding to the extensions.
+            # Do it manually.
+            if (GGML_AVX512_VBMI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
+            endif()
+            if (GGML_AVX512_VNNI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
+            endif()
+            if (GGML_AVX512_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
+            endif()
+            if (GGML_AMX_TILE)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_TILE__>)
+            endif()
+            if (GGML_AMX_INT8)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_INT8__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_INT8__>)
+            endif()
+            if (GGML_AMX_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AMX_BF16__>)
+            endif()
+        elseif (GGML_AVX2)
+            list(APPEND ARCH_FLAGS /arch:AVX2)
+        elseif (GGML_AVX)
+            list(APPEND ARCH_FLAGS /arch:AVX)
+        endif()
+    else()
+        if (GGML_NATIVE)
+            list(APPEND ARCH_FLAGS -march=native)
+        endif()
+        if (GGML_F16C)
+            list(APPEND ARCH_FLAGS -mf16c)
+        endif()
+        if (GGML_FMA)
+            list(APPEND ARCH_FLAGS -mfma)
+        endif()
+        if (GGML_AVX)
+            list(APPEND ARCH_FLAGS -mavx)
+        endif()
+        if (GGML_AVX2)
+            list(APPEND ARCH_FLAGS -mavx2)
+        endif()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS -mavx512f)
+            list(APPEND ARCH_FLAGS -mavx512dq)
+            list(APPEND ARCH_FLAGS -mavx512bw)
+        endif()
+        if (GGML_AVX512_VBMI)
+            list(APPEND ARCH_FLAGS -mavx512vbmi)
+        endif()
+        if (GGML_AVX512_VNNI)
+            list(APPEND ARCH_FLAGS -mavx512vnni)
+        endif()
+        if (GGML_AVX512_BF16)
+            list(APPEND ARCH_FLAGS -mavx512bf16)
+        endif()
+        if (GGML_AMX_TILE)
+            list(APPEND ARCH_FLAGS -mamx-tile)
+        endif()
+        if (GGML_AMX_INT8)
+            list(APPEND ARCH_FLAGS -mamx-int8)
+        endif()
+        if (GGML_AMX_BF16)
+            list(APPEND ARCH_FLAGS -mamx-bf16)
+        endif()
+    endif()
+
+    target_compile_options(ggml-amx PRIVATE ${ARCH_FLAGS})
+else()
+    set(GGML_AMX OFF PARENT_SCOPE)
+    message(WARNING "AMX requires x86 and gcc version > 11.0. Turning off GGML_AMX.")
+endif()
diff --git a/ggml/src/ggml-amx/common.h b/ggml/src/ggml-amx/common.h
index 2b6c6352704..5db8ce30d9b 100644
--- a/ggml/src/ggml-amx/common.h
+++ b/ggml/src/ggml-amx/common.h
@@ -1,7 +1,8 @@
 #pragma once
 
 #include "ggml.h"
-#include "ggml-cpu-impl.h" // <immintrin.h>
+// hack until AMX is moved into the CPU backend
+#include "../ggml-cpu/ggml-cpu-impl.h" // <immintrin.h>
 
 #include <algorithm>
 #include <memory>
diff --git a/ggml/src/ggml-amx.cpp b/ggml/src/ggml-amx/ggml-amx.cpp
similarity index 98%
rename from ggml/src/ggml-amx.cpp
rename to ggml/src/ggml-amx/ggml-amx.cpp
index 144dc9d8a50..8568e7965fd 100644
--- a/ggml/src/ggml-amx.cpp
+++ b/ggml/src/ggml-amx/ggml-amx.cpp
@@ -317,8 +317,6 @@ static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const st
             const enum ggml_type type = src0->type;
             const int64_t ne0 = op->ne[0];
 
-            bool is_training = src0->grad || src1->grad;
-
             // amx kernels enables for Q4_0, Q4_1, Q8_0, F16
             // Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256
             bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16);
@@ -326,7 +324,6 @@ static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const st
             bool can_use_amx =
                 is_contiguous_2d(src0) &&       // src0 must be contiguous
                 is_contiguous_2d(src1) &&       // src1 must be contiguous
-                !is_training &&                 // inference only
                 src1->type == GGML_TYPE_F32 &&  // src1 must be float32
                 has_amx_kernels &&              // with amx kernel impls
                 ne0 % (TILE_N * 2) == 0;        // out_features is 32x
@@ -421,9 +418,18 @@ ggml_backend_reg_t ggml_backend_amx_reg(void) {
 
 #else // if defined(__AMX_INT8__)
 
+ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void) {
+    return nullptr;
+}
+
+bool ggml_backend_is_amx(ggml_backend_t backend) {
+    GGML_UNUSED(backend);
+    return false;
+}
+
 ggml_backend_t ggml_backend_amx_init(void) {
     fprintf(stderr, "GGML is not compiled with AMX support!\n");
-    return ggml_backend_t{};
+    return nullptr;
 }
 
 void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
@@ -433,4 +439,8 @@ void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads) {
     GGML_UNUSED(n_threads);
 }
 
+ggml_backend_reg_t ggml_backend_amx_reg(void) {
+    return nullptr;
+}
+
 #endif
diff --git a/ggml/src/ggml-amx/mmq.cpp b/ggml/src/ggml-amx/mmq.cpp
index 239d15121a6..529bee25b5e 100644
--- a/ggml/src/ggml-amx/mmq.cpp
+++ b/ggml/src/ggml-amx/mmq.cpp
@@ -496,19 +496,20 @@ inline void from_float(const float * x, char * vy, int64_t k);
 
 template <>
 inline void from_float<block_q8_0>(const float * x, char * vy, int64_t k) {
-    quantize_row_q8_0(x, vy, k);
+    // FIXME: using unoptimized reference impl until moved to CPU backend
+    quantize_row_q8_0_ref(x, (block_q8_0 *)vy, k);
 }
 
 template <>
 inline void from_float<block_q8_1>(const float * x, char * vy, int64_t k) {
-    quantize_row_q8_1(x, vy, k);
+    quantize_row_q8_1_ref(x, (block_q8_1 *)vy, k);
 }
 
 template <>
 inline void from_float<block_q8_K>(const float * x, char * vy, int64_t k) {
 #if 1
     // TODO: this is reference impl!
-    quantize_row_q8_K(x, vy, k);
+    quantize_row_q8_K_ref(x, (block_q8_K *)vy, k);
 #else
     quantize_row_q8_K_vnni(x, vy, k);
 #endif
diff --git a/ggml/src/ggml-backend-impl.h b/ggml/src/ggml-backend-impl.h
index fa8d5b7fb68..36d72e95f02 100644
--- a/ggml/src/ggml-backend-impl.h
+++ b/ggml/src/ggml-backend-impl.h
@@ -8,6 +8,8 @@
 extern "C" {
 #endif
 
+    #define GGML_BACKEND_API_VERSION 1
+
     //
     // Backend buffer type
     //
@@ -63,20 +65,20 @@ extern "C" {
         enum ggml_backend_buffer_usage usage;
     };
 
-    ggml_backend_buffer_t ggml_backend_buffer_init(
+    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
                    ggml_backend_buffer_type_t buft,
             struct ggml_backend_buffer_i      iface,
                    void *                     context,
                    size_t                     size);
 
     // do not use directly, use ggml_backend_tensor_copy instead
-    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
+    GGML_API bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
 
     // multi-buffer
     // buffer that contains a collection of buffers
-    ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
-    bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
-    void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+    GGML_API ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
+    GGML_API bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
+    GGML_API void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
 
     //
     // Backend (stream)
@@ -199,17 +201,55 @@ extern "C" {
     };
 
     struct ggml_backend_reg {
-        // int api_version; // TODO: for dynamic loading
+        int api_version; // initialize to GGML_BACKEND_API_VERSION
         struct ggml_backend_reg_i iface;
         void * context;
     };
 
-
     // Internal backend registry API
-    void ggml_backend_register(ggml_backend_reg_t reg);
-    void ggml_backend_device_register(ggml_backend_dev_t device);
-    // TODO: backends can be loaded as a dynamic library, in which case it needs to export this function
-    // typedef ggml_backend_register_t * (*ggml_backend_init)(void);
+    GGML_API void ggml_backend_register(ggml_backend_reg_t reg);
+    GGML_API void ggml_backend_device_register(ggml_backend_dev_t device);
+
+    // Add backend dynamic loading support to the backend
+
+    // Initialize the backend
+    typedef ggml_backend_reg_t (*ggml_backend_init_t)(void);
+    // Optional: obtain a score for the backend based on the system configuration
+    // Higher scores are preferred, 0 means the backend is not supported in the current system
+    typedef int                (*ggml_backend_score_t)(void);
+
+#ifdef GGML_BACKEND_DL
+#    ifdef __cplusplus
+#        define GGML_BACKEND_DL_IMPL(reg_fn)                             \
+            extern "C" {                                                 \
+            GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
+            }                                                            \
+            ggml_backend_reg_t ggml_backend_init(void) {                 \
+                return reg_fn();                                         \
+            }
+#        define GGML_BACKEND_DL_SCORE_IMPL(score_fn)       \
+            extern "C" {                                   \
+            GGML_BACKEND_API int ggml_backend_score(void); \
+            }                                              \
+            int ggml_backend_score(void) {                 \
+                return score_fn();                         \
+            }
+#    else
+#        define GGML_BACKEND_DL_IMPL(reg_fn)                              \
+            GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void);  \
+            ggml_backend_reg_t                  ggml_backend_init(void) { \
+                return reg_fn();                                          \
+            }
+#        define GGML_BACKEND_DL_SCORE_IMPL(score_fn)        \
+            GGML_BACKEND_API int ggml_backend_score(void);  \
+            int                  ggml_backend_score(void) { \
+                return score_fn();                          \
+            }
+#    endif
+#else
+#    define GGML_BACKEND_DL_IMPL(reg_fn)
+#    define GGML_BACKEND_DL_SCORE_IMPL(score_fn)
+#endif
 
 #ifdef  __cplusplus
 }
diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
new file mode 100644
index 00000000000..66927148abc
--- /dev/null
+++ b/ggml/src/ggml-backend-reg.cpp
@@ -0,0 +1,552 @@
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+#include <algorithm>
+#include <codecvt>
+#include <cstring>
+#include <filesystem>
+#include <locale>
+#include <memory>
+#include <string>
+#include <type_traits>
+#include <vector>
+
+#ifdef _WIN32
+#    define WIN32_LEAN_AND_MEAN
+#    ifndef NOMINMAX
+#        define NOMINMAX
+#    endif
+#    include <windows.h>
+#elif defined(__APPLE__)
+#    include <mach-o/dyld.h>
+#    include <dlfcn.h>
+#else
+#    include <dlfcn.h>
+#    include <unistd.h>
+#endif
+
+// Backend registry
+#ifdef GGML_USE_CPU
+#include "ggml-cpu.h"
+#endif
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#ifdef GGML_USE_SYCL
+#include "ggml-sycl.h"
+#endif
+
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
+#ifdef GGML_USE_OPENCL
+#include "ggml-opencl.h"
+#endif
+
+#ifdef GGML_USE_BLAS
+#include "ggml-blas.h"
+#endif
+
+#ifdef GGML_USE_RPC
+#include "ggml-rpc.h"
+#endif
+
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
+
+#ifdef GGML_USE_KOMPUTE
+#include "ggml-kompute.h"
+#endif
+
+#ifdef _WIN32
+
+using dl_handle = std::remove_pointer_t<HMODULE>;
+
+struct dl_handle_deleter {
+    void operator()(HMODULE handle) {
+        FreeLibrary(handle);
+    }
+};
+
+static dl_handle * dl_load_library(const std::wstring & path) {
+    // suppress error dialogs for missing DLLs
+    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+    HMODULE handle = LoadLibraryW(path.c_str());
+
+    SetErrorMode(old_mode);
+
+    return handle;
+}
+
+static dl_handle * dl_load_library(const std::string & path) {
+    std::wstring_convert<std::codecvt_utf8_utf16<wchar_t>> converter;
+    return dl_load_library(converter.from_bytes(path));
+}
+
+static void * dl_get_sym(dl_handle * handle, const char * name) {
+    DWORD old_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
+    SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
+
+    void * p = (void *) GetProcAddress(handle, name);
+
+    SetErrorMode(old_mode);
+
+    return p;
+}
+
+#else
+
+using dl_handle = void;
+
+struct dl_handle_deleter {
+    void operator()(void * handle) {
+        dlclose(handle);
+    }
+};
+
+static void * dl_load_library(const std::string & path) {
+    dl_handle * handle = dlopen(path.c_str(), RTLD_NOW | RTLD_LOCAL);
+
+    return handle;
+}
+
+static void * dl_get_sym(dl_handle * handle, const char * name) {
+    return dlsym(handle, name);
+}
+
+#endif
+
+using dl_handle_ptr = std::unique_ptr<dl_handle, dl_handle_deleter>;
+
+struct ggml_backend_reg_entry {
+    ggml_backend_reg_t reg;
+    dl_handle_ptr handle;
+};
+
+struct ggml_backend_registry {
+    std::vector<ggml_backend_reg_entry> backends;
+    std::vector<ggml_backend_dev_t> devices;
+
+    ggml_backend_registry() {
+#ifdef GGML_USE_CUDA
+        register_backend(ggml_backend_cuda_reg());
+#endif
+#ifdef GGML_USE_METAL
+        register_backend(ggml_backend_metal_reg());
+#endif
+#ifdef GGML_USE_SYCL
+        register_backend(ggml_backend_sycl_reg());
+#endif
+#ifdef GGML_USE_VULKAN
+        register_backend(ggml_backend_vk_reg());
+#endif
+#ifdef GGML_USE_OPENCL
+        register_backend(ggml_backend_opencl_reg());
+#endif
+#ifdef GGML_USE_CANN
+        register_backend(ggml_backend_cann_reg());
+#endif
+#ifdef GGML_USE_BLAS
+        register_backend(ggml_backend_blas_reg());
+#endif
+#ifdef GGML_USE_RPC
+        register_backend(ggml_backend_rpc_reg());
+#endif
+#ifdef GGML_USE_KOMPUTE
+        register_backend(ggml_backend_kompute_reg());
+#endif
+#ifdef GGML_USE_CPU
+        register_backend(ggml_backend_cpu_reg());
+#endif
+    }
+
+    ~ggml_backend_registry() {
+        // FIXME: backends cannot be safely unloaded without a function to destroy all the backend resources,
+        // since backend threads may still be running and accessing resources from the dynamic library
+        for (auto & entry : backends) {
+            if (entry.handle) {
+                entry.handle.release(); // NOLINT
+            }
+        }
+    }
+
+    void register_backend(ggml_backend_reg_t reg, dl_handle_ptr handle = nullptr) {
+        if (!reg) {
+            return;
+        }
+
+#ifndef NDEBUG
+        GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
+            __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
+#endif
+        backends.push_back({ reg, std::move(handle) });
+        for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
+            register_device(ggml_backend_reg_dev_get(reg, i));
+        }
+    }
+
+    void register_device(ggml_backend_dev_t device) {
+#ifndef NDEBUG
+        GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
+#endif
+        devices.push_back(device);
+    }
+
+    ggml_backend_reg_t load_backend(const char * path, bool silent) {
+        dl_handle_ptr handle { dl_load_library(path) };
+        if (!handle) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: failed to load %s\n", __func__, path);
+            }
+            return nullptr;
+        }
+
+        auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
+        if (score_fn && score_fn() == 0) {
+            if (!silent) {
+                GGML_LOG_INFO("%s: backend %s is not supported on this system\n", __func__, path);
+            }
+            return nullptr;
+        }
+
+        auto backend_init_fn = (ggml_backend_init_t) dl_get_sym(handle.get(), "ggml_backend_init");
+        if (!backend_init_fn) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: failed to find ggml_backend_init in %s\n", __func__, path);
+            }
+            return nullptr;
+        }
+
+        ggml_backend_reg_t reg = backend_init_fn();
+        if (!reg || reg->api_version != GGML_BACKEND_API_VERSION) {
+            if (!silent) {
+                if (!reg) {
+                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: ggml_backend_init returned NULL\n", __func__, path);
+                } else {
+                    GGML_LOG_ERROR("%s: failed to initialize backend from %s: incompatible API version (backend: %d, current: %d)\n",
+                        __func__, path, reg->api_version, GGML_BACKEND_API_VERSION);
+                }
+            }
+            return nullptr;
+        }
+
+        GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path);
+
+        register_backend(reg, std::move(handle));
+
+        return reg;
+    }
+
+    void unload_backend(ggml_backend_reg_t reg, bool silent) {
+        auto it = std::find_if(backends.begin(), backends.end(),
+                               [reg](const ggml_backend_reg_entry & entry) { return entry.reg == reg; });
+
+        if (it == backends.end()) {
+            if (!silent) {
+                GGML_LOG_ERROR("%s: backend not found\n", __func__);
+            }
+            return;
+        }
+
+        if (!silent) {
+            GGML_LOG_DEBUG("%s: unloading %s backend\n", __func__, ggml_backend_reg_name(reg));
+        }
+
+        // remove devices
+        devices.erase(
+            std::remove_if(devices.begin(), devices.end(),
+                            [reg](ggml_backend_dev_t dev) { return ggml_backend_dev_backend_reg(dev) == reg; }),
+            devices.end());
+
+        // remove backend
+        backends.erase(it);
+    }
+};
+
+static ggml_backend_registry & get_reg() {
+    static ggml_backend_registry reg;
+    return reg;
+}
+
+// Internal API
+void ggml_backend_register(ggml_backend_reg_t reg) {
+    get_reg().register_backend(reg);
+}
+
+void ggml_backend_device_register(ggml_backend_dev_t device) {
+    get_reg().register_device(device);
+}
+
+// Backend (reg) enumeration
+static bool striequals(const char * a, const char * b) {
+    for (; *a && *b; a++, b++) {
+        if (std::tolower(*a) != std::tolower(*b)) {
+            return false;
+        }
+    }
+    return *a == *b;
+}
+
+size_t ggml_backend_reg_count() {
+    return get_reg().backends.size();
+}
+
+ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
+    GGML_ASSERT(index < ggml_backend_reg_count());
+    return get_reg().backends[index].reg;
+}
+
+ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
+    for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
+        ggml_backend_reg_t reg = ggml_backend_reg_get(i);
+        if (striequals(ggml_backend_reg_name(reg), name)) {
+            return reg;
+        }
+    }
+    return nullptr;
+}
+
+// Device enumeration
+size_t ggml_backend_dev_count() {
+    return get_reg().devices.size();
+}
+
+ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
+    GGML_ASSERT(index < ggml_backend_dev_count());
+    return get_reg().devices[index];
+}
+
+ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        if (striequals(ggml_backend_dev_name(dev), name)) {
+            return dev;
+        }
+    }
+    return nullptr;
+}
+
+ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        if (ggml_backend_dev_type(dev) == type) {
+            return dev;
+        }
+    }
+    return nullptr;
+}
+
+// Convenience functions
+ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
+    if (!dev) {
+        return nullptr;
+    }
+    return ggml_backend_dev_init(dev, params);
+}
+
+ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
+    if (!dev) {
+        return nullptr;
+    }
+    return ggml_backend_dev_init(dev, params);
+}
+
+ggml_backend_t ggml_backend_init_best(void) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
+    if (!dev) {
+        dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+    }
+    if (!dev) {
+        return nullptr;
+    }
+    return ggml_backend_dev_init(dev, nullptr);
+}
+
+// Dynamic loading
+ggml_backend_reg_t ggml_backend_load(const char * path) {
+    return get_reg().load_backend(path, false);
+}
+
+void ggml_backend_unload(ggml_backend_reg_t reg) {
+    get_reg().unload_backend(reg, true);
+}
+
+static std::string get_executable_path() {
+#if defined(__APPLE__)
+    // get executable path
+    std::vector<char> path;
+    uint32_t size;
+    while (true) {
+        size = path.size();
+        if (_NSGetExecutablePath(path.data(), &size) == 0) {
+            break;
+        }
+        path.resize(size);
+    }
+    std::string base_path(path.data(), size);
+    // remove executable name
+    auto last_slash = base_path.find_last_of('/');
+    if (last_slash != std::string::npos) {
+        base_path = base_path.substr(0, last_slash);
+    }
+    return base_path + "/";
+#elif defined(__linux__)
+    std::string base_path = ".";
+    std::vector<char> path(1024);
+    while (true) {
+        // get executable path
+        ssize_t len = readlink("/proc/self/exe", path.data(), path.size());
+        if (len == -1) {
+            break;
+        }
+        if (len < (ssize_t) path.size()) {
+            base_path = std::string(path.data(), len);
+            // remove executable name
+            auto last_slash = base_path.find_last_of('/');
+            if (last_slash != std::string::npos) {
+                base_path = base_path.substr(0, last_slash);
+            }
+            break;
+        }
+        path.resize(path.size() * 2);
+    }
+
+    return base_path + "/";
+#elif defined(_WIN32)
+    std::vector<char> path(MAX_PATH);
+    DWORD len = GetModuleFileNameA(NULL, path.data(), path.size());
+    if (len == 0) {
+        return "";
+    }
+    std::string base_path(path.data(), len);
+    // remove executable name
+    auto last_slash = base_path.find_last_of('\\');
+    if (last_slash != std::string::npos) {
+        base_path = base_path.substr(0, last_slash);
+    }
+    return base_path + "\\";
+#endif
+}
+
+static std::string backend_filename_prefix() {
+#ifdef _WIN32
+    return "ggml-";
+#else
+    return "libggml-";
+#endif
+}
+
+static std::string backend_filename_suffix() {
+#ifdef _WIN32
+    return ".dll";
+#else
+    return ".so";
+#endif
+}
+
+static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent, const char * user_search_path) {
+    // enumerate all the files that match [lib]ggml-name-*.[so|dll] in the search paths
+     // TODO: search system paths
+    std::string file_prefix = backend_filename_prefix() + name + "-";
+    std::vector<std::string> search_paths;
+    if (user_search_path == nullptr) {
+        search_paths.push_back("./");
+        search_paths.push_back(get_executable_path());
+    } else {
+#if defined(_WIN32)
+        search_paths.push_back(std::string(user_search_path) + "\\");
+#else
+        search_paths.push_back(std::string(user_search_path) + "/");
+#endif
+    }
+
+    int best_score = 0;
+    std::string best_path;
+
+    namespace fs = std::filesystem;
+    for (const auto & search_path : search_paths) {
+        if (!fs::exists(search_path)) {
+            continue;
+        }
+        fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
+        for (const auto & entry : dir_it) {
+            if (entry.is_regular_file()) {
+                std::string filename = entry.path().filename().string();
+                std::string ext = entry.path().extension().string();
+                if (filename.find(file_prefix) == 0 && ext == backend_filename_suffix()) {
+                    dl_handle_ptr handle { dl_load_library(entry.path().c_str()) };
+                    if (!handle && !silent) {
+                        GGML_LOG_ERROR("%s: failed to load %s\n", __func__, entry.path().string().c_str());
+                    }
+                    if (handle) {
+                        auto score_fn = (ggml_backend_score_t) dl_get_sym(handle.get(), "ggml_backend_score");
+                        if (score_fn) {
+                            int s = score_fn();
+#ifndef NDEBUG
+                            GGML_LOG_DEBUG("%s: %s score: %d\n", __func__, entry.path().string().c_str(), s);
+#endif
+                            if (s > best_score) {
+                                best_score = s;
+                                best_path = entry.path().string();
+                            }
+                        } else {
+                            if (!silent) {
+                                GGML_LOG_INFO("%s: failed to find ggml_backend_score in %s\n", __func__, entry.path().string().c_str());
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    if (best_score == 0) {
+        // try to load the base backend
+        for (const auto & search_path : search_paths) {
+            std::string path = search_path + backend_filename_prefix() + name + backend_filename_suffix();
+            if (fs::exists(path)) {
+                return get_reg().load_backend(path.c_str(), silent);
+            }
+        }
+        return nullptr;
+    }
+
+    return get_reg().load_backend(best_path.c_str(), silent);
+}
+
+void ggml_backend_load_all() {
+    ggml_backend_load_all_from_path(nullptr);
+}
+
+void ggml_backend_load_all_from_path(const char * dir_path) {
+#ifdef NDEBUG
+    bool silent = true;
+#else
+    bool silent = false;
+#endif
+
+    ggml_backend_load_best("blas", silent, dir_path);
+    ggml_backend_load_best("cann", silent, dir_path);
+    ggml_backend_load_best("cuda", silent, dir_path);
+    ggml_backend_load_best("hip", silent, dir_path);
+    ggml_backend_load_best("kompute", silent, dir_path);
+    ggml_backend_load_best("metal", silent, dir_path);
+    ggml_backend_load_best("rpc", silent, dir_path);
+    ggml_backend_load_best("sycl", silent, dir_path);
+    ggml_backend_load_best("vulkan", silent, dir_path);
+    ggml_backend_load_best("opencl", silent, dir_path);
+    ggml_backend_load_best("musa", silent, dir_path);
+    ggml_backend_load_best("cpu", silent, dir_path);
+}
diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index af92d63544d..fdb4b986f61 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -252,6 +252,7 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten
 }
 
 void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor);
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     if (size == 0) {
@@ -266,6 +267,7 @@ void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, siz
 }
 
 void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor);
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     if (size == 0) {
@@ -279,7 +281,7 @@ void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, siz
     buf->iface.get_tensor(buf, tensor, data, offset, size);
 }
 
-GGML_API void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
     if (size == 0) {
@@ -525,197 +527,6 @@ void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * na
     return reg->iface.get_proc_address(reg, name);
 }
 
-// Backend registry
-
-#ifdef GGML_USE_CUDA
-#include "ggml-cuda.h"
-#endif
-
-#ifdef GGML_USE_METAL
-#include "ggml-metal.h"
-#endif
-
-#ifdef GGML_USE_SYCL
-#include "ggml-sycl.h"
-#endif
-
-#ifdef GGML_USE_VULKAN
-#include "ggml-vulkan.h"
-#endif
-
-#ifdef GGML_USE_BLAS
-#include "ggml-blas.h"
-#endif
-
-#ifdef GGML_USE_RPC
-#include "ggml-rpc.h"
-#endif
-
-#ifndef __AMX_INT8__
-#undef GGML_USE_AMX
-#endif
-
-#ifdef GGML_USE_AMX
-#  include "ggml-amx.h"
-#endif
-
-#ifdef GGML_USE_CANN
-#include "ggml-cann.h"
-#endif
-
-#ifdef GGML_USE_KOMPUTE
-#include "ggml-kompute.h"
-#endif
-
-#include "ggml-cpu.h"
-
-struct ggml_backend_registry {
-    std::vector<ggml_backend_reg_t> backends;
-    std::vector<ggml_backend_dev_t> devices;
-
-    ggml_backend_registry() {
-#ifdef GGML_USE_CUDA
-        register_backend(ggml_backend_cuda_reg());
-#endif
-#ifdef GGML_USE_METAL
-        register_backend(ggml_backend_metal_reg());
-#endif
-#ifdef GGML_USE_SYCL
-        register_backend(ggml_backend_sycl_reg());
-#endif
-#ifdef GGML_USE_VULKAN
-        register_backend(ggml_backend_vk_reg());
-#endif
-#ifdef GGML_USE_CANN
-        register_backend(ggml_backend_cann_reg());
-#endif
-#ifdef GGML_USE_BLAS
-        register_backend(ggml_backend_blas_reg());
-#endif
-#ifdef GGML_USE_RPC
-        register_backend(ggml_backend_rpc_reg());
-#endif
-#ifdef GGML_USE_AMX
-        register_backend(ggml_backend_amx_reg());
-#endif
-#ifdef GGML_USE_KOMPUTE
-        register_backend(ggml_backend_kompute_reg());
-#endif
-
-        register_backend(ggml_backend_cpu_reg());
-    }
-
-    void register_backend(ggml_backend_reg_t reg) {
-#ifndef NDEBUG
-        GGML_LOG_DEBUG("%s: registered backend %s (%zu devices)\n",
-            __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
-#endif
-        backends.push_back(reg);
-        for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
-            register_device(ggml_backend_reg_dev_get(reg, i));
-        }
-    }
-
-    void register_device(ggml_backend_dev_t device) {
-#ifndef NDEBUG
-        GGML_LOG_DEBUG("%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
-#endif
-        devices.push_back(device);
-    }
-};
-
-static ggml_backend_registry & get_reg() {
-    static ggml_backend_registry reg;
-    return reg;
-}
-
-// Internal API
-void ggml_backend_register(ggml_backend_reg_t reg) {
-    get_reg().register_backend(reg);
-}
-
-void ggml_backend_device_register(ggml_backend_dev_t device) {
-    get_reg().register_device(device);
-}
-
-// Backend (reg) enumeration
-size_t ggml_backend_reg_count() {
-    return get_reg().backends.size();
-}
-
-ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
-    GGML_ASSERT(index < ggml_backend_reg_count());
-    return get_reg().backends[index];
-}
-
-ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
-    for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
-        ggml_backend_reg_t reg = ggml_backend_reg_get(i);
-        if (strcmp(ggml_backend_reg_name(reg), name) == 0) {
-            return reg;
-        }
-    }
-    return NULL;
-}
-
-// Device enumeration
-size_t ggml_backend_dev_count() {
-    return get_reg().devices.size();
-}
-
-ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
-    GGML_ASSERT(index < ggml_backend_dev_count());
-    return get_reg().devices[index];
-}
-
-ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
-    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
-        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
-        if (strcmp(ggml_backend_dev_name(dev), name) == 0) {
-            return dev;
-        }
-    }
-    return NULL;
-}
-
-ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
-    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
-        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
-        if (ggml_backend_dev_type(dev) == type) {
-            return dev;
-        }
-    }
-    return NULL;
-}
-
-// Convenience functions
-ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
-    ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
-    if (!dev) {
-        return NULL;
-    }
-    return ggml_backend_dev_init(dev, params);
-}
-
-ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
-    ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
-    if (!dev) {
-        return NULL;
-    }
-    return ggml_backend_dev_init(dev, params);
-}
-
-ggml_backend_t ggml_backend_init_best(void) {
-    ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
-    if (!dev) {
-        dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-    }
-    if (!dev) {
-        return NULL;
-    }
-    return ggml_backend_dev_init(dev, NULL);
-}
-
 // multi-buffer buffer
 
 struct ggml_backend_multi_buffer_context {
@@ -880,7 +691,7 @@ static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backen
 }
 
 static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor, const struct ggml_tensor * op) {
-    ggml_backend_buffer_t buffer = tensor->buffer;
+    ggml_backend_buffer_t buffer = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
     if (buffer == NULL) {
         return -1;
     }
@@ -913,8 +724,6 @@ static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML
 
 // returns the backend that should be used for the node based on the current locations
 static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * tensor) {
-    // TODO: use supports_op to check if the backend supports the op
-
     // assign pre-allocated nodes to their backend
     int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor, tensor);
     if (cur_backend_id != -1) {
@@ -933,7 +742,8 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
 
     if (tensor->buffer || (tensor->view_src && tensor->view_src->buffer)) {
         // since the tensor is pre-allocated, it cannot be moved to another backend
-        GGML_ABORT("pre-allocated tensor in a backend that cannot run the operation");
+        ggml_backend_buffer_t buffer = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+        GGML_ABORT("pre-allocated tensor (%s) in a buffer (%s) that cannot run the operation (%s)", tensor->name, ggml_backend_buffer_name(buffer), ggml_op_name(tensor->op));
     }
 
     // graph input
@@ -1640,7 +1450,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
         bool parallel) {
     GGML_ASSERT(n_backends > 0);
     GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
-    GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
+    GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
 
     struct ggml_backend_sched * sched = (ggml_backend_sched *) calloc(1, sizeof(struct ggml_backend_sched));
 
@@ -1729,12 +1539,13 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
 
     ggml_backend_sched_split_graph(sched, measure_graph);
 
+    ggml_backend_sched_synchronize(sched);
+
     if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
         return false;
     }
 
     ggml_backend_sched_reset(sched);
-    ggml_backend_sched_synchronize(sched);
 
     return true;
 }
@@ -2036,17 +1847,6 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
     return true;
 }
 
-
-
-#include "ggml-backend.h"
-#include "ggml-backend-impl.h"
-#include "ggml-cpu.h"
-#include "ggml-impl.h"
-#include <cctype>
-#include <string>
-
-// ggml-backend interface
-
 // CPU backend - buffer
 
 static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
@@ -2120,7 +1920,9 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
     /* .reset           = */ NULL,
 };
 
-// CPU backend - buffer type
+// CPU backend buffer type
+
+// this buffer type is defined here to make it available to all backends
 
 static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
     return "CPU";
@@ -2161,7 +1963,7 @@ ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
         },
-        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .device  = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
         /* .context = */ NULL,
     };
 
@@ -2184,479 +1986,14 @@ static ggml_backend_buffer_type_t ggml_backend_cpu_buffer_from_ptr_type(void) {
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
         },
-        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .device  = */ NULL, // FIXME ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
         /* .context = */ NULL,
     };
 
     return &ggml_backend_cpu_buffer_type;
 }
 
-#ifdef GGML_USE_CPU_HBM
-
-// buffer type HBM
-
-#include <hbwmalloc.h>
-
-static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
-    return "CPU_HBM";
-
-    GGML_UNUSED(buft);
-}
-
-static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    hbw_free(buffer->context);
-}
-
-static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    void * ptr;
-    int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
-    if (result != 0) {
-        GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
-        return NULL;
-    }
-
-    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
-    buffer->buft = buft;
-    buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
-
-    return buffer;
-}
-
-ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
-    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
-        /* .iface    = */ {
-            /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
-            /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
-            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
-            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
-        },
-        /* .context  = */ NULL,
-    };
-
-    return &ggml_backend_cpu_buffer_type_hbm;
-}
-#endif
-
-static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backend_dev_t device) {
-    static ggml_backend_buffer_type_t bufts[] = {
-#ifdef GGML_USE_CPU_HBM
-        ggml_backend_cpu_hbm_buffer_type(),
-#endif
-        NULL
-    };
-
-    return bufts;
-
-    GGML_UNUSED(device);
-}
-
-// CPU backend - backend (stream)
-
-struct ggml_backend_cpu_context {
-    int                 n_threads;
-    ggml_threadpool_t   threadpool;
-
-    uint8_t *           work_data;
-    size_t              work_size;
-
-    ggml_abort_callback abort_callback;
-    void *              abort_callback_data;
-};
-
-static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
-    return "CPU";
-
-    GGML_UNUSED(backend);
-}
-
-static void ggml_backend_cpu_free(ggml_backend_t backend) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-    delete[] cpu_ctx->work_data;
-    delete cpu_ctx;
-    delete backend;
-}
-
-struct ggml_backend_plan_cpu {
-    struct ggml_cplan cplan;
-    struct ggml_cgraph cgraph;
-};
-
-static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-
-    struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;
-
-    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
-    cpu_plan->cgraph = *cgraph; // FIXME: deep copy
-
-    if (cpu_plan->cplan.work_size > 0) {
-        cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
-        if (cpu_plan->cplan.work_data == NULL) {
-            delete cpu_plan;
-            return NULL;
-        }
-    }
-
-    cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
-    cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
-
-    return cpu_plan;
-}
-
-static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
-
-    delete[] cpu_plan->cplan.work_data;
-    delete cpu_plan;
-
-    GGML_UNUSED(backend);
-}
-
-static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
-
-    return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
-
-    GGML_UNUSED(backend);
-}
-
-static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-
-    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
-
-    if (cpu_ctx->work_size < cplan.work_size) {
-        delete[] cpu_ctx->work_data;
-        cpu_ctx->work_data = new uint8_t[cplan.work_size];
-        if (cpu_ctx->work_data == NULL) {
-            cpu_ctx->work_size = 0;
-            return GGML_STATUS_ALLOC_FAILED;
-        }
-        cpu_ctx->work_size = cplan.work_size;
-    }
-    cplan.work_data = (uint8_t *)cpu_ctx->work_data;
-
-    cplan.abort_callback      = cpu_ctx->abort_callback;
-    cplan.abort_callback_data = cpu_ctx->abort_callback_data;
-
-    return ggml_graph_compute(cgraph, &cplan);
-}
-
-static const struct ggml_backend_i ggml_backend_cpu_i = {
-    /* .get_name                = */ ggml_backend_cpu_get_name,
-    /* .free                    = */ ggml_backend_cpu_free,
-    /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_async        = */ NULL,
-    /* .cpy_tensor_async        = */ NULL,
-    /* .synchronize             = */ NULL,
-    /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,
-    /* .graph_plan_free         = */ ggml_backend_cpu_graph_plan_free,
-    /* .graph_plan_update       = */ NULL,
-    /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
-    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
-    /* .event_record            = */ NULL,
-    /* .event_wait              = */ NULL,
-};
-
-static ggml_guid_t ggml_backend_cpu_guid(void) {
-    static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
-    return &guid;
-}
-
-ggml_backend_t ggml_backend_cpu_init(void) {
-    // initialize CPU backend now to avoid slowing the first graph computation
-    ggml_cpu_init();
-
-    struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
-    if (ctx == NULL) {
-        return NULL;
-    }
-
-    ctx->n_threads           = GGML_DEFAULT_N_THREADS;
-    ctx->threadpool          = NULL;
-    ctx->work_data           = NULL;
-    ctx->work_size           = 0;
-    ctx->abort_callback      = NULL;
-    ctx->abort_callback_data = NULL;
-
-    ggml_backend_t cpu_backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_cpu_guid(),
-        /* .interface = */ ggml_backend_cpu_i,
-        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
-        /* .context   = */ ctx,
-    };
-
-    if (cpu_backend == NULL) {
-        delete ctx;
-        return NULL;
-    }
-
-    return cpu_backend;
-}
-
-bool ggml_backend_is_cpu(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
-}
-
-void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
-    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
-
-    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
-    ctx->n_threads = n_threads;
-}
-
-void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
-    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
-
-    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
-
-    if (ctx->threadpool && ctx->threadpool != threadpool) {
-        // already had a different threadpool, pause/suspend it before switching
-        ggml_threadpool_pause(ctx->threadpool);
-    }
-    ctx->threadpool = threadpool;
-}
-
-void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
-    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
-
-    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
-    ctx->abort_callback = abort_callback;
-    ctx->abort_callback_data = abort_callback_data;
-}
-
 ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
     GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned");
     return ggml_backend_buffer_init(ggml_backend_cpu_buffer_from_ptr_type(), ggml_backend_cpu_buffer_from_ptr_i, ptr, size);
 }
-
-// CPU backend - device
-
-struct ggml_backend_cpu_device_context {
-    std::string description = "CPU";
-
-    ggml_backend_cpu_device_context() {
-#ifdef __APPLE__
-        size_t len = 0;
-        if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) {
-            description.resize(len);
-            sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT
-        }
-#elif defined(__linux__)
-        FILE * f = fopen("/proc/cpuinfo", "r");
-        if (f) {
-            char buf[1024];
-            while (fgets(buf, sizeof(buf), f)) {
-                if (strncmp(buf, "model name", 10) == 0) {
-                    char * p = strchr(buf, ':');
-                    if (p) {
-                        p++;
-                        while (std::isspace(*p)) {
-                            p++;
-                        }
-                        while (std::isspace(p[strlen(p) - 1])) {
-                            p[strlen(p) - 1] = '\0';
-                        }
-                        description = p;
-                        break;
-                    }
-                }
-            }
-            fclose(f);
-        }
-#elif defined(_WIN32)
-        HKEY hKey;
-        if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
-                        TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
-                        0,
-                        KEY_READ,
-                        &hKey) == ERROR_SUCCESS) {
-            DWORD cpu_brand_size = 0;
-            if (RegQueryValueExA(hKey,
-                                TEXT("ProcessorNameString"),
-                                NULL,
-                                NULL,
-                                NULL,
-                                &cpu_brand_size) == ERROR_SUCCESS) {
-                description.resize(cpu_brand_size);
-                if (RegQueryValueExA(hKey,
-                                    TEXT("ProcessorNameString"),
-                                    NULL,
-                                    NULL,
-                                    (LPBYTE)&description[0], // NOLINT
-                                    &cpu_brand_size) == ERROR_SUCCESS) {
-                    if (description.find('\0') != std::string::npos) {
-                        description.resize(description.find('\0'));
-                    }
-                }
-            }
-            RegCloseKey(hKey);
-        }
-#endif
-    }
-};
-
-static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
-    return "CPU";
-
-    GGML_UNUSED(dev);
-}
-
-static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
-    struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context;
-
-    return ctx->description.c_str();
-}
-
-static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    // TODO
-    *free = 0;
-    *total = 0;
-
-    GGML_UNUSED(dev);
-}
-
-static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) {
-    return GGML_BACKEND_DEVICE_TYPE_CPU;
-
-    GGML_UNUSED(dev);
-}
-
-static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
-    props->name        = ggml_backend_cpu_device_get_name(dev);
-    props->description = ggml_backend_cpu_device_get_description(dev);
-    props->type        = ggml_backend_cpu_device_get_type(dev);
-    ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
-    props->caps = {
-        /* .async                 = */ false,
-        /* .host_buffer           = */ false,
-        /* .buffer_from_host_ptr  = */ true,
-        /* .events                = */ false,
-    };
-}
-
-static ggml_backend_t ggml_backend_cpu_device_init_backend(ggml_backend_dev_t dev, const char * params) {
-    return ggml_backend_cpu_init();
-
-    GGML_UNUSED(dev);
-    GGML_UNUSED(params);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_backend_dev_t dev) {
-    return ggml_backend_cpu_buffer_type();
-
-    GGML_UNUSED(dev);
-}
-
-static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
-    return ggml_backend_cpu_buffer_from_ptr(ptr, size);
-
-    GGML_UNUSED(dev);
-    GGML_UNUSED(max_tensor_size);
-}
-
-static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
-    switch (op->op) {
-        case GGML_OP_CPY:
-            return
-                op->type != GGML_TYPE_IQ2_XXS &&
-                op->type != GGML_TYPE_IQ2_XS  &&
-                op->type != GGML_TYPE_IQ1_S   &&
-                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
-        case GGML_OP_MUL_MAT:
-            //return op->src[1]->type == GGML_TYPE_F32; // TMP: workaround until sync with latest ggml
-            return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_get_type_traits_cpu(op->src[0]->type)->vec_dot_type;
-        case GGML_OP_ROPE_BACK:
-            return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
-        case GGML_OP_IM2COL_BACK:
-            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
-        case GGML_OP_OUT_PROD:
-            return (op->src[0]->type == GGML_TYPE_F32 || ggml_is_quantized(op->src[0]->type)) && op->src[1]->type == GGML_TYPE_F32;
-        default:
-            return true;
-    }
-
-    GGML_UNUSED(dev);
-}
-
-static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    return ggml_backend_buft_is_host(buft);
-
-    GGML_UNUSED(dev);
-}
-
-static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
-    /* .get_name             = */ ggml_backend_cpu_device_get_name,
-    /* .get_description      = */ ggml_backend_cpu_device_get_description,
-    /* .get_memory           = */ ggml_backend_cpu_device_get_memory,
-    /* .get_type             = */ ggml_backend_cpu_device_get_type,
-    /* .get_props            = */ ggml_backend_cpu_device_get_props,
-    /* .init_backend         = */ ggml_backend_cpu_device_init_backend,
-    /* .get_buffer_type      = */ ggml_backend_cpu_device_get_buffer_type,
-    /* .get_host_buffer_type = */ NULL,
-    /* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,
-    /* .supports_op          = */ ggml_backend_cpu_device_supports_op,
-    /* .supports_buft        = */ ggml_backend_cpu_device_supports_buft,
-    /* .offload_op           = */ NULL,
-    /* .event_new            = */ NULL,
-    /* .event_free           = */ NULL,
-    /* .event_synchronize    = */ NULL,
-};
-
-// CPU backend - backend (reg)
-
-static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
-    return "CPU";
-
-    GGML_UNUSED(reg);
-}
-
-static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
-    return 1;
-
-    GGML_UNUSED(reg);
-}
-
-static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
-    GGML_ASSERT(index == 0);
-
-    static ggml_backend_cpu_device_context ctx;
-    static ggml_backend_device ggml_backend_cpu_device = {
-        /* .iface   = */ ggml_backend_cpu_device_i,
-        /* .reg     = */ reg,
-        /* .context = */ &ctx,
-    };
-
-    return &ggml_backend_cpu_device;
-}
-
-static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) {
-    if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
-        return (void *)ggml_backend_cpu_set_n_threads;
-    }
-    if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
-        return (void *)ggml_backend_cpu_get_extra_bufts;
-    }
-
-    return NULL;
-
-    GGML_UNUSED(reg);
-}
-
-static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
-    /* .get_name         = */ ggml_backend_cpu_reg_get_name,
-    /* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
-    /* .get_device       = */ ggml_backend_cpu_reg_get_device,
-    /* .get_proc_address = */ ggml_backend_cpu_get_proc_address,
-};
-
-ggml_backend_reg_t ggml_backend_cpu_reg(void) {
-    static struct ggml_backend_reg ggml_backend_cpu_reg = {
-        /* .iface   = */ ggml_backend_cpu_reg_i,
-        /* .context = */ NULL,
-    };
-
-    return &ggml_backend_cpu_reg;
-}
diff --git a/ggml/src/ggml-blas/CMakeLists.txt b/ggml/src/ggml-blas/CMakeLists.txt
new file mode 100644
index 00000000000..0bf3c05d93a
--- /dev/null
+++ b/ggml/src/ggml-blas/CMakeLists.txt
@@ -0,0 +1,87 @@
+if (GGML_STATIC)
+    set(BLA_STATIC ON)
+endif()
+#if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
+#    set(BLA_SIZEOF_INTEGER 8)
+#endif()
+
+set(BLA_VENDOR ${GGML_BLAS_VENDOR})
+find_package(BLAS)
+
+if (BLAS_FOUND)
+    message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
+
+    ggml_add_backend_library(ggml-blas
+                             ggml-blas.cpp
+                            )
+
+    if (${GGML_BLAS_VENDOR} MATCHES "Apple")
+        add_compile_definitions(ACCELERATE_NEW_LAPACK)
+        add_compile_definitions(ACCELERATE_LAPACK_ILP64)
+        add_compile_definitions(GGML_BLAS_USE_ACCELERATE)
+    elseif ("${BLAS_INCLUDE_DIRS}" STREQUAL "")
+        # BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
+        # see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
+        find_package(PkgConfig REQUIRED)
+        if (${GGML_BLAS_VENDOR} MATCHES "Generic")
+            pkg_check_modules(DepBLAS blas)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "OpenBLAS")
+            # As of openblas v0.3.22, the 64-bit is named openblas64.pc
+            pkg_check_modules(DepBLAS openblas64)
+            if (NOT DepBLAS_FOUND)
+                pkg_check_modules(DepBLAS openblas)
+            endif()
+        elseif (${GGML_BLAS_VENDOR} MATCHES "FLAME")
+            add_compile_definitions(GGML_BLAS_USE_BLIS)
+            pkg_check_modules(DepBLAS blis)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "ATLAS")
+            pkg_check_modules(DepBLAS blas-atlas)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "FlexiBLAS")
+            pkg_check_modules(DepBLAS flexiblas_api)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "Intel")
+            add_compile_definitions(GGML_BLAS_USE_MKL)
+            # all Intel* libraries share the same include path
+            pkg_check_modules(DepBLAS mkl-sdl)
+        elseif (${GGML_BLAS_VENDOR} MATCHES "NVHPC")
+            # this doesn't provide pkg-config
+            # suggest to assign BLAS_INCLUDE_DIRS on your own
+            if ("${NVHPC_VERSION}" STREQUAL "")
+                message(WARNING "Better to set NVHPC_VERSION")
+            else()
+                set(DepBLAS_FOUND ON)
+                set(DepBLAS_INCLUDE_DIRS "/opt/nvidia/hpc_sdk/${CMAKE_SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}/${NVHPC_VERSION}/math_libs/include")
+            endif()
+        endif()
+        if (DepBLAS_FOUND)
+            set(BLAS_INCLUDE_DIRS ${DepBLAS_INCLUDE_DIRS})
+        else()
+            message(WARNING "BLAS_INCLUDE_DIRS neither been provided nor been automatically"
+            " detected by pkgconfig, trying to find cblas.h from possible paths...")
+            find_path(BLAS_INCLUDE_DIRS
+                NAMES cblas.h
+                HINTS
+                    /usr/include
+                    /usr/local/include
+                    /usr/include/openblas
+                    /opt/homebrew/opt/openblas/include
+                    /usr/local/opt/openblas/include
+                    /usr/include/x86_64-linux-gnu/openblas/include
+            )
+        endif()
+    endif()
+
+    message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
+
+    target_compile_options(ggml-blas PRIVATE ${BLAS_LINKER_FLAGS})
+
+    if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
+        add_compile_definitions(GGML_BLAS_USE_MKL)
+    endif()
+
+    target_link_libraries     (ggml-blas PRIVATE ${BLAS_LIBRARIES})
+    target_include_directories(ggml-blas PRIVATE ${BLAS_INCLUDE_DIRS})
+else()
+    message(ERROR "BLAS not found, please refer to "
+                  "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
+                  " to set correct GGML_BLAS_VENDOR")
+endif()
diff --git a/ggml/src/ggml-blas.cpp b/ggml/src/ggml-blas/ggml-blas.cpp
similarity index 98%
rename from ggml/src/ggml-blas.cpp
rename to ggml/src/ggml-blas/ggml-blas.cpp
index 8d96220b9f4..ec158dfac6e 100644
--- a/ggml/src/ggml-blas.cpp
+++ b/ggml/src/ggml-blas/ggml-blas.cpp
@@ -6,7 +6,7 @@
 #include <vector>
 #include <cstring>
 
-#if defined(GGML_USE_ACCELERATE)
+#if defined(GGML_BLAS_USE_ACCELERATE)
 #   include <Accelerate/Accelerate.h>
 #elif defined(GGML_BLAS_USE_MKL)
 #   include <mkl.h>
@@ -320,7 +320,7 @@ static const char * ggml_backend_blas_device_get_name(ggml_backend_dev_t dev) {
 }
 
 static const char * ggml_backend_blas_device_get_description(ggml_backend_dev_t dev) {
-    #if defined(GGML_USE_ACCELERATE)
+    #if defined(GGML_BLAS_USE_ACCELERATE)
         return "Accelerate";
     #elif defined(GGML_BLAS_USE_MKL)
         return "MKL";
@@ -506,9 +506,12 @@ static const struct ggml_backend_reg_i ggml_backend_blas_reg_i = {
 
 ggml_backend_reg_t ggml_backend_blas_reg(void) {
     static struct ggml_backend_reg ggml_backend_blas_reg = {
-        /* .iface   = */ ggml_backend_blas_reg_i,
-        /* .context = */ NULL,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_blas_reg_i,
+        /* .context     = */ NULL,
     };
 
     return &ggml_backend_blas_reg;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_blas_reg)
diff --git a/ggml/src/ggml-cann/CMakeLists.txt b/ggml/src/ggml-cann/CMakeLists.txt
new file mode 100644
index 00000000000..05cf06bfab4
--- /dev/null
+++ b/ggml/src/ggml-cann/CMakeLists.txt
@@ -0,0 +1,76 @@
+if ("cann${CANN_INSTALL_DIR}" STREQUAL "cann" AND DEFINED ENV{ASCEND_TOOLKIT_HOME})
+    set(CANN_INSTALL_DIR $ENV{ASCEND_TOOLKIT_HOME})
+    message(STATUS "CANN: updated CANN_INSTALL_DIR from ASCEND_TOOLKIT_HOME=$ENV{ASCEND_TOOLKIT_HOME}")
+endif()
+
+# Auto-detech Soc type and Soc version, if detect failed, will abort build
+set(SOC_VERSION "")
+function(detect_ascend_soc_type SOC_VERSION)
+    execute_process(
+        COMMAND bash -c "npu-smi info|awk -F' ' 'NF > 0 && NR==7 {print $3}'"
+        OUTPUT_VARIABLE npu_info
+        RESULT_VARIABLE npu_result
+        OUTPUT_STRIP_TRAILING_WHITESPACE
+    )
+    if("${npu_info}" STREQUAL "" OR ${npu_result})
+        message(FATAL_ERROR "Auto-detech ascend soc type failed, please specify manually or check ascend device working normally.")
+    endif()
+    set(${SOC_VERSION} "Ascend${npu_info}" PARENT_SCOPE)
+endfunction()
+
+if(NOT SOC_TYPE)
+    detect_ascend_soc_type(SOC_VERSION)
+    set(SOC_TYPE "${SOC_VERSION}")
+    message(STATUS "CANN: SOC_VERSION auto-detected is:${SOC_VERSION}")
+endif()
+
+string(TOLOWER ${SOC_TYPE} SOC_VERSION) # SOC_VERSION need lower
+
+# Construct Soc specify compile option: ASCEND_#Soc_Major_SN. Such as ASCEND_910B, ASCEND_310P.
+string(REGEX MATCH "[0-9]+[a-zA-Z]" SOC_TYPE_MAJOR_SN "${SOC_VERSION}")
+set(SOC_TYPE_COMPILE_OPTION "ASCEND_${SOC_TYPE_MAJOR_SN}")
+string(TOUPPER ${SOC_TYPE_COMPILE_OPTION} SOC_TYPE_COMPILE_OPTION)
+
+if (CANN_INSTALL_DIR)
+    # Only Support Linux.
+    if (NOT UNIX)
+        message(FATAL_ERROR "CANN: CANN toolkit supports unix but not ${CMAKE_SYSTEM_NAME}")
+    endif()
+
+    # Supported platforms: x86-64, arm64
+    if (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
+    elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "amd64")
+    else()
+        message(FATAL_ERROR "CANN: CANN toolkit supports x86-64 and arm64 but not ${CMAKE_SYSTEM_PROCESSOR}")
+    endif()
+
+    # Set header and libs
+    set(CANN_INCLUDE_DIRS
+        ${CANN_INSTALL_DIR}/include
+        ${CANN_INSTALL_DIR}/include/aclnn
+        ${CANN_INSTALL_DIR}/acllib/include
+    )
+
+    add_subdirectory(kernels)
+    list(APPEND CANN_LIBRARIES
+        ascendcl
+        nnopbase
+        opapi
+        acl_op_compiler
+        ascendc_kernels
+    )
+
+    file(GLOB GGML_SOURCES_CANN "*.cpp")
+
+    ggml_add_backend_library(ggml-cann ${GGML_SOURCES_CANN})
+    target_link_libraries(ggml-cann PRIVATE ${CANN_LIBRARIES})
+    target_include_directories(ggml-cann PRIVATE ${CANN_INCLUDE_DIRS})
+    target_link_directories(ggml-cann PRIVATE ${CANN_INSTALL_DIR}/lib64)
+
+    target_compile_definitions(ggml-cann PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
+
+    message(STATUS "CANN: CANN_INCLUDE_DIRS =  ${CANN_INCLUDE_DIRS}")
+    message(STATUS "CANN: CANN_LIBRARIES =  ${CANN_LIBRARIES}")
+else()
+    message(FATAL_ERROR "CANN: Can't find CANN_INSTALL_DIR, did you forget to source set_var.sh?")
+endif()
diff --git a/ggml/src/ggml-cann/aclnn_ops.cpp b/ggml/src/ggml-cann/aclnn_ops.cpp
index a4ec8418e2a..b2d857e1e54 100644
--- a/ggml/src/ggml-cann/aclnn_ops.cpp
+++ b/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -22,11 +22,14 @@
 
 #include "aclnn_ops.h"
 
+#include <aclnnop/aclnn_addcdiv.h>
 #include <aclnnop/aclnn_avgpool2d.h>
+#include <aclnnop/aclnn_batch_matmul.h>
 #include <aclnnop/aclnn_cast.h>
 #include <aclnnop/aclnn_constant_pad_nd.h>
 #include <aclnnop/aclnn_copy.h>
 #include <aclnnop/aclnn_cos.h>
+#include <aclnnop/aclnn_div.h>
 #include <aclnnop/aclnn_exp.h>
 #include <aclnnop/aclnn_fill_scalar.h>
 #include <aclnnop/aclnn_group_norm.h>
@@ -34,6 +37,7 @@
 #include <aclnnop/aclnn_layer_norm.h>
 #include <aclnnop/aclnn_matmul.h>
 #include <aclnnop/aclnn_max_pool.h>
+#include <aclnnop/aclnn_mm.h>
 #include <aclnnop/aclnn_permute.h>
 #include <aclnnop/aclnn_pow_tensor_tensor.h>
 #include <aclnnop/aclnn_reduce_sum.h>
@@ -53,6 +57,7 @@
 #include <exception>
 #include <vector>
 
+#include "ggml-impl.h"
 #include "kernels/ascendc_kernels.h"
 
 #define GGML_COMMON_DECL_C
@@ -241,10 +246,14 @@ void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
     aclTensor* acl_dst = ggml_cann_create_tensor(dst);
 
-    int64_t concat_dim = 1;
+    const int32_t dim = ggml_get_op_params_i32(dst, 0);
+
+    GGML_ASSERT(dim >= 0 && dim < 4);
+    int32_t acl_dim = 3 - dim;
+
     aclTensor* tensors[] = {acl_src0, acl_src1};
     aclTensorList* tensorList = aclCreateTensorList(tensors, 2);
-    aclnn_concat(ctx, tensorList, acl_dst, concat_dim);
+    aclnn_concat(ctx, tensorList, acl_dst, acl_dim);
 
     ACL_CHECK(aclDestroyTensorList(tensorList));
     ACL_CHECK(aclDestroyTensor(acl_dst));
@@ -1096,9 +1105,9 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
 }
 
 /**
- * @brief Creates an ACL tensor initialized with ones using a provided buffer.
+ * @brief Creates an ACL tensor initialized with value using a provided buffer.
  *
- * This function initializes a tensor with ones using the specified buffer and
+ * This function initializes a tensor with value using the specified buffer and
  * tensor parameters.
  *
  * @param ctx The context for the CANN backend operations.
@@ -1111,12 +1120,12 @@ static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
  * @param type_size The size of each element in the tensor data type.
  * @param value The value to be used for initializing the tensor (default
  * is 1.0).
- * @return An ACL tensor initialized with ones.
+ * @return An ACL tensor initialized with value.
  */
-static aclTensor* aclnn_ones(ggml_backend_cann_context& ctx, void* buffer,
-                             size_t n_bytes, int64_t* ne, int64_t dims,
-                             aclDataType type, size_t type_size,
-                             float value = 1.0f) {
+static aclTensor* aclnn_values(ggml_backend_cann_context& ctx, void* buffer,
+                               size_t n_bytes, int64_t* ne, int64_t dims,
+                               aclDataType type, size_t type_size,
+                               float value = 1.0f) {
     aclTensor* acl_tensor =
         aclnn_zero(ctx, buffer, n_bytes, ne, dims, type, type_size);
     float alpha_host = 1.0f;
@@ -1158,7 +1167,7 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     size_t one_tensor_n_bytes = src->ne[0] * ggml_element_size(src);
     ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
 
-    aclTensor* acl_gamma = aclnn_ones(
+    aclTensor* acl_gamma = aclnn_values(
         ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne, 1,
         ggml_cann_type_mapping(src->type), ggml_element_size(src));
 
@@ -1202,9 +1211,9 @@ void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
 
     aclTensor* mask_tensor =
-        aclnn_ones(ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne,
-                   GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
-                   ggml_element_size(src), value);
+        aclnn_values(ctx, one_tensor_allocator.get(), one_tensor_n_bytes,
+                     src->ne, GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
+                     ggml_element_size(src), value);
 
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
@@ -1437,10 +1446,6 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     ggml_tensor* src0 = dst->src[0];  // kernel
     ggml_tensor* src1 = dst->src[1];  // input
 
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
-
     GGML_TENSOR_BINARY_OP_LOCALS;
 
     // aclnnIm2col only works on 2D. set s1, p1, d1 to 1 to perform 2D
@@ -1462,9 +1467,6 @@ void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     const int64_t OH = is_2D ? ne2 : 1;
     const int64_t OW = ne1;
 
-    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nb10 == sizeof(float));
-
     // memory allocated increased to 3x when is_2D == false
     const int64_t n_bytes_factor = is_2D ? 1 : 3;
 
@@ -1768,6 +1770,92 @@ static void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
     ACL_CHECK(aclnnSin(workspaceAddr, workspaceSize, executor, ctx.stream()));
 }
 
+/**
+ * @brief Performs element-wise division of tensor1 by tensor2 , multiplies the
+ result by the scalar value and adds it to self .
+ *
+ * Performs element-wise division of tensor1 by tensor2,
+ * multiplies the result by the scalar value and adds it to self .
+ * The operation is defined as:
+ * \f[
+ *     \text{out}_i = \text{selft}_i + \text{value} \times
+ \frac{\text{tensor1}_i}{\text{tensor2}_i}
+ * \f]
+
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_self The source tensor on which the addcdiv function will be
+ applied.
+ * @param tensor1 Numerator tensor.
+ * @param tensor2 Denominator tensor.
+ * @param value The value to be used for coefficient.
+ */
+static void aclnn_inplace_addcdiv(ggml_backend_cann_context& ctx,
+                                  aclTensor* acl_self, aclTensor* tensor1,
+                                  aclTensor* tensor2, float value) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+    aclScalar* acl_value = aclCreateScalar(&value, aclDataType::ACL_FLOAT);
+
+    ACL_CHECK(aclnnInplaceAddcdivGetWorkspaceSize(
+        acl_self, tensor1, tensor2, acl_value, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnInplaceAddcdiv(workspaceAddr, workspaceSize, executor,
+                                  ctx.stream()));
+}
+
+/**
+ * @brief Matrix division, optionally in-place.
+ *
+ * This function division each element of the source tensor `acl_src` by the
+ * tensor `acl_other` and stores the result in the destination tensor `acl_dst`.
+ * If `inplace` is true, `acl_dst` will not be used and the operation is
+ * performed in-place on `acl_src`. The operation is defined as: \f[
+ *     \text{dst}_i = \frac{\text{acl_src}_i}{\text{acl_other}_i}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src Numerator tensor..
+ * @param acl_other Denominator tensor.
+ * @param acl_dst The destination tensor where the result will be stored if
+ * `inplace` is false.
+ * @param inplace Flag indicating whether to perform the operation in-place on
+ * `acl_src`.
+ */
+static void aclnn_div_tensor(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                             aclTensor* acl_other, aclTensor* acl_dst,
+                             bool inplace) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    if (inplace) {
+        ACL_CHECK(aclnnInplaceDivGetWorkspaceSize(acl_src, acl_other,
+                                                  &workspaceSize, &executor));
+        if (workspaceSize > 0) {
+            ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+            workspaceAddr = workspace_allocator.get();
+        }
+
+        ACL_CHECK(aclnnInplaceDiv(workspaceAddr, workspaceSize, executor,
+                                  ctx.stream()));
+    } else {
+        ACL_CHECK(aclnnDivGetWorkspaceSize(acl_src, acl_other, acl_dst,
+                                           &workspaceSize, &executor));
+        if (workspaceSize > 0) {
+            ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+            workspaceAddr = workspace_allocator.get();
+        }
+
+        ACL_CHECK(
+            aclnnDiv(workspaceAddr, workspaceSize, executor, ctx.stream()));
+    }
+}
+
 void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx,
                                   ggml_tensor* dst) {
     const ggml_tensor* src = dst->src[0];
@@ -2311,7 +2399,16 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                                ctx.stream()));
 
     switch (src0->type) {
-        case GGML_TYPE_F32:
+        case GGML_TYPE_F32: {
+#ifdef ASCEND_310P
+            // Special operation for get_row_f32 kernel of 310P: clear the
+            // content of dest data buffer when row is not aligned to 32 bytes
+            if ((src0->ne[0] % 8) != 0) {
+                size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] *
+                                 src0->ne[0] * ggml_type_size(GGML_TYPE_F32);
+                ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
+            }
+#endif
             aclrtlaunch_ascendc_get_row_f32(
                 24, ctx.stream(), src0->data, src1->data, dst->data,
                 ((ggml_tensor*)src0->extra)->ne,
@@ -2320,7 +2417,19 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
                 ((ggml_tensor*)dst->extra)->nb);
             break;
-        case GGML_TYPE_F16:
+        }
+        case GGML_TYPE_F16: {
+#ifdef ASCEND_310P
+            // Special operation for get_row_f16 kernel of 310P: clear the
+            // content of dest data buffer when row is not aligned to 32 bytes
+            if ((src0->ne[0] % 16) != 0) {
+                size_t dst_len =
+                    src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] *
+                    ggml_type_size(
+                        GGML_TYPE_F32);  // out is also f32, even input is f16
+                ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
+            }
+#endif
             aclrtlaunch_ascendc_get_row_f16(
                 24, ctx.stream(), src0->data, src1->data, dst->data,
                 ((ggml_tensor*)src0->extra)->ne,
@@ -2329,6 +2438,7 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
                 ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
                 ((ggml_tensor*)dst->extra)->nb);
             break;
+        }
         case GGML_TYPE_Q4_0:
             aclrtlaunch_ascendc_get_row_q4_0(
                 24, ctx.stream(), src0->data, src1->data, dst->data,
@@ -2407,7 +2517,6 @@ static void aclnn_mat_mul(ggml_backend_cann_context& ctx, aclTensor* acl_input,
                           aclTensor* acl_weight, aclTensor* acl_dst) {
     int8_t cube_math_type = 1;  // ALLOW_FP32_DOWN_PRECISION, when input is
                                 // fp32, atlas a2 will transpose it to HFLOAT32.
-
     uint64_t workspaceSize = 0;
     aclOpExecutor* executor;
     void* workspaceAddr = nullptr;
@@ -2425,6 +2534,81 @@ static void aclnn_mat_mul(ggml_backend_cann_context& ctx, aclTensor* acl_input,
         aclnnMatmul(workspaceAddr, workspaceSize, executor, ctx.stream()));
 }
 
+/**
+ * @brief Performs matrix multiplication of two 2D tensors.
+ *
+ * This function computes the matrix multiplication of the input tensor
+ * `acl_input` and the weight tensor `acl_weight`, and stores the result in the
+ * destination tensor `acl_dst`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst}=\text {acl_input@acl_weight}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_input The input tensor for the matrix multiplication.
+ * @param acl_weight The weight tensor for the matrix multiplication.
+ * @param acl_dst The destination tensor where the result of the matrix
+ * multiplication will be stored.
+ */
+static void aclnn_mat_mul_2d(ggml_backend_cann_context& ctx,
+                             aclTensor* acl_input, aclTensor* acl_weight,
+                             aclTensor* acl_dst) {
+    int8_t cube_math_type = 2;
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnMmGetWorkspaceSize(acl_input, acl_weight, acl_dst,
+                                      cube_math_type, &workspaceSize,
+                                      &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnMm(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
+/**
+ * @brief Performs matrix multiplication of two 3D tensors.
+ *
+ * This function computes the matrix multiplication of the input tensor
+ * `acl_input` and the weight tensor `acl_weight`, and stores the result in the
+ * destination tensor `acl_dst`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst}=\text {acl_input@acl_weight}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_input The input tensor for the matrix multiplication.
+ * @param acl_weight The weight tensor for the matrix multiplication.
+ * @param acl_dst The destination tensor where the result of the matrix
+ * multiplication will be stored.
+ */
+static void aclnn_mat_mul_3d(ggml_backend_cann_context& ctx,
+                             aclTensor* acl_input, aclTensor* acl_weight,
+                             aclTensor* acl_dst) {
+    int8_t cube_math_type = 2;
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnBatchMatMulGetWorkspaceSize(acl_input, acl_weight, acl_dst,
+                                               cube_math_type, &workspaceSize,
+                                               &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnBatchMatMul(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
 /**
  * @brief Performs matrix multiplication with floating-point precision on
  * tensors using the CANN backend.
@@ -2446,20 +2630,39 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
     // broadcast, when weight ne2 or ne3 is not 1, weight need repeat.
     BCAST_MUL_MAT_SHAPE(input, weight, dst);
 
-    // transpose weight: [1,2,3,4] -> [1,2,4,3]
+    int64_t n_dims = bcast_dims;
+    if (bcast_input_ne[3] == bcast_weight_ne[3] && bcast_input_ne[3] == 1) {
+        if (bcast_input_ne[2] == 1 && bcast_weight_ne[2] == 1) {
+            n_dims = 2;
+        } else if (bcast_input_ne[2] == 1) {
+            n_dims = 3;
+        }
+    }
+
+    aclTensor* acl_input_tensor =
+        ggml_cann_create_tensor(input, bcast_input_ne, bcast_input_nb, n_dims);
     int64_t transpose_ne[] = {bcast_weight_ne[1], bcast_weight_ne[0],
                               bcast_weight_ne[2], bcast_weight_ne[3],
                               bcast_weight_ne[4], bcast_weight_ne[5]};
     size_t transpose_nb[] = {bcast_weight_nb[1], bcast_weight_nb[0],
                              bcast_weight_nb[2], bcast_weight_nb[3],
                              bcast_weight_nb[4], bcast_weight_nb[5]};
-
     aclTensor* acl_weight_tensor =
-        ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, bcast_dims);
-    aclTensor* acl_input_tensor =
-        ggml_cann_create_tensor(input, BCAST_MUL_MAT_PARAM(input));
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst, BCAST_MUL_MAT_PARAM(dst));
-    aclnn_mat_mul(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+        ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims);
+    aclTensor* acl_dst =
+        ggml_cann_create_tensor(dst, bcast_dst_ne, bcast_dst_nb, n_dims);
+
+    switch (n_dims) {
+        case 2:
+            aclnn_mat_mul_2d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+            break;
+        case 3:
+            aclnn_mat_mul_3d(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+            break;
+        default:
+            aclnn_mat_mul(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+            break;
+    }
 
     ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
     ACL_CHECK(aclDestroyTensor(acl_input_tensor));
@@ -2480,51 +2683,47 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
  * multiplication will be stored.
  */
 static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
-                                   ggml_tensor* dst,
-                                   const enum ggml_type type) {
+                                    ggml_tensor* dst,
+                                    const enum ggml_type type) {
     ggml_tensor* src0 = dst->src[0];  // weight
     ggml_tensor* src1 = dst->src[1];  // input
 
-    // The shape of the weight is NCHW. Matrix multiplication uses HW dims. HC
-    // is regarded as batch. weight need transpose.
-    int64_t weight_ne[] = {src0->ne[1], src0->ne[0]};
+    // The shape of the weight is NCHW.
+    // Matrix multiplication uses HW dims.
+    // HC is regarded as batch.
+    // weight need transpose.
     float weight_elem_size;
     if (type == GGML_TYPE_Q4_0) {
         weight_elem_size = float(sizeof(uint8_t)) / 2;
-    }
-    else if (type == GGML_TYPE_Q8_0) {
+    } else if (type == GGML_TYPE_Q8_0) {
         weight_elem_size = float(sizeof(uint8_t));
-    }
-    else {
+    } else {
         GGML_ABORT("Only support Q4_0 and Q8_0 MUL_MAT");
     }
-    float weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
-
-    // size of one matrix is element_size * height * width.
-    size_t weight_stride = weight_elem_size * src0->ne[0] * src0->ne[1];
+    float weight_nb[] = {src0->ne[0] * weight_elem_size, weight_elem_size};
+    size_t weight_stride = src0->ne[1] * src0->ne[0] * weight_elem_size;
     size_t weight_size = weight_stride * src0->ne[2] * src0->ne[3];
 
     // scale stored at the end of weight. Also need transpose.
-    GGML_ASSERT(QK4_0 == QK8_0);
-    int64_t scale_ne[] = {src0->ne[1], src0->ne[0] / QK8_0};
     size_t scale_elem_size = sizeof(uint16_t);
     size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size,
                          scale_elem_size};
-    size_t scale_stride = scale_elem_size * src0->ne[0] * src0->ne[1] / QK8_0;
+    size_t scale_stride = src0->ne[1] * src0->ne[0] / QK8_0 * scale_elem_size;
     char* scale_offset = (char*)src0->data + weight_size;
 
     // input
-    void* input_buffer;
     size_t input_elem_size = sizeof(uint16_t);
     int64_t input_ne[] = {src1->ne[0], src1->ne[1]};
-    size_t input_nb[] = {input_elem_size, input_elem_size * src1->ne[0]};
-    size_t input_stride = input_elem_size * src1->ne[0] * src1->ne[1];
-
+    size_t input_nb[] = {input_elem_size, input_ne[0] * input_elem_size};
+    size_t input_stride = input_ne[0] * input_ne[1] * input_elem_size;
     ggml_cann_pool_alloc input_alloctor(ctx.pool());
+    void* input_buffer = src1->data;
+
+    // case in
     if (src1->type != GGML_TYPE_F16) {
         aclTensor* acl_src1_tensor = ggml_cann_create_tensor(src1);
-        input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
-        input_buffer = input_alloctor.get();
+        input_buffer =
+            input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
 
         int64_t* input_cast_ne = src1->ne;
         size_t input_cast_nb[GGML_MAX_DIMS];
@@ -2537,85 +2736,136 @@ static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
             input_buffer, ACL_FLOAT16, input_elem_size, input_cast_ne,
             input_cast_nb, GGML_MAX_DIMS);
         aclnn_cast(ctx, acl_src1_tensor, acl_input_tensor, ACL_FLOAT16);
+
         ACL_CHECK(aclDestroyTensor(acl_input_tensor));
         ACL_CHECK(aclDestroyTensor(acl_src1_tensor));
-    } else {
-        input_buffer = src1->data;
     }
 
     // output
     size_t output_elem_size = sizeof(uint16_t);
-    int64_t output_ne[] = {dst->ne[0], dst->ne[1]};
-    size_t output_nb[] = {output_elem_size, output_elem_size * dst->ne[0]};
-    ggml_cann_pool_alloc output_alloctor(
-        ctx.pool(), ggml_nelements(dst) * output_elem_size);
-    void* output_buffer = output_alloctor.get();
-    size_t output_stride = output_elem_size * dst->ne[0] * dst->ne[1];
+    size_t output_nb[] = {output_elem_size, dst->ne[0] * output_elem_size};
+    ggml_cann_pool_alloc output_allocator(ctx.pool());
+    void* output_buffer =
+        output_allocator.alloc(ggml_nelements(dst) * output_elem_size);
+    size_t output_stride = dst->ne[0] * dst->ne[1] * output_elem_size;
 
     // aclnn
+    int64_t max_elem_size = 65535;
+    int64_t split_size = (src0->ne[1] / max_elem_size) + 1;
+    ggml_cann_pool_alloc workspace_allocator(ctx.pool());
+    aclOpExecutor* executor = nullptr;
     uint64_t workspaceSize = 0;
-    aclOpExecutor* executor;
     void* workspaceAddr = nullptr;
-
     for (int64_t n1 = 0; n1 < src1->ne[3]; n1++) {
         for (int64_t c1 = 0; c1 < src1->ne[2]; c1++) {
             int64_t n0 = n1 / (src1->ne[3] / src0->ne[3]);
             int64_t c0 = c1 / (src1->ne[2] / src0->ne[2]);
 
-            int64_t batch1 = n1 * src1->ne[2] + c1;
-            int64_t batch0 = n0 * src0->ne[2] + c0;
+            int64_t batch1 = (n1 * src1->ne[2]) + c1;
+            int64_t batch0 = (n0 * src0->ne[2]) + c0;
 
             aclTensor* acl_input_tensor = ggml_cann_create_tensor(
                 (char*)input_buffer + batch1 * input_stride, ACL_FLOAT16,
                 input_elem_size, input_ne, input_nb, 2);
+
+            // first split
+            int64_t weight_ne_offset = 0;
+            int64_t weight_ne[2] = {
+                max_elem_size > src0->ne[1] ? src0->ne[1] : max_elem_size,
+                src0->ne[0]};
+            int64_t scale_ne_offset = 0;
+            int64_t scale_ne[2] = {weight_ne[0], weight_ne[1] / QK8_0};
+            int64_t output_ne_offset = 0;
+            int64_t output_ne[2] = {weight_ne[0], dst->ne[1]};
+
             aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
                 (char*)src0->data + batch0 * weight_stride,
                 ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
-                weight_nb, 2);
+                weight_nb, 2, ACL_FORMAT_ND, weight_ne_offset);
             aclTensor* acl_scale_tensor = ggml_cann_create_tensor(
                 scale_offset + batch0 * scale_stride, ACL_FLOAT16,
-                scale_elem_size, scale_ne, scale_nb, 2);
+                scale_elem_size, scale_ne, scale_nb, 2, ACL_FORMAT_ND,
+                scale_ne_offset);
             aclTensor* acl_output_tensor = ggml_cann_create_tensor(
                 (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16,
-                output_elem_size, output_ne, output_nb, 2);
+                output_elem_size, output_ne, output_nb, 2, ACL_FORMAT_ND,
+                output_ne_offset);
 
             ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
                 acl_input_tensor, acl_weight_tensor, acl_scale_tensor, nullptr,
                 nullptr, nullptr, nullptr, QK8_0, acl_output_tensor,
                 &workspaceSize, &executor));
-
-            if (workspaceSize > 0 && workspaceAddr == nullptr) {
-                ggml_cann_pool_alloc workspace_allocator(ctx.pool(),
-                                                         workspaceSize);
-                workspaceAddr = workspace_allocator.get();
+            if (workspaceAddr == nullptr) {
+                workspaceAddr = workspace_allocator.alloc(workspaceSize);
             }
-
             ACL_CHECK(aclnnWeightQuantBatchMatmulV2(
                 workspaceAddr, workspaceSize, executor, ctx.stream()));
 
-            ACL_CHECK(aclDestroyTensor(acl_input_tensor));
             ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
             ACL_CHECK(aclDestroyTensor(acl_scale_tensor));
             ACL_CHECK(aclDestroyTensor(acl_output_tensor));
+
+            // other splits
+            for (int64_t split = 1; split < split_size; split++) {
+                weight_ne_offset +=
+                    weight_elem_size * weight_ne[0] * weight_ne[1];
+                weight_ne[0] = max_elem_size * (split + 1) > src0->ne[1]
+                                   ? src0->ne[1] - (max_elem_size * split)
+                                   : max_elem_size;
+                scale_ne_offset += scale_elem_size * scale_ne[0] * scale_ne[1];
+                scale_ne[0] = weight_ne[0];
+                output_ne_offset +=
+                    output_elem_size * output_ne[0] * output_ne[1];
+                output_ne[0] = weight_ne[0];
+
+                acl_weight_tensor = ggml_cann_create_tensor(
+                    (char*)src0->data + batch0 * weight_stride,
+                    ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
+                    weight_nb, 2, ACL_FORMAT_ND, weight_ne_offset);
+                acl_scale_tensor = ggml_cann_create_tensor(
+                    scale_offset + batch0 * scale_stride, ACL_FLOAT16,
+                    scale_elem_size, scale_ne, scale_nb, 2, ACL_FORMAT_ND,
+                    scale_ne_offset);
+                acl_output_tensor = ggml_cann_create_tensor(
+                    (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16,
+                    output_elem_size, output_ne, output_nb, 2, ACL_FORMAT_ND,
+                    output_ne_offset);
+
+                ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
+                    acl_input_tensor, acl_weight_tensor, acl_scale_tensor,
+                    nullptr, nullptr, nullptr, nullptr, QK8_0,
+                    acl_output_tensor, &workspaceSize, &executor));
+                ACL_CHECK(aclnnWeightQuantBatchMatmulV2(
+                    workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+                ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
+                ACL_CHECK(aclDestroyTensor(acl_scale_tensor));
+                ACL_CHECK(aclDestroyTensor(acl_output_tensor));
+            }
+
+            ACL_CHECK(aclDestroyTensor(acl_input_tensor));
         }
     }
 
     // cast out
-    int64_t* output_cast_ne = dst->ne;
-    size_t output_cast_nb[GGML_MAX_DIMS];
-    output_cast_nb[0] = sizeof(uint16_t);
-    for (int i = 1; i < GGML_MAX_DIMS; i++) {
-        output_cast_nb[i] = output_cast_nb[i - 1] * output_cast_ne[i - 1];
-    }
+    if (dst->type != GGML_TYPE_F16) {
+        int64_t* output_cast_ne = dst->ne;
+        size_t output_cast_nb[GGML_MAX_DIMS];
+        output_cast_nb[0] = sizeof(uint16_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            output_cast_nb[i] = output_cast_nb[i - 1] * output_cast_ne[i - 1];
+        }
 
-    aclTensor* acl_output_tensor =
-        ggml_cann_create_tensor(output_buffer, ACL_FLOAT16, output_elem_size,
-                                output_cast_ne, output_cast_nb, GGML_MAX_DIMS);
-    aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
-    aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor, ACL_FLOAT);
+        aclTensor* acl_output_tensor = ggml_cann_create_tensor(
+            output_buffer, ACL_FLOAT16, output_elem_size, output_cast_ne,
+            output_cast_nb, GGML_MAX_DIMS);
+        aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
+        aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor,
+                   ggml_cann_type_mapping(dst->type));
 
-    ACL_CHECK(aclDestroyTensor(acl_output_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_dst_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_output_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_dst_tensor));
+    }
 }
 
 void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
@@ -2714,12 +2964,14 @@ static void aclnn_index_fill_tensor(ggml_backend_cann_context& ctx,
 static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
                              aclTensor* acl_cos_repeat_tensor,
                              aclTensor* acl_sin_repeat_tensor,
-                             float theta_scale, bool is_neox) {
+                             float theta_scale, float freq_scale,
+                             float attn_factor, bool is_neox) {
     // int sin/cos cache, cache has different repeat method depond on
     // @param.is_neox
 
     ggml_tensor* src0 = dst->src[0];  // input
     ggml_tensor* src1 = dst->src[1];  // position
+    ggml_tensor* src2 = dst->src[2];  // freq_factors
 
     // arange, [0,1,...,ne0/2]
     int64_t arange_length = src0->ne[0] / 2;
@@ -2748,11 +3000,26 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     ggml_cann_pool_alloc theta_scale_allocator(ctx.pool(),
                                                arange_length * sizeof(float_t));
     void* theta_scale_buffer = theta_scale_allocator.get();
-    aclTensor* acl_theta_scale_tensor = aclnn_ones(
+    aclTensor* acl_theta_scale_tensor = aclnn_values(
         ctx, theta_scale_buffer, arange_length * sizeof(float_t), arange_ne,
         GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), theta_scale);
     aclnn_pow_tensor_tensor(ctx, acl_theta_scale_tensor, acl_arange_tensor);
 
+    // freq_scale
+    if (freq_scale != 1) {
+        aclnn_muls(ctx, acl_theta_scale_tensor, freq_scale, nullptr, true);
+    }
+
+    // freq_factors
+    if (src2) {
+        aclTensor* acl_freq_factors_tensor = ggml_cann_create_tensor(
+            src2->data, ggml_cann_type_mapping(src2->type),
+            ggml_type_size(src2->type), arange_ne, arange_nb, GGML_MAX_DIMS);
+        aclnn_div_tensor(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor,
+                         nullptr, true);
+        ACL_CHECK(aclDestroyTensor(acl_freq_factors_tensor));
+    }
+
     // position
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
     int64_t position_length = src1->ne[0];
@@ -2816,6 +3083,12 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
         GGML_MAX_DIMS, ACL_FORMAT_ND);
     aclnn_cos(ctx, acl_permute_tensor, acl_cos_tensor);
 
+    // attn_factor
+    if (attn_factor != 1) {
+        aclnn_muls(ctx, acl_sin_tensor, attn_factor, nullptr, true);
+        aclnn_muls(ctx, acl_cos_tensor, attn_factor, nullptr, true);
+    }
+
     // repeat
     if (is_neox) {
         int64_t repeatsArray[] = {1, 1, 1, 2};
@@ -2841,15 +3114,27 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
     ACL_CHECK(aclDestroyTensor(acl_cos_tensor));
 }
 
+#ifdef __cplusplus
+extern "C" {
+#endif
+aclnnStatus aclnnRotaryPositionEmbeddingGetWorkspaceSize(
+    const aclTensor* x, const aclTensor* cos, const aclTensor* sin,
+    int64_t mode, const aclTensor* yOut, uint64_t* workspaceSize,
+    aclOpExecutor** executor);
+aclnnStatus aclnnRotaryPositionEmbedding(void* workspace,
+                                         uint64_t workspaceSize,
+                                         aclOpExecutor* executor,
+                                         aclrtStream stream);
+#ifdef __cplusplus
+}
+#endif
+
 void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     // TODO: use ascendc
     // Only test with LLAMA model.
     ggml_tensor* src0 = dst->src[0];  // input
     ggml_tensor* src2 = dst->src[2];  // freq_factors
 
-    // TODO: with freq_factors
-    GGML_ASSERT(src2 == NULL);
-
     // param
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
     // const int n_past     = ((int32_t *) dst->op_params)[0];
@@ -2867,13 +3152,11 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
     memcpy(&beta_fast, (int32_t*)dst->op_params + 9, sizeof(float));
     memcpy(&beta_slow, (int32_t*)dst->op_params + 10, sizeof(float));
 
-    GGML_ASSERT(n_dims <= ne0);
+    // TODO: n_dims <= ne0
+    GGML_ASSERT(n_dims == ne0);
     GGML_ASSERT(n_dims % 2 == 0);
-
     // TODO: ext_factor != 0
     GGML_ASSERT(ext_factor == 0);
-    // TODO: freq_scale != 1
-    GGML_ASSERT(freq_scale == 1);
 
     const float theta_scale = powf(freq_base, -2.0f / n_dims);
 
@@ -2904,7 +3187,13 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float_t),
                                 sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
     aclnn_cache_init(ctx, dst, acl_cos_reshape_tensor, acl_sin_reshape_tensor,
-                     theta_scale, is_neox);
+                     theta_scale, freq_scale, attn_factor, is_neox);
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src0);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+#ifdef ASCEND_310P
+    // Special ROPE operation for 310P
 
     // roll input
     void* input_roll_buffer;
@@ -2947,7 +3236,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         for (int i = 1; i < GGML_MAX_DIMS; i++) {
             minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
         }
-        acl_minus_one_tensor = aclnn_ones(
+        acl_minus_one_tensor = aclnn_values(
             ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
             minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
         int64_t dim = 3;
@@ -2974,17 +3263,15 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
 
         ACL_CHECK(aclDestroyTensor(acl_input_roll_tensor));
         ACL_CHECK(aclDestroyTensor(acl_input_tensor));
-
         // init [-1, -1, -1, 1, 1，1，...]
         minus_one_scale_buffer = minus_one_scale_allocator.get();
-
         int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1};
         size_t minus_one_nb[GGML_MAX_DIMS];
         minus_one_nb[0] = sizeof(float_t);
         for (int i = 1; i < GGML_MAX_DIMS; i++) {
             minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
         }
-        acl_minus_one_tensor = aclnn_ones(
+        acl_minus_one_tensor = aclnn_values(
             ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
             minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
         // -1 * first half
@@ -3026,14 +3313,12 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
               acl_input_roll_mul_scale_tensor);
 
     // output
-    aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
-    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
     void* output_fp32_buffer;
     if (src0->type == GGML_TYPE_F32) {
-        aclnn_inplace_mul(ctx, acl_src0, acl_cos_reshape_tensor);
+        aclnn_inplace_mul(ctx, acl_src, acl_cos_reshape_tensor);
         aclnn_inplace_mul(ctx, acl_input_roll_mul_scale_tensor,
                           acl_sin_reshape_tensor);
-        aclnn_add(ctx, acl_src0, acl_input_roll_mul_scale_tensor, acl_dst);
+        aclnn_add(ctx, acl_src, acl_input_roll_mul_scale_tensor, acl_dst);
         // TODO: ne0 != n_dims in mode2
     } else if (src0->type == GGML_TYPE_F16) {
         size_t input_fp32_nb[GGML_MAX_DIMS];
@@ -3060,7 +3345,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         aclTensor* output_fp32_tensor = ggml_cann_create_tensor(
             output_fp32_buffer, ACL_FLOAT, sizeof(float_t), dst->ne,
             input_fp32_nb, GGML_MAX_DIMS);
-        aclnn_mul(ctx, acl_src0, acl_cos_reshape_tensor, input_fp32_tensor1);
+        aclnn_mul(ctx, acl_src, acl_cos_reshape_tensor, input_fp32_tensor1);
         aclnn_mul(ctx, acl_input_roll_mul_scale_tensor, acl_sin_reshape_tensor,
                   input_fp32_tensor2);
         aclnn_add(ctx, input_fp32_tensor1, input_fp32_tensor2,
@@ -3070,13 +3355,73 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
         ACL_CHECK(aclDestroyTensor(input_fp32_tensor1));
         ACL_CHECK(aclDestroyTensor(input_fp32_tensor2));
         ACL_CHECK(aclDestroyTensor(output_fp32_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_minus_one_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_input_roll_mul_scale_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_input_roll_reshape_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_src));
     }
+    return;
+#endif
 
-    ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
+    // src0 == GGML_TYPE_F16
+    // TODO: optimization this `if` code
+    if (src0->type == GGML_TYPE_F16) {
+        ggml_cann_pool_alloc sin_final_allocator(
+            ctx.pool(), src0->ne[0] * src0->ne[2] * ggml_type_size(src0->type));
+        ggml_cann_pool_alloc cos_final_allocator(
+            ctx.pool(), src0->ne[0] * src0->ne[2] * ggml_type_size(src0->type));
+        void* sin_final_buffer = sin_final_allocator.get();
+        void* cos_final_buffer = cos_final_allocator.get();
+
+        int64_t sin_final_ne[4] = {src0->ne[0], 1, src0->ne[2], 1};
+        size_t sin_final_nb[GGML_MAX_DIMS];
+        sin_final_nb[0] = ggml_type_size(src0->type);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            sin_final_nb[i] = sin_final_nb[i - 1] * sin_final_ne[i - 1];
+        }
+        aclTensor* acl_sin_final_tensor = ggml_cann_create_tensor(
+            sin_final_buffer, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), sin_final_ne, sin_final_nb,
+            GGML_MAX_DIMS);
+        aclTensor* acl_cos_final_tensor = ggml_cann_create_tensor(
+            cos_final_buffer, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), sin_final_ne, sin_final_nb,
+            GGML_MAX_DIMS);
+
+        aclnn_cast(ctx, acl_sin_reshape_tensor, acl_sin_final_tensor,
+                   ggml_cann_type_mapping(src0->type));
+        aclnn_cast(ctx, acl_cos_reshape_tensor, acl_cos_final_tensor,
+                   ggml_cann_type_mapping(src0->type));
+        ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
+        acl_sin_reshape_tensor = acl_sin_final_tensor;
+        acl_cos_reshape_tensor = acl_cos_final_tensor;
+    }
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+
+    void* workspaceAddr = nullptr;
+
+    int acl_mode = mode;
+    if (mode == 0) {
+        acl_mode = 1;
+    }
+
+    ACL_CHECK(aclnnRotaryPositionEmbeddingGetWorkspaceSize(
+        acl_src, acl_cos_reshape_tensor, acl_sin_reshape_tensor, acl_mode,
+        acl_dst, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnRotaryPositionEmbedding(workspaceAddr, workspaceSize,
+                                           executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyTensor(acl_src));
     ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_minus_one_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_input_roll_mul_scale_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_input_roll_reshape_tensor));
-    ACL_CHECK(aclDestroyTensor(acl_src0));
+    ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
     ACL_CHECK(aclDestroyTensor(acl_dst));
 }
diff --git a/ggml/src/ggml-cann/common.h b/ggml/src/ggml-cann/common.h
index edfa496148f..5164cb74ec9 100644
--- a/ggml/src/ggml-cann/common.h
+++ b/ggml/src/ggml-cann/common.h
@@ -211,17 +211,20 @@ struct ggml_cann_pool_alloc {
 struct ggml_backend_cann_context {
     int32_t device;                  /**< Device ID. */
     std::string name;                /**< Name of the device. */
+    std::string description;         /**< Description of the device. */
     aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
 
-    aclrtStream streams[GGML_CANN_MAX_STREAMS] = {
-        {nullptr}}; /**< Array of streams for the device. */
+    aclrtStream streams[GGML_CANN_MAX_STREAMS] = {nullptr}; /**< Array of streams for the device. */
 
     /**
      * @brief Constructor for initializing the context with a given device.
      * @param device Device ID.
      */
     explicit ggml_backend_cann_context(int device)
-        : device(device), name("CANN" + std::to_string(device)) {}
+        : device(device), name("CANN" + std::to_string(device)) {
+        ggml_cann_set_device(device);
+        description = aclrtGetSocName();
+    }
 
     /**
      * @brief Destructor for cleaning up resources.
diff --git a/ggml/src/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
similarity index 95%
rename from ggml/src/ggml-cann.cpp
rename to ggml/src/ggml-cann/ggml-cann.cpp
index 77634088143..d410c02445c 100644
--- a/ggml/src/ggml-cann.cpp
+++ b/ggml/src/ggml-cann/ggml-cann.cpp
@@ -122,6 +122,10 @@ static ggml_cann_device_info ggml_cann_init() {
         ACL_CHECK(aclrtMemGetAllocationGranularity(
             &prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED,
             &info.devices[id].vmm_granularity));
+
+        size_t free, total;
+        ggml_backend_cann_get_device_memory(id, &free, &total);
+        info.devices[id].total_vram = free;
     }
 
     // TODO: add more device info later.
@@ -208,6 +212,11 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
      * @return A pointer to the allocated buffer.
      */
     void* alloc(size_t size, size_t* actual_size) override {
+        const size_t alignment = 128;
+        size = GGML_PAD(size, alignment);
+        if (size == 0) {
+            size = alignment;
+        }
 #ifdef DEBUG_CANN_MALLOC
         int nnz = 0;
         size_t max_size = 0;
@@ -246,13 +255,11 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
             return ptr;
         }
         void* ptr;
-        size_t look_ahead_size = (size_t)(1.05 * size);
-        look_ahead_size = 256 * ((look_ahead_size + 255) / 256);
         ggml_cann_set_device(device);
         ACL_CHECK(
-            aclrtMalloc(&ptr, look_ahead_size, ACL_MEM_MALLOC_HUGE_FIRST));
-        *actual_size = look_ahead_size;
-        pool_size += look_ahead_size;
+            aclrtMalloc(&ptr, size, ACL_MEM_MALLOC_HUGE_FIRST));
+        *actual_size = size;
+        pool_size += size;
 #ifdef DEBUG_CANN_MALLOC
         GGML_LOG_INFO(
             "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, "
@@ -296,7 +303,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
     /**
      * @brief The maximum size of the virtual memory pool (32 GB).
      */
-    static const size_t CANN_POOL_VMM_MAX_SIZE = 1ull << 35;  // 32 GB
+    size_t max_size;
 
     /**
      * @brief The device ID associated with this buffer pool.
@@ -341,7 +348,11 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
      */
     explicit ggml_cann_pool_vmm(int device)
         : device(device),
-          granularity(ggml_cann_info().devices[device].vmm_granularity) {}
+          granularity(ggml_cann_info().devices[device].vmm_granularity) {
+        auto dev = ggml_cann_info().devices[device];
+        granularity = dev.vmm_granularity;
+        max_size = dev.total_vram;
+    }
 
     /**
      * @brief Destructor to free all buffers in the virtual memory pool.
@@ -370,17 +381,19 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
         // round up the allocation size to the alignment to ensure that all
         // allocations are aligned for all data types
         const size_t alignment = 128;
-        size = alignment * ((size + alignment - 1) / alignment);
+        size = GGML_PAD(size, alignment);
+        if (size == 0) {
+            size = alignment;
+        }
 
         size_t avail = pool_size - pool_used;
 
         if (size > avail) {
             // round up to the next multiple of the granularity
             size_t reserve_size = size - avail;
-            reserve_size =
-                granularity * ((reserve_size + granularity - 1) / granularity);
+            reserve_size = GGML_PAD(reserve_size, granularity);
 
-            GGML_ASSERT(pool_size + reserve_size <= CANN_POOL_VMM_MAX_SIZE);
+            GGML_ASSERT(pool_size + reserve_size <= max_size);
 
             // allocate more physical memory
             aclrtPhysicalMemProp prop = {};
@@ -396,7 +409,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
             // reserve virtual address space (if not already reserved)
             if (pool_addr == 0) {
                 ACL_CHECK(aclrtReserveMemAddress(
-                    &pool_addr, CANN_POOL_VMM_MAX_SIZE, 0, NULL, 1));
+                    &pool_addr, max_size, 0, NULL, 1));
             }
 
             // map at the end of the pool
@@ -409,10 +422,11 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
             // add to the pool
             pool_size += reserve_size;
 
-            // GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB (
-            // reserved %llu MB)\n",
-            //       device, (unsigned long long) (pool_size/1024/1024),
-            //       (unsigned long long) (reserve_size/1024/1024));
+#ifdef DEBUG_CANN_MALLOC
+             GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB (reserved %llu MB)\n",
+                   device, (unsigned long long) (pool_size/1024/1024),
+                   (unsigned long long) (reserve_size/1024/1024));
+#endif
         }
 
         GGML_ASSERT(pool_addr != 0);
@@ -457,7 +471,6 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
  */
 std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
     int device) {
-    // return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_leg(device));
     return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
 }
 
@@ -1130,10 +1143,10 @@ ggml_backend_cann_buffer_type(int32_t device) {
     static bool ggml_backend_cann_buffer_type_initialized = false;
 
     if (!ggml_backend_cann_buffer_type_initialized) {
-        for (int32_t i = 0; i < GGML_CANN_MAX_DEVICES; i++) {
+        for (int32_t i = 0; i < ggml_cann_info().device_count; i++) {
             ggml_backend_cann_buffer_types[i] = {
                 /* .iface    = */ ggml_backend_cann_buffer_type_interface,
-                /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), device),
+                /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cann_reg(), i),
                 /* .context  = */
                  new ggml_backend_cann_buffer_type_context{
                     i, "CANN" + std::to_string(i)},
@@ -1199,10 +1212,15 @@ static void * ggml_cann_host_malloc(size_t size) {
         return nullptr;
     }
 
+    const size_t alignment = 128;
+    size = GGML_PAD(size, alignment);
+    if (size == 0) {
+        size = alignment;
+    }
+
     void * hostPtr = nullptr;
     aclError err = aclrtMallocHost((void **) &hostPtr, size);
     if (err != ACL_SUCCESS) {
-
         GGML_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
                            size / 1024.0 / 1024.0, aclGetRecentErrMsg());
         return nullptr;
@@ -1669,12 +1687,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
             }
         case GGML_OP_MUL_MAT: {
             switch (op->src[0]->type) {
-                case GGML_TYPE_F16:
-                case GGML_TYPE_F32:
                 case GGML_TYPE_Q8_0:
-                    // TODO: fix me
                     // Current groupsize should not be greater than k-1 in
-                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize().
+                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize
+                    if (op->src[0]->ne[0] <= QK8_0) {
+                        return false;
+                    }
+                case GGML_TYPE_F16:
+                case GGML_TYPE_F32:
                 case GGML_TYPE_Q4_0:
                     return true;
                 default:
@@ -1706,9 +1726,50 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
                     return false;
             }
         }
+        case GGML_OP_CONT: {
+            // TODO: support GGML_TYPE_BF16
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                    return true;
+                default:
+                    return false;
+            }
+        }
+        case GGML_OP_ROPE: {
+            // TODO: with ops-test v == 1
+            float * ext_factor = (float*)((int32_t*)op->op_params + 7);
+            // TODO: n_dims <= ne0
+            if (op->src[0]->ne[0] != op->op_params[1]) {
+                return false;
+            }
+            // TODO: ext_factor != 0
+            if (*ext_factor != 0) {
+                return false;
+            }
+
+            const int mode = ((const int32_t *) op->op_params)[2];
+            if (mode & GGML_ROPE_TYPE_MROPE) {
+                return false;
+            }
+            if (mode & GGML_ROPE_TYPE_VISION) {
+                return false;
+            }
+
+            return true;
+        }
+        case GGML_OP_UPSCALE: {
+            // aclnnUpsampleNearest2dGetWorkspaceSize not support
+            // selfDimN[2]/outDimN[2] or selfDimC[3]/outDimC[3] not equal
+            if (op->src[0]->ne[2] * op->ne[3] != op->src[0]->ne[3] * op->ne[2]) {
+                return false;
+            }
+            return true;
+        }
+        case GGML_OP_IM2COL:
+        case GGML_OP_CONCAT:
         case GGML_OP_DUP:
         case GGML_OP_REPEAT:
-        case GGML_OP_CONCAT:
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
@@ -1722,17 +1783,13 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
         case GGML_OP_SCALE:
         case GGML_OP_SQR:
         case GGML_OP_CLAMP:
-        case GGML_OP_CONT:
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_SOFT_MAX:
-        case GGML_OP_ROPE:
-        case GGML_OP_IM2COL:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
         case GGML_OP_GROUP_NORM:
-        case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
@@ -2041,7 +2098,7 @@ static void * ggml_backend_cann_reg_get_proc_address(ggml_backend_reg_t reg, con
 static const ggml_backend_reg_i ggml_backend_cann_reg_interface = {
     /* .get_name          = */ ggml_backend_cann_reg_get_name,
     /* .get_device_count  = */ ggml_backend_cann_reg_get_device_count,
-    /* .get_device_get    = */ ggml_backend_cann_reg_get_device,
+    /* .get_device        = */ ggml_backend_cann_reg_get_device,
     /* .get_proc_address  = */ ggml_backend_cann_reg_get_proc_address,
 };
 
@@ -2064,16 +2121,17 @@ ggml_backend_reg_t ggml_backend_cann_reg() {
                 dev_ctx->name = GGML_CANN_NAME + std::to_string(i);
                 ggml_cann_set_device(i);
                 ggml_backend_dev_t dev = new ggml_backend_device {
-                    /* .interface = */ ggml_backend_cann_device_interface,
-                    /* .reg       = */ &reg,
-                    /* .context   = */ dev_ctx
+                    /* .iface   = */ ggml_backend_cann_device_interface,
+                    /* .reg     = */ &reg,
+                    /* .context = */ dev_ctx
                 };
                 ctx->devices.push_back(dev);
             }
 
             reg = ggml_backend_reg {
-                /* .interface = */ ggml_backend_cann_reg_interface,
-                /* .context   = */ ctx
+                /* .api_version = */ GGML_BACKEND_API_VERSION,
+                /* .iface       = */ ggml_backend_cann_reg_interface,
+                /* .context     = */ ctx
             };
         }
 
@@ -2126,3 +2184,5 @@ void ggml_backend_cann_get_device_memory(int32_t device, size_t* free,
     ggml_cann_set_device(device);
     ACL_CHECK(aclrtGetMemInfo(ACL_HBM_MEM, free, total));
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_cann_reg)
diff --git a/ggml/src/ggml-cann/kernels/CMakeLists.txt b/ggml/src/ggml-cann/kernels/CMakeLists.txt
index 5b4fef91b58..d687220c3c5 100644
--- a/ggml/src/ggml-cann/kernels/CMakeLists.txt
+++ b/ggml/src/ggml-cann/kernels/CMakeLists.txt
@@ -1,7 +1,3 @@
-if (NOT SOC_TYPE)
-    set (SOC_TYPE "Ascend910B3")
-endif()
-
 file(GLOB SRC_FILES
     get_row_f32.cpp
     get_row_f16.cpp
@@ -13,7 +9,6 @@ file(GLOB SRC_FILES
     dup.cpp
 )
 
-string(TOLOWER ${SOC_TYPE} SOC_VERSION)
 set(ASCEND_CANN_PACKAGE_PATH ${CANN_INSTALL_DIR})
 set(RUN_MODE "npu" CACHE STRING "run mode: npu/sim")
 
@@ -30,4 +25,6 @@ ascendc_library(ascendc_kernels STATIC
     ${SRC_FILES}
 )
 
+message(STATUS "CANN: compile ascend kernels witch SOC_TYPE:${SOC_TYPE}, SOC_VERSION:${SOC_VERSION}, compile macro:-D${SOC_TYPE_COMPILE_OPTION}.")
+ascendc_compile_definitions(ascendc_kernels PRIVATE "-D${SOC_TYPE_COMPILE_OPTION}")
 # ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)
diff --git a/ggml/src/ggml-cann/kernels/dup.cpp b/ggml/src/ggml-cann/kernels/dup.cpp
index e2c651152f4..c7ba38d10a0 100644
--- a/ggml/src/ggml-cann/kernels/dup.cpp
+++ b/ggml/src/ggml-cann/kernels/dup.cpp
@@ -5,6 +5,7 @@
 using namespace AscendC;
 
 #define BUFFER_NUM 2
+const int64_t SUPPORTED_MAX_DIM = 65535;  // currently the limit of max block dim supportted by dup kernel is 65535template <typename SRC_T, typename DST_T>
 
 template <typename SRC_T, typename DST_T>
 class DupByRows {
@@ -51,24 +52,36 @@ class DupByRows {
 
     __aicore__ inline void copy_in() {
         LocalTensor<SRC_T> src_local = src_queue.AllocTensor<SRC_T>();
-
-        DataCopyExtParams dataCopyParams;
-        dataCopyParams.blockCount = 1;
-        dataCopyParams.blockLen = num_elem * sizeof(SRC_T);
-        DataCopyPadExtParams<SRC_T> padParams;
-        DataCopyPad(src_local, src_gm, dataCopyParams, padParams);
-
+        const size_t elem_per_block = 32 / sizeof(SRC_T);
+        size_t tail = num_elem % elem_per_block;
+        size_t cpy_elements_len = tail > 0 ? num_elem + 1 : num_elem;
+        DataCopy(src_local, src_gm, cpy_elements_len);
         src_queue.EnQue(src_local);
     }
 
     __aicore__ inline void copy_out() {
         LocalTensor<DST_T> dst_local = dst_queue.DeQue<DST_T>();
-
+#ifdef ASCEND_310P
+        const size_t elem_per_block = 32 / sizeof(DST_T);
+        size_t tail = num_elem % elem_per_block;
+        size_t len = num_elem & ~(elem_per_block - 1);
+        if (len > 0) {
+            DataCopy(dst_gm, dst_local, len);
+        }
+        if(tail != 0) {
+            for (size_t i = tail; i < elem_per_block; i++) {
+                dst_local[len + i].SetValue(0, 0);
+            }
+            SetAtomicAdd<float>();
+            DataCopy(dst_gm[len], dst_local[len], elem_per_block);
+            SetAtomicNone();
+        }
+#else
         DataCopyExtParams dataCopyParams;
         dataCopyParams.blockCount = 1;
         dataCopyParams.blockLen = num_elem * sizeof(DST_T);
         DataCopyPad(dst_gm, dst_local, dataCopyParams);
-
+#endif
         dst_queue.FreeTensor(dst_local);
     }
 
diff --git a/ggml/src/ggml-cann/kernels/get_row_f16.cpp b/ggml/src/ggml-cann/kernels/get_row_f16.cpp
index c704b5b2ec0..416b45104de 100644
--- a/ggml/src/ggml-cann/kernels/get_row_f16.cpp
+++ b/ggml/src/ggml-cann/kernels/get_row_f16.cpp
@@ -14,7 +14,7 @@ class GET_ROW_F16 {
                                 int64_t *output_ne_ub, size_t *output_nb_ub) {
         // TODO, use template for F16/f32
         int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
+        op_block_idx = GetBlockIdx();
 
         for (int i = 0; i < 4; i++) {
             input_ne[i] = input_ne_ub[i];
@@ -59,32 +59,42 @@ class GET_ROW_F16 {
     }
 
     __aicore__ inline void copy_in(uint32_t offset, size_t len) {
+        size_t origin_len = len;
         LocalTensor<half> input_local = input_queue.AllocTensor<half>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(input_local, input_gm[offset], len);
+        const size_t elem_per_block = 32 / sizeof(half);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
         if(tail != 0) {
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = tail * sizeof(half);
-            DataCopyPadExtParams<half> padParams;
-            DataCopyPad(input_local[len], input_gm[offset + len],
-                        dataCopyParams, padParams);
+            len += elem_per_block;
         }
+        DataCopy(input_local, input_gm[offset], len);
         input_queue.EnQue(input_local);
     }
 
     __aicore__ inline void copy_out(uint32_t offset, size_t len) {
         LocalTensor<float> output_local = output_queue.DeQue<float>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(output_gm[offset], output_local, len);
+        const size_t elem_per_block = 32 / sizeof(float);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
+        if (len > 0) {
+            DataCopy(output_gm[offset], output_local, len);
+        }
+
         if(tail != 0) {
+#ifdef ASCEND_310P
+            for (size_t i = tail; i < elem_per_block; i++) {
+                output_local[len + i].SetValue(0, 0);
+            }
+            SetAtomicAdd<float>();
+            DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
+            SetAtomicNone();
+#else
             DataCopyExtParams dataCopyParams;
             dataCopyParams.blockCount = 1;
             dataCopyParams.blockLen = tail * sizeof(float);
             DataCopyPad(output_gm[offset + len], output_local[len],
                         dataCopyParams);
+#endif
         }
         output_queue.FreeTensor(output_local);
     }
@@ -150,6 +160,7 @@ class GET_ROW_F16 {
     GlobalTensor<float> output_gm;
     TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
     TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    int64_t op_block_idx;
 };
 
 template <typename T>
diff --git a/ggml/src/ggml-cann/kernels/get_row_f32.cpp b/ggml/src/ggml-cann/kernels/get_row_f32.cpp
index 9db080af369..02116905b18 100644
--- a/ggml/src/ggml-cann/kernels/get_row_f32.cpp
+++ b/ggml/src/ggml-cann/kernels/get_row_f32.cpp
@@ -13,7 +13,7 @@ class GET_ROW_F32 {
                                 int64_t *indices_ne_ub, size_t *indices_nb_ub,
                                 int64_t *output_ne_ub, size_t *output_nb_ub) {
         int64_t op_block_num = GetBlockNum();
-        int64_t op_block_idx = GetBlockIdx();
+        op_block_idx = GetBlockIdx();
 
         for (int i = 0; i < 4; i++) {
             input_ne[i] = input_ne_ub[i];
@@ -55,31 +55,40 @@ class GET_ROW_F32 {
 
     __aicore__ inline void copy_in(uint32_t offset, size_t len) {
         LocalTensor<float> input_local = input_queue.AllocTensor<float>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(input_local, input_gm[offset], len);
+        const size_t elem_per_block = 32 / sizeof(float);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
         if(tail != 0) {
-            DataCopyExtParams dataCopyParams;
-            dataCopyParams.blockCount = 1;
-            dataCopyParams.blockLen = tail * sizeof(float);
-            DataCopyPadExtParams<float> padParams;
-            DataCopyPad(input_local[len], input_gm[offset + len],
-                        dataCopyParams, padParams);
+            len += elem_per_block;
         }
+        DataCopy(input_local, input_gm[offset], len);
         input_queue.EnQue(input_local);
     }
 
     __aicore__ inline void copy_out(uint32_t offset, size_t len) {
         LocalTensor<float> output_local = output_queue.DeQue<float>();
-        size_t tail = len % 32;
-        len = len & ~31;
-        DataCopy(output_gm[offset], output_local, len);
+        const size_t elem_per_block = 32 / sizeof(float);
+        size_t tail = len % elem_per_block;
+        len = len & ~(elem_per_block - 1);
+        if (len > 0) {
+            DataCopy(output_gm[offset], output_local, len);
+        }
+
         if(tail != 0) {
+#ifdef ASCEND_310P
+            for (size_t i = tail; i < elem_per_block; i++) {
+                output_local[len + i].SetValue(0, 0);
+            }
+            SetAtomicAdd<float>();
+            DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
+            SetAtomicNone();
+#else
             DataCopyExtParams dataCopyParams;
             dataCopyParams.blockCount = 1;
             dataCopyParams.blockLen = tail * sizeof(float);
             DataCopyPad(output_gm[offset + len], output_local[len],
                         dataCopyParams);
+#endif
         }
         output_queue.FreeTensor(output_local);
     }
@@ -144,6 +153,7 @@ class GET_ROW_F32 {
     GlobalTensor<float> output_gm;
     TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
     TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    int64_t op_block_idx;
 };
 
 template <typename T>
diff --git a/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp b/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp
index a80bfeec241..4fbe722086c 100644
--- a/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp
+++ b/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp
@@ -2,6 +2,15 @@
 
 // optimize me. Use template to avoid copy code.
 using namespace AscendC;
+#ifdef ASCEND_310P // 310P not support 4bit get row
+    extern "C" __global__ __aicore__ void ascendc_get_row_q4_0(
+        GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
+        GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm,
+        GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support 4bit get row.\n");
+    }
+#else
 
 #define BUFFER_NUM 2
 
@@ -191,3 +200,5 @@ extern "C" __global__ __aicore__ void ascendc_get_row_q4_0(
             indices_nb_ub, output_ne_ub, output_nb_ub);
     op.calculate();
 }
+
+#endif // #ifdef ASCEND_310P
diff --git a/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp b/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp
index 8423b3f02a8..504b43afaa1 100644
--- a/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp
+++ b/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp
@@ -1,6 +1,14 @@
 #include "kernel_operator.h"
 
 using namespace AscendC;
+#ifdef ASCEND_310P
+    extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0(
+        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support f16->8bit quantization.\n");
+    }
+#else
 
 #define BUFFER_NUM 2
 #define QK8_0 32
@@ -206,3 +214,5 @@ extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0(
     op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
     op.calculate();
 }
+
+#endif // #ifdef ASCEND_310P
diff --git a/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp b/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp
index b7c575093e9..05b0bc1df59 100644
--- a/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp
+++ b/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp
@@ -1,6 +1,14 @@
 #include "kernel_operator.h"
 
 using namespace AscendC;
+#ifdef ASCEND_310P // 310P not support f32->8bit quantization
+    extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0(
+        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support f32->8bit quantization.\n");
+    }
+#else
 
 #define BUFFER_NUM 2
 #define QK8_0 32
@@ -204,3 +212,5 @@ extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0(
     op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
     op.calculate();
 }
+
+#endif // #ifdef ASCEND_310P
diff --git a/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp b/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp
index 9c8c86b66ad..1188937b744 100644
--- a/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp
+++ b/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp
@@ -1,6 +1,21 @@
 #include "kernel_operator.h"
 
 using namespace AscendC;
+#ifdef ASCEND_310P // 310P not support float->4bit quantization
+    extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
+        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support f32->4bit quantization.\n");
+    }
+
+    extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
+        GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+        GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+        // let following test cases can continue run, here just print error information. Of Cource the test case that call this operator is failed.
+        printf("Ascend310P not support f16->4bit quantization.\n");
+    }
+#else
 
 #define BUFFER_NUM 2
 #define Group_Size 32
@@ -276,3 +291,5 @@ extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
     op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
     op.calculate();
 }
+
+#endif // #ifdef ASCEND_310P
diff --git a/ggml/src/ggml-common.h b/ggml/src/ggml-common.h
index 05016139345..f13fd4dea6f 100644
--- a/ggml/src/ggml-common.h
+++ b/ggml/src/ggml-common.h
@@ -6,7 +6,20 @@
 typedef uint16_t ggml_half;
 typedef uint32_t ggml_half2;
 
-#define GGML_COMMON_AGGR
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_CPP)
+#include <cstdint>
+
+typedef uint16_t ggml_half;
+typedef uint32_t ggml_half2;
+
+// std-c++ allow anonymous unions but some compiler warn on it
+#define GGML_COMMON_AGGR_U data
+// std-c++ do not allow it.
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_METAL)
@@ -15,7 +28,8 @@ typedef uint32_t ggml_half2;
 typedef half  ggml_half;
 typedef half2 ggml_half2;
 
-#define GGML_COMMON_AGGR
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_CUDA)
@@ -29,7 +43,8 @@ typedef half2 ggml_half2;
 typedef half  ggml_half;
 typedef half2 ggml_half2;
 
-#define GGML_COMMON_AGGR data
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_HIP)
@@ -39,7 +54,8 @@ typedef half2 ggml_half2;
 typedef half  ggml_half;
 typedef half2 ggml_half2;
 
-#define GGML_COMMON_AGGR data
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #elif defined(GGML_COMMON_DECL_SYCL)
@@ -49,7 +65,8 @@ typedef half2 ggml_half2;
 typedef sycl::half  ggml_half;
 typedef sycl::half2 ggml_half2;
 
-#define GGML_COMMON_AGGR data
+#define GGML_COMMON_AGGR_U
+#define GGML_COMMON_AGGR_S data
 
 #define GGML_COMMON_DECL
 #endif
@@ -154,9 +171,9 @@ typedef struct {
         struct {
             ggml_half d; // delta
             ggml_half m; // min
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t qs[QK4_1 / 2]; // nibbles / quants
 } block_q4_1;
 static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
@@ -175,9 +192,9 @@ typedef struct {
         struct {
             ggml_half d; // delta
             ggml_half m; // min
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t qh[4];         // 5-th bit of quants
     uint8_t qs[QK5_1 / 2]; // nibbles / quants
 } block_q5_1;
@@ -196,37 +213,13 @@ typedef struct {
         struct {
             ggml_half d; // delta
             ggml_half s; // d * sum(qs[i])
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 ds;
-    };
+    } GGML_COMMON_AGGR_U;
     int8_t qs[QK8_1]; // quants
 } block_q8_1;
 static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding");
 
-typedef struct {
-    ggml_half d[4];        // deltas for 4 q4_0 blocks
-    uint8_t qs[QK4_0 * 2]; // nibbles / quants for 4 q4_0 blocks
-} block_q4_0x4;
-static_assert(sizeof(block_q4_0x4) == 4 * sizeof(ggml_half) + QK4_0 * 2, "wrong q4_0x4 block size/padding");
-
-typedef struct {
-    ggml_half d[8];        // deltas for 8 q4_0 blocks
-    uint8_t qs[QK4_0 * 4]; // nibbles / quants for 8 q4_0 blocks
-} block_q4_0x8;
-static_assert(sizeof(block_q4_0x8) == 8 * sizeof(ggml_half) + QK4_0 * 4, "wrong q4_0x8 block size/padding");
-
-typedef struct {
-    ggml_half d[4];        // deltas for 4 q8_0 blocks
-    int8_t qs[QK8_0 * 4];  // quants for 4 q8_0 blocks
-} block_q8_0x4;
-static_assert(sizeof(block_q8_0x4) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong q8_0x4 block size/padding");
-
-typedef struct {
-    ggml_half d[8];        // deltas for 8 q8_0 blocks
-    int8_t qs[QK8_0 * 8];  // quants for 8 q8_0 blocks
-} block_q8_0x8;
-static_assert(sizeof(block_q8_0x8) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong q8_0x8 block size/padding");
-
 //
 // Ternary quantization
 //
@@ -261,9 +254,9 @@ typedef struct {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
 } block_q2_K;
 static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
 
@@ -288,9 +281,9 @@ typedef struct {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
     uint8_t qs[QK_K/2];           // 4--bit quants
 } block_q4_K;
@@ -305,9 +298,9 @@ typedef struct {
         struct {
             ggml_half d;    // super-block scale for quantized scales
             ggml_half dmin; // super-block scale for quantized mins
-        } GGML_COMMON_AGGR;
+        } GGML_COMMON_AGGR_S;
         ggml_half2 dm;
-    };
+    } GGML_COMMON_AGGR_U;
     uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
     uint8_t qh[QK_K/8];           // quants, high bit
     uint8_t qs[QK_K/2];           // quants, low 4 bits
@@ -431,6 +424,13 @@ static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_
 #define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
 #define GGML_TABLE_END() };
 
+#define GGML_COMMON_IMPL
+#elif defined(GGML_COMMON_IMPL_CPP)
+#include <cstdint>
+
+#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
+#define GGML_TABLE_END() };
+
 #define GGML_COMMON_IMPL
 #elif defined(GGML_COMMON_IMPL_METAL)
 #include <metal_stdlib>
@@ -473,7 +473,7 @@ GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128)
     240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
 GGML_TABLE_END()
 
-//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+//#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A // lowest compute capability for integer intrinsics
 GGML_TABLE_BEGIN(uint64_t, ksigns64, 128)
     0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
     0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,
diff --git a/ggml/src/ggml-cpu/CMakeLists.txt b/ggml/src/ggml-cpu/CMakeLists.txt
new file mode 100644
index 00000000000..5d47323370a
--- /dev/null
+++ b/ggml/src/ggml-cpu/CMakeLists.txt
@@ -0,0 +1,358 @@
+function(ggml_add_cpu_backend_variant_impl tag_name)
+    if (tag_name)
+        set(GGML_CPU_NAME ggml-cpu-${tag_name})
+    else()
+        set(GGML_CPU_NAME ggml-cpu)
+    endif()
+
+    ggml_add_backend_library(${GGML_CPU_NAME})
+
+    list (APPEND GGML_CPU_SOURCES
+        ggml-cpu/ggml-cpu.c
+        ggml-cpu/ggml-cpu.cpp
+        ggml-cpu/ggml-cpu-aarch64.cpp
+        ggml-cpu/ggml-cpu-aarch64.h
+        ggml-cpu/ggml-cpu-hbm.cpp
+        ggml-cpu/ggml-cpu-hbm.h
+        ggml-cpu/ggml-cpu-quants.c
+        ggml-cpu/ggml-cpu-quants.h
+        ggml-cpu/ggml-cpu-traits.cpp
+        ggml-cpu/ggml-cpu-traits.h
+        ggml-cpu/amx/amx.cpp
+        ggml-cpu/amx/amx.h
+        ggml-cpu/amx/mmq.cpp
+        ggml-cpu/amx/mmq.h
+        ggml-cpu/ggml-cpu-impl.h
+        )
+
+    target_compile_features(${GGML_CPU_NAME} PRIVATE c_std_11 cxx_std_17)
+    target_include_directories(${GGML_CPU_NAME} PRIVATE . ggml-cpu)
+
+    if (APPLE AND GGML_ACCELERATE)
+        find_library(ACCELERATE_FRAMEWORK Accelerate)
+        if (ACCELERATE_FRAMEWORK)
+            message(STATUS "Accelerate framework found")
+
+            target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_ACCELERATE)
+            target_compile_definitions(${GGML_CPU_NAME} PRIVATE ACCELERATE_NEW_LAPACK)
+            target_compile_definitions(${GGML_CPU_NAME} PRIVATE ACCELERATE_LAPACK_ILP64)
+
+            target_link_libraries(${GGML_CPU_NAME} PRIVATE ${ACCELERATE_FRAMEWORK})
+        else()
+            message(WARNING "Accelerate framework not found")
+        endif()
+    endif()
+
+    if (GGML_OPENMP)
+        find_package(OpenMP)
+        if (OpenMP_FOUND)
+            target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_OPENMP)
+
+            target_link_libraries(${GGML_CPU_NAME} PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+        else()
+            message(WARNING "OpenMP not found")
+        endif()
+    endif()
+
+    if (GGML_LLAMAFILE)
+        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_LLAMAFILE)
+
+        list(APPEND GGML_CPU_SOURCES
+                    ggml-cpu/llamafile/sgemm.cpp
+                    ggml-cpu/llamafile/sgemm.h)
+    endif()
+
+    if (GGML_CPU_HBM)
+        find_library(memkind memkind REQUIRED)
+
+        message(STATUS "Using memkind for CPU HBM")
+
+        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_HBM)
+
+        target_link_libraries(${GGML_CPU_NAME} PUBLIC memkind)
+    endif()
+
+    if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
+        CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
+        (NOT CMAKE_OSX_ARCHITECTURES      AND
+        NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+            CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
+
+        message(STATUS "ARM detected")
+
+        if (MSVC)
+            list(APPEND ARCH_DEFINITIONS __aarch64__) # MSVC defines _M_ARM64 instead
+            list(APPEND ARCH_DEFINITIONS __ARM_NEON)
+            list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_FMA)
+
+            set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
+            string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
+
+            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
+            if (GGML_COMPILER_SUPPORT_DOTPROD)
+                list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_DOTPROD)
+
+                message(STATUS "ARM feature DOTPROD enabled")
+            endif ()
+
+            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
+
+            if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
+                list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_MATMUL_INT8)
+
+                message(STATUS "ARM feature MATMUL_INT8 enabled")
+            endif ()
+
+            check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+            if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+                list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+
+                message(STATUS "ARM feature FP16_VECTOR_ARITHMETIC enabled")
+            endif ()
+
+            set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
+        elseif (APPLE)
+            if (GGML_NATIVE)
+                set(USER_PROVIDED_MARCH FALSE)
+                foreach(flag_var IN ITEMS CMAKE_C_FLAGS CMAKE_CXX_FLAGS CMAKE_REQUIRED_FLAGS)
+                    if ("${${flag_var}}" MATCHES "-march=[a-zA-Z0-9+._-]+")
+                        set(USER_PROVIDED_MARCH TRUE)
+                        break()
+                    endif()
+                endforeach()
+
+                if (NOT USER_PROVIDED_MARCH)
+                    set(MARCH_FLAGS "-march=armv8.2a")
+
+                    check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
+                    if (GGML_COMPILER_SUPPORT_DOTPROD)
+                        set(MARCH_FLAGS "${MARCH_FLAGS}+dotprod")
+                        list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_DOTPROD)
+
+                        message(STATUS "ARM feature DOTPROD enabled")
+                    endif ()
+
+                    set(TEST_I8MM_FLAGS "-march=armv8.2a+i8mm")
+
+                    set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
+                    set(CMAKE_REQUIRED_FLAGS     "${CMAKE_REQUIRED_FLAGS} ${TEST_I8MM_FLAGS}")
+
+                    check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmmlaq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
+                    if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
+                        set(MARCH_FLAGS "${MARCH_FLAGS}+i8mm")
+                        list(APPEND ARCH_DEFINITIONS __ARM_FEATURE_MATMUL_INT8)
+
+                        message(STATUS "ARM feature MATMUL_INT8 enabled")
+                    endif ()
+
+                    set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
+
+                    list(APPEND ARCH_FLAGS "${MARCH_FLAGS}")
+                endif ()
+            endif ()
+        else()
+            check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
+            if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
+                list(APPEND ARCH_FLAGS -mfp16-format=ieee)
+            endif()
+            if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
+                # Raspberry Pi 1, Zero
+                list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
+            endif()
+            if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
+                if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
+                    # Android armeabi-v7a
+                    list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
+                else()
+                    # Raspberry Pi 2
+                    list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+                endif()
+            endif()
+            if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
+                # Android arm64-v8a
+                # Raspberry Pi 3, 4, Zero 2 (32-bit)
+                list(APPEND ARCH_FLAGS -mno-unaligned-access)
+            endif()
+            if (GGML_SVE)
+                list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
+            endif()
+        endif()
+    elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
+            (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+            CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
+        if (MSVC)
+            # instruction set detection for MSVC only
+            if (GGML_NATIVE)
+                include(ggml-cpu/cmake/FindSIMD.cmake)
+            endif ()
+            if (GGML_AVX512)
+                list(APPEND ARCH_FLAGS /arch:AVX512)
+                # /arch:AVX512 includes: __AVX512F__, __AVX512CD__, __AVX512BW__, __AVX512DQ__, and __AVX512VL__
+                # MSVC has no compile-time flags enabling specific
+                # AVX512 extensions, neither it defines the
+                # macros corresponding to the extensions.
+                # Do it manually.
+                list(APPEND ARCH_DEFINITIONS GGML_AVX512)
+                if (GGML_AVX512_VBMI)
+                    list(APPEND ARCH_DEFINITIONS __AVX512VBMI__)
+                    if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+                        list(APPEND ARCH_FLAGS -mavx512vbmi)
+                    endif()
+                endif()
+                if (GGML_AVX512_VNNI)
+                    list(APPEND ARCH_DEFINITIONS __AVX512VNNI__ GGML_AVX512_VNNI)
+                    if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+                        list(APPEND ARCH_FLAGS -mavx512vnni)
+                    endif()
+                endif()
+                if (GGML_AVX512_BF16)
+                    list(APPEND ARCH_DEFINITIONS __AVX512BF16__ GGML_AVX512_BF16)
+                    if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+                        list(APPEND ARCH_FLAGS -mavx512bf16)
+                    endif()
+                endif()
+                if (GGML_AMX_TILE)
+                    list(APPEND ARCH_DEFINITIONS __AMX_TILE__ GGML_AMX_TILE)
+                endif()
+                if (GGML_AMX_INT8)
+                    list(APPEND ARCH_DEFINITIONS __AMX_INT8__ GGML_AMX_INT8)
+                endif()
+                if (GGML_AMX_BF16)
+                    list(APPEND ARCH_DEFINITIONS __AMX_BF16__ GGML_AMX_BF16)
+                endif()
+            elseif (GGML_AVX2)
+                list(APPEND ARCH_FLAGS /arch:AVX2)
+                list(APPEND ARCH_DEFINITIONS GGML_AVX2 GGML_FMA GGML_F16C)
+            elseif (GGML_AVX)
+                list(APPEND ARCH_FLAGS /arch:AVX)
+                list(APPEND ARCH_DEFINITIONS GGML_AVX)
+            else ()
+                list(APPEND ARCH_FLAGS /arch:SSE4.2)
+                list(APPEND ARCH_DEFINITIONS GGML_SSE42)
+            endif()
+            if (GGML_AVX_VNNI)
+                # MSVC generates AVX512 with AVX-VNNI intrinsics even with /arch:AVX2
+                #list(APPEND ARCH_DEFINITIONS __AVXVNNI__ GGML_AVX_VNNI)
+            endif()
+        else ()
+            if (GGML_NATIVE)
+                list(APPEND ARCH_FLAGS -march=native)
+            else ()
+                list(APPEND ARCH_FLAGS -msse4.2)
+                list(APPEND ARCH_DEFINITIONS GGML_SSE42)
+                if (GGML_F16C)
+                    list(APPEND ARCH_FLAGS -mf16c)
+                    list(APPEND ARCH_DEFINITIONS GGML_F16C)
+                endif()
+                if (GGML_FMA)
+                    list(APPEND ARCH_FLAGS -mfma)
+                    list(APPEND ARCH_DEFINITIONS GGML_FMA)
+                endif()
+                if (GGML_AVX)
+                    list(APPEND ARCH_FLAGS -mavx)
+                    list(APPEND ARCH_DEFINITIONS GGML_AVX)
+                endif()
+                if (GGML_AVX2)
+                    list(APPEND ARCH_FLAGS -mavx2)
+                    list(APPEND ARCH_DEFINITIONS GGML_AVX2)
+                endif()
+                if (GGML_AVX_VNNI)
+                    list(APPEND ARCH_FLAGS -mavxvnni)
+                    list(APPEND ARCH_DEFINITIONS GGML_AVX_VNNI)
+                endif()
+                if (GGML_AVX512)
+                    list(APPEND ARCH_FLAGS -mavx512f)
+                    list(APPEND ARCH_FLAGS -mavx512cd)
+                    list(APPEND ARCH_FLAGS -mavx512vl)
+                    list(APPEND ARCH_FLAGS -mavx512dq)
+                    list(APPEND ARCH_FLAGS -mavx512bw)
+                    list(APPEND ARCH_DEFINITIONS GGML_AVX512)
+                endif()
+                if (GGML_AVX512_VBMI)
+                    list(APPEND ARCH_FLAGS -mavx512vbmi)
+                    list(APPEND ARCH_DEFINITIONS GGML_AVX512_VBMI)
+                endif()
+                if (GGML_AVX512_VNNI)
+                    list(APPEND ARCH_FLAGS -mavx512vnni)
+                    list(APPEND ARCH_DEFINITIONS GGML_AVX512_VNNI)
+                endif()
+                if (GGML_AVX512_BF16)
+                    list(APPEND ARCH_FLAGS -mavx512bf16)
+                    list(APPEND ARCH_DEFINITIONS GGML_AVX512_BF16)
+                endif()
+                if (GGML_AMX_TILE)
+                    list(APPEND ARCH_FLAGS -mamx-tile)
+                    list(APPEND ARCH_DEFINITIONS GGML_AMX_TILE)
+                endif()
+                if (GGML_AMX_INT8)
+                    list(APPEND ARCH_FLAGS -mamx-int8)
+                    list(APPEND ARCH_DEFINITIONS GGML_AMX_INT8)
+                endif()
+                if (GGML_AMX_BF16)
+                    list(APPEND ARCH_FLAGS -mamx-bf16)
+                    list(APPEND ARCH_DEFINITIONS GGML_AMX_BF16)
+                endif()
+            endif()
+        endif()
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
+        message(STATUS "PowerPC detected")
+        execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1" OUTPUT_VARIABLE POWER10_M)
+        string(FIND "${POWER10_M}" "POWER10" substring_index)
+        if (NOT DEFINED substring_index OR "${substring_index}" STREQUAL "")
+            set(substring_index -1)
+        endif()
+
+        if (${substring_index} GREATER_EQUAL 0)
+        list(APPEND ARCH_FLAGS -mcpu=power10)
+        elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+        list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
+        else()
+            list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
+            # TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
+        endif()
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
+        message(STATUS "loongarch64 detected")
+
+        list(APPEND ARCH_FLAGS -march=loongarch64)
+        if (GGML_LASX)
+            list(APPEND ARCH_FLAGS -mlasx)
+        endif()
+        if (GGML_LSX)
+            list(APPEND ARCH_FLAGS -mlsx)
+        endif()
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64")
+        message(STATUS "RISC-V detected")
+        if (GGML_RVV)
+            list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
+        endif()
+    else()
+        message(STATUS "Unknown architecture")
+    endif()
+
+    if (GGML_CPU_AARCH64)
+        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_AARCH64)
+    endif()
+
+    message(STATUS "Adding CPU backend variant ${GGML_CPU_NAME}: ${ARCH_FLAGS} ${ARCH_DEFINITIONS}")
+    target_sources(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_SOURCES})
+    target_compile_options(${GGML_CPU_NAME} PRIVATE ${ARCH_FLAGS})
+    target_compile_definitions(${GGML_CPU_NAME} PRIVATE ${ARCH_DEFINITIONS})
+
+    if (GGML_BACKEND_DL)
+        # The feature detection code is compiled as a separate target so that
+        # it can be built without the architecture flags
+        # Since multiple variants of the CPU backend may be included in the same
+        # build, using set_source_files_properties() to set the arch flags is not possible
+        set(GGML_CPU_FEATS_NAME ${GGML_CPU_NAME}-feats)
+        add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/cpu-feats-x86.cpp)
+        target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
+        target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARCH_DEFINITIONS})
+        target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
+        set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        target_link_libraries(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_FEATS_NAME})
+    endif()
+
+    if (EMSCRIPTEN)
+        set_target_properties(${GGML_CPU_NAME} PROPERTIES COMPILE_FLAGS "-msimd128")
+    endif()
+endfunction()
diff --git a/ggml/src/ggml-cpu/amx/amx.cpp b/ggml/src/ggml-cpu/amx/amx.cpp
new file mode 100644
index 00000000000..5ec5263ceb4
--- /dev/null
+++ b/ggml/src/ggml-cpu/amx/amx.cpp
@@ -0,0 +1,220 @@
+#include "amx.h"
+#include "common.h"
+#include "mmq.h"
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-cpu-traits.h"
+
+#if defined(__gnu_linux__)
+#include <sys/syscall.h>
+#include <unistd.h>
+#endif
+
+#include <cstdlib>
+#include <cstring>
+#include <memory>
+
+#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
+
+// AMX type_trais
+namespace ggml::cpu::amx {
+class tensor_traits : public ggml::cpu::tensor_traits {
+    bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
+        size = ggml_backend_amx_desired_wsize(op);
+        return true;
+    }
+
+    bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) override {
+        if (op->op == GGML_OP_MUL_MAT) {
+            ggml_backend_amx_mul_mat(params, op);
+            return true;
+        }
+        return false;
+    }
+};
+
+static ggml::cpu::tensor_traits * get_tensor_traits(ggml_backend_buffer_t, struct ggml_tensor *) {
+    static tensor_traits traits;
+    return &traits;
+}
+}  // namespace ggml::cpu::amx
+
+// AMX buffer interface
+static void ggml_backend_amx_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+}
+
+static void * ggml_backend_amx_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return (void *) (buffer->context);
+}
+
+static void ggml_backend_amx_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    tensor->extra = (void *) ggml::cpu::amx::get_tensor_traits(buffer, tensor);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
+                                                  uint8_t value, size_t offset, size_t size) {
+    memset((char *) tensor->data + offset, value, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_amx_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
+                                               const void * data, size_t offset, size_t size) {
+    if (qtype_has_amx_kernels(tensor->type)) {
+        GGML_LOG_DEBUG("%s: amx repack tensor %s of type %s\n", __func__, tensor->name, ggml_type_name(tensor->type));
+        ggml_backend_amx_convert_weight(tensor, data, offset, size);
+    } else {
+        memcpy((char *) tensor->data + offset, data, size);
+    }
+
+    GGML_UNUSED(buffer);
+}
+
+/*
+// need to figure what we need to do with buffer->extra.
+static void ggml_backend_amx_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(!qtype_has_amx_kernels(tensor->type));
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static bool ggml_backend_amx_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    if (ggml_backend_buffer_is_host(src->buffer)) {
+        if (qtype_has_amx_kernels(src->type)) {
+            ggml_backend_amx_convert_weight(dst, src->data, 0, ggml_nbytes(dst));
+        } else {
+            memcpy(dst->data, src->data, ggml_nbytes(src));
+        }
+        return true;
+    }
+    return false;
+
+    GGML_UNUSED(buffer);
+}
+*/
+
+static void ggml_backend_amx_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    memset(buffer->context, value, buffer->size);
+}
+
+static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
+    /* .free_buffer     = */ ggml_backend_amx_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_amx_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_amx_buffer_init_tensor,
+    /* .memset_tensor   = */ ggml_backend_amx_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_amx_buffer_set_tensor,
+    /* .get_tensor      = */ nullptr,
+    /* .cpy_tensor      = */ nullptr,
+    /* .clear           = */ ggml_backend_amx_buffer_clear,
+    /* .reset           = */ nullptr,
+};
+
+static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "AMX";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_amx_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * data = ggml_aligned_malloc(size);
+    if (data == NULL) {
+        fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
+        return NULL;
+    }
+
+    return ggml_backend_buffer_init(buft, ggml_backend_amx_buffer_interface, data, size);
+}
+
+static size_t ggml_backend_amx_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+namespace ggml::cpu::amx {
+class extra_buffer_type : ggml::cpu::extra_buffer_type {
+    bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
+        // handle only 2d gemm for now
+        auto is_contiguous_2d = [](const struct ggml_tensor * t) {
+            return ggml_is_contiguous(t) && t->ne[3] == 1 && t->ne[2] == 1;
+        };
+
+        if (op->op == GGML_OP_MUL_MAT && is_contiguous_2d(op->src[0]) &&  // src0 must be contiguous
+            is_contiguous_2d(op->src[1]) &&                               // src1 must be contiguous
+            op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_amx_buffer_type() &&
+            op->ne[0] % (TILE_N * 2) == 0 &&                              // out_features is 32x
+            (qtype_has_amx_kernels(op->src[0]->type) || (op->src[0]->type == GGML_TYPE_F16))) {
+            // src1 must be host buffer
+            if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
+                return false;
+            }
+            // src1 must be float32
+            if (op->src[1]->type == GGML_TYPE_F32) {
+                return true;
+            }
+        }
+        return false;
+    }
+
+    ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
+        if (op->op == GGML_OP_MUL_MAT && op->src[0]->buffer &&
+            op->src[0]->buffer->buft == ggml_backend_amx_buffer_type()) {
+            return (ggml::cpu::tensor_traits *) op->src[0]->extra;
+        }
+
+        return nullptr;
+    }
+};
+}  // namespace ggml::cpu::amx
+
+static size_t ggml_backend_amx_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+    return ggml_backend_amx_get_alloc_size(tensor);
+
+    GGML_UNUSED(buft);
+}
+
+#define ARCH_GET_XCOMP_PERM     0x1022
+#define ARCH_REQ_XCOMP_PERM     0x1023
+#define XFEATURE_XTILECFG       17
+#define XFEATURE_XTILEDATA      18
+
+static bool ggml_amx_init() {
+#if defined(__gnu_linux__)
+    if (syscall(SYS_arch_prctl, ARCH_REQ_XCOMP_PERM, XFEATURE_XTILEDATA)) {
+        fprintf(stderr, "AMX is not ready to be used!\n");
+        return false;
+    }
+    return true;
+#elif defined(_WIN32)
+    return true;
+#endif
+}
+
+ggml_backend_buffer_type_t ggml_backend_amx_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_buffer_type_amx = {
+        /* .iface = */ {
+                        /* .get_name         = */ ggml_backend_amx_buffer_type_get_name,
+                        /* .alloc_buffer     = */ ggml_backend_amx_buffer_type_alloc_buffer,
+                        /* .get_alignment    = */ ggml_backend_amx_buffer_type_get_alignment,
+                        /* .get_max_size     = */ nullptr,  // defaults to SIZE_MAX
+                        /* .get_alloc_size   = */ ggml_backend_amx_buffer_type_get_alloc_size,
+                        /* .is_host          = */ nullptr,
+                        },
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context = */ new ggml::cpu::amx::extra_buffer_type(),
+    };
+
+    if (!ggml_amx_init()) {
+        return nullptr;
+    }
+
+    return &ggml_backend_buffer_type_amx;
+}
+
+#endif  // defined(__AMX_INT8__) && defined(__AVX512VNNI__)
diff --git a/ggml/src/ggml-cpu/amx/amx.h b/ggml/src/ggml-cpu/amx/amx.h
new file mode 100644
index 00000000000..5b65d76bdc8
--- /dev/null
+++ b/ggml/src/ggml-cpu/amx/amx.h
@@ -0,0 +1,8 @@
+#include "ggml-backend.h"
+#include "ggml-cpu-impl.h"
+
+// GGML internal header
+
+#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
+ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
+#endif
diff --git a/ggml/src/ggml-cpu/amx/common.h b/ggml/src/ggml-cpu/amx/common.h
new file mode 100644
index 00000000000..f392e898518
--- /dev/null
+++ b/ggml/src/ggml-cpu/amx/common.h
@@ -0,0 +1,91 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-cpu-impl.h"
+
+#include <algorithm>
+#include <memory>
+#include <type_traits>
+
+#if defined(GGML_USE_OPENMP)
+#include <omp.h>
+#endif
+
+#define TILE_M 16
+#define TILE_N 16
+#define TILE_K 32
+#define VNNI_BLK 4
+
+#define AMX_BLK_SIZE 32
+
+#define TMM0 0
+#define TMM1 1
+#define TMM2 2
+#define TMM3 3
+#define TMM4 4
+#define TMM5 5
+#define TMM6 6
+#define TMM7 7
+
+// parallel routines
+template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
+inline T div_up(T x, T y) { return (x + y - 1) / y; }
+
+template <typename T>
+inline void balance211(T n, T nth, T ith, T& n_start, T& n_end) {
+#if 0
+    // onednn partition pattern
+    T& n_my = n_end;
+    if (nth <= 1 || n == 0) {
+        n_start = 0;
+        n_my = n;
+    } else {
+        T n1 = div_up(n, nth);
+        T n2 = n1 - 1;
+        T T1 = n - n2 * nth;
+        n_my = ith < T1 ? n1 : n2;
+        n_start = ith <= T1 ? ith*n1 : T1 * n1 + (ith - T1) * n2;
+    }
+    n_end += n_start;
+#else
+    // pytorch aten partition pattern
+    T n_my = div_up(n, nth);
+    n_start = ith * n_my;
+    n_end = std::min(n_start + n_my, n);
+#endif
+}
+
+template <typename func_t>
+inline void parallel_for(int n, const func_t& f) {
+#if defined(GGML_USE_OPENMP)
+#pragma omp parallel
+{
+    int nth = omp_get_num_threads();
+    int ith = omp_get_thread_num();
+    int tbegin, tend;
+    balance211(n, nth, ith, tbegin, tend);
+    f(tbegin, tend);
+}
+#else
+    f(0, n);
+#endif
+}
+
+template <typename func_t>
+inline void parallel_for_ggml(const ggml_compute_params * params, int n, const func_t & f) {
+    int tbegin, tend;
+    balance211(n, params->nth, params->ith, tbegin, tend);
+    f(tbegin, tend);
+}
+
+// quantized types that have AMX support
+inline bool qtype_has_amx_kernels(const enum ggml_type type) {
+    // TODO: fix padding for vnni format
+    return (type == GGML_TYPE_Q4_0) ||
+        (type == GGML_TYPE_Q4_1) ||
+        (type == GGML_TYPE_Q8_0) ||
+        (type == GGML_TYPE_Q4_K) ||
+        (type == GGML_TYPE_Q5_K) ||
+        (type == GGML_TYPE_Q6_K) ||
+        (type == GGML_TYPE_IQ4_XS);
+}
diff --git a/ggml/src/ggml-cpu/amx/mmq.cpp b/ggml/src/ggml-cpu/amx/mmq.cpp
new file mode 100644
index 00000000000..0ea91596bc7
--- /dev/null
+++ b/ggml/src/ggml-cpu/amx/mmq.cpp
@@ -0,0 +1,2511 @@
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Wpedantic"
+#pragma GCC diagnostic ignored "-Wunused-local-typedefs"
+#endif
+
+#include "amx.h"
+#include "mmq.h"
+#include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
+#include "ggml-cpu-quants.h"
+#include "ggml-quants.h"
+#include <algorithm>
+#include <type_traits>
+
+#if defined(__gnu_linux__)
+#include <sys/syscall.h>
+#include <unistd.h>
+#endif
+
+#if (defined(_WIN32) || defined(_WIN64))
+#define RESTRICT __restrict
+#else
+#define RESTRICT __restrict__
+#endif
+
+#if (defined(_WIN32) || defined(_WIN64))
+#define ALWAYS_INLINE __forceinline
+#elif __has_attribute(always_inline) || defined(__GNUC__)
+#define ALWAYS_INLINE __attribute__((__always_inline__)) inline
+#else
+#define ALWAYS_INLINE inline
+#endif
+
+#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
+
+namespace {
+
+// Forced unrolling
+template <int n>
+struct Unroll {
+    template <typename Func, typename... Args>
+    ALWAYS_INLINE void operator()(const Func& f, Args... args) const {
+        Unroll<n - 1>{}(f, args...);
+        f(std::integral_constant<int, n - 1>{}, args...);
+    }
+};
+
+template <>
+struct Unroll<1> {
+    template <typename Func, typename... Args>
+    ALWAYS_INLINE void operator()(const Func& f, Args... args) const {
+        f(std::integral_constant<int, 0>{}, args...);
+    }
+};
+
+// type traits
+template <typename T> struct PackedTypes {};
+template <> struct PackedTypes<block_q4_0> { using type = int8_t; };
+template <> struct PackedTypes<block_q4_1> { using type = uint8_t; };
+template <> struct PackedTypes<block_q8_0> { using type = int8_t; };
+template <typename T> using packed_B_type = typename PackedTypes<T>::type;
+
+template <typename T>
+struct do_compensate : std::integral_constant<bool,
+    std::is_same<T, block_q8_0>::value> {};
+
+template <typename T>
+struct do_unpack : std::integral_constant<bool,
+    std::is_same<T, block_q4_0>::value ||
+    std::is_same<T, block_q4_1>::value> {};
+
+template <typename T>
+struct is_type_qkk : std::integral_constant<bool,
+    std::is_same<T, block_q4_K>::value ||
+    std::is_same<T, block_q5_K>::value ||
+    std::is_same<T, block_q6_K>::value ||
+    std::is_same<T, block_iq4_xs>::value> {};
+
+#define GGML_DISPATCH_FLOATING_TYPES(TYPE, ...)                                        \
+    [&] {                                                                              \
+        switch (TYPE) {                                                                \
+            case GGML_TYPE_F16: {                                                      \
+                using type = ggml_fp16_t;                                              \
+                constexpr int blck_size = 16;                                          \
+                return __VA_ARGS__();                                                  \
+            }                                                                          \
+            case GGML_TYPE_BF16: {                                                     \
+                using type = ggml_bf16_t;                                              \
+                constexpr int blck_size = 32;                                          \
+                return __VA_ARGS__();                                                  \
+            }                                                                          \
+            default:                                                                   \
+                fprintf(stderr, "Unsupported floating data type\n");                   \
+        }                                                                              \
+    }()
+
+#define GGML_DISPATCH_QTYPES(QT, ...)                                                  \
+    [&] {                                                                              \
+        switch (QT) {                                                                  \
+            case GGML_TYPE_Q4_0: {                                                     \
+                using type = block_q4_0;                                               \
+                using vec_dot_type = block_q8_0;                                       \
+                constexpr int blck_size = QK4_0;                                       \
+                return __VA_ARGS__();                                                  \
+            }                                                                          \
+            case GGML_TYPE_Q4_1: {                                                     \
+                using type = block_q4_1;                                               \
+                using vec_dot_type = block_q8_1;                                       \
+                constexpr int blck_size = QK4_1;                                       \
+                return __VA_ARGS__();                                                  \
+            }                                                                          \
+            case GGML_TYPE_Q8_0: {                                                     \
+                using type = block_q8_0;                                               \
+                using vec_dot_type = block_q8_0;                                       \
+                constexpr int blck_size = QK8_0;                                       \
+                return __VA_ARGS__();                                                  \
+            }                                                                          \
+            case GGML_TYPE_Q4_K: {                                                     \
+                using type = block_q4_K;                                               \
+                using vec_dot_type = block_q8_K;                                       \
+                constexpr int blck_size = QK_K;                                        \
+                return __VA_ARGS__();                                                  \
+            }                                                                          \
+            case GGML_TYPE_Q5_K: {                                                     \
+                using type = block_q5_K;                                               \
+                using vec_dot_type = block_q8_K;                                       \
+                constexpr int blck_size = QK_K;                                        \
+                return __VA_ARGS__();                                                  \
+            }                                                                          \
+            case GGML_TYPE_Q6_K: {                                                     \
+                using type = block_q6_K;                                               \
+                using vec_dot_type = block_q8_K;                                       \
+                constexpr int blck_size = QK_K;                                        \
+                return __VA_ARGS__();                                                  \
+            }                                                                          \
+            case GGML_TYPE_IQ4_XS: {                                                   \
+                using type = block_iq4_xs;                                             \
+                using vec_dot_type = block_q8_K;                                       \
+                constexpr int blck_size = QK_K;                                        \
+                return __VA_ARGS__();                                                  \
+            }                                                                          \
+            default:                                                                   \
+                fprintf(stderr, "Unsupported quantized data type: %d\n", int(TYPE));   \
+        }                                                                              \
+    }()
+
+#define GGML_DISPATCH_BOOL(BOOL_V, BOOL_NAME, ...)                                     \
+    [&] {                                                                              \
+        if (BOOL_V) {                                                                  \
+            constexpr bool BOOL_NAME = true;                                           \
+            return __VA_ARGS__();                                                      \
+        } else {                                                                       \
+            constexpr bool BOOL_NAME = false;                                          \
+            return __VA_ARGS__();                                                      \
+        }                                                                              \
+    }()
+
+// define amx tile config data structure
+struct tile_config_t{
+    uint8_t palette_id = 0;
+    uint8_t start_row = 0;
+    uint8_t reserved_0[14] = {0};
+    uint16_t colsb[16] = {0};
+    uint8_t rows[16] = {0};
+};
+
+// Notes: amx tile config
+//
+// Typically, TMUL calculates A and B of size 16 x 64 containing INT8 values,
+// and accumulate the result to a 16 x 16 matrix C containing INT32 values,
+//
+// As many GGUF quantized types as `block_size` of 32, so a 16-16-32 config is used
+// instead of the normally used 16-16-64 config.
+//
+//    Block A: {16, 32}, dtype = int8_t
+//    Block B: {16, 32}, dtype = uint8_t/int8_t
+//    Block C: {16, 16}, dtype = int32_t
+//
+// Block B needs to be prepacked to vnni format before feeding into  TMUL:
+//    packed_B: from {n, k} to {k/vnni_blk, n, vnni_blck}, viewed in 2d, we get {8, 64}
+//
+// Therefore, we get tileconfig:
+//             A    B    C
+//    rows    16    8   16
+//    colsb   32   64   16
+//
+// For tile distribution, follow a 2-2-4 pattern, e.g. A used TMM2-TMM3, B used TMM0-TMM1,
+// C used TMM4-TMM7:
+//            B TMM0  B TMM1
+//    A TMM2  C TMM4  C TMM6
+//    A TMM3  C TMM5  C TMM7
+//
+// Each `amx` kernel handles 4 blocks at a time: 2MB * 2NB, when m < 2 * BLOCK_M, unpack A
+// will be needed.
+//
+// Here another commonly used pattern 1-3-3 is skipped, as it is mostly used when m <=16;
+// and the sinlge batch gemm (m=1) has a special fast path with `avx512-vnni`.
+//
+// ref: https://www.intel.com/content/www/us/en/developer/articles/code-sample/
+//    advanced-matrix-extensions-intrinsics-functions.html
+//
+
+#define TC_CONFIG_TILE(i, r, cb) tc.rows[i] = r; tc.colsb[i] = cb
+void ggml_tile_config_init(void) {
+    static thread_local bool is_first_time = true;
+
+    if (!is_first_time) {
+        return;
+    }
+
+    static thread_local tile_config_t tc;
+    tile_config_t current_tc;
+    _tile_storeconfig(&current_tc);
+
+    // load only when config changes
+    if (tc.palette_id == 0 || (memcmp(&current_tc.colsb, &tc.colsb, sizeof(uint16_t) * 8) != 0 &&
+                               memcmp(&current_tc.rows, &tc.rows, sizeof(uint8_t) * 8) != 0)) {
+        tc.palette_id = 1;
+        tc.start_row = 0;
+        TC_CONFIG_TILE(TMM0, 8, 64);
+        TC_CONFIG_TILE(TMM1, 8, 64);
+        TC_CONFIG_TILE(TMM2, 16, 32);
+        TC_CONFIG_TILE(TMM3, 16, 32);
+        TC_CONFIG_TILE(TMM4, 16, 64);
+        TC_CONFIG_TILE(TMM5, 16, 64);
+        TC_CONFIG_TILE(TMM6, 16, 64);
+        TC_CONFIG_TILE(TMM7, 16, 64);
+        _tile_loadconfig(&tc);
+    }
+
+    is_first_time = false;
+}
+
+// we need an extra 16 * 4B (TILE_N * int32_t) for each NB/KB block for compensation.
+// See the notes `s8s8 igemm compensation in avx512-vnni` for detail.
+template <typename TB>
+int get_tile_size() {
+    int tile_size = TILE_N * sizeof(TB);
+    if (do_compensate<TB>::value) {
+        tile_size += TILE_N * sizeof(int32_t);
+    }
+    if (std::is_same<TB, block_q4_K>::value ||
+        std::is_same<TB, block_q5_K>::value) {
+        tile_size += TILE_N * 4;
+    }
+    if (std::is_same<TB, block_iq4_xs>::value) {
+        tile_size += TILE_N * 2;
+    }
+    return tile_size;
+}
+
+template <typename TB, int BLOCK_K>
+int get_row_size(int K) {
+    int KB = K / BLOCK_K;
+    int row_size = KB * sizeof(TB);
+    if (do_compensate<TB>::value) {
+        row_size += KB * sizeof(int32_t);
+    }
+    if (std::is_same<TB, block_q4_K>::value ||
+        std::is_same<TB, block_q5_K>::value) {
+        row_size += KB * 4;
+    }
+    if (std::is_same<TB, block_iq4_xs>::value) {
+        row_size += KB * 2;
+    }
+    return row_size;
+}
+
+// vectorized dtype conversion
+inline float FP16_TO_FP32(ggml_half val) {
+    __m256i v = _mm256_setr_epi16(
+        val, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
+    __m512 o = _mm512_cvtph_ps(v);
+    return _mm512_cvtss_f32(o);
+}
+
+inline __m512 FP16_TO_FP32_VEC(ggml_half val) {
+    __m256i v = _mm256_set1_epi16(val);
+    return _mm512_cvtph_ps(v);
+}
+
+// horizontal reduce
+inline float _mm512_reduce_max_ps(const __m512 x) {
+    __m512 v = x;
+    __m512 v1 = _mm512_shuffle_f32x4(v, v, 0x4E);
+    v = _mm512_max_ps(v, v1);
+    v1 = _mm512_shuffle_f32x4(v, v, 0xB1);
+    v = _mm512_max_ps(v, v1);
+    v1 = _mm512_shuffle_ps(v, v, 0x4E);
+    v = _mm512_max_ps(v, v1);
+    v1 = _mm512_shuffle_ps(v, v, 0xB1);
+    v = _mm512_max_ps(v, v1);
+    return _mm512_cvtss_f32(v);
+}
+
+// transpose utils
+#define SHUFFLE_EPI32(a, b, mask) \
+    _mm256_castps_si256(_mm256_shuffle_ps(_mm256_castsi256_ps(a), _mm256_castsi256_ps(b), mask))
+inline void transpose_8x8_32bit(__m256i * v, __m256i * v1) {
+    // unpacking and 32-bit elements
+    v1[0] = _mm256_unpacklo_epi32(v[0], v[1]);
+    v1[1] = _mm256_unpackhi_epi32(v[0], v[1]);
+    v1[2] = _mm256_unpacklo_epi32(v[2], v[3]);
+    v1[3] = _mm256_unpackhi_epi32(v[2], v[3]);
+    v1[4] = _mm256_unpacklo_epi32(v[4], v[5]);
+    v1[5] = _mm256_unpackhi_epi32(v[4], v[5]);
+    v1[6] = _mm256_unpacklo_epi32(v[6], v[7]);
+    v1[7] = _mm256_unpackhi_epi32(v[6], v[7]);
+
+    // shuffling the 32-bit elements
+    v[0] = SHUFFLE_EPI32(v1[0], v1[2], 0x44);
+    v[1] = SHUFFLE_EPI32(v1[0], v1[2], 0xee);
+    v[2] = SHUFFLE_EPI32(v1[4], v1[6], 0x44);
+    v[3] = SHUFFLE_EPI32(v1[4], v1[6], 0xee);
+    v[4] = SHUFFLE_EPI32(v1[1], v1[3], 0x44);
+    v[5] = SHUFFLE_EPI32(v1[1], v1[3], 0xee);
+    v[6] = SHUFFLE_EPI32(v1[5], v1[7], 0x44);
+    v[7] = SHUFFLE_EPI32(v1[5], v1[7], 0xee);
+
+    // shuffling 128-bit elements
+    v1[0] = _mm256_permute2f128_si256(v[2], v[0], 0x02);
+    v1[1] = _mm256_permute2f128_si256(v[3], v[1], 0x02);
+    v1[2] = _mm256_permute2f128_si256(v[6], v[4], 0x02);
+    v1[3] = _mm256_permute2f128_si256(v[7], v[5], 0x02);
+    v1[4] = _mm256_permute2f128_si256(v[2], v[0], 0x13);
+    v1[5] = _mm256_permute2f128_si256(v[3], v[1], 0x13);
+    v1[6] = _mm256_permute2f128_si256(v[6], v[4], 0x13);
+    v1[7] = _mm256_permute2f128_si256(v[7], v[5], 0x13);
+}
+
+inline void transpose_16x4_32bit(__m512i * r, __m512i * d) {
+
+    static const __m512i index1 = _mm512_set_epi32(
+        0x0f, 0x0b, 0x07, 0x03,
+        0x0e, 0x0a, 0x06, 0x02,
+        0x0d, 0x09, 0x05, 0x01,
+        0x0c, 0x08, 0x04, 0x00);
+
+    d[0] = _mm512_permutexvar_epi32(index1, r[0]);
+    d[1] = _mm512_permutexvar_epi32(index1, r[1]);
+    d[2] = _mm512_permutexvar_epi32(index1, r[2]);
+    d[3] = _mm512_permutexvar_epi32(index1, r[3]);
+
+    r[0] = _mm512_shuffle_i32x4(d[0], d[1], 0x44);
+    r[1] = _mm512_shuffle_i32x4(d[0], d[1], 0xee);
+    r[2] = _mm512_shuffle_i32x4(d[2], d[3], 0x44);
+    r[3] = _mm512_shuffle_i32x4(d[2], d[3], 0xee);
+
+    d[0] = _mm512_shuffle_i32x4(r[0], r[2], 0x88);
+    d[1] = _mm512_shuffle_i32x4(r[0], r[2], 0xdd);
+    d[2] = _mm512_shuffle_i32x4(r[1], r[3], 0x88);
+    d[3] = _mm512_shuffle_i32x4(r[1], r[3], 0xdd);
+}
+
+inline void transpose_16x16_32bit(__m512i * v) {
+    __m512i v1[16];
+    v1[0] = _mm512_unpacklo_epi32(v[0], v[1]);
+    v1[1] = _mm512_unpackhi_epi32(v[0], v[1]);
+    v1[2] = _mm512_unpacklo_epi32(v[2], v[3]);
+    v1[3] = _mm512_unpackhi_epi32(v[2], v[3]);
+    v1[4] = _mm512_unpacklo_epi32(v[4], v[5]);
+    v1[5] = _mm512_unpackhi_epi32(v[4], v[5]);
+    v1[6] = _mm512_unpacklo_epi32(v[6], v[7]);
+    v1[7] = _mm512_unpackhi_epi32(v[6], v[7]);
+    v1[8] = _mm512_unpacklo_epi32(v[8], v[9]);
+    v1[9] = _mm512_unpackhi_epi32(v[8], v[9]);
+    v1[10] = _mm512_unpacklo_epi32(v[10], v[11]);
+    v1[11] = _mm512_unpackhi_epi32(v[10], v[11]);
+    v1[12] = _mm512_unpacklo_epi32(v[12], v[13]);
+    v1[13] = _mm512_unpackhi_epi32(v[12], v[13]);
+    v1[14] = _mm512_unpacklo_epi32(v[14], v[15]);
+    v1[15] = _mm512_unpackhi_epi32(v[14], v[15]);
+
+    v[0] = _mm512_unpacklo_epi64(v1[0], v1[2]);
+    v[1] = _mm512_unpackhi_epi64(v1[0], v1[2]);
+    v[2] = _mm512_unpacklo_epi64(v1[1], v1[3]);
+    v[3] = _mm512_unpackhi_epi64(v1[1], v1[3]);
+    v[4] = _mm512_unpacklo_epi64(v1[4], v1[6]);
+    v[5] = _mm512_unpackhi_epi64(v1[4], v1[6]);
+    v[6] = _mm512_unpacklo_epi64(v1[5], v1[7]);
+    v[7] = _mm512_unpackhi_epi64(v1[5], v1[7]);
+    v[8] = _mm512_unpacklo_epi64(v1[8], v1[10]);
+    v[9] = _mm512_unpackhi_epi64(v1[8], v1[10]);
+    v[10] = _mm512_unpacklo_epi64(v1[9], v1[11]);
+    v[11] = _mm512_unpackhi_epi64(v1[9], v1[11]);
+    v[12] = _mm512_unpacklo_epi64(v1[12], v1[14]);
+    v[13] = _mm512_unpackhi_epi64(v1[12], v1[14]);
+    v[14] = _mm512_unpacklo_epi64(v1[13], v1[15]);
+    v[15] = _mm512_unpackhi_epi64(v1[13], v1[15]);
+
+    v1[0] = _mm512_shuffle_i32x4(v[0], v[4], 0x88);
+    v1[1] = _mm512_shuffle_i32x4(v[1], v[5], 0x88);
+    v1[2] = _mm512_shuffle_i32x4(v[2], v[6], 0x88);
+    v1[3] = _mm512_shuffle_i32x4(v[3], v[7], 0x88);
+    v1[4] = _mm512_shuffle_i32x4(v[0], v[4], 0xdd);
+    v1[5] = _mm512_shuffle_i32x4(v[1], v[5], 0xdd);
+    v1[6] = _mm512_shuffle_i32x4(v[2], v[6], 0xdd);
+    v1[7] = _mm512_shuffle_i32x4(v[3], v[7], 0xdd);
+    v1[8] = _mm512_shuffle_i32x4(v[8], v[12], 0x88);
+    v1[9] = _mm512_shuffle_i32x4(v[9], v[13], 0x88);
+    v1[10] = _mm512_shuffle_i32x4(v[10], v[14], 0x88);
+    v1[11] = _mm512_shuffle_i32x4(v[11], v[15], 0x88);
+    v1[12] = _mm512_shuffle_i32x4(v[8], v[12], 0xdd);
+    v1[13] = _mm512_shuffle_i32x4(v[9], v[13], 0xdd);
+    v1[14] = _mm512_shuffle_i32x4(v[10], v[14], 0xdd);
+    v1[15] = _mm512_shuffle_i32x4(v[11], v[15], 0xdd);
+
+    v[0] = _mm512_shuffle_i32x4(v1[0], v1[8], 0x88);
+    v[1] = _mm512_shuffle_i32x4(v1[1], v1[9], 0x88);
+    v[2] = _mm512_shuffle_i32x4(v1[2], v1[10], 0x88);
+    v[3] = _mm512_shuffle_i32x4(v1[3], v1[11], 0x88);
+    v[4] = _mm512_shuffle_i32x4(v1[4], v1[12], 0x88);
+    v[5] = _mm512_shuffle_i32x4(v1[5], v1[13], 0x88);
+    v[6] = _mm512_shuffle_i32x4(v1[6], v1[14], 0x88);
+    v[7] = _mm512_shuffle_i32x4(v1[7], v1[15], 0x88);
+    v[8] = _mm512_shuffle_i32x4(v1[0], v1[8], 0xdd);
+    v[9] = _mm512_shuffle_i32x4(v1[1], v1[9], 0xdd);
+    v[10] = _mm512_shuffle_i32x4(v1[2], v1[10], 0xdd);
+    v[11] = _mm512_shuffle_i32x4(v1[3], v1[11], 0xdd);
+    v[12] = _mm512_shuffle_i32x4(v1[4], v1[12], 0xdd);
+    v[13] = _mm512_shuffle_i32x4(v1[5], v1[13], 0xdd);
+    v[14] = _mm512_shuffle_i32x4(v1[6], v1[14], 0xdd);
+    v[15] = _mm512_shuffle_i32x4(v1[7], v1[15], 0xdd);
+}
+
+void quantize_row_q8_K_vnni(const float * RESTRICT x, void * RESTRICT vy, int64_t k) {
+    assert(k % QK_K == 0);
+    const int KB = k / QK_K;
+    constexpr int kVecs = QK_K / 16;
+
+    block_q8_K * y = reinterpret_cast<block_q8_K *>(vy);
+
+    // hold 16 float vecs from x
+    __m512  v[kVecs];
+
+    // hold the quants vecs
+    __m512i vq[kVecs / 4];
+
+    // hold the packed quants vecs
+    __m512i vq_packed[kVecs / 4];
+
+    const __m512 signBit = _mm512_set1_ps(-0.f);
+
+    for (int i = 0; i < KB; ++i) {
+        // Compute max(abs(e)) for the block
+        __m512 vamax = _mm512_set1_ps(0.f);
+        for (int j = 0; j < kVecs; ++j) {
+            v[j] = _mm512_loadu_ps(x); x += 16;
+            vamax = _mm512_max_ps(vamax, _mm512_andnot_ps(signBit, v[j]));
+        }
+        const float amax = _mm512_reduce_max_ps(vamax);
+
+        // Quantize these floats
+        const float iscale = 127.f / amax;
+        y[i].d = GGML_FP32_TO_FP16(1 / iscale);
+        const float id = ( amax != 0.0f ) ? iscale : 0.f;
+        const __m512 vscale = _mm512_set1_ps(id);
+
+        // Apply multiplier and round to nearest integer
+        for (int j = 0; j < kVecs; ++j) {
+            v[j] = _mm512_mul_ps(v[j], vscale);
+            v[j] = _mm512_roundscale_ps(v[j], (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC));
+        }
+
+        // Pack to epi8 vecs
+        for (int j = 0; j < kVecs / 4; ++j) {
+            __m128i q8_0 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 0]));
+            __m128i q8_1 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 1]));
+            __m128i q8_2 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 2]));
+            __m128i q8_3 = _mm512_cvtepi32_epi8(_mm512_cvtps_epi32(v[j * 4 + 3]));
+
+            __m256i q8_01 = _mm256_insertf128_si256(_mm256_castsi128_si256(q8_0), (q8_1), 1);
+            __m256i q8_23 = _mm256_insertf128_si256(_mm256_castsi128_si256(q8_2), (q8_3), 1);
+
+            vq[j] = _mm512_inserti32x8(_mm512_castsi256_si512(q8_01), q8_23, 1);
+            _mm512_storeu_si512((__m512i *)(y[i].qs + j * 64), vq[j]);
+        }
+
+        // Compute the bsums with vnni
+        transpose_16x4_32bit(vq, vq_packed);
+
+        const __m512i one = _mm512_set1_epi8(1);
+        __m512i sum = _mm512_setzero_si512();
+        for (int k = 0; k < 4; ++k) {
+            sum = _mm512_dpbusd_epi32(sum, one, vq_packed[k]);
+        }
+        _mm256_storeu_si256((__m256i *)(y[i].bsums), _mm512_cvtepi32_epi16(sum));
+    }
+}
+
+// quantize A from float to `vec_dot_type`
+template <typename T>
+inline void from_float(const float * x, char * vy, int64_t k);
+
+template <>
+inline void from_float<block_q8_0>(const float * x, char * vy, int64_t k) {
+    quantize_row_q8_0(x, (block_q8_0 *)vy, k);
+}
+
+template <>
+inline void from_float<block_q8_1>(const float * x, char * vy, int64_t k) {
+    quantize_row_q8_1(x, (block_q8_1 *)vy, k);
+}
+
+template <>
+inline void from_float<block_q8_K>(const float * x, char * vy, int64_t k) {
+#if 1
+    // TODO: this is reference impl!
+    quantize_row_q8_K_ref(x, (block_q8_K *)vy, k);
+#else
+    quantize_row_q8_K_vnni(x, vy, k);
+#endif
+}
+
+// load A from memory to array when nrows can not fill in whole tile
+void unpack_A(int8_t * RESTRICT tile, const block_q8_0 * RESTRICT A, int lda, int nr) {
+    assert(nr != TILE_M);
+    for (int m = 0; m < nr; ++m) {
+        const __m256i v = _mm256_loadu_si256((const __m256i *)(A[m * lda].qs));
+        _mm256_storeu_si256((__m256i *)(tile + m * TILE_K), v);
+    }
+}
+
+void unpack_A(int8_t * RESTRICT tile, const block_q8_1 * RESTRICT A, int lda, int nr) {
+    assert(nr != TILE_M);
+    for (int m = 0; m < nr; ++m) {
+        const __m256i v = _mm256_loadu_si256((const __m256i *)(A[m * lda].qs));
+        _mm256_storeu_si256((__m256i *)(tile + m * TILE_K), v);
+    }
+}
+
+template <typename TB>
+void unpack_A(int8_t * RESTRICT tile, const block_q8_K * RESTRICT A, int lda, int k, int nr) {
+    assert(nr <= TILE_M);
+    for (int m = 0; m < nr; ++m) {
+        const __m256i v = _mm256_loadu_si256((const __m256i *)(A[m * lda].qs + k * 32));
+        _mm256_storeu_si256((__m256i *)(tile + m * TILE_K), v);
+    }
+}
+
+template <>
+void unpack_A<block_q6_K>(int8_t * RESTRICT tile, const block_q8_K * RESTRICT A, int lda, int k, int nr) {
+    assert(nr <= TILE_M);
+    // zero padding k from 16 to 32, so that we don't have to re-config amx
+    const __m128i zero = _mm_setzero_si128();
+    for (int m = 0; m < nr; ++m) {
+        const __m128i v = _mm_loadu_si128((const __m128i *)(A[m * lda].qs + k * 16));
+        const __m256i r = _mm256_insertf128_si256(_mm256_castsi128_si256(v), zero, 1);
+        _mm256_storeu_si256((__m256i *)(tile + m * TILE_K), r);
+    }
+}
+
+#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
+inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) {
+    const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi);
+    const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp);
+    const __m256i lowMask = _mm256_set1_epi8(0xF);
+    return _mm256_and_si256(lowMask, bytes);
+}
+
+// used for block_q4_K
+inline __m512i bytes_from_nibbles_64(const uint8_t * rsi) {
+    const __m256i tmp = _mm256_loadu_si256((const __m256i *)rsi);
+    const __m256i lowMask = _mm256_set1_epi8(0xF);
+    const __m256i q4l = _mm256_and_si256(tmp, lowMask);
+    const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(tmp, 4), lowMask);
+    return _mm512_inserti32x8(_mm512_castsi256_si512(q4l), q4h, 1);
+}
+
+// used for block_q5_K
+inline __m512i bytes_from_nibbles_64(const uint8_t * qs, const uint8_t * qh, int k) {
+    const __m256i lowMask = _mm256_set1_epi8(0xF);
+    __m256i hmask = _mm256_set1_epi8(1);
+    hmask = _mm256_slli_epi16(hmask, k);
+
+    const __m256i q5bits = _mm256_loadu_si256((const __m256i *)qs);
+    const __m256i hbits = _mm256_loadu_si256((const __m256i *)qh);
+
+    const __m256i q5l_0 = _mm256_and_si256(q5bits, lowMask);
+    const __m256i q5h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), k + 0), 4);
+    const __m256i q5_0  = _mm256_add_epi8(q5l_0, q5h_0);
+    hmask = _mm256_slli_epi16(hmask, 1);
+
+    const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), lowMask);
+    const __m256i q5h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), k + 1), 4);
+    const __m256i q5_1  = _mm256_add_epi8(q5l_1, q5h_1);
+
+    return _mm512_inserti32x8(_mm512_castsi256_si512(q5_0), q5_1, 1);
+}
+
+// used for block_q6_K
+inline void bytes_from_nibbles_128(__m512i& r0, __m512i& r1, const uint8_t * qs, const uint8_t * qh) {
+    const __m256i m4 = _mm256_set1_epi8(0xF);
+    const __m256i m2 = _mm256_set1_epi8(0x3);
+
+    const __m256i q6bits1 = _mm256_loadu_si256((const __m256i *)qs);
+    const __m256i q6bits2 = _mm256_loadu_si256((const __m256i *)(qs + 32));
+    const __m256i q6bitsH = _mm256_loadu_si256((const __m256i *)qh);
+
+    const __m256i q6h_0 = _mm256_slli_epi16(_mm256_and_si256(                  q6bitsH,     m2), 4);
+    const __m256i q6h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q6bitsH, 2), m2), 4);
+    const __m256i q6h_2 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q6bitsH, 4), m2), 4);
+    const __m256i q6h_3 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q6bitsH, 6), m2), 4);
+
+    const __m256i q6_0 = _mm256_or_si256(_mm256_and_si256(q6bits1, m4), q6h_0);
+    const __m256i q6_1 = _mm256_or_si256(_mm256_and_si256(q6bits2, m4), q6h_1);
+    const __m256i q6_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q6bits1, 4), m4), q6h_2);
+    const __m256i q6_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q6bits2, 4), m4), q6h_3);
+
+    r0 = _mm512_inserti32x8(_mm512_castsi256_si512(q6_0), q6_1, 1);
+    r1 = _mm512_inserti32x8(_mm512_castsi256_si512(q6_2), q6_3, 1);
+}
+
+inline __m512i packNibbles(__m512i r0, __m512i r1) {
+    return _mm512_or_si512(r0, _mm512_slli_epi16(r1, 4));
+}
+
+template <typename TB>
+inline void pack_qs(void * RESTRICT packed_B, const TB * RESTRICT B, int KB) {
+    int8_t tmp[8 * 64];
+    __m256i v[8], v2[8];
+    for (int n = 0; n < 8; ++n) {
+        v[n] = bytes_from_nibbles_32(B[n * KB].qs);
+    }
+    transpose_8x8_32bit(v, v2);
+    for (int n = 0; n < 8; ++n) {
+        _mm256_storeu_si256((__m256i *)(tmp + n * 64), v2[n]);
+    }
+    for (int n = 0; n < 8; ++n) {
+        v[n] = bytes_from_nibbles_32(B[(n + 8) * KB].qs);
+    }
+    transpose_8x8_32bit(v, v2);
+    for (int n = 0; n < 8; ++n) {
+        _mm256_storeu_si256((__m256i *)(tmp + n * 64 + 32), v2[n]);
+    }
+
+    // pack again with 128 to fully utilize vector length
+    for (int n = 0; n < 8; n += 2) {
+        __m512i r0 = _mm512_loadu_si512((const __m512i *)(tmp + n * 64));
+        __m512i r1 = _mm512_loadu_si512((const __m512i *)(tmp + n * 64 + 64));
+        __m512i r1r0 = packNibbles(r0, r1);
+        _mm512_storeu_si512((__m512i *)((char *)packed_B + n * 32), r1r0);
+    }
+}
+
+template <>
+inline void pack_qs<block_q8_0>(void * RESTRICT packed_B, const block_q8_0 * RESTRICT B, int KB) {
+    __m256i v[8], v2[8];
+    for (int n = 0; n < 8; ++n) {
+        v[n] = _mm256_loadu_si256((const __m256i *)(B[n * KB].qs));
+    }
+    transpose_8x8_32bit(v, v2);
+    for (int n = 0; n < 8; ++n) {
+        _mm256_storeu_si256((__m256i *)((char *)packed_B + n * 64), v2[n]);
+    }
+    for (int n = 0; n < 8; ++n) {
+        v[n] = _mm256_loadu_si256((const __m256i *)(B[(n + 8) * KB].qs));
+    }
+    transpose_8x8_32bit(v, v2);
+    for (int n = 0; n < 8; ++n) {
+        _mm256_storeu_si256((__m256i *)((char *)packed_B + n * 64 + 32), v2[n]);
+    }
+}
+
+template <>
+inline void pack_qs<block_q4_K>(void * RESTRICT packed_B, const block_q4_K * RESTRICT B, int KB) {
+    __m512i v[16];
+    // QK_K 256 with 8 groups, handle 2 groups at a time
+    char * pb = (char *)packed_B;
+    for (int k = 0; k < QK_K / 64; ++k) {
+        // pack 2 groups { n, g,  k} to {g, k/4, 4n}
+        //          e.g. {16, 2, 32} to {2,   8, 64}
+        for (int n = 0; n < TILE_N; ++n) {
+            v[n] = bytes_from_nibbles_64(B[n * KB].qs + k * 32);
+        }
+
+        transpose_16x16_32bit(v);
+
+        // pack again with 128 to fully utilize vector length
+        for (int n = 0; n < TILE_N; n += 2) {
+            _mm512_storeu_si512((__m512i *)pb, packNibbles(v[n], v[n + 1]));
+            pb += 64;
+        }
+    }
+}
+
+template <>
+inline void pack_qs<block_q5_K>(void * RESTRICT packed_B, const block_q5_K * RESTRICT B, int KB) {
+    __m512i v[16];
+    const __m512i lowMask = _mm512_set1_epi8(0xF);
+    // QK_K 256 with 8 groups, handle 2 groups at a time
+    char * pb = (char *)packed_B;
+    char * ph = (char *)packed_B + (QK_K / 2) * TILE_N;
+    for (int k = 0; k < QK_K / 64; ++k) {
+        // pack 2 groups { n, g,  k} to {g, k/4, 4n}
+        //          e.g. {16, 2, 32} to {2,   8, 64}
+        for (int n = 0; n < TILE_N; ++n) {
+            v[n] = bytes_from_nibbles_64(B[n * KB].qs + k * 32, B[n * KB].qh, /* group */2 * k);
+        }
+
+        transpose_16x16_32bit(v);
+
+        // 1. pack lower 4bits with 2 groups
+        for (int n = 0; n < TILE_N; n += 2) {
+            // get lower 4 bits
+            const __m512i r0 = _mm512_and_si512(v[n], lowMask);
+            const __m512i r1 = _mm512_and_si512(v[n + 1], lowMask);
+            _mm512_storeu_si512((__m512i *)pb, packNibbles(r0, r1)); pb += 64;
+        }
+
+        // 2. pack higher 1bit with 2 groups
+        const __m512i hmask = _mm512_set1_epi8(0x10);
+        for (int g = 0; g < 2; ++g) {
+            __m512i hbits = _mm512_setzero_si512();
+            hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 0], hmask), 4));
+            hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 1], hmask), 3));
+            hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 2], hmask), 2));
+            hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 8 + 3], hmask), 1));
+            hbits = _mm512_add_epi8(hbits,                   _mm512_and_si512(v[g * 8 + 4], hmask)    );
+            hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 8 + 5], hmask), 1));
+            hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 8 + 6], hmask), 2));
+            hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 8 + 7], hmask), 3));
+            _mm512_storeu_si512((__m512i *)ph, hbits); ph += 64;
+        }
+    }
+}
+
+template <>
+inline void pack_qs<block_q6_K>(void * RESTRICT packed_B, const block_q6_K * RESTRICT B, int KB) {
+    __m512i v[32];
+    const __m512i lowMask = _mm512_set1_epi8(0xF);
+    // QK_K 256 with 8 groups, handle 4 groups at a time
+    char * pb = (char *)packed_B;
+    char * ph = (char *)packed_B + (QK_K / 2) * TILE_N;
+    for (int k = 0; k < QK_K / 128; ++k) {
+        for (int n = 0; n < TILE_N; ++n) {
+            bytes_from_nibbles_128(v[n], v[n + 16], B[n * KB].ql + k * 64, B[n * KB].qh + k * 32);
+        }
+
+        // top half: group 0,1 or 4,5; bottom half: group 2,3 or 6,7
+        transpose_16x16_32bit(v);
+        transpose_16x16_32bit(v + 16);
+
+        // 1. pack lower 4bits with 4 groups
+        for (int n = 0; n < 32; n += 2) {
+            const __m512i r0 = _mm512_and_si512(v[n], lowMask);
+            const __m512i r1 = _mm512_and_si512(v[n + 1], lowMask);
+            _mm512_storeu_si512((__m512i *)pb, packNibbles(r0, r1)); pb += 64;
+        }
+
+        // 2. pack higher 2bit with 4 groups
+        const __m512i hmask = _mm512_set1_epi8(0x30);
+        for (int g = 0; g < 8; ++g) {
+            __m512i hbits = _mm512_setzero_si512();
+            hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 4 + 0], hmask), 4));
+            hbits = _mm512_add_epi8(hbits, _mm512_srli_epi16(_mm512_and_si512(v[g * 4 + 1], hmask), 2));
+            hbits = _mm512_add_epi8(hbits,                   _mm512_and_si512(v[g * 4 + 2], hmask)    );
+            hbits = _mm512_add_epi8(hbits, _mm512_slli_epi16(_mm512_and_si512(v[g * 4 + 3], hmask), 2));
+            _mm512_storeu_si512((__m512i *)ph, hbits); ph += 64;
+        }
+    }
+}
+
+template <>
+inline void pack_qs<block_iq4_xs>(void * RESTRICT packed_B, const block_iq4_xs * RESTRICT B, int KB) {
+    __m512i v[16];
+    char * pb = (char *)packed_B;
+    for (int k = 0; k < QK_K / 64; ++k) {
+        for (int n = 0; n < TILE_N; ++n) {
+            __m256i r0 = bytes_from_nibbles_32(B[n * KB].qs + k * 32 +  0);
+            __m256i r1 = bytes_from_nibbles_32(B[n * KB].qs + k * 32 + 16);
+            v[n] = _mm512_inserti32x8(_mm512_castsi256_si512(r0), r1, 1);
+        }
+
+        transpose_16x16_32bit(v);
+
+        // pack again with 128 to fully utilize vector length
+        for (int n = 0; n < TILE_N; n += 2) {
+            _mm512_storeu_si512((__m512i *)pb, packNibbles(v[n], v[n + 1]));
+            pb += 64;
+        }
+    }
+}
+
+// pack B to vnni formats in 4bits or 8 bits
+void pack_B(void * RESTRICT packed_B, const block_q4_0 * RESTRICT B, int KB) {
+    pack_qs(packed_B, B, KB);
+    ggml_half * d0 = reinterpret_cast<ggml_half *>((char *)packed_B + TILE_N * TILE_K / 2);
+    for (int n = 0; n < TILE_N; ++n) {
+        d0[n] = B[n * KB].d;
+    }
+}
+
+void pack_B(void * RESTRICT packed_B, const block_q4_1 * RESTRICT B, int KB) {
+    pack_qs(packed_B, B, KB);
+    ggml_half * d0 = reinterpret_cast<ggml_half *>((char *)packed_B + TILE_N * TILE_K / 2);
+    ggml_half * m0 = d0 + TILE_N;
+    for (int n = 0; n < TILE_N; ++n) {
+        d0[n] = B[n * KB].d;
+        m0[n] = B[n * KB].m;
+    }
+}
+
+inline void s8s8_compensation(void * RESTRICT packed_B) {
+    // packed_B layout:
+    //   quants {TILE_N, TILEK}  int8_t
+    //   d0     {TILE_N}      ggml_half
+    //   comp   {TILE_N}        int32_t
+    const int offset = TILE_N * TILE_K + TILE_N * sizeof(ggml_half);
+    __m512i vcomp = _mm512_setzero_si512();
+    const __m512i off = _mm512_set1_epi8(static_cast<char>(0x80));
+    for (int k = 0; k < 8; ++k) {
+        __m512i vb = _mm512_loadu_si512((const __m512i *)((const char *)packed_B + k * 64));
+        vcomp = _mm512_dpbusd_epi32(vcomp, off, vb);
+    }
+    _mm512_storeu_si512((__m512i *)((char *)(packed_B) + offset), vcomp);
+}
+
+void pack_B(void * RESTRICT packed_B, const block_q8_0 * RESTRICT B, int KB) {
+    pack_qs(packed_B, B, KB);
+    ggml_half * d0 = reinterpret_cast<ggml_half *>((char *)packed_B + TILE_N * TILE_K);
+    for (int n = 0; n < TILE_N; ++n) {
+        d0[n] = B[n * KB].d;
+    }
+    s8s8_compensation(packed_B);
+}
+
+// convert 8 * {min, scale} from int6 to int8
+inline void unpack_mins_and_scales(const uint8_t * scales, uint32_t * utmp) {
+    const uint32_t kmask1 = 0x3f3f3f3f;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+    const uint32_t kmask3 = 0x03030303;
+
+    memcpy(utmp, scales, 12);
+    utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+    const uint32_t uaux = utmp[1] & kmask1;
+    utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+    utmp[2] = uaux;
+    utmp[0] &= kmask1;
+}
+
+// packed_B layout:
+//   quants {8, TILE_N, 16}  uint8
+//   scales {8, TILE_N}      uint8
+//   mins   {8, TILE_N}      uint8
+//   d      {TILE_N}     ggml_half
+//   dmin   {TILE_N}     ggml_half
+void pack_B(void * RESTRICT packed_B, const block_q4_K * RESTRICT B, int KB) {
+    pack_qs(packed_B, B, KB);
+
+    uint8_t * scales = reinterpret_cast<uint8_t *>((char *)packed_B + (QK_K / 2) * TILE_N);
+    uint8_t * mins = scales + 8 * TILE_N;
+    ggml_half * d = reinterpret_cast<ggml_half *>(mins + 8 * TILE_N);
+    ggml_half * dmin = d + TILE_N;
+
+    union {
+        uint32_t u32[4];
+        uint8_t  u8[16];
+    } s;
+
+    for (int n = 0; n < TILE_N; ++n) {
+        unpack_mins_and_scales(B[n * KB].scales, s.u32);
+        for (int k = 0; k < 8; ++k) {
+            scales[k * TILE_N + n] = s.u8[k];
+            mins[(k >> 1) * TILE_N * 2 + n * 2 + (k & 0x1)] = s.u8[k + 8];
+        }
+        d[n] = B[n * KB].d;
+        dmin[n] = B[n * KB].dmin;
+    }
+}
+
+// packed_B layout:
+//   quants {8, TILE_N, 16}  uint8
+//   qh     {8, TILE_N,  4}  uint8
+//   scales {8, TILE_N}      uint8
+//   mins   {8, TILE_N}      uint8
+//   d      {TILE_N}     ggml_half
+//   dmin   {TILE_N}     ggml_half
+void pack_B(void * RESTRICT packed_B, const block_q5_K * RESTRICT B, int KB) {
+    pack_qs(packed_B, B, KB);
+
+    uint8_t * scales = reinterpret_cast<uint8_t *>((char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N);
+    uint8_t * mins = scales + 8 * TILE_N;
+    ggml_half * d = reinterpret_cast<ggml_half *>(mins + 8 * TILE_N);
+    ggml_half * dmin = d + TILE_N;
+
+    union {
+        uint32_t u32[4];
+        uint8_t  u8[16];
+    } s;
+
+    for (int n = 0; n < TILE_N; ++n) {
+        unpack_mins_and_scales(B[n * KB].scales, s.u32);
+        for (int k = 0; k < 8; ++k) {
+            scales[k * TILE_N + n] = s.u8[k];
+            mins[(k >> 1) * TILE_N * 2 + n * 2 + (k & 0x1)] = s.u8[k + 8];
+        }
+        d[n] = B[n * KB].d;
+        dmin[n] = B[n * KB].dmin;
+    }
+}
+
+// packed_B layout:
+//   quants {16, TILE_N, 8}  uint8
+//   qh     {16, TILE_N, 4}  uint8
+//   scales {16, TILE_N}      uint8
+//   d      {TILE_N}     ggml_half
+void pack_B(void * RESTRICT packed_B, const block_q6_K * RESTRICT B, int KB) {
+    pack_qs(packed_B, B, KB);
+
+    uint8_t * scales = reinterpret_cast<uint8_t *>((char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N);
+    ggml_half * d = reinterpret_cast<ggml_half *>(scales + 16 * TILE_N);
+    for (int n = 0; n < TILE_N; ++n) {
+        const int8_t * ps = B[n * KB].scales;
+        for (int k = 0; k < 16; ++k) {
+            scales[k * TILE_N + n] = ps[k];
+        }
+        d[n] = B[n * KB].d;
+    }
+}
+
+// packed_B layout:
+//   quants {8, TILE_N, 16}  uint8
+//   scales {8, TILE_N}       int8
+//   d      {TILE_N}     ggml_half
+void pack_B(void * RESTRICT packed_B, const block_iq4_xs * RESTRICT B, int KB) {
+    pack_qs(packed_B, B, KB);
+
+    int8_t * scales = reinterpret_cast<int8_t *>((char *)packed_B + (QK_K / 2) * TILE_N);
+    ggml_half * d = reinterpret_cast<ggml_half *>(scales + 8 * TILE_N);
+
+    // pack the scales
+    for (int n = 0; n < TILE_N; ++n) {
+        uint16_t sh = B[n * KB].scales_h;
+        for (int k = 0; k < 8; k += 2) {
+            const int16_t ls1 = ((B[n * KB].scales_l[k / 2] & 0xf) | ((sh << 4) & 0x30)) - 32;
+            const int16_t ls2 = ((B[n * KB].scales_l[k / 2] >>  4) | ((sh << 2) & 0x30)) - 32;
+            scales[(k + 0) * TILE_N + n] = ls1;
+            scales[(k + 1) * TILE_N + n] = ls2;
+            sh >>= 4;
+        }
+        d[n] = B[n * KB].d;
+    }
+}
+
+template<typename TB, typename packed_B_t = packed_B_type<TB>>
+void unpack_B(packed_B_t * RESTRICT tile, const void * RESTRICT packed_B) {
+    GGML_UNUSED(tile);
+    GGML_UNUSED(packed_B);
+}
+
+template <>
+void unpack_B<block_q4_0>(int8_t * RESTRICT tile, const void * RESTRICT packed_B) {
+  const __m512i off = _mm512_set1_epi8(8);
+  const __m512i lowMask = _mm512_set1_epi8(0xF);
+  for (int n = 0; n < 8; n += 2) {
+    __m512i bytes = _mm512_loadu_si512((const __m512i *)((const char *)packed_B + n * 32));
+    const __m512i r0 = _mm512_sub_epi8(_mm512_and_si512(bytes, lowMask), off);
+    const __m512i r1 = _mm512_sub_epi8(_mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask), off);
+    _mm512_storeu_si512((__m512i *)(tile + n * 64 +  0), r0);
+    _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1);
+  }
+}
+
+template <>
+void unpack_B<block_q4_1>(uint8_t * RESTRICT tile, const void * RESTRICT packed_B) {
+    const __m512i lowMask = _mm512_set1_epi8(0xF);
+    for (int n = 0; n < 8; n += 2) {
+        __m512i bytes = _mm512_loadu_si512((const __m512i *)((const char *)packed_B + n * 32));
+        const __m512i r0 = _mm512_and_si512(bytes, lowMask);
+        const __m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
+        _mm512_storeu_si512((__m512i *)(tile + n * 64 +  0), r0);
+        _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1);
+    }
+}
+
+// packed_B_t for QKK is int8_t
+template <typename TB>
+void unpack_B(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) {
+    const int packed_B_group_size = QK_K / 2 * TILE_N / 8;
+    const char * packed_B_group = (const char *)packed_B + k * packed_B_group_size;
+    const __m512i lowMask = _mm512_set1_epi8(0xF);
+    for (int n = 0; n < 8; n += 2) {
+        __m512i bytes = _mm512_loadu_si512(packed_B_group + n * 32);
+        const __m512i r0 = _mm512_and_si512(bytes, lowMask);
+        const __m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
+        _mm512_storeu_si512((__m512i *)(tile + n * 64 +  0), r0);
+        _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1);
+    }
+}
+
+template <>
+void unpack_B<block_q5_K>(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) {
+    // lower 4bits, stride 256 bytes
+    const int packed_l4_group_size = QK_K / 2 * TILE_N / 8;
+    const char * pb = (const char *)packed_B + k * packed_l4_group_size;
+
+    // higher 1bit, stride 64 bytes
+    const int packed_h1_group_size = QK_K / 8 * TILE_N / 8;
+    const char * ph = (const char *)packed_B + (QK_K / 2) * TILE_N + k * packed_h1_group_size;
+    const __m512i hbits = _mm512_loadu_si512(ph);
+
+    const __m512i lowMask = _mm512_set1_epi8(0xF);
+    __m512i hmask0 = _mm512_set1_epi8(0x1);
+    __m512i hmask1 = _mm512_set1_epi8(0x2);
+
+    for (int n = 0; n < 8; n += 2) {
+        __m512i bytes = _mm512_loadu_si512(pb + n * 32);
+        __m512i r0 = _mm512_and_si512(bytes, lowMask);
+        __m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
+        __m512i h0 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask0), n), 4);
+        __m512i h1 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask1), n + 1), 4);
+
+        hmask0 = _mm512_slli_epi16(hmask0, 2);
+        hmask1 = _mm512_slli_epi16(hmask1, 2);
+        r0 = _mm512_add_epi8(r0, h0);
+        r1 = _mm512_add_epi8(r1, h1);
+        _mm512_storeu_si512((__m512i *)(tile + n * 64 +  0), r0);
+        _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1);
+    }
+}
+
+template <>
+void unpack_B<block_q6_K>(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) {
+    // lower 4bits, stride 128 bytes
+    const int packed_l4_group_size = QK_K / 2 * TILE_N / 16;
+    const char * pb = (const char *)packed_B + k * packed_l4_group_size;
+
+    // higher 2bits, stride 64 bytes
+    const int packed_h2_group_size = QK_K / 4 * TILE_N / 16;
+    const char * ph = (const char *)packed_B + (QK_K / 2) * TILE_N + k * packed_h2_group_size;
+    const __m512i hbits = _mm512_loadu_si512(ph);
+
+    const __m512i off = _mm512_set1_epi8(32);
+    const __m512i lowMask = _mm512_set1_epi8(0xF);
+    __m512i hmask0 = _mm512_set1_epi8(0x3); // 0011
+    __m512i hmask1 = _mm512_set1_epi8(0xC); // 1100
+
+    // notes: skip zero padding from row4 to row7 as we have done so in `unpack_A`
+    __m512i bytes = _mm512_loadu_si512(pb);
+    __m512i r0 = _mm512_and_si512(bytes, lowMask);
+    __m512i r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
+    __m512i h0 = _mm512_slli_epi16(_mm512_and_si512(hbits, hmask0), 4);
+    __m512i h1 = _mm512_slli_epi16(_mm512_and_si512(hbits, hmask1), 2);
+    _mm512_storeu_si512((__m512i *)(tile +  0), _mm512_sub_epi8(_mm512_add_epi8(r0, h0), off));
+    _mm512_storeu_si512((__m512i *)(tile + 64), _mm512_sub_epi8(_mm512_add_epi8(r1, h1), off));
+
+    hmask0 = _mm512_slli_epi16(hmask0, 4);
+    hmask1 = _mm512_slli_epi16(hmask1, 4);
+
+    bytes = _mm512_loadu_si512(pb + 64);
+    r0 = _mm512_and_si512(bytes, lowMask);
+    r1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
+    h0 =                   _mm512_and_si512(hbits, hmask0);
+    h1 = _mm512_srli_epi16(_mm512_and_si512(hbits, hmask1), 2);
+    _mm512_storeu_si512((__m512i *)(tile + 128), _mm512_sub_epi8(_mm512_add_epi8(r0, h0), off));
+    _mm512_storeu_si512((__m512i *)(tile + 192), _mm512_sub_epi8(_mm512_add_epi8(r1, h1), off));
+}
+
+template <>
+void unpack_B<block_iq4_xs>(int8_t * RESTRICT tile, const void * RESTRICT packed_B, int k) {
+    static const __m512i values128 = _mm512_set_epi8(
+        113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127,
+        113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127,
+        113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127,
+        113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127
+    );
+
+    const int packed_B_group_size = QK_K / 2 * TILE_N / 8;
+    const char * pb = (const char *)packed_B + k * packed_B_group_size;
+    const __m512i lowMask = _mm512_set1_epi8(0xF);
+
+    for (int n = 0; n < 8; n += 2) {
+        __m512i bytes = _mm512_loadu_si512(pb + n * 32);
+        const __m512i r0 = _mm512_shuffle_epi8(values128, _mm512_and_si512(bytes, lowMask));
+        const __m512i r1 = _mm512_shuffle_epi8(values128, _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask));
+        _mm512_storeu_si512((__m512i *)(tile + n * 64 +  0), r0);
+        _mm512_storeu_si512((__m512i *)(tile + n * 64 + 64), r1);
+    }
+}
+
+template <typename TA, typename TB, bool is_acc>
+struct acc_C {};
+
+template <bool is_acc>
+struct acc_C<block_q8_0, block_q4_0, is_acc> {
+    static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_0 * A, int lda, const void * packed_B, int nr) {
+        const int offset = TILE_N * TILE_K / 2;
+        const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
+
+        for (int m = 0; m < nr; ++m) {
+            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
+
+            __m512 vsum;
+            if (is_acc) {
+                vsum = _mm512_loadu_ps(C + m * ldc);
+            } else {
+                vsum = _mm512_set1_ps(0.f);
+            }
+            vsum = _mm512_fmadd_ps(vtile, _mm512_mul_ps(vd0, vd1), vsum);
+            _mm512_storeu_ps(C + m * ldc, vsum);
+        }
+    }
+};
+
+template <bool is_acc>
+struct acc_C<block_q8_1, block_q4_1, is_acc> {
+    static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_1 * A, int lda, const void * packed_B, int nr) {
+        const int offset = TILE_N * TILE_K / 2;
+        const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
+        const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset + TILE_N * sizeof(ggml_half))));
+
+        for (int m = 0; m < nr; ++m) {
+            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].s));
+            const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
+
+            __m512 vsum;
+            if (is_acc) {
+                vsum = _mm512_loadu_ps(C + m * ldc);
+            } else {
+                vsum = _mm512_set1_ps(0.f);
+            }
+            vsum = _mm512_fmadd_ps(vtile, _mm512_mul_ps(vd0, vd1), vsum);
+            vsum = _mm512_fmadd_ps(vm0, vs1, vsum);
+            _mm512_storeu_ps(C + m * ldc, vsum);
+        }
+    }
+};
+
+template <bool is_acc>
+struct acc_C<block_q8_0, block_q8_0, is_acc> {
+    static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_0 * A, int lda, const void * packed_B, int nr) {
+        const int offset = TILE_N * TILE_K;
+        const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
+
+        for (int m = 0; m < nr; ++m) {
+            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
+
+            __m512 vsum;
+            if (is_acc) {
+                vsum = _mm512_loadu_ps(C + m * ldc);
+            } else {
+                vsum = _mm512_set1_ps(0.f);
+            }
+            vsum = _mm512_fmadd_ps(vtile, _mm512_mul_ps(vd0, vd1), vsum);
+            _mm512_storeu_ps(C + m * ldc, vsum);
+        }
+    }
+};
+
+template <bool is_acc>
+struct acc_C<block_q8_K, block_q4_K, is_acc> {
+    static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) {
+        const uint8_t * scales = reinterpret_cast<const uint8_t *>((const char *)packed_B + (QK_K / 2) * TILE_N);
+        const uint8_t * mins = scales + 8 * TILE_N;
+        const ggml_half * d0 = reinterpret_cast<const ggml_half *>(mins + 8 * TILE_N);
+        const ggml_half * dmin = d0 + TILE_N;
+
+        const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0));
+        const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)dmin));
+
+        for (int m = 0; m < nr; ++m) {
+            const float d1 = A[m * lda].d;
+            const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0);
+            const __m512 vdm = _mm512_mul_ps(_mm512_set1_ps(-d1), vdmin);
+            const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
+
+            __m512 vsum;
+            if (is_acc) {
+                vsum = _mm512_loadu_ps(C + m * ldc);
+            } else {
+                vsum = _mm512_set1_ps(0.f);
+            }
+
+            const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[m * lda].bsums);
+            const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
+
+            __m512i acc_m = _mm512_setzero_si512();
+            for (int k = 0; k < 4; ++k) {
+                __m512i vmask = _mm512_set1_epi32(k);
+                __m512i va = _mm512_permutexvar_epi32(vmask, _mm512_castsi128_si512(q8s));
+                __m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(mins + k * 32)));
+                acc_m = _mm512_dpwssds_epi32(acc_m, va, vb);
+            }
+
+            vsum = _mm512_fmadd_ps(vtile, vd, vsum);
+            vsum = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc_m), vdm, vsum);
+            _mm512_storeu_ps(C + m * ldc, vsum);
+        }
+    }
+};
+
+template <bool is_acc>
+struct acc_C<block_q8_K, block_q5_K, is_acc> {
+    static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) {
+        const uint8_t * scales = reinterpret_cast<const uint8_t *>((const char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N);
+        const uint8_t * mins = scales + 8 * TILE_N;
+        const ggml_half * d0 = reinterpret_cast<const ggml_half *>(mins + 8 * TILE_N);
+        const ggml_half * dmin = d0 + TILE_N;
+
+        const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0));
+        const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)dmin));
+
+        for (int m = 0; m < nr; ++m) {
+            const float d1 = A[m * lda].d;
+            const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0);
+            const __m512 vdm = _mm512_mul_ps(_mm512_set1_ps(-d1), vdmin);
+            const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
+
+            __m512 vsum;
+            if (is_acc) {
+                vsum = _mm512_loadu_ps(C + m * ldc);
+            } else {
+                vsum = _mm512_set1_ps(0.f);
+            }
+
+            const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[m * lda].bsums);
+            const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
+
+            __m512i acc_m = _mm512_setzero_si512();
+            for (int k = 0; k < 4; ++k) {
+                __m512i vmask = _mm512_set1_epi32(k);
+                __m512i va = _mm512_permutexvar_epi32(vmask, _mm512_castsi128_si512(q8s));
+                __m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(mins + k * 32)));
+                acc_m = _mm512_dpwssds_epi32(acc_m, va, vb);
+            }
+
+            vsum = _mm512_fmadd_ps(vtile, vd, vsum);
+            vsum = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc_m), vdm, vsum);
+            _mm512_storeu_ps(C + m * ldc, vsum);
+        }
+    }
+};
+
+template <bool is_acc>
+struct acc_C<block_q8_K, block_q6_K, is_acc> {
+    static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) {
+        const uint8_t * scales = reinterpret_cast<const uint8_t *>((const char *)packed_B + (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N);
+        const ggml_half * d0 = reinterpret_cast<const ggml_half *>(scales + 16 * TILE_N);
+
+        const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0));
+
+        for (int m = 0; m < nr; ++m) {
+            const float d1 = A[m * lda].d;
+            const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0);
+            const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
+
+            __m512 vsum;
+            if (is_acc) {
+                vsum = _mm512_loadu_ps(C + m * ldc);
+            } else {
+                vsum = _mm512_set1_ps(0.f);
+            }
+
+            vsum = _mm512_fmadd_ps(vtile, vd, vsum);
+            _mm512_storeu_ps(C + m * ldc, vsum);
+        }
+    }
+};
+
+template <bool is_acc>
+struct acc_C<block_q8_K, block_iq4_xs, is_acc> {
+    static void apply(float * RESTRICT C, int ldc, const int32_t * RESTRICT tile, const block_q8_K * A, int lda, const void * packed_B, int nr) {
+        const int8_t * scales = reinterpret_cast<const int8_t *>((const char *)packed_B + (QK_K / 2) * TILE_N);
+        const ggml_half * d0 = reinterpret_cast<const ggml_half *>(scales + 8 * TILE_N);
+
+        const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)d0));
+
+        for (int m = 0; m < nr; ++m) {
+            const float d1 = A[m * lda].d;
+            const __m512 vd = _mm512_mul_ps(_mm512_set1_ps(d1), vd0);
+            const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
+
+            __m512 vsum;
+            if (is_acc) {
+                vsum = _mm512_loadu_ps(C + m * ldc);
+            } else {
+                vsum = _mm512_set1_ps(0.f);
+            }
+
+            vsum = _mm512_fmadd_ps(vtile, vd, vsum);
+            _mm512_storeu_ps(C + m * ldc, vsum);
+        }
+    }
+};
+
+template <typename TB> constexpr int get_quants_size();
+template <> constexpr int get_quants_size<block_q4_K>() { return (QK_K / 2) * TILE_N; }
+template <> constexpr int get_quants_size<block_q5_K>() { return (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N; }
+template <> constexpr int get_quants_size<block_q6_K>() { return (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N; }
+template <> constexpr int get_quants_size<block_iq4_xs>() { return (QK_K / 2) * TILE_N; }
+
+// used for QKK format
+template <typename TB, bool is_acc,
+          typename std::enable_if<is_type_qkk<TB>::value, int>::type = 0>
+inline void scale_C(const int32_t * RESTRICT tile, int32_t * RESTRICT sumi, const void * packed_B, int k, int nr) {
+    const uint8_t * scales = reinterpret_cast<const uint8_t *>((const char *)packed_B + get_quants_size<TB>());
+    const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(scales + k * TILE_N)));
+
+    for (int m = 0; m < nr; ++m) {
+        __m512i vsumi;
+        if (is_acc) {
+            vsumi = _mm512_loadu_si512(sumi + m * TILE_N);
+        } else {
+            vsumi = _mm512_setzero_si512();
+        }
+        __m512i vtile = _mm512_loadu_si512(tile + m * TILE_N);
+        vsumi = _mm512_add_epi32(vsumi, _mm512_mullo_epi32(vtile, vscale));
+        _mm512_storeu_si512((__m512i *)(sumi + m * TILE_N), vsumi);
+    }
+}
+
+template <typename TA, typename TB, typename TC, int BLOCK_M, int BLOCK_N, int BLOCK_K>
+struct tinygemm_kernel_avx {
+    static void apply(int K, const TA * RESTRICT A, const TB * RESTRICT B, TC * RESTRICT C, int ldc) {
+        GGML_UNUSED(K);
+        GGML_UNUSED(A);
+        GGML_UNUSED(B);
+        GGML_UNUSED(C);
+        GGML_UNUSED(ldc);
+    }
+};
+
+template <int BLOCK_M, int BLOCK_N, int BLOCK_K>
+struct tinygemm_kernel_avx<float, ggml_fp16_t, float, BLOCK_M, BLOCK_N, BLOCK_K> {
+    static void apply(int K, const float * RESTRICT A, const ggml_fp16_t * RESTRICT B, float * RESTRICT C, int ldc) {
+        constexpr int ROWS = BLOCK_M;
+        constexpr int COLS = BLOCK_N;
+        assert(BLOCK_K == 16);
+
+        __m512 va;
+        __m512 vb[COLS];
+        __m512 vc[ROWS * COLS];
+
+        auto loadc = [&](auto idx) {
+            vc[idx] = _mm512_setzero_ps();
+        };
+        Unroll<ROWS * COLS>{}(loadc);
+
+        auto compute = [&](auto idx, auto k) {
+            constexpr int row = idx / COLS;
+            constexpr int col = idx % COLS;
+
+            if constexpr (col == 0) {
+                va = _mm512_loadu_ps(A + row * K + k);
+            }
+            if constexpr (row == 0) {
+                vb[col] =  _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(B + col * K + k)));
+            }
+            vc[idx] = _mm512_fmadd_ps(va, vb[col], vc[idx]);
+        };
+
+        for (int k = 0; k < K; k += 16) {
+            Unroll<ROWS * COLS>{}(compute, k);
+        }
+
+        auto storec = [&](auto idx) {
+            constexpr int row = idx / COLS;
+            constexpr int col = idx % COLS;
+            C[row * ldc + col] = _mm512_reduce_add_ps(vc[idx]);
+        };
+        Unroll<ROWS * COLS>{}(storec);
+    }
+};
+
+#define LAUNCH_TINYGEMM_KERNEL_AVX(MB_SIZE, NB_SIZE)                                \
+    tinygemm_kernel_avx<float, type, float, MB_SIZE, NB_SIZE, blck_size>::apply(    \
+        K, (const float *)src1->data + mb_start * K,                                \
+        (const type *)src0->data + nb_start * K,                                    \
+        (float *)dst->data + mb_start * ldc + nb_start, ldc);
+
+
+// re-organize in the format {NB, KB, TILE_SIZE}:
+#define PACKED_INDEX(n, k, KB, tile_size) (n * KB + k) * tile_size
+
+template<typename TB, int BLOCK_K>
+void convert_B_packed_format(void * RESTRICT packed_B, const TB * RESTRICT B, int N, int K) {
+    const int NB = N / TILE_N;
+    const int KB = K / BLOCK_K;
+    const int TILE_SIZE = get_tile_size<TB>();
+
+    // parallel on NB should be enough
+    parallel_for(NB, [&](int begin, int end) {
+        for (int n = begin; n < end; ++n) {
+            for (int k = 0; k < KB; ++k) {
+                int n0 = n * TILE_N;
+                pack_B((char *)packed_B + PACKED_INDEX(n, k, KB, TILE_SIZE), &B[n0 * KB + k], KB);
+            }
+        }
+    });
+}
+
+template <typename TA, typename TB, typename TC, int BLOCK_M, int BLOCK_N, int BLOCK_K>
+struct tinygemm_kernel_vnni {};
+
+template <int BLOCK_M, int BLOCK_N, int BLOCK_K>
+struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLOCK_K> {
+    static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
+
+        constexpr int COLS = BLOCK_N / 16;
+        const int TILE_SIZE = TILE_N * sizeof(block_q4_0);
+
+        const block_q8_0 * RESTRICT A = static_cast<const block_q8_0 *>(_A);
+        const char * RESTRICT B = static_cast<const char *>(_B);
+
+        __m512i va[8];
+        __m512 vc[COLS];
+        __m512 vd1;
+
+        // sum of offsets, shared across COLS
+        //
+        // avx512-vnni does not have `_mm512_dpbssd_epi32`,
+        // need to transfrom ss to us:
+        //   a * (b - 8) is equavilent to b * a - 8 * a
+        //   s    u   u                   u   s   u   s
+        //
+        __m512i vcomp;
+
+        const __m512i off = _mm512_set1_epi8(8);
+        const __m512i lowMask = _mm512_set1_epi8(0xF);
+
+        auto loadc = [&](auto col) {
+            vc[col] = _mm512_setzero_ps();
+        };
+        Unroll<COLS>{}(loadc);
+
+        auto compute = [&](auto col, auto i) {
+            // load a and compute compensation
+            if constexpr (col == 0) {
+                const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
+                vcomp = _mm512_setzero_si512();
+                for (int k = 0; k < 8; ++k) {
+                    va[k] = _mm512_set1_epi32(a_ptr[k]);
+                    vcomp = _mm512_dpbusd_epi32(vcomp, off, va[k]);
+                }
+                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
+            }
+
+            // load b
+            __m512i vsum = _mm512_setzero_si512();
+            const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE);
+            for (int k = 0; k < 8; k += 2) {
+                __m512i bytes = _mm512_loadu_si512((const __m512i *)(b_ptr + k * 32));
+                __m512i vb0 = _mm512_and_si512(bytes, lowMask);
+                vsum = _mm512_dpbusd_epi32(vsum, vb0, va[k + 0]);
+                __m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
+                vsum = _mm512_dpbusd_epi32(vsum, vb1, va[k + 1]);
+            }
+            const int offset = TILE_N * TILE_K / 2;
+            const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset)));
+            vsum = _mm512_sub_epi32(vsum, vcomp);
+
+            vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(vsum), _mm512_mul_ps(vd0, vd1), vc[col]);
+        };
+
+        for (int i = 0; i < KB; ++i) {
+            Unroll<COLS>{}(compute, i);
+        }
+
+        //store to C
+        auto storec = [&](auto col) {
+            _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
+        };
+        Unroll<COLS>{}(storec);
+    }
+};
+
+template <int BLOCK_N, int BLOCK_K>
+struct tinygemm_kernel_vnni<block_q8_1, block_q4_1, float, 1, BLOCK_N, BLOCK_K> {
+    static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
+
+        constexpr int COLS = BLOCK_N / 16;
+        const int TILE_SIZE = TILE_N * sizeof(block_q4_1);
+
+        const block_q8_1 * RESTRICT A = static_cast<const block_q8_1 *>(_A);
+        const char * RESTRICT B = static_cast<const char *>(_B);
+
+        __m512i va[8];
+        __m512i vb[8];
+        __m512 vc[COLS];
+        __m512 vd1, vs1;
+
+        const __m512i lowMask = _mm512_set1_epi8(0xF);
+
+        auto loadc = [&](auto col) {
+            vc[col] = _mm512_setzero_ps();
+        };
+        Unroll<COLS>{}(loadc);
+
+        auto compute = [&](auto col, auto i) {
+            // load a
+            if constexpr (col == 0) {
+                const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
+                for (int k = 0; k < 8; ++k) {
+                    va[k] = _mm512_set1_epi32(a_ptr[k]);
+                }
+                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
+                vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].s));
+            }
+
+            // load b
+            const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE);
+            for (int k = 0; k < 8; k += 2) {
+                __m512i bytes = _mm512_loadu_si512((const __m512i *)(b_ptr + k * 32));
+                vb[k + 0] = _mm512_and_si512(bytes, lowMask);
+                vb[k + 1] = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
+            }
+            const int offset = TILE_N * TILE_K / 2;
+            const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset)));
+            const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset + TILE_N * sizeof(ggml_half))));
+
+            __m512i vsum = _mm512_setzero_si512();
+            for (int k = 0; k < 8; ++k) {
+                vsum = _mm512_dpbusd_epi32(vsum, vb[k], va[k]);
+            }
+
+            vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(vsum), _mm512_mul_ps(vd0, vd1), vc[col]);
+            vc[col] = _mm512_fmadd_ps(vm0, vs1, vc[col]);
+        };
+
+        for (int i = 0; i < KB; ++i) {
+            Unroll<COLS>{}(compute, i);
+        }
+
+        //store to C
+        auto storec = [&](auto col) {
+            _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
+        };
+        Unroll<COLS>{}(storec);
+    }
+};
+
+template <int BLOCK_M, int BLOCK_N, int BLOCK_K>
+struct tinygemm_kernel_vnni<block_q8_0, block_q8_0, float, BLOCK_M, BLOCK_N, BLOCK_K> {
+    static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
+
+        constexpr int COLS = BLOCK_N / 16;
+        const int TILE_SIZE = TILE_N * sizeof(block_q8_0) + TILE_N * sizeof(int32_t);
+
+        const block_q8_0 * RESTRICT A = static_cast<const block_q8_0 *>(_A);
+        const char * RESTRICT B = static_cast<const char *>(_B);
+
+        __m512i va[8];
+        __m512i vb[8];
+        __m512 vc[COLS];
+        __m512 vd1;
+
+        // Notes: s8s8 igemm compensation in avx512-vnni
+        // change s8s8 to u8s8 with compensate
+        //   a * b = (a + 128) * b - 128 * b
+        //   s   s       u       s    u    s
+        //
+        // (128 * b is pre-computed when packing B to vnni formats)
+        //
+        const __m512i off = _mm512_set1_epi8(static_cast<char>(0x80));
+
+        auto loadc = [&](auto col) {
+            vc[col] = _mm512_setzero_ps();
+        };
+        Unroll<COLS>{}(loadc);
+
+        auto compute = [&](auto col, auto i) {
+            // load a and add offset 128
+            if constexpr (col == 0) {
+                const int32_t * a_ptr = reinterpret_cast<const int32_t *>(A[0 * KB + i].qs);
+                for (int k = 0; k < 8; ++k) {
+                    va[k] = _mm512_set1_epi32(a_ptr[k]);
+                    va[k] = _mm512_add_epi8(va[k], off);
+                }
+                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
+            }
+
+            // load b
+            const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE);
+            for (int k = 0; k < 8; ++k) {
+                vb[k] = _mm512_loadu_si512((const __m512i *)(b_ptr + k * 64));
+            }
+            const int offset = TILE_N * TILE_K;
+            const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset)));
+            const int offset2 = TILE_N * TILE_K + TILE_N * sizeof(ggml_half);
+            const __m512i vcomp = _mm512_loadu_si512((const __m512i *)(b_ptr + offset2));
+
+            __m512i vsum = _mm512_setzero_si512();
+            for (int k = 0; k < 8; ++k) {
+                vsum = _mm512_dpbusd_epi32(vsum, va[k], vb[k]);
+            }
+            vsum = _mm512_sub_epi32(vsum, vcomp);
+
+            vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(vsum), _mm512_mul_ps(vd0, vd1), vc[col]);
+        };
+
+        for (int i = 0; i < KB; ++i) {
+            Unroll<COLS>{}(compute, i);
+        }
+
+        //store to C
+        auto storec = [&](auto col) {
+            _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
+        };
+        Unroll<COLS>{}(storec);
+    }
+};
+
+template <int BLOCK_M, int BLOCK_N, int BLOCK_K>
+struct tinygemm_kernel_vnni<block_q8_K, block_q4_K, float, BLOCK_M, BLOCK_N, BLOCK_K> {
+    static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
+
+        constexpr int COLS = BLOCK_N / 16;
+        const int TILE_SIZE = TILE_N * sizeof(block_q4_K) + TILE_N * 4;
+
+        const block_q8_K * RESTRICT A = static_cast<const block_q8_K *>(_A);
+        const char * RESTRICT B = static_cast<const char *>(_B);
+
+        // a.qs:   8 groups, 32 bytes each group (m256i)
+        __m512i va[8];
+        // a.bsum: 8 groups,  2 bytes each group (m128i)
+        __m512i va_bsum;
+        __m512 vc[COLS];
+        __m512 vd1;
+
+        // packed_B:
+        const int offset_scales = (QK_K / 2) * TILE_N;
+        const int offset_mins   = (QK_K / 2) * TILE_N +  8 * TILE_N;
+        const int offset_d0     = (QK_K / 2) * TILE_N + 16 * TILE_N;
+        const int offset_dmin   = (QK_K / 2) * TILE_N + 16 * TILE_N + TILE_N * sizeof(ggml_half);
+
+        const __m512i lowMask = _mm512_set1_epi8(0xF);
+
+        auto loadc = [&](auto col) {
+            vc[col] = _mm512_setzero_ps();
+        };
+        Unroll<COLS>{}(loadc);
+
+        // Notes: vnni formats in QK_K
+        //   a) quants vnni format
+        //     int8  {k/4, n, 4}, viewed as 2d {k/4, 4n}, k = 32
+        //     from {16, 32} to {8, 64}
+        //
+        //   b) min vnni format
+        //     int16 {k/2, n, 2}, viewed as 2d {k/2, 2n}, k = 8
+        //     from {16,  8} to {4, 32}
+        //
+        auto compute = [&](auto col, auto i) {
+            // load a
+            if constexpr (col == 0) {
+                for (int k_group = 0; k_group < QK_K / 32; ++k_group) {
+                    va[k_group] = _mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)(A[0 * KB + i].qs + k_group * 32)));
+                }
+                const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums);
+                const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
+                va_bsum = _mm512_castsi128_si512(q8s);
+                vd1 = _mm512_set1_ps(A[0 * KB + i].d);
+            }
+
+            // step 1: accumultate the quants
+            __m512i acc = _mm512_setzero_si512();
+            const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE);
+            const char * b_qs  = b_ptr;
+            for (int k_group = 0; k_group < QK_K / 32; ++k_group) {
+                __m512i vsum = _mm512_setzero_si512();
+                for (int k = 0; k < 8; k += 2) {
+                    __m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 0), va[k_group]);
+                    __m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 1), va[k_group]);
+
+                    __m512i bytes = _mm512_loadu_si512((const __m512i *)b_qs);
+                    __m512i vb0 = _mm512_and_si512(bytes, lowMask);
+                    vsum = _mm512_dpbusd_epi32(vsum, vb0, va0);
+                    __m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
+                    vsum = _mm512_dpbusd_epi32(vsum, vb1, va1);
+
+                    b_qs += 64;
+                }
+                // vacc += scale * (q8 @ q4)
+                const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N)));
+                acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale));
+            }
+            const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0)));
+            vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]);
+
+            // step 2: accumulate the mins
+            __m512i acc_m = _mm512_setzero_si512();
+            for (int k = 0; k < 4; ++k) {
+                __m512i vmask = _mm512_set1_epi32(k);
+                __m512i va = _mm512_permutexvar_epi32(vmask, va_bsum);
+                __m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_mins + k * 32)));
+                acc_m = _mm512_dpwssds_epi32(acc_m, va, vb);
+            }
+            const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_dmin)));
+            vc[col] = _mm512_fnmadd_ps(_mm512_cvtepi32_ps(acc_m), _mm512_mul_ps(vdmin, vd1), vc[col]);
+        };
+
+        for (int i = 0; i < KB; ++i) {
+            Unroll<COLS>{}(compute, i);
+        }
+
+        //store to C
+        auto storec = [&](auto col) {
+            _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
+        };
+        Unroll<COLS>{}(storec);
+    }
+};
+
+template <int BLOCK_M, int BLOCK_N, int BLOCK_K>
+struct tinygemm_kernel_vnni<block_q8_K, block_q5_K, float, BLOCK_M, BLOCK_N, BLOCK_K> {
+    static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
+
+        constexpr int COLS = BLOCK_N / 16;
+        const int TILE_SIZE = TILE_N * sizeof(block_q5_K) + TILE_N * 4;
+
+        const block_q8_K * RESTRICT A = static_cast<const block_q8_K *>(_A);
+        const char * RESTRICT B = static_cast<const char *>(_B);
+
+        // a.qs:   8 groups, 32 bytes each group (m256i)
+        __m512i va[8];
+        // a.bsum: 8 groups,  2 bytes each group (m128i)
+        __m512i va_bsum;
+        __m512 vc[COLS];
+        __m512 vd1;
+
+        // packed_B:
+        const int offset_qh     = (QK_K / 2) * TILE_N;
+        const int offset_scales = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N;
+        const int offset_mins   = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N +  8 * TILE_N;
+        const int offset_d0     = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N + 16 * TILE_N;
+        const int offset_dmin   = (QK_K / 2) * TILE_N + (QK_K / 8) * TILE_N + 16 * TILE_N + TILE_N * sizeof(ggml_half);
+
+        const __m512i lowMask = _mm512_set1_epi8(0xF);
+
+        auto loadc = [&](auto col) {
+            vc[col] = _mm512_setzero_ps();
+        };
+        Unroll<COLS>{}(loadc);
+
+        // Q5_K and Q4_K shares the same vnni formats, refer to notes above.
+        auto compute = [&](auto col, auto i) {
+            // load a
+            if constexpr (col == 0) {
+                for (int k_group = 0; k_group < QK_K / 32; ++k_group) {
+                    va[k_group] = _mm512_castsi256_si512(_mm256_loadu_si256((const __m256i *)(A[0 * KB + i].qs + k_group * 32)));
+                }
+                const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums);
+                const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
+                va_bsum = _mm512_castsi128_si512(q8s);
+                vd1 = _mm512_set1_ps(A[0 * KB + i].d);
+            }
+
+            // step 1: accumultate the quants
+            __m512i acc = _mm512_setzero_si512();
+            const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE);
+            const char * b_qs  = b_ptr;
+            const char * b_qh  = b_ptr + offset_qh;
+            for (int k_group = 0; k_group < QK_K / 32; ++k_group) {
+                __m512i vsum = _mm512_setzero_si512();
+                __m512i hmask0 = _mm512_set1_epi8(0x1);
+                __m512i hmask1 = _mm512_set1_epi8(0x2);
+                __m512i hbits = _mm512_loadu_si512((const __m512i *)(b_qh + k_group * 64));
+                for (int k = 0; k < 8; k += 2) {
+                    __m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 0), va[k_group]);
+                    __m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(k + 1), va[k_group]);
+
+                    __m512i bytes = _mm512_loadu_si512((const __m512i *)b_qs);
+                    __m512i vb0 = _mm512_and_si512(bytes, lowMask);
+                    __m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
+
+                    __m512i vh0 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask0), k), 4);
+                    __m512i vh1 = _mm512_slli_epi16(_mm512_srli_epi16(_mm512_and_si512(hbits, hmask1), k + 1), 4);
+
+                    hmask0 = _mm512_slli_epi16(hmask0, 2);
+                    hmask1 = _mm512_slli_epi16(hmask1, 2);
+                    vb0 = _mm512_add_epi8(vb0, vh0);
+                    vb1 = _mm512_add_epi8(vb1, vh1);
+
+                    vsum = _mm512_dpbusd_epi32(vsum, vb0, va0);
+                    vsum = _mm512_dpbusd_epi32(vsum, vb1, va1);
+
+                    b_qs += 64;
+                }
+                // vacc += scale * (q8 @ q5)
+                const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N)));
+                acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale));
+            }
+            const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0)));
+            vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]);
+
+            // step 2: accumulate the mins
+            __m512i acc_m = _mm512_setzero_si512();
+            for (int k = 0; k < 4; ++k) {
+                __m512i vmask = _mm512_set1_epi32(k);
+                __m512i va = _mm512_permutexvar_epi32(vmask, va_bsum);
+                __m512i vb = _mm512_cvtepi8_epi16(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_mins + k * 32)));
+                acc_m = _mm512_dpwssds_epi32(acc_m, va, vb);
+            }
+            const __m512 vdmin = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_dmin)));
+            vc[col] = _mm512_fnmadd_ps(_mm512_cvtepi32_ps(acc_m), _mm512_mul_ps(vdmin, vd1), vc[col]);
+        };
+
+        for (int i = 0; i < KB; ++i) {
+            Unroll<COLS>{}(compute, i);
+        }
+
+        //store to C
+        auto storec = [&](auto col) {
+            _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
+        };
+        Unroll<COLS>{}(storec);
+    }
+};
+
+template <int BLOCK_M, int BLOCK_N, int BLOCK_K>
+struct tinygemm_kernel_vnni<block_q8_K, block_q6_K, float, BLOCK_M, BLOCK_N, BLOCK_K> {
+    static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
+
+        constexpr int COLS = BLOCK_N / 16;
+        const int TILE_SIZE = TILE_N * sizeof(block_q6_K);
+
+        const block_q8_K * RESTRICT A = static_cast<const block_q8_K *>(_A);
+        const char * RESTRICT B = static_cast<const char *>(_B);
+
+        // load the 256 bytes from A to 4 avx512 vectors
+        __m512i va[4];
+        __m512 vc[COLS];
+        __m512 vd1;
+
+        // packed_B:
+        const int offset_qh     = (QK_K / 2) * TILE_N;
+        const int offset_scales = (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N;
+        const int offset_d0     = (QK_K / 2) * TILE_N + (QK_K / 4) * TILE_N + 16 * TILE_N;
+
+        // compensation
+        __m512i vcomp;
+
+        const __m512i m32s = _mm512_set1_epi32(32);
+        const __m512i lowMask = _mm512_set1_epi8(0xF);
+
+        auto loadc = [&](auto col) {
+            vc[col] = _mm512_setzero_ps();
+        };
+        Unroll<COLS>{}(loadc);
+
+        auto compute = [&](auto col, auto i) {
+            if constexpr (col == 0) {
+                // load a
+                va[0] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs +   0));
+                va[1] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs +  64));
+                va[2] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 128));
+                va[3] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 192));
+
+                const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums);
+                vcomp = _mm512_mullo_epi32(_mm512_cvtepi16_epi32(q8sums), m32s);
+                vd1 = _mm512_set1_ps(A[0 * KB + i].d);
+            }
+
+            // accmulate the quants
+            __m512i acc = _mm512_setzero_si512();
+            const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE);
+            const char * b_qs = b_ptr;
+            const char * b_qh = b_ptr + offset_qh;
+            int mask = 0;
+            for (int k_group = 0; k_group < QK_K / 16; ++k_group) {
+                int r = k_group >> 2;
+                __m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]);
+                __m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]);
+
+                __m512i vsum = _mm512_setzero_si512();
+                __m512i hmask = _mm512_set1_epi8(0x3);
+
+                __m512i bytes = _mm512_loadu_si512(b_qs);
+                __m512i hbits = _mm512_loadu_si512(b_qh);
+                __m512i vb0 = _mm512_and_si512(bytes, lowMask);
+                __m512i vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
+                __m512i vh0 = _mm512_slli_epi16(_mm512_and_si512(hbits, hmask), 4);
+                __m512i vh1 = _mm512_slli_epi16(_mm512_and_si512(hbits, _mm512_slli_epi16(hmask, 2)), 2);
+
+                vb0 = _mm512_add_epi8(vb0, vh0);
+                vb1 = _mm512_add_epi8(vb1, vh1);
+                vsum = _mm512_dpbusd_epi32(vsum, vb0, va0);
+                vsum = _mm512_dpbusd_epi32(vsum, vb1, va1);
+                b_qs += 64;
+
+                va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]);
+                va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]);
+
+                bytes = _mm512_loadu_si512(b_qs);
+                vb0 = _mm512_and_si512(bytes, lowMask);
+                vb1 = _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask);
+                vh0 =                   _mm512_and_si512(hbits, _mm512_slli_epi16(hmask, 4));
+                vh1 = _mm512_srli_epi16(_mm512_and_si512(hbits, _mm512_slli_epi16(hmask, 6)), 2);
+                vb0 = _mm512_add_epi8(vb0, vh0);
+                vb1 = _mm512_add_epi8(vb1, vh1);
+                vsum = _mm512_dpbusd_epi32(vsum, vb0, va0);
+                vsum = _mm512_dpbusd_epi32(vsum, vb1, va1);
+                b_qs += 64;
+                b_qh += 64;
+
+                // B * A - 32 * A
+                __m512i vmask = _mm512_set1_epi32(k_group);
+                vsum = _mm512_sub_epi32(vsum, _mm512_permutexvar_epi32(vmask, vcomp));
+
+                // vacc += scale * (q8 @ q6)
+                const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N)));
+                acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale));
+            }
+            const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0)));
+            vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]);
+        };
+
+        for (int i = 0; i < KB; ++i) {
+            Unroll<COLS>{}(compute, i);
+        }
+
+        //store to C
+        auto storec = [&](int col) {
+            _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
+        };
+        Unroll<COLS>{}(storec);
+    }
+};
+
+template <int BLOCK_M, int BLOCK_N, int BLOCK_K>
+struct tinygemm_kernel_vnni<block_q8_K, block_iq4_xs, float, BLOCK_M, BLOCK_N, BLOCK_K> {
+    static void apply(int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
+
+        constexpr int COLS = BLOCK_N / 16;
+        const int TILE_SIZE = TILE_N * sizeof(block_iq4_xs) + TILE_N * 2;
+
+        const block_q8_K * RESTRICT A = static_cast<const block_q8_K *>(_A);
+        const char * RESTRICT B = static_cast<const char *>(_B);
+
+        // load the 256 bytes from A to 4 avx512 vectors
+        __m512i va[4];
+        __m512 vc[COLS];
+        __m512 vd1;
+
+        // packed_B:
+        const int offset_scales = (QK_K / 2) * TILE_N ;
+        const int offset_d0     = (QK_K / 2) * TILE_N + 8 * TILE_N;
+
+        // compensation
+        __m512i vcomp;
+
+        const __m256i m128s = _mm256_set1_epi16(128);
+        const __m512i lowMask = _mm512_set1_epi8(0xF);
+
+        const __m512i values128 = _mm512_set_epi8(
+            113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127,
+            113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127,
+            113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127,
+            113, 89, 69, 53, 38, 25, 13, 1, -10, -22, -35, -49, -65, -83, -104, -127
+        );
+        const __m512i off = _mm512_set1_epi8(static_cast<char>(0x80));
+        const __m512i values256 = _mm512_add_epi8(values128, off);
+
+        auto loadc = [&](auto col) {
+            vc[col] = _mm512_setzero_ps();
+        };
+        Unroll<COLS>{}(loadc);
+
+        auto compute = [&](auto col, auto i) {
+            if constexpr (col == 0) {
+                // load a
+                va[0] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs +   0));
+                va[1] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs +  64));
+                va[2] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 128));
+                va[3] = _mm512_loadu_si512((const __m512i *)(A[0 * KB + i].qs + 192));
+
+                // compensation: 128 * A
+                const __m256i q8sums = _mm256_loadu_si256((const __m256i *)A[0 * KB + i].bsums);
+                vcomp = _mm512_castsi256_si512(_mm256_madd_epi16(q8sums, m128s));
+                vd1 = _mm512_set1_ps(A[0 * KB + i].d);
+            }
+
+            // accmulate the quants
+            __m512i acc = _mm512_setzero_si512();
+            const char * b_ptr = B + PACKED_INDEX(col, i, KB, TILE_SIZE);
+            const char * b_qs = b_ptr;
+            int mask = 0;
+            for (int k_group = 0; k_group < QK_K / 32; ++k_group) {
+                int r = k_group >> 1;
+                __m512i vmask = _mm512_set1_epi32(k_group);
+                __m512i vsum = _mm512_setzero_si512();
+                for (int k = 0; k < 8; k += 2) {
+                    __m512i va0 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]);
+                    __m512i va1 = _mm512_permutexvar_epi32(_mm512_set1_epi32(mask++), va[r]);
+
+                    __m512i bytes = _mm512_loadu_si512(b_qs);
+                    __m512i vb0 = _mm512_shuffle_epi8(values256, _mm512_and_si512(bytes, lowMask));
+                    __m512i vb1 = _mm512_shuffle_epi8(values256, _mm512_and_si512(_mm512_srli_epi16(bytes, 4), lowMask));
+
+                    vsum = _mm512_dpbusd_epi32(vsum, vb0, va0);
+                    vsum = _mm512_dpbusd_epi32(vsum, vb1, va1);
+                    b_qs += 64;
+                }
+                // (B + 128) * A - 128 * A
+                vsum = _mm512_sub_epi32(vsum, _mm512_permutexvar_epi32(vmask, vcomp));
+
+                // vacc += scale * (q8 @ q4)
+                const __m512i vscale = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i *)(b_ptr + offset_scales + k_group * TILE_N)));
+                acc = _mm512_add_epi32(acc, _mm512_mullo_epi32(vsum, vscale));
+            }
+            const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(b_ptr + offset_d0)));
+            vc[col] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(acc), _mm512_mul_ps(vd0, vd1), vc[col]);
+        };
+
+        for (int i = 0; i < KB; ++i) {
+            Unroll<COLS>{}(compute, i);
+        }
+
+        //store to C
+        auto storec = [&](auto col) {
+            _mm512_storeu_ps((__m512i*)(C + 0 * ldc + col * 16), vc[col]);
+        };
+        Unroll<COLS>{}(storec);
+    }
+};
+
+#define LAUNCH_TINYGEMM_KERNEL_VNNI(NB_SIZE)                                         \
+    tinygemm_kernel_vnni<vec_dot_type, type, float, 1, NB_SIZE, blck_size>::apply(   \
+        KB, (const char *)wdata + 0 * row_size_A,                                    \
+        (const char *)src0->data + PACKED_INDEX(nb * kTilesN, 0, KB, TILE_SIZE),     \
+        (float *) dst->data + 0 * N + nb_start, ldc)
+
+template <typename TA, typename TB, typename TC, int BLOCK_K,
+          typename std::enable_if<!is_type_qkk<TB>::value, int>::type = 0>
+void tinygemm_kernel_amx(int M, int N, int KB, const void * RESTRICT _A, const void * RESTRICT _B, TC * RESTRICT C, int ldc) {
+    using packed_B_t = packed_B_type<TB>;
+    const int TILE_SIZE = get_tile_size<TB>();
+    const bool need_unpack = do_unpack<TB>::value;
+
+    GGML_ASSERT(M <= 2 * TILE_M && N == 2 * TILE_N);
+    const TA * RESTRICT A = static_cast<const TA *>(_A);
+    const char * RESTRICT B = static_cast<const char *>(_B);
+
+    const int m0 = std::min(M, TILE_M);
+    const int m1 = std::max(M - TILE_M, 0);
+    const int lda = KB * sizeof(TA);
+    //const int ldb = KB * sizeof(TB);
+
+    static thread_local packed_B_t Tile0[TILE_N * TILE_K];
+    static thread_local packed_B_t Tile1[TILE_N * TILE_K];
+    static thread_local int8_t Tile23[TILE_M * TILE_K];
+
+    static thread_local int32_t TileC0[TILE_M * TILE_N * 4];
+    static thread_local int32_t TileC1[TILE_M * TILE_N * 4];
+
+    // double buffering C to interleave avx512 and amx
+    int32_t * C_cur = TileC0;
+    int32_t * C_pre = TileC1;
+
+    auto Tile4 = [&](int32_t * base) { return base; };
+    auto Tile5 = [&](int32_t * base) { return base + TILE_M * TILE_N; };
+    auto Tile6 = [&](int32_t * base) { return base + 2 * TILE_M * TILE_N; };
+    auto Tile7 = [&](int32_t * base) { return base + 3 * TILE_M * TILE_N; };
+
+    if (M == 2 * TILE_M) {
+        // i = 0
+        const char * B_blk0 = B + PACKED_INDEX(0, 0, KB, TILE_SIZE);
+        const char * B_blk1 = B + PACKED_INDEX(1, 0, KB, TILE_SIZE);
+        if (need_unpack) {
+            unpack_B<TB>(Tile0, B_blk0);
+            _tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK);
+        } else {
+            _tile_loadd(TMM0, B_blk0, TILE_N * VNNI_BLK);
+        }
+
+        _tile_zero(TMM4);
+        _tile_loadd(TMM2, A[0].qs, lda);
+        _tile_dpbssd(TMM4, TMM2, TMM0);
+        _tile_stored(TMM4, Tile4(C_pre), TILE_N * sizeof(int32_t));
+
+        _tile_zero(TMM5);
+        _tile_loadd(TMM3, A[TILE_M * KB + 0].qs, lda);
+        _tile_dpbssd(TMM5, TMM3, TMM0);
+        _tile_stored(TMM5, Tile5(C_pre), TILE_N * sizeof(int32_t));
+
+        if (need_unpack) {
+            unpack_B<TB>(Tile1, B_blk0);
+            _tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK);
+        } else {
+            _tile_loadd(TMM1, B_blk1, TILE_N * VNNI_BLK);
+        }
+
+        _tile_zero(TMM6);
+        _tile_dpbssd(TMM6, TMM2, TMM1);
+        _tile_stored(TMM6, Tile6(C_pre), TILE_N * sizeof(int32_t));
+
+        _tile_zero(TMM7);
+        _tile_dpbssd(TMM7, TMM3, TMM1);
+        _tile_stored(TMM7, Tile7(C_pre), TILE_N * sizeof(int32_t));
+
+        for (int i = 1; i < KB; ++i) {
+            // index of previous iter
+            const int ii = i - 1;
+            const char * B_blk0 = B + PACKED_INDEX(0, i, KB, TILE_SIZE);
+            const char * B_blk1 = B + PACKED_INDEX(1, i, KB, TILE_SIZE);
+            GGML_DISPATCH_BOOL(ii > 0, is_acc, [&] {
+                if (need_unpack) {
+                    unpack_B<TB>(Tile0, B_blk0);
+                    _tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK);
+                } else {
+                    _tile_loadd(TMM0, B_blk0, TILE_N * VNNI_BLK);
+                }
+                _tile_zero(TMM4);
+                _tile_loadd(TMM2, A[i].qs, lda);
+                acc_C<TA, TB, is_acc>::apply(C, ldc, Tile4(C_pre), &A[ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M);
+
+                _tile_dpbssd(TMM4, TMM2, TMM0);
+                _tile_stored(TMM4, Tile4(C_cur), TILE_N * sizeof(int32_t));
+
+                _tile_zero(TMM5);
+                _tile_loadd(TMM3, A[TILE_M * KB + i].qs, lda);
+                acc_C<TA, TB, is_acc>::apply(C + TILE_M * ldc, ldc, Tile5(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M);
+
+                _tile_dpbssd(TMM5, TMM3, TMM0);
+                _tile_stored(TMM5, Tile5(C_cur), TILE_N * sizeof(int32_t));
+
+                if (need_unpack) {
+                    unpack_B<TB>(Tile1, B_blk1);
+                    _tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK);
+                } else {
+                    _tile_loadd(TMM1, B_blk1, TILE_N * VNNI_BLK);
+                }
+                _tile_zero(TMM6);
+                acc_C<TA, TB, is_acc>::apply(C + TILE_N, ldc, Tile6(C_pre), &A[ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M);
+
+                _tile_dpbssd(TMM6, TMM2, TMM1);
+                _tile_stored(TMM6, Tile6(C_cur), TILE_N * sizeof(int32_t));
+
+                _tile_zero(TMM7);
+                acc_C<TA, TB, is_acc>::apply(C + TILE_M * ldc + TILE_N, ldc, Tile7(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M);
+
+                _tile_dpbssd(TMM7, TMM3, TMM1);
+                _tile_stored(TMM7, Tile7(C_cur), TILE_N * sizeof(int32_t));
+
+                std::swap(C_cur, C_pre);
+            });
+        }
+        // final accumulation
+        {
+            int ii = KB - 1;
+            acc_C<TA, TB, true>::apply(C, ldc, Tile4(C_pre), &A[ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M);
+            acc_C<TA, TB, true>::apply(C + TILE_M * ldc, ldc, Tile5(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(0, ii, KB, TILE_SIZE), TILE_M);
+            acc_C<TA, TB, true>::apply(C + TILE_N, ldc, Tile6(C_pre), &A[ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M);
+            acc_C<TA, TB, true>::apply(C + TILE_M * ldc + TILE_N, ldc, Tile7(C_pre), &A[TILE_M * KB + ii], KB, B + PACKED_INDEX(1, ii, KB, TILE_SIZE), TILE_M);
+        }
+    } else {
+        for (int i = 0; i < KB; ++i) {
+            _tile_zero(TMM4);
+            _tile_zero(TMM6);
+            if (m1 != 0) {
+                _tile_zero(TMM5);
+                _tile_zero(TMM7);
+            }
+
+            const char * B_blk0 = B + PACKED_INDEX(0, i, KB, TILE_SIZE);
+            const char * B_blk1 = B + PACKED_INDEX(1, i, KB, TILE_SIZE);
+            if (need_unpack) {
+                unpack_B<TB>(Tile0, B_blk0);
+                _tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK);
+            } else {
+                _tile_loadd(TMM0, B_blk0, TILE_N * VNNI_BLK);
+            }
+
+            if (need_unpack) {
+                unpack_B<TB>(Tile1, B_blk1);
+                _tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK);
+            } else {
+                _tile_loadd(TMM1, B_blk1, TILE_N * VNNI_BLK);
+            }
+
+            if (m0 == TILE_M) {
+                _tile_loadd(TMM2, A[i].qs, lda);
+            } else {
+                unpack_A(Tile23, &A[i], KB, m0);
+                _tile_loadd(TMM2, Tile23, TILE_K);
+            }
+
+            _tile_dpbssd(TMM4, TMM2, TMM0);
+            _tile_dpbssd(TMM6, TMM2, TMM1);
+
+            _tile_stored(TMM4, Tile4(C_cur), TILE_N * sizeof(int32_t));
+            _tile_stored(TMM6, Tile6(C_cur), TILE_N * sizeof(int32_t));
+
+            GGML_DISPATCH_BOOL(i > 0, is_acc, [&] {
+                acc_C<TA, TB, is_acc>::apply(C,          ldc, Tile4(C_cur), &A[i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m0);
+                acc_C<TA, TB, is_acc>::apply(C + TILE_N, ldc, Tile6(C_cur), &A[i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m0);
+            });
+
+            if (m1 != 0) {
+                unpack_A(Tile23, &A[TILE_M * KB + i], KB, m1);
+                _tile_loadd(TMM3, Tile23, TILE_K);
+
+                _tile_dpbssd(TMM5, TMM3, TMM0);
+                _tile_dpbssd(TMM7, TMM3, TMM1);
+                _tile_stored(TMM5, Tile5(C_cur), TILE_N * sizeof(int32_t));
+                _tile_stored(TMM7, Tile7(C_cur), TILE_N * sizeof(int32_t));
+                GGML_DISPATCH_BOOL(i > 0, is_acc, [&] {
+                    acc_C<TA, TB, is_acc>::apply(C + TILE_M * ldc,          ldc, Tile5(C_cur), &A[TILE_M * KB + i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m1);
+                    acc_C<TA, TB, is_acc>::apply(C + TILE_M * ldc + TILE_N, ldc, Tile7(C_cur), &A[TILE_M * KB + i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m1);
+                });
+            }
+        }
+    }
+    return;
+}
+
+template <typename TA, typename TB, typename TC, int BLOCK_K,
+          typename std::enable_if<is_type_qkk<TB>::value, int>::type = 0>
+void tinygemm_kernel_amx(int M, int N, int KB, const void * RESTRICT _A, const void * RESTRICT _B, float * RESTRICT C, int ldc) {
+    static_assert(std::is_same<TA, block_q8_K>::value);
+    const int TILE_SIZE = get_tile_size<TB>();
+
+    GGML_ASSERT(M <= 2 * TILE_M && N == 2 * TILE_N);
+    const TA * RESTRICT A = static_cast<const TA *>(_A);
+    const char * RESTRICT B = static_cast<const char *>(_B);
+
+    const int m0 = std::min(M, TILE_M);
+    const int m1 = std::max(M - TILE_M, 0);
+    //const int lda = KB * sizeof(TA);
+
+    static thread_local int8_t Tile0[TILE_N * TILE_K];
+    static thread_local int8_t Tile1[TILE_N * TILE_K];
+    static thread_local int8_t Tile23[TILE_M * TILE_K];
+
+    // mat mul result for each group
+    static thread_local int32_t Tile4[TILE_M * TILE_N];
+    static thread_local int32_t Tile5[TILE_M * TILE_N];
+    static thread_local int32_t Tile6[TILE_M * TILE_N];
+    static thread_local int32_t Tile7[TILE_M * TILE_N];
+
+    // sum of each QK_K block, contains 8 groups, int32
+    static thread_local int32_t Sumi4[TILE_M * TILE_N];
+    static thread_local int32_t Sumi5[TILE_M * TILE_N];
+    static thread_local int32_t Sumi6[TILE_M * TILE_N];
+    static thread_local int32_t Sumi7[TILE_M * TILE_N];
+
+    const int k_group_size = std::is_same<TB, block_q6_K>::value ? 16 : 32;
+    for (int i = 0; i < KB; ++i) {
+        // step 1: accumulate the quants across 8 groups, each group with 32
+        for (int k = 0; k < QK_K / k_group_size; ++k) {
+            GGML_DISPATCH_BOOL(k > 0, is_acc, [&] {
+                _tile_zero(TMM4);
+                _tile_zero(TMM6);
+
+                unpack_B<TB>(Tile0, B + PACKED_INDEX(0, i, KB, TILE_SIZE), k);
+                _tile_loadd(TMM0, Tile0, TILE_N * VNNI_BLK);
+
+                unpack_B<TB>(Tile1, B + PACKED_INDEX(1, i, KB, TILE_SIZE), k);
+                _tile_loadd(TMM1, Tile1, TILE_N * VNNI_BLK);
+
+                unpack_A<TB>(Tile23, &A[i], KB, k, m0);
+                _tile_loadd(TMM2, Tile23, TILE_K);
+
+                _tile_dpbssd(TMM4, TMM2, TMM0);
+                _tile_dpbssd(TMM6, TMM2, TMM1);
+
+                _tile_stored(TMM4, Tile4, TILE_N * sizeof(int32_t));
+                _tile_stored(TMM6, Tile6, TILE_N * sizeof(int32_t));
+
+                scale_C<TB, is_acc>(Tile4, Sumi4, B + PACKED_INDEX(0, i, KB, TILE_SIZE), k, m0);
+                scale_C<TB, is_acc>(Tile6, Sumi6, B + PACKED_INDEX(1, i, KB, TILE_SIZE), k, m0);
+
+                if (m1 != 0) {
+                    _tile_zero(TMM5);
+                    _tile_zero(TMM7);
+
+                    unpack_A<TB>(Tile23, &A[TILE_M * KB + i], KB, k, m1);
+                    _tile_loadd(TMM3, Tile23, TILE_K);
+
+                    _tile_dpbssd(TMM5, TMM3, TMM0);
+                    _tile_dpbssd(TMM7, TMM3, TMM1);
+
+                    _tile_stored(TMM5, Tile5, TILE_N * sizeof(int32_t));
+                    _tile_stored(TMM7, Tile7, TILE_N * sizeof(int32_t));
+
+                    scale_C<TB, is_acc>(Tile5, Sumi5, B + PACKED_INDEX(0, i, KB, TILE_SIZE), k, m1);
+                    scale_C<TB, is_acc>(Tile7, Sumi7, B + PACKED_INDEX(1, i, KB, TILE_SIZE), k, m1);
+                }
+            });
+        }
+
+        // step 2: accmulate the mins
+        GGML_DISPATCH_BOOL(i > 0, is_acc, [&] {
+            acc_C<TA, TB, is_acc>::apply(C,          ldc, Sumi4, &A[i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m0);
+            acc_C<TA, TB, is_acc>::apply(C + TILE_N, ldc, Sumi6, &A[i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m0);
+            if (m1 != 0) {
+                acc_C<TA, TB, is_acc>::apply(C + TILE_M * ldc,          ldc, Sumi5, &A[TILE_M * KB + i], KB, B + PACKED_INDEX(0, i, KB, TILE_SIZE), m1);
+                acc_C<TA, TB, is_acc>::apply(C + TILE_M * ldc + TILE_N, ldc, Sumi7, &A[TILE_M * KB + i], KB, B + PACKED_INDEX(1, i, KB, TILE_SIZE), m1);
+            }
+        });
+    }
+    return;
+}
+
+} // anonymous namespace
+
+// get the packed tensor size for quantized weights
+size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor) {
+    const enum ggml_type TYPE = tensor->type;
+
+    const int K = tensor->ne[0]; // ne0: in_features
+    const int N = tensor->ne[1]; // ne1: out_features
+
+    auto get_tensor_size = [&] {
+        size_t row_size_B{0};
+        GGML_DISPATCH_QTYPES(TYPE, [&] {
+            row_size_B = get_row_size<type, blck_size>(K);
+        });
+        return N * row_size_B;
+    };
+
+    if (qtype_has_amx_kernels(TYPE)) {
+        return get_tensor_size();
+    } else {
+        // for f16, bf16 we don't do packing
+        return ggml_nbytes(tensor);
+    }
+}
+
+// pack weight to vnni format
+void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset == 0 && size == ggml_nbytes(tensor)); // only full tensor conversion is supported for now
+
+    const enum ggml_type TYPE = tensor->type;
+
+    const int K = tensor->ne[0]; // ne0: in_features
+    const int N = tensor->ne[1]; // ne1: out_features
+
+    GGML_DISPATCH_QTYPES(TYPE, [&] {
+        convert_B_packed_format<type, blck_size>((void *)((char *)tensor->data + offset), (const type *)data, N, K);
+    });
+}
+
+size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst) {
+    struct ggml_tensor * src0 = dst->src[0];
+
+    const enum ggml_type TYPE = src0->type;
+
+    const bool is_floating_type = TYPE == GGML_TYPE_F16;
+    if (is_floating_type) {
+        return 0;
+    }
+
+    const int M = dst->ne[1];
+    const int K = src0->ne[0];
+
+    size_t desired_wsize = 0;
+
+    GGML_DISPATCH_QTYPES(TYPE, [&] {
+        const size_t row_size_A = K / blck_size * sizeof(vec_dot_type);
+        desired_wsize = M * row_size_A;
+    });
+
+    return desired_wsize;
+}
+
+// NB: mixed dtype gemm with Advanced Matrix Extensions (Intel AMX)
+//
+// src0: weight in shape of {N, K}, quantized
+// src1: input  in shape of {M, K}, float32
+// dst:  output in shape of {M, N}, float32
+//
+// the function performs: dst = src1 @ src0.T
+//
+void ggml_backend_amx_mul_mat(const ggml_compute_params * params, struct ggml_tensor * dst) {
+    struct ggml_tensor * src0 = dst->src[0];
+    struct ggml_tensor * src1 = dst->src[1];
+
+    const enum ggml_type TYPE = src0->type;
+
+    // f16 only has avx512 kernels for now,
+    // amx kernels will be added once 6th gen xeon is released.
+    const bool is_floating_type = TYPE == GGML_TYPE_F16;
+
+    const int M = dst->ne[1];
+    const int N = dst->ne[0];
+    const int K = src0->ne[0];
+    const int ldc = dst->nb[1] / dst->nb[0];
+
+    if (is_floating_type) {
+        constexpr int BLOCK_M = 4;
+        constexpr int BLOCK_N = 6;
+        const int MB = div_up(M, BLOCK_M);
+        const int NB = div_up(N, BLOCK_N);
+
+        parallel_for_ggml(params, MB * NB, [&](int begin, int end) {
+            GGML_DISPATCH_FLOATING_TYPES(TYPE, [&] {
+                for (int i = begin; i < end; ++i) {
+                    int mb = i / NB;
+                    int nb = i % NB;
+
+                    int mb_start = mb * BLOCK_M;
+                    int mb_size = std::min(BLOCK_M, M - mb_start);
+                    int nb_start = nb * BLOCK_N;
+                    int nb_size = std::min(BLOCK_N, N - nb_start);
+
+                    switch (mb_size << 4 | nb_size) {
+                        case 0x12: LAUNCH_TINYGEMM_KERNEL_AVX(1, 2); break;
+                        case 0x14: LAUNCH_TINYGEMM_KERNEL_AVX(1, 4); break;
+                        case 0x16: LAUNCH_TINYGEMM_KERNEL_AVX(1, 6); break;
+                        case 0x22: LAUNCH_TINYGEMM_KERNEL_AVX(2, 2); break;
+                        case 0x24: LAUNCH_TINYGEMM_KERNEL_AVX(2, 4); break;
+                        case 0x26: LAUNCH_TINYGEMM_KERNEL_AVX(2, 6); break;
+                        case 0x32: LAUNCH_TINYGEMM_KERNEL_AVX(3, 2); break;
+                        case 0x34: LAUNCH_TINYGEMM_KERNEL_AVX(3, 4); break;
+                        case 0x36: LAUNCH_TINYGEMM_KERNEL_AVX(3, 6); break;
+                        case 0x42: LAUNCH_TINYGEMM_KERNEL_AVX(4, 2); break;
+                        case 0x44: LAUNCH_TINYGEMM_KERNEL_AVX(4, 4); break;
+                        case 0x46: LAUNCH_TINYGEMM_KERNEL_AVX(4, 6); break;
+                        default: fprintf(stderr, "Unexpected block size!\n");
+                    }
+                }
+            });
+        });
+        return;
+    }
+
+    // pointer to work space, used convert A from float to quantized type
+    void * wdata = params->wdata;
+
+    //TODO: performance improvement: merge quant A
+    if (params->ith == 0) {
+        GGML_DISPATCH_QTYPES(TYPE, [&] {
+            const size_t row_size_A = K / blck_size * sizeof(vec_dot_type);
+            const size_t desired_wsize = M * row_size_A;
+            if (params->wsize < desired_wsize) {
+                GGML_ABORT("insufficient work space size");
+            }
+
+            // Q4_0, Q4_1, Q8_0 handles 1 TILE_K per blck_size
+            // Q4_K, Q5_K, Q6_K, IQ4_XS handles 8 TILE_K per blck_size
+            GGML_ASSERT(TILE_K == blck_size || TILE_K * 8 == blck_size);
+
+            const float * A_data = static_cast<const float *>(src1->data);
+            for (int m = 0; m < M; ++m) {
+                from_float<vec_dot_type>(A_data + m * K, (char *)wdata + m * row_size_A, K);
+            }
+        });
+    }
+
+    ggml_barrier(params->threadpool);
+
+    if (M == 1) {
+        // MB = 1 and handle 8 tiles in each block
+        constexpr int kTilesN = 4;
+        constexpr int BLOCK_N = TILE_N * kTilesN;
+        const int NB = div_up(N, BLOCK_N);
+
+        parallel_for_ggml(params, NB, [&](int begin, int end) {
+            GGML_DISPATCH_QTYPES(TYPE, [&] {
+                const int KB = K / blck_size;
+                const int TILE_SIZE = get_tile_size<type>();
+                const int row_size_A = KB * sizeof(vec_dot_type);
+                for (int i = begin; i < end; ++i) {
+                    int nb = i;
+                    int nb_start = nb * BLOCK_N;
+                    int nb_size = std::min(BLOCK_N, N - nb_start); // 32, 64, 96
+
+                    switch (nb_size) {
+                        //case 160: LAUNCH_TINYGEMM_KERNEL_VNNI(160); break;
+                        case 128: LAUNCH_TINYGEMM_KERNEL_VNNI(128); break;
+                        case 96: LAUNCH_TINYGEMM_KERNEL_VNNI(96); break;
+                        case 64: LAUNCH_TINYGEMM_KERNEL_VNNI(64); break;
+                        case 32: LAUNCH_TINYGEMM_KERNEL_VNNI(32); break;
+                        default: fprintf(stderr, "Unexpected n block size!\n");
+                    }
+                }
+            });
+        });
+        return;
+    }
+
+    // handle 4 tiles at a tile
+    constexpr int BLOCK_M = TILE_M * 2;
+    constexpr int BLOCK_N = TILE_N * 2;
+    const int MB = div_up(M, BLOCK_M);
+    const int NB = div_up(N, BLOCK_N);
+
+    parallel_for_ggml(params, MB * NB, [&](int begin, int end) {
+        // init tile config for each thread
+        ggml_tile_config_init();
+
+        GGML_DISPATCH_QTYPES(TYPE, [&] {
+            const int KB = K / blck_size;
+            const int TILE_SIZE = get_tile_size<type>();
+            const int row_size_A = KB * sizeof(vec_dot_type);
+
+            for (int i = begin; i < end; ++i) {
+                int mb = i / NB;
+                int nb = i % NB;
+
+                int mb_start = mb * BLOCK_M;
+                int mb_size = std::min(BLOCK_M, M - mb_start);
+                int nb_start = nb * BLOCK_N;
+                int nb_size = BLOCK_N;
+
+                tinygemm_kernel_amx<vec_dot_type, type, float, blck_size>(
+                    mb_size, nb_size, KB,
+                    (const char *)wdata + mb_start * row_size_A,
+                    (const char *)src0->data + PACKED_INDEX(nb * 2, 0, KB, TILE_SIZE),
+                    (float *) dst->data + mb_start * N + nb_start, ldc);
+            }
+        });
+    });
+}
+
+#endif // if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
diff --git a/ggml/src/ggml-cpu/amx/mmq.h b/ggml/src/ggml-cpu/amx/mmq.h
new file mode 100644
index 00000000000..baf76847734
--- /dev/null
+++ b/ggml/src/ggml-cpu/amx/mmq.h
@@ -0,0 +1,10 @@
+#pragma once
+#include "common.h"
+
+size_t ggml_backend_amx_desired_wsize(const struct ggml_tensor * dst);
+
+size_t ggml_backend_amx_get_alloc_size(const struct ggml_tensor * tensor);
+
+void ggml_backend_amx_convert_weight(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+
+void ggml_backend_amx_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
diff --git a/ggml/cmake/FindSIMD.cmake b/ggml/src/ggml-cpu/cmake/FindSIMD.cmake
similarity index 100%
rename from ggml/cmake/FindSIMD.cmake
rename to ggml/src/ggml-cpu/cmake/FindSIMD.cmake
diff --git a/ggml/src/ggml-cpu/cpu-feats-x86.cpp b/ggml/src/ggml-cpu/cpu-feats-x86.cpp
new file mode 100644
index 00000000000..e8133d411fd
--- /dev/null
+++ b/ggml/src/ggml-cpu/cpu-feats-x86.cpp
@@ -0,0 +1,323 @@
+#include "ggml-backend-impl.h"
+
+#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))
+
+#ifdef _MSC_VER
+#include <intrin.h>
+#endif
+
+#include <cstring>
+#include <vector>
+#include <bitset>
+#include <array>
+#include <string>
+
+// ref: https://cdrdv2-public.intel.com/782156/325383-sdm-vol-2abcd.pdf
+struct cpuid_x86 {
+    bool SSE3(void) { return f_1_ecx[0]; }
+    bool PCLMULQDQ(void) { return f_1_ecx[1]; }
+    bool MONITOR(void) { return f_1_ecx[3]; }
+    bool SSSE3(void) { return f_1_ecx[9]; }
+    bool FMA(void) { return f_1_ecx[12]; }
+    bool CMPXCHG16B(void) { return f_1_ecx[13]; }
+    bool SSE41(void) { return f_1_ecx[19]; }
+    bool SSE42(void) { return f_1_ecx[20]; }
+    bool MOVBE(void) { return f_1_ecx[22]; }
+    bool POPCNT(void) { return f_1_ecx[23]; }
+    bool AES(void) { return f_1_ecx[25]; }
+    bool XSAVE(void) { return f_1_ecx[26]; }
+    bool OSXSAVE(void) { return f_1_ecx[27]; }
+    bool AVX(void) { return f_1_ecx[28]; }
+    bool F16C(void) { return f_1_ecx[29]; }
+    bool RDRAND(void) { return f_1_ecx[30]; }
+
+    bool MSR(void) { return f_1_edx[5]; }
+    bool CX8(void) { return f_1_edx[8]; }
+    bool SEP(void) { return f_1_edx[11]; }
+    bool CMOV(void) { return f_1_edx[15]; }
+    bool CLFSH(void) { return f_1_edx[19]; }
+    bool MMX(void) { return f_1_edx[23]; }
+    bool FXSR(void) { return f_1_edx[24]; }
+    bool SSE(void) { return f_1_edx[25]; }
+    bool SSE2(void) { return f_1_edx[26]; }
+
+    bool FSGSBASE(void) { return f_7_ebx[0]; }
+    bool BMI1(void) { return f_7_ebx[3]; }
+    bool HLE(void) { return is_intel && f_7_ebx[4]; }
+    bool AVX2(void) { return f_7_ebx[5]; }
+    bool BMI2(void) { return f_7_ebx[8]; }
+    bool ERMS(void) { return f_7_ebx[9]; }
+    bool INVPCID(void) { return f_7_ebx[10]; }
+    bool RTM(void) { return is_intel && f_7_ebx[11]; }
+    bool AVX512F(void) { return f_7_ebx[16]; }
+    bool AVX512DQ(void) { return f_7_ebx[17]; }
+    bool RDSEED(void) { return f_7_ebx[18]; }
+    bool ADX(void) { return f_7_ebx[19]; }
+    bool AVX512PF(void) { return f_7_ebx[26]; }
+    bool AVX512ER(void) { return f_7_ebx[27]; }
+    bool AVX512CD(void) { return f_7_ebx[28]; }
+    bool AVX512BW(void) { return f_7_ebx[30]; }
+    bool AVX512VL(void) { return f_7_ebx[31]; }
+
+    bool SHA(void) { return f_7_ebx[29]; }
+
+    bool PREFETCHWT1(void) { return f_7_ecx[0]; }
+
+    bool LAHF(void) { return f_81_ecx[0]; }
+    bool LZCNT(void) { return is_intel && f_81_ecx[5]; }
+    bool ABM(void) { return is_amd && f_81_ecx[5]; }
+    bool SSE4a(void) { return is_amd && f_81_ecx[6]; }
+    bool XOP(void) { return is_amd && f_81_ecx[11]; }
+    bool TBM(void) { return is_amd && f_81_ecx[21]; }
+
+    bool SYSCALL(void) { return is_intel && f_81_edx[11]; }
+    bool MMXEXT(void) { return is_amd && f_81_edx[22]; }
+    bool RDTSCP(void) { return is_intel && f_81_edx[27]; }
+    bool _3DNOWEXT(void) { return is_amd && f_81_edx[30]; }
+    bool _3DNOW(void) { return is_amd && f_81_edx[31]; }
+
+    bool AVX512_VBMI(void) { return f_7_ecx[1]; }
+    bool AVX512_VNNI(void) { return f_7_ecx[11]; }
+    bool AVX512_FP16(void) { return f_7_edx[23]; }
+    bool AVX512_BF16(void) { return f_7_1_eax[5]; }
+    bool AVX_VNNI(void) { return f_7_1_eax[4]; }
+
+    bool AMX_TILE(void) { return f_7_edx[24]; }
+    bool AMX_INT8(void) { return f_7_edx[25]; }
+    bool AMX_FP16(void) { return f_7_1_eax[21]; }
+    bool AMX_BF16(void) { return f_7_edx[22]; }
+
+#ifdef _MSC_VER
+    static void cpuid(int cpu_info[4], int eax) {
+        __cpuid(cpu_info, eax);
+    }
+    static void cpuidex(int cpu_info[4], int eax, int ecx) {
+        __cpuidex(cpu_info, eax, ecx);
+    }
+#else
+    static void cpuid(int cpu_info[4], int eax) {
+        __asm__ __volatile__(
+            "cpuid"
+            : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
+            : "a"(eax), "c"(0));
+    }
+    static void cpuidex(int cpu_info[4], int eax, int ecx) {
+        __asm__ __volatile__(
+            "cpuid"
+            : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
+            : "a"(eax), "c"(ecx));
+    }
+#endif
+
+    cpuid_x86() {
+        std::array<int, 4> cpui;
+        std::vector<std::array<int, 4>> data;
+
+        // calling __cpuid with 0x0 as the function_id argument
+        // gets the number of the highest valid function ID.
+        cpuid(cpui.data(), 0);
+        int n_ids = cpui[0];
+
+        for (int i = 0; i <= n_ids; ++i) {
+            cpuidex(cpui.data(), i, 0);
+            data.push_back(cpui);
+        }
+
+        // capture vendor string
+        char vendor[0x20] = {};
+        *reinterpret_cast<int *>(vendor)     = data[0][1];
+        *reinterpret_cast<int *>(vendor + 4) = data[0][3];
+        *reinterpret_cast<int *>(vendor + 8) = data[0][2];
+        this->vendor = vendor;
+        if (this->vendor == "GenuineIntel") {
+            is_intel = true;
+        } else if (this->vendor == "AuthenticAMD") {
+            is_amd = true;
+        }
+
+        // load bitset with flags for function 0x00000001
+        if (n_ids >= 1) {
+            f_1_ecx = data[1][2];
+            f_1_edx = data[1][3];
+        }
+
+        // load bitset with flags for function 0x00000007
+        if (n_ids >= 7) {
+            f_7_ebx = data[7][1];
+            f_7_ecx = data[7][2];
+            f_7_edx = data[7][3];
+            cpuidex(cpui.data(), 7, 1);
+            f_7_1_eax = cpui[0];
+        }
+
+        // calling __cpuid with 0x80000000 as the function_id argument
+        // gets the number of the highest valid extended ID.
+        cpuid(cpui.data(), 0x80000000);
+        unsigned int n_ex_ids = cpui[0];
+
+        std::vector<std::array<int, 4>> ext_data;
+        for (unsigned int i = 0x80000000; i <= n_ex_ids; ++i) {
+            cpuidex(cpui.data(), i, 0);
+            ext_data.push_back(cpui);
+        }
+
+        // load bitset with flags for function 0x80000001
+        if (n_ex_ids >= 0x80000001) {
+            f_81_ecx = ext_data[1][2];
+            f_81_edx = ext_data[1][3];
+        }
+
+        // interpret CPU brand string if reported
+        char brand[0x40] = {};
+        if (n_ex_ids >= 0x80000004) {
+            std::memcpy(brand, ext_data[2].data(), sizeof(cpui));
+            std::memcpy(brand + 16, ext_data[3].data(), sizeof(cpui));
+            std::memcpy(brand + 32, ext_data[4].data(), sizeof(cpui));
+            this->brand = brand;
+        }
+    }
+
+    bool is_intel = false;
+    bool is_amd = false;
+    std::string vendor;
+    std::string brand;
+    std::bitset<32> f_1_ecx;
+    std::bitset<32> f_1_edx;
+    std::bitset<32> f_7_ebx;
+    std::bitset<32> f_7_ecx;
+    std::bitset<32> f_7_edx;
+    std::bitset<32> f_7_1_eax;
+    std::bitset<32> f_81_ecx;
+    std::bitset<32> f_81_edx;
+};
+
+#if 0
+void test_x86_is() {
+    cpuid_x86 is;
+    printf("CPU Vendor: %s\n", is.vendor.c_str());
+    printf("Brand: %s\n", is.brand.c_str());
+    printf("is_intel: %d\n", is.is_intel);
+    printf("is_amd: %d\n", is.is_amd);
+    printf("sse3: %d\n", is.SSE3());
+    printf("pclmulqdq: %d\n", is.PCLMULQDQ());
+    printf("ssse3: %d\n", is.SSSE3());
+    printf("fma: %d\n", is.FMA());
+    printf("cmpxchg16b: %d\n", is.CMPXCHG16B());
+    printf("sse41: %d\n", is.SSE41());
+    printf("sse42: %d\n", is.SSE42());
+    printf("movbe: %d\n", is.MOVBE());
+    printf("popcnt: %d\n", is.POPCNT());
+    printf("aes: %d\n", is.AES());
+    printf("xsave: %d\n", is.XSAVE());
+    printf("osxsave: %d\n", is.OSXSAVE());
+    printf("avx: %d\n", is.AVX());
+    printf("f16c: %d\n", is.F16C());
+    printf("rdrand: %d\n", is.RDRAND());
+    printf("msr: %d\n", is.MSR());
+    printf("cx8: %d\n", is.CX8());
+    printf("sep: %d\n", is.SEP());
+    printf("cmov: %d\n", is.CMOV());
+    printf("clflush: %d\n", is.CLFSH());
+    printf("mmx: %d\n", is.MMX());
+    printf("fxsr: %d\n", is.FXSR());
+    printf("sse: %d\n", is.SSE());
+    printf("sse2: %d\n", is.SSE2());
+    printf("fsgsbase: %d\n", is.FSGSBASE());
+    printf("bmi1: %d\n", is.BMI1());
+    printf("hle: %d\n", is.HLE());
+    printf("avx2: %d\n", is.AVX2());
+    printf("bmi2: %d\n", is.BMI2());
+    printf("erms: %d\n", is.ERMS());
+    printf("invpcid: %d\n", is.INVPCID());
+    printf("rtm: %d\n", is.RTM());
+    printf("avx512f: %d\n", is.AVX512F());
+    printf("rdseed: %d\n", is.RDSEED());
+    printf("adx: %d\n", is.ADX());
+    printf("avx512pf: %d\n", is.AVX512PF());
+    printf("avx512er: %d\n", is.AVX512ER());
+    printf("avx512cd: %d\n", is.AVX512CD());
+    printf("sha: %d\n", is.SHA());
+    printf("prefetchwt1: %d\n", is.PREFETCHWT1());
+    printf("lahf: %d\n", is.LAHF());
+    printf("lzcnt: %d\n", is.LZCNT());
+    printf("abm: %d\n", is.ABM());
+    printf("sse4a: %d\n", is.SSE4a());
+    printf("xop: %d\n", is.XOP());
+    printf("tbm: %d\n", is.TBM());
+    printf("syscall: %d\n", is.SYSCALL());
+    printf("mmxext: %d\n", is.MMXEXT());
+    printf("rdtscp: %d\n", is.RDTSCP());
+    printf("3dnowext: %d\n", is._3DNOWEXT());
+    printf("3dnow: %d\n", is._3DNOW());
+    printf("avx512_vbmi: %d\n", is.AVX512_VBMI());
+    printf("avx512_vnni: %d\n", is.AVX512_VNNI());
+    printf("avx512_fp16: %d\n", is.AVX512_FP16());
+    printf("avx512_bf16: %d\n", is.AVX512_BF16());
+    printf("amx_tile: %d\n", is.AMX_TILE());
+    printf("amx_int8: %d\n", is.AMX_INT8());
+    printf("amx_fp16: %d\n", is.AMX_FP16());
+    printf("amx_bf16: %d\n", is.AMX_BF16());
+}
+#endif
+
+static int ggml_backend_cpu_x86_score() {
+    // FIXME: this does not check for OS support
+
+    int score = 0;
+    cpuid_x86 is;
+
+#ifdef GGML_FMA
+    if (!is.FMA()) { return 0; }
+    score += 1;
+#endif
+#ifdef GGML_F16C
+    if (!is.F16C()) { return 0; }
+    score += 1<<1;
+#endif
+#ifdef GGML_SSE42
+    if (!is.SSE42()) { return 0; }
+    score += 1<<2;
+#endif
+#ifdef GGML_AVX
+    if (!is.AVX()) { return 0; }
+    score += 1<<4;
+#endif
+#ifdef GGML_AVX2
+    if (!is.AVX2()) { return 0; }
+    score += 1<<5;
+#endif
+#ifdef GGML_AVX_VNNI
+    if (!is.AVX_VNNI()) { return 0; }
+    score += 1<<6;
+#endif
+#ifdef GGML_AVX512
+    if (!is.AVX512F()) { return 0; }
+    if (!is.AVX512CD()) { return 0; }
+    if (!is.AVX512VL()) { return 0; }
+    if (!is.AVX512DQ()) { return 0; }
+    if (!is.AVX512BW()) { return 0; }
+    score += 1<<7;
+#endif
+#ifdef GGML_AVX512_VBMI
+    if (!is.AVX512_VBMI()) { return 0; }
+    score += 1<<8;
+#endif
+#ifdef GGML_AVX512_BF16
+    if (!is.AVX512_BF16()) { return 0; }
+    score += 1<<9;
+#endif
+#ifdef GGML_AVX512_VNNI
+    if (!is.AVX512_VNNI()) { return 0; }
+    score += 1<<10;
+#endif
+#ifdef GGML_AMX_INT8
+    if (!is.AMX_INT8()) { return 0; }
+    score += 1<<11;
+#endif
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_x86_score)
+
+#endif // defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))
diff --git a/ggml/src/ggml-aarch64.c b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
similarity index 78%
rename from ggml/src/ggml-aarch64.c
rename to ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
index 81f62ff4f32..a51d1a6c595 100644
--- a/ggml/src/ggml-aarch64.c
+++ b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
@@ -1,23 +1,56 @@
-// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
-// SPDX-License-Identifier: MIT
-//
-
-#define GGML_COMMON_IMPL_C
+#define GGML_COMMON_IMPL_CPP
+#define GGML_COMMON_DECL_CPP
 #include "ggml-common.h"
+#include "ggml-backend-impl.h"
 
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-cpu-impl.h"
+#include "ggml-cpu-traits.h"
+
+#include <cmath>
+#include <cstring>
+#include <cassert>
+#include <cfloat>
+#include <cstdlib> // for qsort
+#include <cstdio>  // for GGML_ASSERT
+
+#include "ggml-cpu-aarch64.h"
+
+// TODO: move to include file?
+template <int K> constexpr int QK_0() {
+    if constexpr (K == 4) {
+        return QK4_0;
+    }
+    if constexpr (K == 8) {
+        return QK8_0;
+    }
+    return -1;
+}
+
+template <int K, int N> struct block {
+    ggml_half d[N];                         // deltas for N qK_0 blocks
+    int8_t    qs[(QK_0<K>() * N * K) / 8];  // quants for N qK_0 blocks
+};
+
+// control size
+static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
+static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
+static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
+static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
+
+using block_q4_0x4 = block<4, 4>;
+using block_q4_0x8 = block<4, 8>;
+using block_q8_0x4 = block<8, 4>;
+using block_q8_0x8 = block<8, 8>;
 
-#include <math.h>
-#include <string.h>
-#include <assert.h>
-#include <float.h>
-#include <stdlib.h> // for qsort
-#include <stdio.h>  // for GGML_ASSERT
+struct block_iq4_nlx4 {
+    ggml_half d[4];            // deltas for 4 iq4_nl blocks
+    uint8_t   qs[QK4_NL * 2];  // nibbles / quants for 4 iq4_nl blocks
+};
 
-#include "ggml-aarch64.h"
+static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
 
 #if defined(__GNUC__)
 #pragma GCC diagnostic ignored "-Woverlength-strings"
@@ -132,7 +165,7 @@ static inline __m512i sum_i16_pairs_int_32x16(const __m512i x) {
 }
 
 static inline __m512i mul_sum_us8_pairs_int32x16(const __m512i ax, const __m512i sy) {
-#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+#if defined(__AVX512VNNI__)
     const __m512i zero = _mm512_setzero_si512();
     return _mm512_dpbusd_epi32(zero, ax, sy);
 #else
@@ -187,52 +220,14 @@ static inline __m256i mul_sum_i8_pairs_int32x8(const __m256i x, const __m256i y)
 }
 #endif
 
-static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
-    block_q4_0x4 out;
-
-    for (int i = 0; i < 4; i++) {
-        out.d[i] = in[i].d;
-    }
-
-    for (int i = 0; i < QK4_0 * 2; i++) {
-        int src_offset = (i / (4 * blck_size_interleave)) * blck_size_interleave;
-        int src_id = (i % (4 * blck_size_interleave)) / blck_size_interleave;
-        src_offset += (i % blck_size_interleave);
-
-        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
-    }
-
-    return out;
-}
+static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
 
-// interleave 8 block_q4_0s in blocks of blck_size_interleave
-// returns an interleaved block_q4_0x8
-// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
-// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
-static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
-    block_q4_0x8 out;
-
-    for (int i = 0; i < 8; i++) {
-        out.d[i] = in[i].d;
-    }
-
-    for (int i = 0; i < QK4_0 * 4; i++) {
-        int src_offset = (i / (8 * blck_size_interleave)) * blck_size_interleave;
-        int src_id = (i % (8 * blck_size_interleave)) / blck_size_interleave;
-        src_offset += (i % blck_size_interleave);
-
-        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
-    }
-
-    return out;
-}
-
-void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k) {
+static void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK8_0 == 32);
     assert(k % QK8_0 == 0);
     const int nb = k / QK8_0;
 
-    block_q8_0x4 * restrict y = (block_q8_0x4 *) vy;
+    block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
 
 #if defined(__ARM_NEON)
     float32x4_t srcv[4][8];
@@ -321,12 +316,12 @@ void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k)
 #endif
 }
 
-void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k) {
+static void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
     assert(QK8_0 == 32);
     assert(k % QK8_0 == 0);
     const int nb = k / QK8_0;
 
-    block_q8_0x4 * restrict y = (block_q8_0x4 *) vy;
+    block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
 
 #if defined(__ARM_NEON)
     float32x4_t srcv[4][8];
@@ -536,7 +531,7 @@ void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k)
 #endif
 }
 
-void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nrow, int64_t n_per_row, int64_t blck_size_interleave) {
+static void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t nrow, int64_t n_per_row, int64_t blck_size_interleave) {
     assert(nrow == 4);
     UNUSED(nrow);
     if (blck_size_interleave == 4) {
@@ -548,58 +543,7 @@ void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nro
     }
 }
 
-static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int blck_size_interleave) {
-    assert(n_per_row % QK4_0 == 0);
-    const int nb = n_per_row / QK4_0;
-
-    void * out_ptr = NULL;
-    if (nrows_interleaved == 8) {
-        out_ptr = (block_q4_0x8 *) dst;
-    }
-    else if (nrows_interleaved == 4) {
-        out_ptr = (block_q4_0x4 *) dst;
-    }
-    assert(nrows_interleaved <= 8);
-    block_q4_0 dst_tmp[8];
-
-    for (int b = 0; b < (nrow * n_per_row); b += nrows_interleaved * n_per_row) {
-
-        for (int64_t x = 0; x < nb; x++) {
-
-            for (int i  = 0; i < nrows_interleaved; i++ ) {
-                quantize_row_q4_0_ref(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0);
-            }
-
-            if (nrows_interleaved == 8) {
-                *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave, 0x88);
-                out_ptr = (block_q4_0x8 *) out_ptr + 1;
-            }
-            else if (nrows_interleaved == 4) {
-                *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave, 0x88);
-                out_ptr = (block_q4_0x4 *) out_ptr + 1;
-            }
-        }
-    }
-
-    return ((nrow * n_per_row) / QK4_0 * sizeof(block_q4_0));
-}
-
-size_t quantize_q4_0_4x4(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    UNUSED(quant_weights);
-    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
-}
-
-size_t quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    UNUSED(quant_weights);
-    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
-}
-
-size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
-    UNUSED(quant_weights);
-    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
-}
-
-void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+static void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
     const int ncols_interleaved = 4;
@@ -618,67 +562,47 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    if (ggml_cpu_has_neon()) {
-        const void * b_ptr = vx;
-        const void * a_ptr = vy;
-        float * res_ptr = s;
-
-        __asm__ __volatile__(
-            "movi v31.16b, #0x4\n"
-            "movi v30.16b, #0xf0\n"
-            "add %x[b_ptr], %x[b_ptr], #0x8\n"
-            "1:"  // Column loop
-            "add x22, %x[a_ptr], #0x2\n"
-            "movi v29.16b, #0x0\n"
-            "mov x21, %x[nb]\n"
-            "2:"  // Block loop
-            "ldr q28, [%x[b_ptr], #0x0]\n"
-            "ldr q27, [x22, #0x0]\n"
-            "movi v26.4s, #0x0\n"
-            "sub x20, x22, #0x2\n"
-            "ldr q25, [x22, #0x10]\n"
-            "ldr q24, [%x[b_ptr], #0x10]\n"
-            "sub x21, x21, #0x1\n"
-            "add x22, x22, #0x22\n"
-            "ldr q23, [%x[b_ptr], #0x20]\n"
-            "ldr q22, [%x[b_ptr], #0x30]\n"
-            "ld1r { v21.8h }, [x20]\n"
-            "ldr q20, [%x[b_ptr], #-0x8]\n"
-            "sshl v16.16b, v28.16b, v31.16b\n"
-            "and v28.16b, v28.16b, v30.16b\n"
-            "sshl v19.16b, v24.16b, v31.16b\n"
-            "and v24.16b, v24.16b, v30.16b\n"
-            "add %x[b_ptr], %x[b_ptr], #0x48\n"
-            "sshl v18.16b, v23.16b, v31.16b\n"
-            "and v23.16b, v23.16b, v30.16b\n"
-            ".inst 0x4f9be21a  // sdot v26.4s, v16.16b, v27.4b[0]\n"
-            "sshl v17.16b, v22.16b, v31.16b\n"
-            "and v22.16b, v22.16b, v30.16b\n"
-            "fcvtl v21.4s, v21.4h\n"
-            "fcvtl v16.4s, v20.4h\n"
-            ".inst 0x4f99e39a  // sdot v26.4s, v28.16b, v25.4b[0]\n"
-            "fmul v16.4s, v16.4s, v21.4s\n"
-            ".inst 0x4fbbe27a  // sdot v26.4s, v19.16b, v27.4b[1]\n"
-            ".inst 0x4fb9e31a  // sdot v26.4s, v24.16b, v25.4b[1]\n"
-            ".inst 0x4f9bea5a  // sdot v26.4s, v18.16b, v27.4b[2]\n"
-            ".inst 0x4f99eafa  // sdot v26.4s, v23.16b, v25.4b[2]\n"
-            ".inst 0x4fbbea3a  // sdot v26.4s, v17.16b, v27.4b[3]\n"
-            ".inst 0x4fb9eada  // sdot v26.4s, v22.16b, v25.4b[3]\n"
-            "scvtf v26.4s, v26.4s, #0x4\n"
-            "fmla v29.4s, v26.4s, v16.4s\n"
-            "cbnz x21, 2b\n"
-            "sub %x[nc], %x[nc], #0x4\n"
-            "str q29, [%x[res_ptr], #0x0]\n"
-            "add %x[res_ptr], %x[res_ptr], #0x10\n"
-            "cbnz %x[nc], 1b\n"
-            : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
-            : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
-            : "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22"
-            );
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *)vx;
+
+        for (int c = 0; c < nc; c += ncols_interleaved) {
+            const block_q8_0 * a_ptr = (const block_q8_0 *)vy;
+            float32x4_t acc = vdupq_n_f32(0);
+            for (int b = 0; b < nb; b++) {
+                int8x16_t b0 = vld1q_s8((const int8_t *)b_ptr->qs);
+                int8x16_t b1 = vld1q_s8((const int8_t *)b_ptr->qs + 16);
+                int8x16_t b2 = vld1q_s8((const int8_t *)b_ptr->qs + 32);
+                int8x16_t b3 = vld1q_s8((const int8_t *)b_ptr->qs + 48);
+                float16x4_t bd = vld1_f16((const __fp16 *)b_ptr->d);
+
+                int8x16_t a0 = vld1q_s8(a_ptr->qs);
+                int8x16_t a1 = vld1q_s8(a_ptr->qs + qk/2);
+                float16x4_t ad = vld1_dup_f16((const __fp16 *)&a_ptr->d);
+
+                int32x4_t ret = vdupq_n_s32(0);
+
+                ret = vdotq_laneq_s32(ret, b0 << 4, a0, 0);
+                ret = vdotq_laneq_s32(ret, b1 << 4, a0, 1);
+                ret = vdotq_laneq_s32(ret, b2 << 4, a0, 2);
+                ret = vdotq_laneq_s32(ret, b3 << 4, a0, 3);
+
+                ret = vdotq_laneq_s32(ret, b0 & 0xf0U, a1, 0);
+                ret = vdotq_laneq_s32(ret, b1 & 0xf0U, a1, 1);
+                ret = vdotq_laneq_s32(ret, b2 & 0xf0U, a1, 2);
+                ret = vdotq_laneq_s32(ret, b3 & 0xf0U, a1, 3);
+
+                acc = vfmaq_f32(acc, vcvtq_n_f32_s32(ret, 4),
+                                vmulq_f32(vcvt_f32_f16(ad), vcvt_f32_f16(bd)));
+                a_ptr++;
+                b_ptr++;
+            }
+            vst1q_f32(s, acc);
+            s += ncols_interleaved;
+        }
         return;
     }
-#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
     float sumf[4];
     int sumi;
 
@@ -704,7 +628,7 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
     }
 }
 
-void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+static void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
     const int ncols_interleaved = 4;
@@ -814,7 +738,7 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
     }
 }
 
-void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+static void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
     const int ncols_interleaved = 8;
@@ -1087,7 +1011,103 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
     }
 }
 
-void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+static void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+        const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl);
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        float * res_ptr = s;
+
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+            float32x4_t sumf = vdupq_n_f32(0);
+            for (int l = 0; l < nb; l++) {
+                uint8x16_t b_0 = vld1q_u8(b_ptr[l].qs + 0);
+                uint8x16_t b_1 = vld1q_u8(b_ptr[l].qs + 16);
+                uint8x16_t b_2 = vld1q_u8(b_ptr[l].qs + 32);
+                uint8x16_t b_3 = vld1q_u8(b_ptr[l].qs + 48);
+
+                int8x16_t b_0_hi = vqtbl1q_s8(kvalues, b_0 >> 4);
+                int8x16_t b_0_lo = vqtbl1q_s8(kvalues, b_0 & 0x0F);
+                int8x16_t b_1_hi = vqtbl1q_s8(kvalues, b_1 >> 4);
+                int8x16_t b_1_lo = vqtbl1q_s8(kvalues, b_1 & 0x0F);
+                int8x16_t b_2_hi = vqtbl1q_s8(kvalues, b_2 >> 4);
+                int8x16_t b_2_lo = vqtbl1q_s8(kvalues, b_2 & 0x0F);
+                int8x16_t b_3_hi = vqtbl1q_s8(kvalues, b_3 >> 4);
+                int8x16_t b_3_lo = vqtbl1q_s8(kvalues, b_3 & 0x0F);
+
+                int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 0);
+                int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16);
+
+                int32x4_t sumi = vdupq_n_s32(0);
+                sumi = vdotq_laneq_s32(sumi, b_0_lo, a_0, 0);
+                sumi = vdotq_laneq_s32(sumi, b_0_hi, a_1, 0);
+                sumi = vdotq_laneq_s32(sumi, b_1_lo, a_0, 1);
+                sumi = vdotq_laneq_s32(sumi, b_1_hi, a_1, 1);
+                sumi = vdotq_laneq_s32(sumi, b_2_lo, a_0, 2);
+                sumi = vdotq_laneq_s32(sumi, b_2_hi, a_1, 2);
+                sumi = vdotq_laneq_s32(sumi, b_3_lo, a_0, 3);
+                sumi = vdotq_laneq_s32(sumi, b_3_hi, a_1, 3);
+
+                float32x4_t a_d = vcvt_f32_f16(vld1_dup_f16((const float16_t *)&a_ptr[l].d));
+                float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d));
+                float32x4_t d = a_d * b_d;
+
+                sumf = vmlaq_f32(sumf, d, vcvtq_f32_s32(sumi));
+            }
+
+            vst1q_f32(res_ptr + x * 4, sumf);
+        }
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    {
+        float sumf[4];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        sumi = 0;
+                        for (int i = 0; i < blocklen; ++i) {
+                            const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                            const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                        }
+                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+static void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
     const int ncols_interleaved = 4;
@@ -1108,7 +1128,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
     UNUSED(blocklen);
 
 #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
-    if (ggml_cpu_has_neon()) {
+    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
         const void * b_ptr = vx;
         const void * a_ptr = vy;
         float * res_ptr = s;
@@ -1603,7 +1623,7 @@ void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void *
     }
 }
 
-void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+static void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
     const int ncols_interleaved = 4;
@@ -2057,7 +2077,7 @@ void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void *
     }
 }
 
-void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+static void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
     const int qk = QK8_0;
     const int nb = n / qk;
     const int ncols_interleaved = 8;
@@ -2577,31 +2597,31 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
                     const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
 
                     // Shuffle pattern one - right side input
-                    const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
-                    const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
+                    const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
+                    const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
 
-                    const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
-                    const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
+                    const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
+                    const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
 
-                    const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
-                    const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
+                    const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
+                    const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
 
-                    const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
-                    const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
+                    const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
+                    const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
 
                     // Shuffle pattern two - right side input
 
-                    const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
-                    const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
+                    const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
+                    const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
 
-                    const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
-                    const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
+                    const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
+                    const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
 
-                    const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
-                    const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
+                    const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
+                    const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
 
-                    const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
-                    const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
+                    const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
+                    const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
 
                     // Scale values - Load the weight scale values of two block_q4_0x8
                     const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
@@ -2635,31 +2655,31 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
 
                         // Shuffle pattern one - left side input
 
-                        const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
-                        const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+                        const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                        const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
 
-                        const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
-                        const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+                        const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                        const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
 
-                        const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
-                        const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+                        const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                        const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
 
-                        const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
-                        const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+                        const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                        const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
 
                         // Shuffle pattern two - left side input
 
-                        const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
-                        const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+                        const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                        const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
 
-                        const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
-                        const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+                        const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                        const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
 
-                        const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
-                        const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+                        const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                        const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
 
-                        const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
-                        const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+                        const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                        const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
 
                         // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
                         // Resembles MMLAs into 2x2 matrices in ARM Version
@@ -2688,10 +2708,10 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
 
 
                         // Straighten out to make 4 row vectors
-                        __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, 78));
-                        __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01);
-                        __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, 78));
-                        __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11);
+                        __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
+                        __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
+                        __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
+                        __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
 
                         // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
                         const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68);
@@ -2770,31 +2790,31 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
                     const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
 
                     // Shuffle pattern one - right side input
-                    const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
-                    const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
+                    const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
+                    const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
 
-                    const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
-                    const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
+                    const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
+                    const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
 
-                    const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
-                    const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
+                    const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
+                    const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
 
-                    const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
-                    const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
+                    const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
+                    const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
 
                     // Shuffle pattern two - right side input
 
-                    const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
-                    const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
+                    const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
+                    const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
 
-                    const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
-                    const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
+                    const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
+                    const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
 
-                    const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
-                    const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
+                    const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
+                    const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
 
-                    const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
-                    const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
+                    const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
+                    const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
 
 
                     // Scale values - Load the weight scale values of two block_q4_0x8
@@ -2826,31 +2846,31 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
 
                     // Shuffle pattern one - left side input
 
-                    const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
-                    const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+                    const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                    const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
 
-                    const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
-                    const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+                    const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                    const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
 
-                    const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
-                    const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+                    const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                    const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
 
-                    const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
-                    const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+                    const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                    const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
 
                     // Shuffle pattern two - left side input
 
-                    const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
-                    const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+                    const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                    const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
 
-                    const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
-                    const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+                    const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                    const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
 
-                    const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
-                    const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+                    const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                    const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
 
-                    const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
-                    const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+                    const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                    const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
 
                     // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
                     // Resembles MMLAs into 2x2 matrices in ARM Version
@@ -2879,10 +2899,10 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
 
 
                     // Straighten out to make 4 row vectors
-                    __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, 78));
-                    __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01);
-                    __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, 78));
-                    __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11);
+                    __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
+                    __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
+                    __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
 
                     // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
                     const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68);
@@ -3476,3 +3496,767 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
         }
     }
 }
+
+static void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_DOTPROD)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+        const int8x16_t kvalues = vld1q_s8(kvalues_iq4nl);
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+
+                float32x4_t sumf[4];
+                for (int m = 0; m < 4; m++) {
+                    sumf[m] = vdupq_n_f32(0);
+                }
+
+                for (int l = 0; l < nb; l++) {
+                    float32x4_t a_d = vcvt_f32_f16(vld1_f16((const float16_t *)a_ptr[l].d));
+                    float32x4_t b_d = vcvt_f32_f16(vld1_f16((const float16_t *)b_ptr[l].d));
+
+                    int32x4_t sumi_0 = vdupq_n_s32(0);
+                    int32x4_t sumi_1 = vdupq_n_s32(0);
+                    int32x4_t sumi_2 = vdupq_n_s32(0);
+                    int32x4_t sumi_3 = vdupq_n_s32(0);
+
+                    for (int k = 0; k < 4; k++) {
+                        int8x16_t a_0 = vld1q_s8(a_ptr[l].qs + 16 * k + 0);
+                        int8x16_t a_1 = vld1q_s8(a_ptr[l].qs + 16 * k + 64);
+
+                        uint8x16_t b = vld1q_u8(b_ptr[l].qs + 16 * k);
+                        int8x16_t b_hi = vqtbl1q_s8(kvalues, b >> 4);
+                        int8x16_t b_lo = vqtbl1q_s8(kvalues, b & 0xF);
+
+                        sumi_0 = vdotq_laneq_s32(sumi_0, b_lo, a_0, 0);
+                        sumi_1 = vdotq_laneq_s32(sumi_1, b_lo, a_0, 1);
+                        sumi_2 = vdotq_laneq_s32(sumi_2, b_lo, a_0, 2);
+                        sumi_3 = vdotq_laneq_s32(sumi_3, b_lo, a_0, 3);
+                        sumi_0 = vdotq_laneq_s32(sumi_0, b_hi, a_1, 0);
+                        sumi_1 = vdotq_laneq_s32(sumi_1, b_hi, a_1, 1);
+                        sumi_2 = vdotq_laneq_s32(sumi_2, b_hi, a_1, 2);
+                        sumi_3 = vdotq_laneq_s32(sumi_3, b_hi, a_1, 3);
+                    }
+
+                    sumf[0] = vmlaq_f32(sumf[0], vmulq_laneq_f32(b_d, a_d, 0), vcvtq_f32_s32(sumi_0));
+                    sumf[1] = vmlaq_f32(sumf[1], vmulq_laneq_f32(b_d, a_d, 1), vcvtq_f32_s32(sumi_1));
+                    sumf[2] = vmlaq_f32(sumf[2], vmulq_laneq_f32(b_d, a_d, 2), vcvtq_f32_s32(sumi_2));
+                    sumf[3] = vmlaq_f32(sumf[3], vmulq_laneq_f32(b_d, a_d, 3), vcvtq_f32_s32(sumi_3));
+                }
+
+                for (int m = 0; m < 4; m++) {
+                    vst1q_f32(s + (y * 4 + m) * bs + x * 4, sumf[m]);
+                }
+            }
+        }
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    {
+        float sumf[4][4];
+        int sumi;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+                }
+                for (int l = 0; l < nb; l++) {
+                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                        for (int m = 0; m < 4; m++) {
+                            for (int j = 0; j < ncols_interleaved; j++) {
+                                sumi = 0;
+                                for (int i = 0; i < blocklen; ++i) {
+                                    const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                    const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
+                                }
+                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            }
+                        }
+                    }
+                }
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++)
+                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+                }
+            }
+        }
+    }
+}
+
+static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
+    block_q4_0x4 out;
+
+    for (int i = 0; i < 4; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_0 * 2 / blck_size_interleave;
+
+    if (blck_size_interleave == 8) {
+        const uint64_t xor_mask = 0x8888888888888888ULL;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            uint64_t elems;
+            // Using memcpy to avoid unaligned memory accesses
+            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+            elems ^= xor_mask;
+            memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+        }
+    } else if (blck_size_interleave == 4) {
+        const uint32_t xor_mask = 0x88888888;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            uint32_t elems;
+            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
+            elems ^= xor_mask;
+            memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+// interleave 8 block_q4_0s in blocks of blck_size_interleave
+// returns an interleaved block_q4_0x8
+// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
+// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
+static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) {
+    block_q4_0x8 out;
+
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_0 * 4 / blck_size_interleave;
+    const uint64_t xor_mask = 0x8888888888888888ULL;
+
+    for (int i = 0; i < end; ++i) {
+        int src_id = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
+
+        uint64_t elems;
+        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+        elems ^= xor_mask;
+        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+    }
+
+    return out;
+}
+
+static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
+    GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
+    constexpr int nrows_interleaved = 4;
+
+    block_q4_0x4 * dst = (block_q4_0x4 *)t->data;
+    const block_q4_0 * src = (const block_q4_0 *)data;
+    block_q4_0 dst_tmp[4];
+    int nrow = ggml_nrows(t);
+    int nblocks = t->ne[0] / QK4_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q4_0x4(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
+    GGML_ASSERT(interleave_block == 8);
+    constexpr int nrows_interleaved = 8;
+
+    block_q4_0x8 * dst = (block_q4_0x8*)t->data;
+    const block_q4_0 * src = (const block_q4_0*) data;
+    block_q4_0 dst_tmp[8];
+    int nrow = ggml_nrows(t);
+    int nblocks = t->ne[0] / QK4_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i  = 0; i < nrows_interleaved; i++ ) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_q4_0x8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_size_interleave) {
+    block_iq4_nlx4 out;
+
+    for (int i = 0; i < 4; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_NL * 2 / blck_size_interleave;
+
+    // TODO: this branch seems wrong
+    //if (blck_size_interleave == 8) {
+    //    for (int i = 0; i < end; ++i) {
+    //        int src_id = i % 4;
+    //        int src_offset = (i / 4) * blck_size_interleave;
+    //        int dst_offset = i * blck_size_interleave;
+
+    //        // Using memcpy to avoid unaligned memory accesses
+    //        memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t));
+    //    }
+    //} else
+    if (blck_size_interleave == 4) {
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint32_t));
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
+    //GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
+    GGML_ASSERT(interleave_block == 4);
+
+    block_iq4_nlx4 * dst = (block_iq4_nlx4 *)t->data;
+    const block_iq4_nl * src = (const block_iq4_nl *)data;
+    block_iq4_nl dst_tmp[4];
+    int nrow = ggml_nrows(t);
+    int nrows_interleaved = 4;
+    int nblocks = t->ne[0] / QK4_0;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
+
+    if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_iq4_nlx4(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
+namespace ggml::cpu::aarch64 {
+// repack
+template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
+int repack(struct ggml_tensor *, const void *, size_t);
+
+// TODO: generalise.
+template <> int repack<block_q4_0, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_0_to_q4_0_4_bl(t, 4, data, data_size);
+}
+
+template <> int repack<block_q4_0, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_0_to_q4_0_4_bl(t, 8, data, data_size);
+}
+
+template <> int repack<block_q4_0, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_q4_0_to_q4_0_8_bl(t, 8, data, data_size);
+}
+
+template <> int repack<block_iq4_nl, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_iq4_nl_to_iq4_nl_4_bl(t, 4, data, data_size);
+}
+
+// TODO: needs to be revisited
+//template <> int repack<block_iq4_nl, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
+//    return repack_iq4_nl_to_iq4_nl_4_bl(t, 8, data, data_size);
+//}
+
+// gemv
+template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
+void gemv(int, float *, size_t, const void *, const void *, int, int);
+
+template <> void gemv<block_q4_0, 4, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q4_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_q4_0, 8, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemv<block_q4_0, 8, 8>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <>
+void gemv<block_iq4_nl, 4, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+// gemm
+template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
+void gemm(int, float *, size_t, const void *, const void *, int, int);
+
+template <> void gemm<block_q4_0, 4, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q4_0, 8, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <> void gemm<block_q4_0, 8, 8>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+template <>
+void gemm<block_iq4_nl, 4, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
+class tensor_traits_base : public ggml::cpu::tensor_traits {
+  public:
+    virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0;
+};
+
+template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> class tensor_traits : public tensor_traits_base {
+
+    bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
+        // not realy a GGML_TYPE_Q8_0 but same size.
+        switch (op->op) {
+        case GGML_OP_MUL_MAT:
+            size = ggml_row_size(GGML_TYPE_Q8_0, ggml_nelements(op->src[1]));
+            return true;
+        case GGML_OP_MUL_MAT_ID:
+            size = ggml_row_size(GGML_TYPE_Q8_0, ggml_nelements(op->src[1]));
+            size = GGML_PAD(size, sizeof(int64_t));  // + padding for next bloc.
+            size += sizeof(int64_t) * (1+op->src[0]->ne[2]) * op->src[1]->ne[2];
+            return true;
+        default:
+            // GGML_ABORT("fatal error");
+            break;
+        }
+        return false;
+    }
+
+    bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) override {
+        switch (op->op) {
+        case GGML_OP_MUL_MAT:
+            forward_mul_mat(params, op);
+            return true;
+        case GGML_OP_MUL_MAT_ID:
+            forward_mul_mat_id(params, op);
+            return true;
+        default:
+            // GGML_ABORT("fatal error");
+            break;
+        }
+        return false;
+    }
+
+    void forward_mul_mat(ggml_compute_params * params, ggml_tensor * op) {
+        const ggml_tensor * src0 = op->src[0];
+        const ggml_tensor * src1 = op->src[1];
+        ggml_tensor *       dst  = op;
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        const int ith = params->ith;
+        const int nth = params->nth;
+
+        GGML_ASSERT(ne0 == ne01);
+        GGML_ASSERT(ne1 == ne11);
+        GGML_ASSERT(ne2 == ne12);
+        GGML_ASSERT(ne3 == ne13);
+
+        // dst cannot be transposed or permuted
+        GGML_ASSERT(nb0 == sizeof(float));
+        GGML_ASSERT(nb0 <= nb1);
+        GGML_ASSERT(nb1 <= nb2);
+        GGML_ASSERT(nb2 <= nb3);
+
+        GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+        GGML_ASSERT(ggml_n_dims(op->src[0]) == 2);
+        // GGML_ASSERT(ggml_n_dims(op->src[1]) == 2);
+
+        char *       wdata = static_cast<char *>(params->wdata);
+        const size_t nbw1  = ggml_row_size(GGML_TYPE_Q8_0, ne10);
+
+        assert(params->wsize >= nbw1 * ne11);
+
+        const ggml_from_float_t from_float = ggml_get_type_traits_cpu(GGML_TYPE_Q8_0)->from_float;
+
+        int64_t i11_processed = 0;
+        for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
+            quantize_mat_q8_0((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), 4, ne10,
+                              INTER_SIZE);
+        }
+        i11_processed = ne11 - ne11 % 4;
+        for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) {
+            from_float((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), ne10);
+        }
+
+        ggml_barrier(params->threadpool);
+
+        const void * src1_wdata      = params->wdata;
+        const size_t src1_col_stride = ggml_row_size(GGML_TYPE_Q8_0, ne10);
+        int64_t      src0_start      = (ith * ne01) / nth;
+        int64_t      src0_end        = ((ith + 1) * ne01) / nth;
+        src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
+        src0_end   = (src0_end % NB_COLS) ? src0_end + NB_COLS - (src0_end % NB_COLS) : src0_end;
+        if (src0_start >= src0_end) {
+            return;
+        }
+
+        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
+        if (ne11 > 3) {
+            gemm<BLOC_TYPE, INTER_SIZE, NB_COLS>(ne00, (float *) ((char *) dst->data) + src0_start, ne01,
+                                                 (const char *) src0->data + src0_start * nb01,
+                                                 (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
+        }
+        for (int iter = ne11 - ne11 % 4; iter < ne11; iter++) {
+            gemv<BLOC_TYPE, INTER_SIZE, NB_COLS>(ne00, (float *) ((char *) dst->data + (iter * nb1)) + src0_start, ne01,
+                                                 (const char *) src0->data + src0_start * nb01,
+                                                 (const char *) src1_wdata + (src1_col_stride * iter), 1,
+                                                 src0_end - src0_start);
+        }
+    }
+
+    void forward_mul_mat_id(ggml_compute_params * params, ggml_tensor * op) {
+        const ggml_tensor * src0 = op->src[0];
+        const ggml_tensor * src1 = op->src[1];
+        const ggml_tensor * ids  = op->src[2];
+        ggml_tensor *       dst  = op;
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        const int ith = params->ith;
+        const int nth = params->nth;
+
+        const ggml_from_float_t from_float = ggml_get_type_traits_cpu(GGML_TYPE_Q8_0)->from_float;
+
+        // we don't support permuted src0 or src1
+        GGML_ASSERT(nb00 == ggml_type_size(src0->type));
+        GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+        // dst cannot be transposed or permuted
+        GGML_ASSERT(nb0 == sizeof(float));
+        GGML_ASSERT(nb0 <= nb1);
+        GGML_ASSERT(nb1 <= nb2);
+        GGML_ASSERT(nb2 <= nb3);
+
+        GGML_ASSERT(ne03 == 1);
+        GGML_ASSERT(ne13 == 1);
+        GGML_ASSERT(ne3  == 1);
+
+        GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+        // row groups
+        const int n_ids = ids->ne[0]; // n_expert_used
+        const int n_as  = ne02;       // n_expert
+
+        const size_t nbw1 = ggml_row_size(GGML_TYPE_Q8_0, ne10);
+        const size_t nbw2 = nbw1*ne11;
+        const size_t nbw3 = nbw2*ne12;
+
+        struct mmid_row_mapping {
+            int32_t i1;
+            int32_t i2;
+        };
+
+        GGML_ASSERT(params->wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as * sizeof(int64_t) +
+                                      n_as * ne12 * sizeof(mmid_row_mapping)));
+
+        auto                      wdata             = (char *) params->wdata;
+        auto                      wdata_src1_end    = (char *) wdata + GGML_PAD(nbw3, sizeof(int64_t));
+        int64_t *                 matrix_row_counts = (int64_t *) (wdata_src1_end);                      // [n_as]
+        struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as);  // [n_as][ne12]
+
+        // src1: float32 => block_q8_0
+        for (int64_t i12 = 0; i12 < ne12; ++i12) {
+            for (int64_t i11 = ith; i11 < ne11; i11 += nth) {
+                from_float((float *)((char *) src1->data + i12 * nb12 + i11 * nb11),
+                           (void *)               (wdata + i12 * nbw2 + i11 * nbw1),
+                           ne10);
+            }
+        }
+
+#define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id) * ne12 + (i1)]
+
+        if (ith == 0) {
+            // initialize matrix_row_counts
+            memset(matrix_row_counts, 0, n_as * sizeof(int64_t));
+
+            // group rows by src0 matrix
+            for (int32_t iid1 = 0; iid1 < ids->ne[1]; ++iid1) {
+                for (int32_t id = 0; id < n_ids; ++id) {
+                    const int32_t i02 =
+                        *(const int32_t *) ((const char *) ids->data + iid1 * ids->nb[1] + id * ids->nb[0]);
+
+                    GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+                    MMID_MATRIX_ROW(i02, matrix_row_counts[i02]) = { id, iid1 };
+                    matrix_row_counts[i02] += 1;
+                }
+            }
+        }
+
+        ggml_barrier(params->threadpool);
+
+        // compute each matrix multiplication in sequence
+        for (int cur_a = 0; cur_a < n_as; ++cur_a) {
+            const int64_t cne1 = matrix_row_counts[cur_a];
+
+            if (cne1 == 0) {
+                continue;
+            }
+
+            auto src0_cur = (const char *) src0->data + cur_a*nb02;
+
+            //const int64_t nr0 = ne01; // src0 rows
+            const int64_t nr1 = cne1; // src1 rows
+
+            int64_t src0_cur_start = (ith * ne01) / nth;
+            int64_t src0_cur_end   = ((ith + 1) * ne01) / nth;
+            src0_cur_start =
+                (src0_cur_start % NB_COLS) ? src0_cur_start + NB_COLS - (src0_cur_start % NB_COLS) : src0_cur_start;
+            src0_cur_end = (src0_cur_end % NB_COLS) ? src0_cur_end + NB_COLS - (src0_cur_end % NB_COLS) : src0_cur_end;
+
+            if (src0_cur_start >= src0_cur_end) return;
+
+            for (int ir1 = 0; ir1 < nr1; ir1++) {
+                struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1);
+                const int id       = row_mapping.i1; // selected expert index
+
+                const int64_t  i11 = id % ne11;
+                const int64_t  i12 = row_mapping.i2; // row index in src1
+
+                const int64_t  i1 = id;  // selected expert index
+                const int64_t  i2 = i12; // row
+
+                auto src1_col = (const char *) wdata + (i11 * nbw1 + i12 * nbw2);
+
+                gemv<BLOC_TYPE, INTER_SIZE, NB_COLS>(
+                        ne00, (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start,
+                        ne01,                    src0_cur + src0_cur_start * nb01,
+                        src1_col, 1, src0_cur_end - src0_cur_start);
+            }
+        }
+#undef MMID_MATRIX_ROW
+    }
+
+    int repack(struct ggml_tensor * t, const void * data, size_t data_size) override {
+        GGML_LOG_DEBUG("%s: repack tensor %s with %s_%dx%d\n", __func__, t->name, ggml_type_name(t->type),
+                       (int) NB_COLS, (int) INTER_SIZE);
+        return ggml::cpu::aarch64::repack<BLOC_TYPE, INTER_SIZE, NB_COLS>(t, data, data_size);
+    }
+};
+
+// instance for Q4
+static const tensor_traits<block_q4_0, 4, 4> q4_0_4x4_q8_0;
+static const tensor_traits<block_q4_0, 8, 4> q4_0_4x8_q8_0;
+static const tensor_traits<block_q4_0, 8, 8> q4_0_8x8_q8_0;
+
+// instance for IQ4
+static const tensor_traits<block_iq4_nl, 4, 4> iq4_nl_4x4_q8_0;
+
+}  // namespace ggml::cpu::aarch64
+
+static const ggml::cpu::tensor_traits * ggml_aarch64_get_optimal_repack_type(const struct ggml_tensor * cur) {
+    if (cur->type == GGML_TYPE_Q4_0) {
+        if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
+            if (cur->ne[1] % 8 == 0) {
+                return &ggml::cpu::aarch64::q4_0_8x8_q8_0;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+            if (cur->ne[1] % 4 == 0) {
+                return &ggml::cpu::aarch64::q4_0_4x8_q8_0;
+            }
+        }
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 4 == 0) {
+                return &ggml::cpu::aarch64::q4_0_4x4_q8_0;
+            }
+        }
+    } else if (cur->type == GGML_TYPE_IQ4_NL) {
+        if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
+            if (cur->ne[1] % 4 == 0) {
+                return &ggml::cpu::aarch64::iq4_nl_4x4_q8_0;
+            }
+        }
+    }
+
+    return nullptr;
+}
+
+static void ggml_backend_cpu_aarch64_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    tensor->extra = (void *) const_cast<ggml::cpu::tensor_traits *>(ggml_aarch64_get_optimal_repack_type(tensor));
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_aarch64_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor,
+                                                       const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(offset == 0);
+    GGML_ASSERT(size == ggml_nbytes(tensor));
+
+    auto tensor_traits = (ggml::cpu::aarch64::tensor_traits_base *) tensor->extra;
+    auto OK            = tensor_traits->repack(tensor, data, size);
+
+    GGML_ASSERT(OK == 0);
+    GGML_UNUSED(buffer);
+}
+
+static const char * ggml_backend_cpu_aarch64_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU_AARCH64";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_aarch64_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
+
+    if (buffer == nullptr) {
+        return nullptr;
+    }
+
+    buffer->buft              = buft;
+    buffer->iface.init_tensor = ggml_backend_cpu_aarch64_buffer_init_tensor;
+    buffer->iface.set_tensor  = ggml_backend_cpu_aarch64_buffer_set_tensor;
+    return buffer;
+}
+
+static size_t ggml_backend_cpu_aarch64_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+namespace ggml::cpu::aarch64 {
+class extra_buffer_type : ggml::cpu::extra_buffer_type {
+    bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override {
+        if (    op->op == GGML_OP_MUL_MAT &&
+                op->src[0]->buffer &&
+                (ggml_n_dims(op->src[0]) == 2) &&
+                op->src[0]->buffer->buft == ggml_backend_cpu_aarch64_buffer_type() &&
+                ggml_aarch64_get_optimal_repack_type(op->src[0])
+                ) {
+            if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
+                return false;
+            }
+            if (op->src[1]->type == GGML_TYPE_F32) {
+                return true;
+            }
+            //if (op->src[1]->type == GGML_TYPE_Q8_0) {
+            //    return true;
+            //}
+            // may be possible if Q8_0 packed...
+        } else if (op->op == GGML_OP_MUL_MAT_ID
+                && op->src[0]->buffer
+                && (ggml_n_dims(op->src[0]) == 3)
+                && op->src[0]->buffer->buft == ggml_backend_cpu_aarch64_buffer_type()
+                && ggml_aarch64_get_optimal_repack_type(op->src[0])
+                ) {
+            if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) {
+                return false;
+            }
+            if (op->src[1]->type == GGML_TYPE_F32) {
+                return true;
+            }
+            //if (op->src[1]->type == GGML_TYPE_Q8_0) {
+            //    return true;
+            //}
+        }
+        return false;
+    }
+
+    ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
+        if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_MUL_MAT_ID) {
+            if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_aarch64_buffer_type()) {
+                return (ggml::cpu::tensor_traits *) op->src[0]->extra;
+            }
+        }
+        return nullptr;
+    }
+};
+}  // namespace ggml::cpu::aarch64
+
+ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_aarch64 = {
+        /* .iface    = */ {
+                           /* .get_name         = */ ggml_backend_cpu_aarch64_buffer_type_get_name,
+                           /* .alloc_buffer     = */ ggml_backend_cpu_aarch64_buffer_type_alloc_buffer,
+                           /* .get_alignment    = */ ggml_backend_cpu_aarch64_buffer_type_get_alignment,
+                           /* .get_max_size     = */ nullptr,  // defaults to SIZE_MAX
+                           /* .get_alloc_size   = */ nullptr,  // defaults to ggml_nbytes
+                           /* .is_host          = */ nullptr,
+                           },
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context = */ new ggml::cpu::aarch64::extra_buffer_type(),
+    };
+
+    return &ggml_backend_cpu_buffer_type_aarch64;
+}
diff --git a/ggml/src/ggml-cpu/ggml-cpu-aarch64.h b/ggml/src/ggml-cpu/ggml-cpu-aarch64.h
new file mode 100644
index 00000000000..6e84c826b40
--- /dev/null
+++ b/ggml/src/ggml-cpu/ggml-cpu-aarch64.h
@@ -0,0 +1,8 @@
+#pragma once
+
+#include "ggml-cpu-traits.h"
+#include "ggml.h"
+
+// GGML internal header
+
+ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);
diff --git a/ggml/src/ggml-cpu/ggml-cpu-hbm.cpp b/ggml/src/ggml-cpu/ggml-cpu-hbm.cpp
new file mode 100644
index 00000000000..fa8dea2af9c
--- /dev/null
+++ b/ggml/src/ggml-cpu/ggml-cpu-hbm.cpp
@@ -0,0 +1,55 @@
+#ifdef GGML_USE_CPU_HBM
+
+#include "ggml-backend.h"
+#include "ggml-backend-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-impl.h"
+
+#include "ggml-cpu-hbm.h"
+
+// buffer type HBM
+
+#include <hbwmalloc.h>
+
+static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU_HBM";
+
+    GGML_UNUSED(buft);
+}
+
+static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    hbw_free(buffer->context);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
+                                                                           size_t                     size) {
+    void * ptr;
+    int    result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
+    if (result != 0) {
+        GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft                 = buft;
+    buffer->iface.free_buffer    = ggml_backend_cpu_hbm_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
+        /* .iface    = */ {
+                           /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
+                           /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
+                           /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+                           /* .get_max_size     = */ nullptr,  // defaults to SIZE_MAX
+                           /* .get_alloc_size   = */ nullptr,  // defaults to ggml_nbytes
+                           /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+                           },
+        /* .context  = */ nullptr,
+    };
+
+    return &ggml_backend_cpu_buffer_type_hbm;
+}
+#endif
diff --git a/ggml/src/ggml-cpu/ggml-cpu-hbm.h b/ggml/src/ggml-cpu/ggml-cpu-hbm.h
new file mode 100644
index 00000000000..09a1f09d72b
--- /dev/null
+++ b/ggml/src/ggml-cpu/ggml-cpu-hbm.h
@@ -0,0 +1,8 @@
+#pragma once
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+// GGML CPU internal header
+
+ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
diff --git a/ggml/src/ggml-cpu-impl.h b/ggml/src/ggml-cpu/ggml-cpu-impl.h
similarity index 53%
rename from ggml/src/ggml-cpu-impl.h
rename to ggml/src/ggml-cpu/ggml-cpu-impl.h
index 5b45155b028..d71076ad12b 100644
--- a/ggml/src/ggml-cpu-impl.h
+++ b/ggml/src/ggml-cpu/ggml-cpu-impl.h
@@ -15,6 +15,18 @@
 extern "C" {
 #endif
 
+struct ggml_compute_params {
+    // ith = thread index, nth = number of threads
+    int ith, nth;
+
+    // work buffer for all threads
+    size_t wsize;
+    void * wdata;
+
+    struct ggml_threadpool * threadpool;
+};
+
+
 #if defined(_MSC_VER)
 
 #define m512bh(p) p
@@ -27,80 +39,6 @@ extern "C" {
 
 #endif
 
-/**
- * Converts brain16 to float32.
- *
- * The bfloat16 floating point format has the following structure:
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───┐
- *     0b0000000000000000 brain16
- *
- * Since bf16 has the same number of exponent bits as a 32bit float,
- * encoding and decoding numbers becomes relatively straightforward.
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───────────────────┐
- *     0b00000000000000000000000000000000 IEEE binary32
- *
- * For comparison, the standard fp16 format has fewer exponent bits.
- *
- *       ┌sign
- *       │
- *       │  ┌exponent
- *       │  │
- *       │  │    ┌mantissa
- *       │  │    │
- *       │┌─┴─┐┌─┴──────┐
- *     0b0000000000000000 IEEE binary16
- *
- * @see IEEE 754-2008
- */
-static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.i = (uint32_t)h.bits << 16;
-    return u.f;
-}
-
-/**
- * Converts float32 to brain16.
- *
- * This is binary identical with Google Brain float conversion.
- * Floats shall round to nearest even, and NANs shall be quiet.
- * Subnormals aren't flushed to zero, except perhaps when used.
- * This code should vectorize nicely if using modern compilers.
- */
-static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
-    ggml_bf16_t h;
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.f = s;
-    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
-        h.bits = (u.i >> 16) | 64; /* force to quiet */
-        return h;
-    }
-    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
-    return h;
-}
-
-#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
-#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
-
 // __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
 #if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
 #ifndef __FMA__
@@ -388,28 +326,6 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 
 #endif // defined(__ARM_NEON)
 
-#if defined(__ARM_NEON) && !defined(_MSC_VER)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    ggml_fp16_internal_t tmp;
-    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-    return (float)tmp;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    ggml_fp16_t res;
-    ggml_fp16_internal_t tmp = f;
-    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-    return res;
-}
-
-#else
-
 #ifdef __wasm_simd128__
 #include <wasm_simd128.h>
 #else
@@ -462,152 +378,8 @@ static __m256 __lasx_xvreplfr2vr_s(float val) {
 }
 #endif
 
-#ifdef __F16C__
-
-#ifdef _MSC_VER
-#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-#else
-#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-#endif
-
-#elif defined(__POWER9_VECTOR__)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-/* the inline asm below is about 12% faster than the lookup method */
-#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    register float f;
-    register double d;
-    __asm__(
-        "mtfprd %0,%2\n"
-        "xscvhpdp %0,%0\n"
-        "frsp %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=f"(f):
-        /* in */   "r"(h));
-    return f;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    register double d;
-    register ggml_fp16_t r;
-    __asm__( /* xscvdphp can work on double or single precision */
-        "xscvdphp %0,%2\n"
-        "mffprd %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=r"(r):
-        /* in */   "f"(f));
-    return r;
-}
-
-#else
-
-// FP16 <-> FP32
-// ref: https://github.com/Maratyszcza/FP16
-
-static inline float fp32_from_bits(uint32_t w) {
-    union {
-        uint32_t as_bits;
-        float as_value;
-    } fp32;
-    fp32.as_bits = w;
-    return fp32.as_value;
-}
-
-static inline uint32_t fp32_to_bits(float f) {
-    union {
-        float as_value;
-        uint32_t as_bits;
-    } fp32;
-    fp32.as_value = f;
-    return fp32.as_bits;
-}
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    const uint32_t w = (uint32_t) h << 16;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    const uint32_t two_w = w + w;
-
-    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float exp_scale = 0x1.0p-112f;
-#else
-    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-#endif
-    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-    const uint32_t magic_mask = UINT32_C(126) << 23;
-    const float magic_bias = 0.5f;
-    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-    const uint32_t result = sign |
-        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-    return fp32_from_bits(result);
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float scale_to_inf = 0x1.0p+112f;
-    const float scale_to_zero = 0x1.0p-110f;
-#else
-    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-#endif
-    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-    const uint32_t w = fp32_to_bits(f);
-    const uint32_t shl1_w = w + w;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-    if (bias < UINT32_C(0x71000000)) {
-        bias = UINT32_C(0x71000000);
-    }
-
-    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-    const uint32_t bits = fp32_to_bits(base);
-    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-    const uint32_t nonsign = exp_bits + mantissa_bits;
-    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-}
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#endif // __F16C__
-
-#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
-
-#ifdef __ARM_FEATURE_SVE
-#include <arm_sve.h>
-#endif // __ARM_FEATURE_SVE
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in ggml_init()
-extern float ggml_table_f32_f16[1 << 16];
-
-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
-// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
-inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
-    uint16_t s;
-    memcpy(&s, &f, sizeof(uint16_t));
-    return ggml_table_f32_f16[s];
-}
-
-#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(GGML_FP32_TO_FP16)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-#endif
+// TODO: move to ggml-threading
+void ggml_barrier(struct ggml_threadpool * tp);
 
 #ifdef __cplusplus
 }
diff --git a/ggml/src/ggml-cpu/ggml-cpu-quants.c b/ggml/src/ggml-cpu/ggml-cpu-quants.c
new file mode 100644
index 00000000000..634c5fa1162
--- /dev/null
+++ b/ggml/src/ggml-cpu/ggml-cpu-quants.c
@@ -0,0 +1,10835 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+
+#include "ggml-quants.h"
+#include "ggml-cpu-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
+#include "ggml-cpu.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#if defined(_MSC_VER)
+// disable "possible loss of data" to avoid warnings for hundreds of casts
+// we should just be careful :)
+#pragma warning(disable: 4244 4267)
+#endif
+
+#define UNUSED GGML_UNUSED
+
+// some compilers don't provide _mm256_set_m128i, e.g. gcc 7
+#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
+// multiply int8_t, add results pairwise twice
+static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
+    // Get absolute values of x vectors
+    const __m128i ax = _mm_sign_epi8(x, x);
+    // Sign the values of the y vectors
+    const __m128i sy = _mm_sign_epi8(y, x);
+    // Perform multiplication and create 16-bit values
+    const __m128i dot = _mm_maddubs_epi16(ax, sy);
+    const __m128i ones = _mm_set1_epi16(1);
+    return _mm_madd_epi16(ones, dot);
+}
+
+#if __AVX__ || __AVX2__ || __AVX512F__
+// horizontally add 8 floats
+static inline float hsum_float_8(const __m256 x) {
+    __m128 res = _mm256_extractf128_ps(x, 1);
+    res = _mm_add_ps(res, _mm256_castps256_ps128(x));
+    res = _mm_add_ps(res, _mm_movehl_ps(res, res));
+    res = _mm_add_ss(res, _mm_movehdup_ps(res));
+    return _mm_cvtss_f32(res);
+}
+
+// horizontally add 8 int32_t
+static inline int hsum_i32_8(const __m256i a) {
+    const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1));
+    const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128);
+    const __m128i sum64 = _mm_add_epi32(hi64, sum128);
+    const __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
+    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
+}
+
+// horizontally add 4 int32_t
+static inline int hsum_i32_4(const __m128i a) {
+    const __m128i hi64 = _mm_unpackhi_epi64(a, a);
+    const __m128i sum64 = _mm_add_epi32(hi64, a);
+    const __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
+    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
+}
+
+#if defined(__AVX2__) || defined(__AVX512F__)
+// spread 32 bits to 32 bytes { 0x00, 0xFF }
+static inline __m256i bytes_from_bits_32(const uint8_t * x) {
+    uint32_t x32;
+    memcpy(&x32, x, sizeof(uint32_t));
+    const __m256i shuf_mask = _mm256_set_epi64x(
+            0x0303030303030303, 0x0202020202020202,
+            0x0101010101010101, 0x0000000000000000);
+    __m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask);
+    const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe);
+    bytes = _mm256_or_si256(bytes, bit_mask);
+    return _mm256_cmpeq_epi8(bytes, _mm256_set1_epi64x(-1));
+}
+
+// Unpack 32 4-bit fields into 32 bytes
+// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
+static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
+{
+    const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi);
+    const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp);
+    const __m256i lowMask = _mm256_set1_epi8( 0xF );
+    return _mm256_and_si256(lowMask, bytes);
+}
+
+// add int16_t pairwise and return as float vector
+static inline __m256 sum_i16_pairs_float(const __m256i x) {
+    const __m256i ones = _mm256_set1_epi16(1);
+    const __m256i summed_pairs = _mm256_madd_epi16(ones, x);
+    return _mm256_cvtepi32_ps(summed_pairs);
+}
+
+static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
+#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+    const __m256i zero = _mm256_setzero_si256();
+    const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy);
+    return _mm256_cvtepi32_ps(summed_pairs);
+#else
+    // Perform multiplication and create 16-bit values
+    const __m256i dot = _mm256_maddubs_epi16(ax, sy);
+    return sum_i16_pairs_float(dot);
+#endif
+}
+
+// multiply int8_t, add results pairwise twice and return as float vector
+static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
+#if __AVXVNNIINT8__
+    const __m256i zero = _mm256_setzero_si256();
+    const __m256i summed_pairs = _mm256_dpbssd_epi32(zero, x, y);
+    return _mm256_cvtepi32_ps(summed_pairs);
+#else
+    // Get absolute values of x vectors
+    const __m256i ax = _mm256_sign_epi8(x, x);
+    // Sign the values of the y vectors
+    const __m256i sy = _mm256_sign_epi8(y, x);
+    return mul_sum_us8_pairs_float(ax, sy);
+#endif
+}
+
+static inline __m128i packNibbles( __m256i bytes )
+{
+    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
+#if __AVX512F__
+    const __m256i bytes_srli_4 = _mm256_srli_epi16(bytes, 4);   // 0000_0000_abcd_0000
+    bytes = _mm256_or_si256(bytes, bytes_srli_4);               // 0000_abcd_abcd_efgh
+    return _mm256_cvtepi16_epi8(bytes);                         // abcd_efgh
+#else
+    const __m256i lowByte = _mm256_set1_epi16( 0xFF );
+    __m256i high = _mm256_andnot_si256( lowByte, bytes );
+    __m256i low = _mm256_and_si256( lowByte, bytes );
+    high = _mm256_srli_epi16( high, 4 );
+    bytes = _mm256_or_si256( low, high );
+
+    // Compress uint16_t lanes into bytes
+    __m128i r0 = _mm256_castsi256_si128( bytes );
+    __m128i r1 = _mm256_extracti128_si256( bytes, 1 );
+    return _mm_packus_epi16( r0, r1 );
+#endif
+}
+#elif defined(__AVX__)
+static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 )
+{
+    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
+    const __m128i lowByte = _mm_set1_epi16( 0xFF );
+    __m128i high = _mm_andnot_si128( lowByte, bytes1 );
+    __m128i low = _mm_and_si128( lowByte, bytes1 );
+    high = _mm_srli_epi16( high, 4 );
+    bytes1 = _mm_or_si128( low, high );
+    high = _mm_andnot_si128( lowByte, bytes2 );
+    low = _mm_and_si128( lowByte, bytes2 );
+    high = _mm_srli_epi16( high, 4 );
+    bytes2 = _mm_or_si128( low, high );
+
+    return _mm_packus_epi16( bytes1, bytes2);
+}
+
+static inline __m128i mul_add_epi8_sse(const __m128i x, const __m128i y) {
+    const __m128i ax = _mm_sign_epi8(x, x);
+    const __m128i sy = _mm_sign_epi8(y, x);
+    return _mm_maddubs_epi16(ax, sy);
+}
+
+// spread 32 bits to 32 bytes { 0x00, 0xFF }
+static inline __m256i bytes_from_bits_32(const uint8_t * x) {
+    uint32_t x32;
+    memcpy(&x32, x, sizeof(uint32_t));
+    const __m128i shuf_maskl = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
+    const __m128i shuf_maskh = _mm_set_epi64x(0x0303030303030303, 0x0202020202020202);
+    __m128i bytesl = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskl);
+    __m128i bytesh = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskh);
+    const __m128i bit_mask = _mm_set1_epi64x(0x7fbfdfeff7fbfdfe);
+    bytesl = _mm_or_si128(bytesl, bit_mask);
+    bytesh = _mm_or_si128(bytesh, bit_mask);
+    bytesl = _mm_cmpeq_epi8(bytesl, _mm_set1_epi64x(-1));
+    bytesh = _mm_cmpeq_epi8(bytesh, _mm_set1_epi64x(-1));
+    return MM256_SET_M128I(bytesh, bytesl);
+}
+
+// Unpack 32 4-bit fields into 32 bytes
+// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
+static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
+{
+    // Load 16 bytes from memory
+    __m128i tmpl = _mm_loadu_si128((const __m128i *)rsi);
+    __m128i tmph = _mm_srli_epi16(tmpl, 4);
+    const __m128i lowMask = _mm_set1_epi8(0xF);
+    tmpl = _mm_and_si128(lowMask, tmpl);
+    tmph = _mm_and_si128(lowMask, tmph);
+    return MM256_SET_M128I(tmph, tmpl);
+}
+
+// add int16_t pairwise and return as float vector
+static inline __m256 sum_i16_pairs_float(const __m128i xh, const __m128i xl) {
+    const __m128i ones = _mm_set1_epi16(1);
+    const __m128i summed_pairsl = _mm_madd_epi16(ones, xl);
+    const __m128i summed_pairsh = _mm_madd_epi16(ones, xh);
+    const __m256i summed_pairs = MM256_SET_M128I(summed_pairsh, summed_pairsl);
+    return _mm256_cvtepi32_ps(summed_pairs);
+}
+
+static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
+    const __m128i axl = _mm256_castsi256_si128(ax);
+    const __m128i axh = _mm256_extractf128_si256(ax, 1);
+    const __m128i syl = _mm256_castsi256_si128(sy);
+    const __m128i syh = _mm256_extractf128_si256(sy, 1);
+    // Perform multiplication and create 16-bit values
+    const __m128i dotl = _mm_maddubs_epi16(axl, syl);
+    const __m128i doth = _mm_maddubs_epi16(axh, syh);
+    return sum_i16_pairs_float(doth, dotl);
+}
+
+// multiply int8_t, add results pairwise twice and return as float vector
+static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
+    const __m128i xl = _mm256_castsi256_si128(x);
+    const __m128i xh = _mm256_extractf128_si256(x, 1);
+    const __m128i yl = _mm256_castsi256_si128(y);
+    const __m128i yh = _mm256_extractf128_si256(y, 1);
+    // Get absolute values of x vectors
+    const __m128i axl = _mm_sign_epi8(xl, xl);
+    const __m128i axh = _mm_sign_epi8(xh, xh);
+    // Sign the values of the y vectors
+    const __m128i syl = _mm_sign_epi8(yl, xl);
+    const __m128i syh = _mm_sign_epi8(yh, xh);
+    // Perform multiplication and create 16-bit values
+    const __m128i dotl = _mm_maddubs_epi16(axl, syl);
+    const __m128i doth = _mm_maddubs_epi16(axh, syh);
+    return sum_i16_pairs_float(doth, dotl);
+}
+
+// larger version of mul_sum_i8_pairs_float where x and y are each represented by four 128-bit vectors
+static inline __m256 mul_sum_i8_quad_float(const __m128i x_1_0, const __m128i x_1_1, const __m128i x_2_0, const __m128i x_2_1,
+                                           const __m128i y_1_0, const __m128i y_1_1, const __m128i y_2_0, const __m128i y_2_1) {
+    const __m128i mone = _mm_set1_epi16(1);
+
+    const __m128i p16_1_0 = mul_add_epi8_sse(x_1_0, y_1_0);
+    const __m128i p16_1_1 = mul_add_epi8_sse(x_1_1, y_1_1);
+    const __m128i p16_2_0 = mul_add_epi8_sse(x_2_0, y_2_0);
+    const __m128i p16_2_1 = mul_add_epi8_sse(x_2_1, y_2_1);
+    const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, mone);
+    const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, mone);
+    const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, mone);
+    const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, mone);
+    const __m128i p_1 = _mm_add_epi32(p_1_0, p_1_1);
+    const __m128i p_2 = _mm_add_epi32(p_2_0, p_2_1);
+    return _mm256_cvtepi32_ps(MM256_SET_M128I(p_2, p_1));
+}
+
+// quad fp16 delta calculation
+static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const float x1, const float y1) {
+    // GGML_FP16_TO_FP32 is faster than Intel F16C
+    return _mm256_set_m128(_mm_set1_ps(GGML_FP16_TO_FP32(x1) * GGML_FP16_TO_FP32(y1)),
+                           _mm_set1_ps(GGML_FP16_TO_FP32(x0) * GGML_FP16_TO_FP32(y0)));
+}
+#endif
+#elif defined(__SSSE3__)
+// horizontally add 4x4 floats
+static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
+    __m128 res_0 =_mm_hadd_ps(a, b);
+    __m128 res_1 =_mm_hadd_ps(c, d);
+    __m128 res =_mm_hadd_ps(res_0, res_1);
+    res =_mm_hadd_ps(res, res);
+    res =_mm_hadd_ps(res, res);
+
+    return _mm_cvtss_f32(res);
+}
+#endif // __AVX__ || __AVX2__ || __AVX512F__
+#endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
+
+#if defined(__ARM_NEON) || defined(__wasm_simd128__) || defined(__POWER9_VECTOR__)
+#define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
+#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
+#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
+#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
+#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
+#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
+#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
+#define B8(c,s  ) B7(c,s,     c), B7(c,s,     s)
+
+// precomputed tables for expanding 8bits to 8 bytes:
+static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4
+static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
+#endif
+
+#if defined(__loongarch_asx)
+
+#ifdef __clang__
+#define VREGS_PREFIX "$vr"
+#define XREGS_PREFIX "$xr"
+#else // GCC
+#define VREGS_PREFIX "$f"
+#define XREGS_PREFIX "$f"
+#endif
+#define __ALL_REGS "0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31"
+// Convert __m128i to __m256i
+static inline __m256i ____m256i(__m128i in) {
+    __m256i out = __lasx_xvldi(0);
+    __asm__ volatile (
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[out], " XREGS_PREFIX"\\i    \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[in], " VREGS_PREFIX "\\j  \n\t"
+        "    xvpermi.q $xr\\i, $xr\\j, 0x20  \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        : [out] "+f" (out) : [in] "f" (in)
+    );
+    return out;
+}
+// Convert two __m128i to __m256i
+static inline __m256i lasx_set_q(__m128i inhi, __m128i inlo) {
+    __m256i out;
+    __asm__ volatile (
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[hi], " VREGS_PREFIX "\\i    \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[lo], " VREGS_PREFIX "\\j  \n\t"
+        "    xvpermi.q $xr\\i, $xr\\j, 0x20  \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        ".ifnc %[out], %[hi]                 \n\t"
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[out], " XREGS_PREFIX "\\i   \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[hi], " VREGS_PREFIX "\\j  \n\t"
+        "    xvori.b $xr\\i, $xr\\j, 0       \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        ".endif                              \n\t"
+        : [out] "=f" (out), [hi] "+f" (inhi)
+        : [lo] "f" (inlo)
+    );
+    return out;
+}
+// Convert __m256i low part to __m128i
+static inline __m128i lasx_extracti128_lo(__m256i in) {
+    __m128i out;
+    __asm__ volatile (
+        ".ifnc %[out], %[in]                 \n\t"
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[out], " VREGS_PREFIX "\\i   \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[in], " XREGS_PREFIX "\\j  \n\t"
+        "    vori.b $vr\\i, $vr\\j, 0        \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        ".endif                              \n\t"
+        : [out] "=f" (out) : [in] "f" (in)
+    );
+    return out;
+}
+// Convert __m256i high part to __m128i
+static inline __m128i lasx_extracti128_hi(__m256i in) {
+    __m128i out;
+    __asm__ volatile (
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[out], " VREGS_PREFIX "\\i   \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[in], " XREGS_PREFIX "\\j  \n\t"
+        "    xvpermi.q $xr\\i, $xr\\j, 0x11  \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        : [out] "=f" (out) : [in] "f" (in)
+    );
+    return out;
+}
+
+static __m256i lasx_set_w(int e7, int e6, int e5, int e4, int e3, int e2, int e1, int e0) {
+    v8i32 __ret = {e0, e1, e2, e3, e4, e5, e6, e7};
+    return (__m256i)__ret;
+}
+
+static __m128i lsx_set_w(int32_t a, int32_t b, int32_t c, int32_t d) {
+    v4i32 __ret = {d, c, b, a};
+    return (__m128i)__ret;
+}
+
+static __m256i lasx_set_d(int64_t a, int64_t b, int64_t c, int64_t d) {
+    v4i64 __ret = {d, c, b, a};
+    return (__m256i)__ret;
+}
+
+static __m256i lasx_insertf128( __m128i x, __m128i y) {
+    return lasx_set_q(x, y);
+}
+
+static __m128i lsx_shuffle_b(__m128i a, __m128i b) {
+    __m128i mask_f, zero, tmp0, tmp2, mask;
+    int f = 0x8f;
+    mask_f = __lsx_vreplgr2vr_b(f);
+    zero = __lsx_vldi(0);
+    tmp0 = __lsx_vand_v(b, mask_f); // get mask with low 4 bit and sign bits
+    tmp0 = __lsx_vori_b(tmp0, 0x10); // make each mask or  with 0x10 prepare for positive
+    mask = __lsx_vsle_b(zero, tmp0); // if mask >= 0, set mask
+    tmp2 = __lsx_vand_v(tmp0, mask); // maskout the in2 < ones
+    return __lsx_vshuf_b(a, zero, tmp2);
+}
+
+static __m256i lasx_shuffle_b(__m256i a, __m256i b) {
+    __m256i mask_f, zero, tmp0, tmp2, mask;
+    int f = 0x8f;
+    mask_f = __lasx_xvreplgr2vr_b(f);
+    zero = __lasx_xvldi(0);
+    tmp0 = __lasx_xvand_v(b, mask_f); // get mask with low 4 bit and sign bits
+    tmp0 = __lasx_xvori_b(tmp0, 0x10); // make each mask or  with 0x10 prepare for positive
+    mask = __lasx_xvsle_b(zero, tmp0); // if mask >= 0, set mask
+    tmp2 = __lasx_xvand_v(tmp0, mask); // maskout the in2 < ones
+    return __lasx_xvshuf_b(a, zero, tmp2);
+}
+
+static __m256i lasx_extu8_16(__m128i a) {
+    __m128i zero = __lsx_vldi(0);
+    __m128i vlo = __lsx_vilvl_b(zero, a);
+    __m128i vhi = __lsx_vilvh_b(zero, a);
+    return lasx_set_q(vhi, vlo);
+}
+
+static __m256i lasx_ext8_16(__m128i a) {
+     __m128i sign = __lsx_vslti_b(a, 0);
+     __m128i vlo = __lsx_vilvl_b(sign, a);
+     __m128i vhi = __lsx_vilvh_b(sign, a);
+     return lasx_set_q(vhi, vlo);
+}
+
+static __m256i lasx_ext16_32(__m128i a) {
+    __m256i tmp1;
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 0), 0);
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 1), 1);
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 2), 2);
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 3), 3);
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 4), 4);
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 5), 5);
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 6), 6);
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 7), 7);
+    return tmp1;
+}
+
+static __m128i lasx_extracti128( __m256i a, int pos) {
+    __m128i ret;
+    if( pos == 0)
+    {
+       ret = lasx_extracti128_lo(a);
+    } else {
+       ret = lasx_extracti128_hi(a);
+    }
+    return ret;
+}
+
+static __m128 lasx_extractf128( __m256 a, int pos) {
+    __m128 ret;
+    if( pos == 0)
+    {
+       ret = (__m128)lasx_extracti128_lo((__m256i)a);
+    } else {
+       ret = (__m128)lasx_extracti128_hi((__m256i)a);
+    }
+    return ret;
+}
+
+static __m128i lsx_hadd_h(__m128i a, __m128i b) {
+    __m128i tmp1 = __lsx_vpickev_h(b, a);
+    __m128i tmp2 = __lsx_vpickod_h(b, a);
+    return __lsx_vadd_h(tmp1, tmp2);
+}
+
+static __m128i lsx_hadd_w(__m128i a, __m128i b) {
+    __m128i tmp1 = __lsx_vpickev_w(b, a);
+    __m128i tmp2 = __lsx_vpickod_w(b, a);
+    return __lsx_vadd_w(tmp1, tmp2);
+}
+
+static __m128 lsx_hadd_s(__m128 a, __m128 b) {
+    __m128 tmp1 = (__m128)__lsx_vpickev_w((__m128i)b, (__m128i)a);
+    __m128 tmp2 = (__m128)__lsx_vpickod_w((__m128i)b, (__m128i)a);
+
+    return __lsx_vfadd_s(tmp1, tmp2);
+}
+
+static __m256i lasx_maddubs_h(__m256i a, __m256i b) {
+    __m256i tmp1, tmp2;
+    tmp1 = __lasx_xvmulwev_h_b(a, b);
+    tmp2 = __lasx_xvmulwod_h_b(a, b);
+    return __lasx_xvsadd_h(tmp1, tmp2);
+}
+
+static __m256i lasx_madd_h(__m256i a, __m256i b) {
+    __m256i tmp1, tmp2;
+    tmp1 = __lasx_xvmulwev_w_h(a, b);
+    tmp2 = __lasx_xvmulwod_w_h(a, b);
+    return __lasx_xvadd_w(tmp1, tmp2);
+}
+
+static __m256i lasx_packs_w(__m256i a, __m256i b) {
+    __m256i tmp, tmp1;
+    tmp = __lasx_xvsat_w(a, 15);
+    tmp1 = __lasx_xvsat_w(b, 15);
+    return __lasx_xvpickev_h(tmp1, tmp);
+}
+
+static __m256i lasx_packs_h(__m256i a, __m256i b) {
+    __m256i tmp, tmp1;
+    tmp = __lasx_xvsat_h(a, 7);
+    tmp1 = __lasx_xvsat_h(b, 7);
+    return __lasx_xvpickev_b(tmp1, tmp);
+}
+
+static __m128i lsx_packs_w(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_w(a, 15);
+    tmp1 = __lsx_vsat_w(b, 15);
+    return __lsx_vpickev_h(tmp1, tmp);
+}
+
+static __m128i lsx_packs_h(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_h(a, 7);
+    tmp1 = __lsx_vsat_h(b, 7);
+    return __lsx_vpickev_b(tmp1, tmp);
+}
+
+static __m128i lsx_packus_h(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_hu(a, 7);
+    tmp1 = __lsx_vsat_hu(b, 7);
+    return __lsx_vpickev_b(tmp1, tmp);
+}
+
+
+static __m128i lsx_maddubs_h(__m128i a, __m128i b) {
+    __m128i tmp1, tmp2;
+    tmp1 = __lsx_vmulwev_h_b(a, b);
+    tmp2 = __lsx_vmulwod_h_b(a, b);
+    return __lsx_vsadd_h(tmp1, tmp2);
+}
+
+static __m128i lsx_madd_h(__m128i a, __m128i b) {
+    __m128i tmp1, tmp2;
+    tmp1 = __lsx_vmulwev_w_h(a, b);
+    tmp2 = __lsx_vmulwod_w_h(a, b);
+    return __lsx_vadd_w(tmp1, tmp2);
+}
+
+// multiply int8_t, add results pairwise twice
+static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
+    // Get absolute values of x vectors
+    const __m128i ax = __lsx_vsigncov_b(x, x);
+    // Sign the values of the y vectors
+    const __m128i sy = __lsx_vsigncov_b(x, y);
+    // Perform multiplication and create 16-bit values
+    const __m128i dot = lsx_maddubs_h(ax, sy);
+    const __m128i ones = __lsx_vreplgr2vr_h(1);
+    return lsx_madd_h(ones, dot);
+}
+
+// horizontally add 8 floats
+static inline float hsum_float_8(const __m256 x) {
+    __m128 res = lasx_extractf128(x, 1);
+    ft_union tmp;
+    res = __lsx_vfadd_s(res, lasx_extractf128(x, 0));
+    res = __lsx_vfadd_s(res, (__m128)__lsx_vpickod_d((__m128i)res, (__m128i)res));
+    res = __lsx_vfadd_s(res, (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w(res, 1), 0));
+    tmp.i = __lsx_vpickve2gr_w(res, 0);
+    return tmp.f;
+}
+
+// horizontally add 8 int32_t
+static inline int hsum_i32_8(const __m256i a) {
+
+    __m256i tmp1 = __lasx_xvpermi_q(a, a, 0x11);
+    __m256i tmp2 = __lasx_xvpermi_q(a, a, 0x00);
+
+    __m128i  tmp1_128 = lasx_extracti128_lo(tmp1);
+    __m128i  tmp2_128 = lasx_extracti128_lo(tmp2);
+
+    __m128i sum128 = __lsx_vadd_w(tmp1_128, tmp2_128);
+
+    __m128i ev = __lsx_vpickev_w(sum128, sum128);
+    __m128i od = __lsx_vpickod_w(sum128, sum128);
+    __m128i sum64 = __lsx_vadd_w(ev, od);
+
+    int sum64_1, sum64_2;
+    sum64_1 = __lsx_vpickve2gr_w(sum64, 0);
+    sum64_2 = __lsx_vpickve2gr_w(sum64, 1);
+
+    return  sum64_1 + sum64_2;
+}
+
+// horizontally add 4 int32_t
+static inline int hsum_i32_4(const __m128i a) {
+    __m128i ev = __lsx_vpickev_w(a, a);
+    __m128i od = __lsx_vpickod_w(a, a);
+    __m128i sum64 = __lsx_vadd_w(ev, od);
+
+    int sum64_1, sum64_2;
+    sum64_1 = __lsx_vpickve2gr_w(sum64, 0);
+    sum64_2 = __lsx_vpickve2gr_w(sum64, 1);
+
+    return  sum64_1 + sum64_2;
+}
+
+// spread 32 bits to 32 bytes { 0x00, 0xFF }
+static inline __m256i bytes_from_bits_32(const uint8_t * x) {
+
+    uint32_t x32;
+    memcpy(&x32, x, sizeof(uint32_t));
+    const __m256i shuf_mask = lasx_set_d(
+            0x0303030303030303, 0x0202020202020202,
+            0x0101010101010101, 0x0000000000000000);
+
+    __m256i bytes = lasx_shuffle_b(__lasx_xvreplgr2vr_w(x32), shuf_mask);
+    const __m256i bit_mask = __lasx_xvreplgr2vr_d(0x7fbfdfeff7fbfdfe);
+    bytes = __lasx_xvor_v(bytes, bit_mask);
+    return __lasx_xvseq_b(bytes, __lasx_xvreplgr2vr_d(-1));
+}
+
+// Unpack 32 4-bit fields into 32 bytes
+// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
+static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) {
+    const __m128i lo = __lsx_vld((const __m128i *)rsi, 0);
+    __m128i hi = __lsx_vsrli_h(lo, 4);
+    return __lasx_xvandi_b(lasx_insertf128(hi, lo), 0xf);
+}
+
+// add int16_t pairwise and return as float vector
+static inline __m256 sum_i16_pairs_float(const __m256i x) {
+    __m256i v = __lasx_xvpackod_h(x, x);
+    __m256i summed_pairs = __lasx_xvaddwev_w_h(x, v);
+    return __lasx_xvffint_s_w(summed_pairs);
+}
+
+static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
+    // Perform multiplication and create 16-bit values
+    const __m256i dot = lasx_maddubs_h(ax, sy);
+    return sum_i16_pairs_float(dot);
+}
+
+// multiply int8_t, add results pairwise twice and return as float vector
+static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
+
+    // Get absolute values of x vectors
+    const __m256i ax = __lasx_xvsigncov_b(x, x);
+    // Sign the values of the y vectors
+    const __m256i sy = __lasx_xvsigncov_b(x, y);
+
+    return mul_sum_us8_pairs_float(ax, sy);
+}
+
+static inline __m128i packNibbles( __m256i bytes ) {
+    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
+    const __m256i lowByte = __lasx_xvreplgr2vr_h(0xFF);
+     __m256i high = __lasx_xvandn_v(lowByte, bytes);
+    __m256i low = __lasx_xvand_v(lowByte, bytes);
+    high = __lasx_xvsrli_h(high, 4);
+    bytes = __lasx_xvor_v(low, high);
+    // Compress uint16_t lanes into bytes
+    __m128i *r0 = (__m128i *)&bytes;
+    __m256i tmp_h128 = __lasx_xvpermi_q(bytes, bytes, 0x11);
+    __m128i *r1 = (__m128i *)&tmp_h128;
+
+    __m128i zero = __lsx_vldi(0);
+    __m128i tmp, tmp2, tmp3;
+
+    tmp = __lsx_vmax_h(zero, *r0);
+    tmp2 = __lsx_vsat_hu(tmp, 7);
+
+    tmp = __lsx_vmax_h(zero, *r1);
+    tmp3 = __lsx_vsat_hu(tmp, 7);
+    return  __lsx_vpickev_b(tmp3, tmp2);
+}
+#endif  //__loongarch_asx
+
+void quantize_row_q4_0(const float * restrict x, void * restrict y, int64_t k) {
+    quantize_row_q4_0_ref(x, y, k);
+}
+
+void quantize_row_q4_1(const float * restrict x, void * restrict y, int64_t k) {
+    quantize_row_q4_1_ref(x, y, k);
+}
+
+void quantize_row_q5_0(const float * restrict x, void * restrict y, int64_t k) {
+    quantize_row_q5_0_ref(x, y, k);
+}
+
+void quantize_row_q5_1(const float * restrict x, void * restrict y, int64_t k) {
+    quantize_row_q5_1_ref(x, y, k);
+}
+
+void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0 * restrict y = vy;
+
+#if defined(__ARM_NEON)
+    for (int i = 0; i < nb; i++) {
+        float32x4_t srcv [8];
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = vld1q_f32(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = vmaxvq_f32(amaxv[0]);
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        for (int j = 0; j < 8; j++) {
+            const float32x4_t v  = vmulq_n_f32(srcv[j], id);
+            const int32x4_t   vi = vcvtnq_s32_f32(v);
+
+            y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
+        }
+    }
+#elif defined(__wasm_simd128__)
+    for (int i = 0; i < nb; i++) {
+        v128_t srcv [8];
+        v128_t asrcv[8];
+        v128_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
+                                   wasm_f32x4_extract_lane(amaxv[0], 1)),
+                               MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
+                                   wasm_f32x4_extract_lane(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        for (int j = 0; j < 8; j++) {
+            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
+            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
+
+            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
+            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
+            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
+            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
+        }
+    }
+#elif defined(__AVX2__) || defined(__AVX__)
+    for (int i = 0; i < nb; i++) {
+        // Load elements into 4 AVX vectors
+        __m256 v0 = _mm256_loadu_ps( x );
+        __m256 v1 = _mm256_loadu_ps( x + 8 );
+        __m256 v2 = _mm256_loadu_ps( x + 16 );
+        __m256 v3 = _mm256_loadu_ps( x + 24 );
+        x += 32;
+
+        // Compute max(abs(e)) for the block
+        const __m256 signBit = _mm256_set1_ps( -0.0f );
+        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
+
+        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
+        max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
+        max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
+        const float maxScalar = _mm_cvtss_f32( max4 );
+
+        // Quantize these floats
+        const float d = maxScalar / 127.f;
+        y[i].d = GGML_FP32_TO_FP16(d);
+        const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
+        const __m256 mul = _mm256_set1_ps( id );
+
+        // Apply the multiplier
+        v0 = _mm256_mul_ps( v0, mul );
+        v1 = _mm256_mul_ps( v1, mul );
+        v2 = _mm256_mul_ps( v2, mul );
+        v3 = _mm256_mul_ps( v3, mul );
+
+        // Round to nearest integer
+        v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
+        v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
+        v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
+        v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
+
+        // Convert floats to integers
+        __m256i i0 = _mm256_cvtps_epi32( v0 );
+        __m256i i1 = _mm256_cvtps_epi32( v1 );
+        __m256i i2 = _mm256_cvtps_epi32( v2 );
+        __m256i i3 = _mm256_cvtps_epi32( v3 );
+
+#if defined(__AVX2__)
+        // Convert int32 to int16
+        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
+        i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
+                                            // Convert int16 to int8
+        i0 = _mm256_packs_epi16( i0, i2 );	// 0, 1, 2, 3,  8, 9, 10, 11,  16, 17, 18, 19,  24, 25, 26, 27,  4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
+
+        // We got our precious signed bytes, but the order is now wrong
+        // These AVX2 pack instructions process 16-byte pieces independently
+        // The following instruction is fixing the order
+        const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
+        i0 = _mm256_permutevar8x32_epi32( i0, perm );
+
+        _mm256_storeu_si256((__m256i *)y[i].qs, i0);
+#else
+        // Since we don't have in AVX some necessary functions,
+        // we split the registers in half and call AVX2 analogs from SSE
+        __m128i ni0 = _mm256_castsi256_si128( i0 );
+        __m128i ni1 = _mm256_extractf128_si256( i0, 1);
+        __m128i ni2 = _mm256_castsi256_si128( i1 );
+        __m128i ni3 = _mm256_extractf128_si256( i1, 1);
+        __m128i ni4 = _mm256_castsi256_si128( i2 );
+        __m128i ni5 = _mm256_extractf128_si256( i2, 1);
+        __m128i ni6 = _mm256_castsi256_si128( i3 );
+        __m128i ni7 = _mm256_extractf128_si256( i3, 1);
+
+        // Convert int32 to int16
+        ni0 = _mm_packs_epi32( ni0, ni1 );
+        ni2 = _mm_packs_epi32( ni2, ni3 );
+        ni4 = _mm_packs_epi32( ni4, ni5 );
+        ni6 = _mm_packs_epi32( ni6, ni7 );
+        // Convert int16 to int8
+        ni0 = _mm_packs_epi16( ni0, ni2 );
+        ni4 = _mm_packs_epi16( ni4, ni6 );
+
+        _mm_storeu_si128((__m128i *)(y[i].qs +  0), ni0);
+        _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
+#endif
+    }
+#elif defined(__riscv_v_intrinsic)
+
+    size_t vl = __riscv_vsetvl_e32m4(QK8_0);
+
+    for (int i = 0; i < nb; i++) {
+        // load elements
+        vfloat32m4_t v_x   = __riscv_vle32_v_f32m4(x+i*QK8_0, vl);
+
+        vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl);
+        vfloat32m1_t tmp   = __riscv_vfmv_v_f_f32m1(0.0f, vl);
+        vfloat32m1_t vmax  = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl);
+        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
+
+        // convert to integer
+        vint16m2_t   vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl);
+        vint8m1_t    vs = __riscv_vncvt_x_x_w_i8m1(vi, vl);
+
+        // store result
+        __riscv_vse8_v_i8m1(y[i].qs , vs, vl);
+    }
+
+#elif defined(__POWER9_VECTOR__)
+    for (int i = 0; i < nb; i++) {
+        vector float srcv [8];
+        vector float asrcv[8];
+        vector float amaxv[8];
+        vector signed int vi[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = vec_xl(0, x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = vec_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = vec_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = vec_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(vec_extract(amaxv[0], 0),
+                                   vec_extract(amaxv[0], 1)),
+                               MAX(vec_extract(amaxv[0], 2),
+                                   vec_extract(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+        const vector float vid = vec_splats(id);
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        for (int j = 0; j < 8; j++) {
+            const vector float v  = vec_round(vec_mul(srcv[j], vid));
+            vi[j] = vec_cts(v, 0);
+        }
+        vec_xst(vec_pack(vec_pack(vi[0], vi[1]), vec_pack(vi[2], vi[3])),  0, &y[i].qs[0]);
+        vec_xst(vec_pack(vec_pack(vi[4], vi[5]), vec_pack(vi[6], vi[7])), 16, &y[i].qs[0]);
+    }
+
+#elif defined(__loongarch_asx)
+    for (int i = 0; i < nb; i++) {
+        ft_union fi;
+        __m256 v0 = (__m256)__lasx_xvld( x , 0);
+        __m256 v1 = (__m256)__lasx_xvld( x , 32);
+        __m256 v2 = (__m256)__lasx_xvld( x , 64);
+        __m256 v3 = (__m256)__lasx_xvld( x , 96);
+        x += 32;
+
+        // Compute max(abs(e)) for the block
+        const __m256 sign_bit = __lasx_xvreplfr2vr_s( -0.0f );
+        __m256 max_abs = (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v0 );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v1 ) );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v2 ) );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v3 ) );
+
+        __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs , 0) );
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
+        __m128 tmp = max4;
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vinsgr2vr_w(tmp, __lsx_vpickve2gr_w( max4, 1 ), 0 ));
+        fi.i = __lsx_vpickve2gr_w( (__m128i)max4, 0 );
+        const float max_scalar = fi.f;
+
+        // Quantize these floats
+        const float d = max_scalar / 127.f;
+        y[i].d = GGML_FP32_TO_FP16(d);
+        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
+        const __m256 mul = (__m256)__lasx_xvreplfr2vr_s( id );
+
+        // Apply the multiplier
+        v0 = __lasx_xvfmul_s( v0, mul );
+        v1 = __lasx_xvfmul_s( v1, mul );
+        v2 = __lasx_xvfmul_s( v2, mul );
+        v3 = __lasx_xvfmul_s( v3, mul );
+
+        // Round to nearest integer
+        __m256i i0 = __lasx_xvftintrne_w_s( v0 );
+        __m256i i1 = __lasx_xvftintrne_w_s( v1 );
+        __m256i i2 = __lasx_xvftintrne_w_s( v2 );
+        __m256i i3 = __lasx_xvftintrne_w_s( v3 );
+
+        __m128i ni0 = lasx_extracti128( i0, 0 );
+        __m128i ni1 = lasx_extracti128( i0, 1);
+        __m128i ni2 = lasx_extracti128( i1, 0);
+        __m128i ni3 = lasx_extracti128( i1, 1);
+        __m128i ni4 = lasx_extracti128( i2, 0);
+        __m128i ni5 = lasx_extracti128( i2, 1);
+        __m128i ni6 = lasx_extracti128( i3, 0);
+        __m128i ni7 = lasx_extracti128( i3, 1);
+
+        // Convert int32 to int16
+        ni0 = lsx_packs_w( ni0, ni1 );
+        ni2 = lsx_packs_w( ni2, ni3 );
+        ni4 = lsx_packs_w( ni4, ni5 );
+        ni6 = lsx_packs_w( ni6, ni7 );
+        // Convert int16 to int8
+        ni0 = lsx_packs_h( ni0, ni2 );
+        ni4 = lsx_packs_h( ni4, ni6 );
+
+        __lsx_vst(ni0, (__m128i *)(y[i].qs +  0), 0);
+        __lsx_vst(ni4, (__m128i *)(y[i].qs + 16), 0);
+
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_0_ref(x, y, k);
+#endif
+}
+
+void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK8_1 == 0);
+    const int nb = k / QK8_1;
+
+    block_q8_1 * restrict y = vy;
+
+#if defined(__ARM_NEON)
+    for (int i = 0; i < nb; i++) {
+        float32x4_t srcv [8];
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = vld1q_f32(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = vmaxvq_f32(amaxv[0]);
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        int32x4_t accv = vdupq_n_s32(0);
+
+        for (int j = 0; j < 8; j++) {
+            const float32x4_t v  = vmulq_n_f32(srcv[j], id);
+            const int32x4_t   vi = vcvtnq_s32_f32(v);
+
+            y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
+
+            accv = vaddq_s32(accv, vi);
+        }
+
+        y[i].s = GGML_FP32_TO_FP16(d * vaddvq_s32(accv));
+    }
+#elif defined(__wasm_simd128__)
+    for (int i = 0; i < nb; i++) {
+        v128_t srcv [8];
+        v128_t asrcv[8];
+        v128_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
+                                   wasm_f32x4_extract_lane(amaxv[0], 1)),
+                               MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
+                                   wasm_f32x4_extract_lane(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        v128_t accv = wasm_i32x4_splat(0);
+
+        for (int j = 0; j < 8; j++) {
+            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
+            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
+
+            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
+            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
+            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
+            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
+
+            accv = wasm_i32x4_add(accv, vi);
+        }
+
+        y[i].s = GGML_FP32_TO_FP16(
+                d * (wasm_i32x4_extract_lane(accv, 0) +
+                     wasm_i32x4_extract_lane(accv, 1) +
+                     wasm_i32x4_extract_lane(accv, 2) +
+                     wasm_i32x4_extract_lane(accv, 3)));
+    }
+#elif defined(__AVX2__) || defined(__AVX__)
+    for (int i = 0; i < nb; i++) {
+        // Load elements into 4 AVX vectors
+        __m256 v0 = _mm256_loadu_ps( x );
+        __m256 v1 = _mm256_loadu_ps( x + 8 );
+        __m256 v2 = _mm256_loadu_ps( x + 16 );
+        __m256 v3 = _mm256_loadu_ps( x + 24 );
+        x += 32;
+
+        // Compute max(abs(e)) for the block
+        const __m256 signBit = _mm256_set1_ps( -0.0f );
+        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
+
+        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
+        max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
+        max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
+        const float max_scalar = _mm_cvtss_f32( max4 );
+
+        // Quantize these floats
+        const float d = max_scalar / 127.f;
+        y[i].d = GGML_FP32_TO_FP16(d);
+        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
+        const __m256 mul = _mm256_set1_ps( id );
+
+        // Apply the multiplier
+        v0 = _mm256_mul_ps( v0, mul );
+        v1 = _mm256_mul_ps( v1, mul );
+        v2 = _mm256_mul_ps( v2, mul );
+        v3 = _mm256_mul_ps( v3, mul );
+
+        // Round to nearest integer
+        v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
+        v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
+        v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
+        v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
+
+        // Convert floats to integers
+        __m256i i0 = _mm256_cvtps_epi32( v0 );
+        __m256i i1 = _mm256_cvtps_epi32( v1 );
+        __m256i i2 = _mm256_cvtps_epi32( v2 );
+        __m256i i3 = _mm256_cvtps_epi32( v3 );
+
+#if defined(__AVX2__)
+        // Compute the sum of the quants and set y[i].s
+        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));
+
+        // Convert int32 to int16
+        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
+        i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
+                                            // Convert int16 to int8
+        i0 = _mm256_packs_epi16( i0, i2 );	// 0, 1, 2, 3,  8, 9, 10, 11,  16, 17, 18, 19,  24, 25, 26, 27,  4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
+
+        // We got our precious signed bytes, but the order is now wrong
+        // These AVX2 pack instructions process 16-byte pieces independently
+        // The following instruction is fixing the order
+        const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
+        i0 = _mm256_permutevar8x32_epi32( i0, perm );
+
+        _mm256_storeu_si256((__m256i *)y[i].qs, i0);
+#else
+        // Since we don't have in AVX some necessary functions,
+        // we split the registers in half and call AVX2 analogs from SSE
+        __m128i ni0 = _mm256_castsi256_si128( i0 );
+        __m128i ni1 = _mm256_extractf128_si256( i0, 1);
+        __m128i ni2 = _mm256_castsi256_si128( i1 );
+        __m128i ni3 = _mm256_extractf128_si256( i1, 1);
+        __m128i ni4 = _mm256_castsi256_si128( i2 );
+        __m128i ni5 = _mm256_extractf128_si256( i2, 1);
+        __m128i ni6 = _mm256_castsi256_si128( i3 );
+        __m128i ni7 = _mm256_extractf128_si256( i3, 1);
+
+        // Compute the sum of the quants and set y[i].s
+        const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
+        const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
+        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
+
+        // Convert int32 to int16
+        ni0 = _mm_packs_epi32( ni0, ni1 );
+        ni2 = _mm_packs_epi32( ni2, ni3 );
+        ni4 = _mm_packs_epi32( ni4, ni5 );
+        ni6 = _mm_packs_epi32( ni6, ni7 );
+        // Convert int16 to int8
+        ni0 = _mm_packs_epi16( ni0, ni2 );
+        ni4 = _mm_packs_epi16( ni4, ni6 );
+
+        _mm_storeu_si128((__m128i *)(y[i].qs +  0), ni0);
+        _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
+#endif
+    }
+#elif defined(__riscv_v_intrinsic)
+
+    size_t vl = __riscv_vsetvl_e32m4(QK8_1);
+
+    for (int i = 0; i < nb; i++) {
+        // load elements
+        vfloat32m4_t v_x   = __riscv_vle32_v_f32m4(x+i*QK8_1, vl);
+
+        vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl);
+        vfloat32m1_t tmp   = __riscv_vfmv_v_f_f32m1(0.0, vl);
+        vfloat32m1_t vmax  = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl);
+        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
+
+        const float d  = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
+
+        // convert to integer
+        vint16m2_t   vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl);
+        vint8m1_t    vs = __riscv_vncvt_x_x_w_i8m1(vi, vl);
+
+        // store result
+        __riscv_vse8_v_i8m1(y[i].qs , vs, vl);
+
+        // compute sum for y[i].s
+        vint16m1_t tmp2 = __riscv_vmv_v_x_i16m1(0, vl);
+        vint16m1_t vwrs = __riscv_vwredsum_vs_i8m1_i16m1(vs, tmp2, vl);
+
+        // set y[i].s
+        int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
+        y[i].s = GGML_FP32_TO_FP16(sum*d);
+    }
+
+#elif defined(__POWER9_VECTOR__)
+    for (int i = 0; i < nb; i++) {
+        vector float srcv [8];
+        vector float asrcv[8];
+        vector float amaxv[8];
+        vector signed int vi[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = vec_xl(0, x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = vec_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = vec_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = vec_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(vec_extract(amaxv[0], 0),
+                                   vec_extract(amaxv[0], 1)),
+                               MAX(vec_extract(amaxv[0], 2),
+                                   vec_extract(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+        const vector float vid = vec_splats(id);
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        vector int accv = vec_splats(0);
+
+        for (int j = 0; j < 8; j++) {
+            const vector float v  = vec_round(vec_mul(srcv[j], vid));
+            vi[j] = vec_cts(v, 0);
+
+            accv = vec_add(accv, vi[j]);
+        }
+        vec_xst(vec_pack(vec_pack(vi[0], vi[1]), vec_pack(vi[2], vi[3])),  0, &y[i].qs[0]);
+        vec_xst(vec_pack(vec_pack(vi[4], vi[5]), vec_pack(vi[6], vi[7])), 16, &y[i].qs[0]);
+
+        accv = vec_add(accv, vec_sld(accv, accv, 4));
+        accv = vec_add(accv, vec_sld(accv, accv, 8));
+        y[i].s = GGML_FP32_TO_FP16(d * vec_extract(accv, 0));
+    }
+
+#elif defined(__loongarch_asx)
+    for (int i = 0; i < nb; i++) {
+        ft_union ft;
+        __m256 v0 = (__m256)__lasx_xvld( x , 0 );
+        __m256 v1 = (__m256)__lasx_xvld( x , 32 );
+        __m256 v2 = (__m256)__lasx_xvld( x , 64 );
+        __m256 v3 = (__m256)__lasx_xvld( x , 96 );
+        x += 32;
+
+        // Compute max(abs(e)) for the block
+        const __m256 sign_bit = __lasx_xvreplfr2vr_s( -0.0f );
+        __m256 max_abs = (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v0 );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v1 ) );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v2 ) );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v3 ) );
+
+        __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs, 0) );
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
+        __m128 tmp = max4;
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x10 ));
+        ft.i = __lsx_vpickve2gr_w( (__m128i)max4, 0 );
+        const float max_scalar = ft.f;
+
+        // Quantize these floats
+        const float d = max_scalar / 127.f;
+        y[i].d = GGML_FP32_TO_FP16(d);
+        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
+        const __m256 mul = __lasx_xvreplfr2vr_s( id );
+
+        // Apply the multiplier
+        v0 = __lasx_xvfmul_s( v0, mul );
+        v1 = __lasx_xvfmul_s( v1, mul );
+        v2 = __lasx_xvfmul_s( v2, mul );
+        v3 = __lasx_xvfmul_s( v3, mul );
+
+        // Round to nearest integer
+        __m256i i0 = __lasx_xvftintrne_w_s( v0 );
+        __m256i i1 = __lasx_xvftintrne_w_s( v1 );
+        __m256i i2 = __lasx_xvftintrne_w_s( v2 );
+        __m256i i3 = __lasx_xvftintrne_w_s( v3 );
+
+        __m128i ni0 = lasx_extracti128(i0, 0);
+        __m128i ni1 = lasx_extracti128( i0, 1);
+        __m128i ni2 = lasx_extracti128( i1, 0);
+        __m128i ni3 = lasx_extracti128( i1, 1);
+        __m128i ni4 = lasx_extracti128( i2, 0 );
+        __m128i ni5 = lasx_extracti128( i2, 1);
+        __m128i ni6 = lasx_extracti128( i3, 0);
+        __m128i ni7 = lasx_extracti128( i3, 1);
+
+        // Compute the sum of the quants and set y[i].s
+        const __m128i s0 = __lsx_vadd_w(__lsx_vadd_w(ni0, ni1), __lsx_vadd_w(ni2, ni3));
+        const __m128i s1 = __lsx_vadd_w(__lsx_vadd_w(ni4, ni5), __lsx_vadd_w(ni6, ni7));
+        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1)));
+
+        // Convert int32 to int16
+        ni0 = lsx_packs_w( ni0, ni1 );
+        ni2 = lsx_packs_w( ni2, ni3 );
+        ni4 = lsx_packs_w( ni4, ni5 );
+        ni6 = lsx_packs_w( ni6, ni7 );
+        // Convert int16 to int8
+        ni0 = lsx_packs_h( ni0, ni2 );
+        ni4 = lsx_packs_h( ni4, ni6 );
+
+        __lsx_vst(ni0, (__m128i *)(y[i].qs +  0), 0);
+        __lsx_vst(ni4, (__m128i *)(y[i].qs + 16), 0);
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_1_ref(x, y, k);
+#endif
+}
+
+//
+// 2-6 bit quantization in super-blocks
+//
+
+//
+// ===================== Helper functions
+//
+static inline int nearest_int(float fval) {
+    assert(fabsf(fval) <= 4194303.f);
+    float val = fval + 12582912.f;
+    int i; memcpy(&i, &val, sizeof(int));
+    return (i & 0x007fffff) - 0x00400000;
+}
+
+static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, int rmse_type,
+        const float * restrict qw) {
+    float max = 0;
+    float amax = 0;
+    for (int i = 0; i < n; ++i) {
+        float ax = fabsf(x[i]);
+        if (ax > amax) { amax = ax; max = x[i]; }
+    }
+    if (amax < GROUP_MAX_EPS) { // all zero
+        for (int i = 0; i < n; ++i) {
+            L[i] = 0;
+        }
+        return 0.f;
+    }
+    float iscale = -nmax / max;
+    if (rmse_type == 0) {
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale * x[i]);
+            L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
+        }
+        return 1/iscale;
+    }
+    bool return_early = false;
+    if (rmse_type < 0) {
+        rmse_type = -rmse_type;
+        return_early = true;
+    }
+    float sumlx = 0;
+    float suml2 = 0;
+#ifdef HAVE_BUGGY_APPLE_LINKER
+    // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7
+    for (volatile int i = 0; i < n; ++i) {
+#else
+    for (int i = 0; i < n; ++i) {
+#endif
+        int l = nearest_int(iscale * x[i]);
+        l = MAX(-nmax, MIN(nmax-1, l));
+        L[i] = l + nmax;
+        float w = qw ? qw[i] : rmse_type == 1 ? x[i] * x[i] : rmse_type == 2 ? 1 : rmse_type == 3 ? fabsf(x[i]) : sqrtf(fabsf(x[i]));
+        sumlx += w*x[i]*l;
+        suml2 += w*l*l;
+    }
+    float scale = suml2 ? sumlx/suml2 : 0.0f;
+    if (return_early) return suml2 > 0 ? 0.5f*(scale + 1/iscale) : 1/iscale;
+    float best = scale * sumlx;
+    for (int is = -9; is <= 9; ++is) {
+        if (is == 0) {
+            continue;
+        }
+        iscale = -(nmax + 0.1f*is) / max;
+        sumlx = suml2 = 0;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale * x[i]);
+            l = MAX(-nmax, MIN(nmax-1, l));
+            float w = qw ? qw[i] : rmse_type == 1 ? x[i] * x[i] : rmse_type == 2 ? 1 : rmse_type == 3 ? fabsf(x[i]) : sqrtf(fabsf(x[i]));
+            sumlx += w*x[i]*l;
+            suml2 += w*l*l;
+        }
+        if (suml2 > 0 && sumlx*sumlx > best*suml2) {
+            for (int i = 0; i < n; ++i) {
+                int l = nearest_int(iscale * x[i]);
+                L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
+            }
+            scale = sumlx/suml2; best = scale*sumlx;
+        }
+    }
+    return scale;
+}
+
+static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, bool do_rmse) {
+    float max = 0;
+    float amax = 0;
+    for (int i = 0; i < n; ++i) {
+        float ax = fabsf(x[i]);
+        if (ax > amax) { amax = ax; max = x[i]; }
+    }
+    if (amax < GROUP_MAX_EPS) { // all zero
+        for (int i = 0; i < n; ++i) { L[i] = 0; }
+        return 0.f;
+    }
+    float iscale = -nmax / max;
+    if (do_rmse) {
+        float sumlx = 0;
+        float suml2 = 0;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale * x[i]);
+            l = MAX(-nmax, MIN(nmax-1, l));
+            L[i] = l;
+            float w = x[i]*x[i];
+            sumlx += w*x[i]*l;
+            suml2 += w*l*l;
+        }
+        for (int itry = 0; itry < 5; ++itry) {
+            int n_changed = 0;
+            for (int i = 0; i < n; ++i) {
+                float w = x[i]*x[i];
+                float slx = sumlx - w*x[i]*L[i];
+                if (slx > 0) {
+                    float sl2 = suml2 - w*L[i]*L[i];
+                    int new_l = nearest_int(x[i] * sl2 / slx);
+                    new_l = MAX(-nmax, MIN(nmax-1, new_l));
+                    if (new_l != L[i]) {
+                        slx += w*x[i]*new_l;
+                        sl2 += w*new_l*new_l;
+                        if (sl2 > 0 && slx*slx*suml2 > sumlx*sumlx*sl2) {
+                            L[i] = new_l; sumlx = slx; suml2 = sl2;
+                            ++n_changed;
+                        }
+                    }
+                }
+            }
+            if (!n_changed) {
+                break;
+            }
+        }
+        for (int i = 0; i < n; ++i) {
+            L[i] += nmax;
+        }
+        return sumlx / suml2;
+    }
+    for (int i = 0; i < n; ++i) {
+        int l = nearest_int(iscale * x[i]);
+        l = MAX(-nmax, MIN(nmax-1, l));
+        L[i] = l + nmax;
+    }
+    return 1/iscale;
+}
+
+static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min,
+        int ntry, float alpha) {
+    float min = x[0];
+    float max = x[0];
+    for (int i = 1; i < n; ++i) {
+        if (x[i] < min) min = x[i];
+        if (x[i] > max) max = x[i];
+    }
+    if (max == min) {
+        for (int i = 0; i < n; ++i) L[i] = 0;
+        *the_min = 0;
+        return 0.f;
+    }
+    if (min > 0) min = 0;
+    float iscale = nmax/(max - min);
+    float scale = 1/iscale;
+    for (int itry = 0; itry < ntry; ++itry) {
+        float sumlx = 0; int suml2 = 0;
+        bool did_change = false;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale*(x[i] - min));
+            l = MAX(0, MIN(nmax, l));
+            if (l != L[i]) {
+                L[i] = l;
+                did_change = true;
+            }
+            sumlx += (x[i] - min)*l;
+            suml2 += l*l;
+        }
+        scale = sumlx/suml2;
+        float sum = 0;
+        for (int i = 0; i < n; ++i) {
+            sum += x[i] - scale*L[i];
+        }
+        min = alpha*min + (1 - alpha)*sum/n;
+        if (min > 0) min = 0;
+        iscale = 1/scale;
+        if (!did_change) break;
+    }
+    *the_min = -min;
+    return scale;
+}
+
+static float make_qkx2_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
+        uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux,
+        float rmin, float rdelta, int nstep, bool use_mad) {
+    float min = x[0];
+    float max = x[0];
+    float sum_w = weights[0];
+    float sum_x = sum_w * x[0];
+#ifdef HAVE_BUGGY_APPLE_LINKER
+    // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7
+    for (volatile int i = 1; i < n; ++i) {
+#else
+    for (int i = 1; i < n; ++i) {
+#endif
+        if (x[i] < min) min = x[i];
+        if (x[i] > max) max = x[i];
+        float w = weights[i];
+        sum_w += w;
+        sum_x += w * x[i];
+    }
+    if (min > 0) min = 0;
+    if (max == min) {
+        for (int i = 0; i < n; ++i) L[i] = 0;
+        *the_min = -min;
+        return 0.f;
+    }
+    float iscale = nmax/(max - min);
+    float scale = 1/iscale;
+    float best_mad = 0;
+    for (int i = 0; i < n; ++i) {
+        int l = nearest_int(iscale*(x[i] - min));
+        L[i] = MAX(0, MIN(nmax, l));
+        float diff = scale * L[i] + min - x[i];
+        diff = use_mad ? fabsf(diff) : diff * diff;
+        float w = weights[i];
+        best_mad += w * diff;
+    }
+    if (nstep < 1) {
+        *the_min = -min;
+        return scale;
+    }
+    for (int is = 0; is <= nstep; ++is) {
+        iscale = (rmin + rdelta*is + nmax)/(max - min);
+        float sum_l = 0, sum_l2 = 0, sum_xl = 0;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale*(x[i] - min));
+            l = MAX(0, MIN(nmax, l));
+            Laux[i] = l;
+            float w = weights[i];
+            sum_l += w*l;
+            sum_l2 += w*l*l;
+            sum_xl += w*l*x[i];
+        }
+        float D = sum_w * sum_l2 - sum_l * sum_l;
+        if (D > 0) {
+            float this_scale = (sum_w * sum_xl - sum_x * sum_l)/D;
+            float this_min   = (sum_l2 * sum_x - sum_l * sum_xl)/D;
+            if (this_min > 0) {
+                this_min = 0;
+                this_scale = sum_xl / sum_l2;
+            }
+            float mad = 0;
+            for (int i = 0; i < n; ++i) {
+                float diff = this_scale * Laux[i] + this_min - x[i];
+                diff = use_mad ? fabsf(diff) : diff * diff;
+                float w = weights[i];
+                mad += w * diff;
+            }
+            if (mad < best_mad) {
+                for (int i = 0; i < n; ++i) {
+                    L[i] = Laux[i];
+                }
+                best_mad = mad;
+                scale = this_scale;
+                min = this_min;
+            }
+        }
+    }
+    *the_min = -min;
+    return scale;
+}
+
+static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * restrict d, uint8_t * restrict m) {
+    if (j < 4) {
+        *d = q[j] & 63; *m = q[j + 4] & 63;
+    } else {
+        *d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
+        *m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
+    }
+}
+
+//========================- 2-bit (de)-quantization
+
+void quantize_row_q2_K(const float * restrict x, void * restrict vy, int64_t k) {
+    quantize_row_q2_K_ref(x, vy, k);
+}
+
+//========================= 3-bit (de)-quantization
+
+void quantize_row_q3_K(const float * restrict x, void * restrict vy, int64_t k) {
+    quantize_row_q3_K_ref(x, vy, k);
+}
+
+// ====================== 4-bit (de)-quantization
+
+void quantize_row_q4_K(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_q4_K * restrict y = vy;
+    quantize_row_q4_K_ref(x, y, k);
+}
+
+// ====================== 5-bit (de)-quantization
+
+void quantize_row_q5_K(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_q5_K * restrict y = vy;
+    quantize_row_q5_K_ref(x, y, k);
+}
+
+// ====================== 6-bit (de)-quantization
+
+void quantize_row_q6_K(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_q6_K * restrict y = vy;
+    quantize_row_q6_K_ref(x, y, k);
+}
+
+// ====================== Ternary (de)-quantization (BitNet b1.58 and TriLMs)
+
+void quantize_row_tq1_0(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_tq1_0 * restrict y = vy;
+    quantize_row_tq1_0_ref(x, y, k);
+}
+
+void quantize_row_tq2_0(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_tq2_0 * restrict y = vy;
+    quantize_row_tq2_0_ref(x, y, k);
+}
+
+static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
+
+//===================================== Q8_K ==============================================
+
+void quantize_row_q8_K(const float * restrict x, void * restrict y, int64_t k) {
+    quantize_row_q8_K_ref(x, y, k);
+}
+
+//===================================== Dot products =================================
+
+//
+// Helper functions
+//
+#if __AVX__ || __AVX2__ || __AVX512F__
+
+// shuffles to pick the required scales in dot products
+static inline __m256i get_scale_shuffle_q3k(int i) {
+    static const uint8_t k_shuffle[128] = {
+         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,     2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,     6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,    10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
+        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,    14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,
+    };
+    return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
+}
+static inline __m256i get_scale_shuffle_k4(int i) {
+    static const uint8_t k_shuffle[256] = {
+         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
+         2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
+         6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,
+        10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
+        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,
+        14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15
+    };
+    return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
+}
+static inline __m128i get_scale_shuffle(int i) {
+    static const uint8_t k_shuffle[128] = {
+         0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
+         2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
+         4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5,
+         6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7,
+         8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
+        10,10,10,10,10,10,10,10, 11,11,11,11,11,11,11,11,
+        12,12,12,12,12,12,12,12, 13,13,13,13,13,13,13,13,
+        14,14,14,14,14,14,14,14, 15,15,15,15,15,15,15,15
+    };
+    return _mm_loadu_si128((const __m128i*)k_shuffle + i);
+}
+#elif defined(__loongarch_asx)
+// shuffles to pick the required scales in dot products
+static inline __m256i get_scale_shuffle_q3k(int i) {
+    static const uint8_t k_shuffle[128] = {
+         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,     2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,     6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,    10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
+        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,    14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,
+    };
+    return __lasx_xvld((const __m256i*)k_shuffle + i, 0);
+}
+static inline __m256i get_scale_shuffle_k4(int i) {
+    static const uint8_t k_shuffle[256] = {
+         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
+         2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
+         6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,
+        10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
+        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,
+        14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15
+    };
+    return __lasx_xvld((const __m256i*)k_shuffle + i, 0);
+}
+static inline __m128i get_scale_shuffle(int i) {
+    static const uint8_t k_shuffle[128] = {
+         0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
+         2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
+         4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5,
+         6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7,
+         8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
+        10,10,10,10,10,10,10,10, 11,11,11,11,11,11,11,11,
+        12,12,12,12,12,12,12,12, 13,13,13,13,13,13,13,13,
+        14,14,14,14,14,14,14,14, 15,15,15,15,15,15,15,15
+    };
+    return __lsx_vld((const __m128i*)k_shuffle + i, 0);
+}
+#endif
+
+void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_0 * restrict x = vx;
+    const block_q8_0 * restrict y = vy;
+
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q4_0 * restrict vx0 = vx;
+        const block_q4_0 * restrict vx1 = (const block_q4_0 *) ((const uint8_t*)vx + bx);
+        const block_q8_0 * restrict vy0 = vy;
+        const block_q8_0 * restrict vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by);
+
+        float32x4_t sumv0 = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; i++) {
+            const block_q4_0 * restrict b_x0 = &vx0[i];
+            const block_q4_0 * restrict b_x1 = &vx1[i];
+            const block_q8_0 * restrict b_y0 = &vy0[i];
+            const block_q8_0 * restrict b_y1 = &vy1[i];
+
+            const uint8x16_t m4b = vdupq_n_u8(0x0F);
+            const int8x16_t  s8b = vdupq_n_s8(0x8);
+
+            const uint8x16_t v0_0 = vld1q_u8(b_x0->qs);
+            const uint8x16_t v0_1 = vld1q_u8(b_x1->qs);
+
+            // 4-bit -> 8-bit
+            const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+            const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+            const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+            const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+            // sub 8
+            const int8x16_t x0_l = vsubq_s8(v0_0l, s8b);
+            const int8x16_t x0_h = vsubq_s8(v0_0h, s8b);
+            const int8x16_t x1_l = vsubq_s8(v0_1l, s8b);
+            const int8x16_t x1_h = vsubq_s8(v0_1h, s8b);
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            float32_t _scale[4] = {
+                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+            };
+            float32x4_t scale = vld1q_f32(_scale);
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+
+        float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2);
+        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
+
+        vst1_f32(s,      vget_low_f32 (sumv2));
+        vst1_f32(s + bs, vget_high_f32(sumv2));
+
+        return;
+    }
+#endif
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__ARM_FEATURE_SVE)
+    svfloat32_t sumv0 = svdup_n_f32(0.0f);
+    svfloat32_t sumv1 = svdup_n_f32(0.0f);
+
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
+
+    // VLA Implementation using switch case
+    switch (vector_length) {
+        case 128:
+            {
+                // predicate for activating higher lanes for 4 float32 elements
+                const svbool_t ph4 = svptrue_pat_b32(SV_VL4);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q4_0 * restrict x0 = &x[ib + 0];
+                    const block_q4_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    // load x
+                    const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
+                    const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
+
+                    // 4-bit -> 8-bit
+                    const svint8_t qx0l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx0r, 0x0F));
+                    const svint8_t qx0h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx0r, 0x04));
+                    const svint8_t qx1l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx1r, 0x0F));
+                    const svint8_t qx1h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx1r, 0x04));
+
+                    // sub 8
+                    const svint8_t qx0ls = svsub_n_s8_x(svptrue_b8(), qx0h, 8);
+                    const svint8_t qx0hs = svsub_n_s8_x(svptrue_b8(), qx0l, 8);
+                    const svint8_t qx1ls = svsub_n_s8_x(svptrue_b8(), qx1h, 8);
+                    const svint8_t qx1hs = svsub_n_s8_x(svptrue_b8(), qx1l, 8);
+
+                    // load y
+                    const svint8_t qy0h = svld1_s8(svptrue_b8(), y0->qs);
+                    const svint8_t qy0l = svld1_s8(svptrue_b8(), y0->qs + 16);
+                    const svint8_t qy1h = svld1_s8(svptrue_b8(), y1->qs);
+                    const svint8_t qy1l = svld1_s8(svptrue_b8(), y1->qs + 16);
+
+                    // dot product
+                    sumv0 = svmla_n_f32_x(ph4, sumv0, svcvt_f32_s32_x(ph4, svadd_x(ph4,
+                                    svdot_s32(svdup_n_s32(0), qx0ls, qy0l),
+                                    svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(ph4, sumv1, svcvt_f32_s32_x(ph4, svadd_x(ph4,
+                                    svdot_s32(svdup_n_s32(0), qx1ls, qy1l),
+                                    svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+            } break;
+        case 256:
+            {
+                // predicate for activating higher lanes for 16 int8 elements
+                const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
+                // predicate for activating lower lanes for  16 int8 elements
+                const svbool_t pl16 = svnot_b_z(svptrue_b8(), ph16);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q4_0 * restrict x0 = &x[ib + 0];
+                    const block_q4_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    // load x
+                    const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
+                    const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
+
+                    // 4-bit -> 8-bit
+                    const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04));
+                    const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04));
+
+                    // sub 8
+                    const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8);
+                    const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8);
+
+                    // load y
+                    const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
+                    const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
+
+                    // dot product
+                    sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+            } break;
+        case 512:
+            {
+                // predicate for activating higher lanes for 32 int8 elements
+                const svbool_t ph32 = svptrue_pat_b8(SV_VL32);
+
+                // predicate for activating higher lanes for 16 int8 elements
+                const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
+                // predicate for activating lower lanes for 16 int8 elements from first 32 int8 activated lanes
+                const svbool_t pl16 = svnot_b_z(ph32, ph16);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q4_0 * restrict x0 = &x[ib + 0];
+                    const block_q4_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    // load x
+                    const svuint8_t qx0r = svld1rq_u8(ph32, x0->qs);
+                    const svuint8_t qx1r = svld1rq_u8(ph32, x1->qs);
+
+                    // 4-bit -> 8-bit
+                    const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04));
+                    const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04));
+
+                    // sub 8
+                    const svint8_t qx0s = svsub_n_s8_x(ph32, qx0, 8);
+                    const svint8_t qx1s = svsub_n_s8_x(ph32, qx1, 8);
+
+                    // load y
+                    const svint8_t qy0 = svld1_s8(ph32, y0->qs);
+                    const svint8_t qy1 = svld1_s8(ph32, y1->qs);
+
+                    // dot product
+                    sumv0 = svmla_n_f32_x(ph32, sumv0, svcvt_f32_s32_x(ph32,
+                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(ph32, sumv1, svcvt_f32_s32_x(ph32,
+                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(ph32, svadd_f32_x(ph32, sumv0, sumv1));
+            } break;
+        default:
+            assert(false && "Unsupported vector length");
+            break;
+    }
+
+#elif defined(__ARM_NEON)
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q4_0 * restrict x0 = &x[ib + 0];
+        const block_q4_0 * restrict x1 = &x[ib + 1];
+        const block_q8_0 * restrict y0 = &y[ib + 0];
+        const block_q8_0 * restrict y1 = &y[ib + 1];
+
+        const uint8x16_t m4b = vdupq_n_u8(0x0F);
+        const int8x16_t  s8b = vdupq_n_s8(0x8);
+
+        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
+        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
+
+        // 4-bit -> 8-bit
+        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+        // sub 8
+        const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b);
+        const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b);
+        const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b);
+        const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b);
+
+        // load y
+        const int8x16_t v1_0l = vld1q_s8(y0->qs);
+        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
+        const int8x16_t v1_1l = vld1q_s8(y1->qs);
+        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
+
+        // dot product into int32x4_t
+        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
+        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+#elif defined(__AVX2__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+
+        // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
+        const __m256i off = _mm256_set1_epi8( 8 );
+        qx = _mm256_sub_epi8( qx, off );
+
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        /* Multiply q with scale and accumulate */
+        acc = _mm256_fmadd_ps( d, q, acc );
+    }
+
+    sumf = hsum_float_8(acc);
+#elif defined(__AVX__)
+    __m256 accum = _mm256_setzero_ps();
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
+        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
+        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
+
+        const __m128i q4b_1_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_1), _mm_set1_epi8(8));
+        const __m128i q4b_1_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_1, 4)), _mm_set1_epi8(8));
+        const __m128i q4b_2_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_2), _mm_set1_epi8(8));
+        const __m128i q4b_2_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_2, 4)), _mm_set1_epi8(8));
+
+        const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+        const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+        const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+        const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+        const __m128i p_1 = _mm_add_epi16(p16_1_0, p16_1_1);
+        const __m128i p_2 = _mm_add_epi16(p16_2_0, p16_2_1);
+        const __m256 p =  sum_i16_pairs_float(p_2, p_1);
+
+        const __m256 deltas = quad_fp16_delta_float(x[ib].d, y[ib].d, x[ib + 1].d, y[ib + 1].d);
+        accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
+    }
+
+    sumf = hsum_float_8(accum);
+#elif defined(__SSSE3__)
+    // set constants
+    const __m128i lowMask = _mm_set1_epi8(0xF);
+    const __m128i off = _mm_set1_epi8(8);
+
+    // Initialize accumulator with zeros
+    __m128 acc_0 = _mm_setzero_ps();
+    __m128 acc_1 = _mm_setzero_ps();
+    __m128 acc_2 = _mm_setzero_ps();
+    __m128 acc_3 = _mm_setzero_ps();
+
+    for (; ib + 1 < nb; ib += 2) {
+        _mm_prefetch(&x[ib] + sizeof(block_q4_0), _MM_HINT_T0);
+        _mm_prefetch(&y[ib] + sizeof(block_q8_0), _MM_HINT_T0);
+
+        // Compute combined scale for the block 0 and 1
+        const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+
+        const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[ib].qs);
+
+        __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1);
+        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
+        bx_0 = _mm_sub_epi8(bx_0, off);
+        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
+
+        __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4));
+        __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[ib].qs + 16));
+        bx_1 = _mm_sub_epi8(bx_1, off);
+        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
+
+        _mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
+        _mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
+
+        // Compute combined scale for the block 2 and 3
+        const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
+
+        const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+
+        __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3);
+        __m128i by_2 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        bx_2 = _mm_sub_epi8(bx_2, off);
+        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
+
+        __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4));
+        __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[ib + 1].qs + 16));
+        bx_3 = _mm_sub_epi8(bx_3, off);
+        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
+
+        // Convert int32_t to float
+        __m128 p0 = _mm_cvtepi32_ps(i32_0);
+        __m128 p1 = _mm_cvtepi32_ps(i32_1);
+        __m128 p2 = _mm_cvtepi32_ps(i32_2);
+        __m128 p3 = _mm_cvtepi32_ps(i32_3);
+
+        // Apply the scale
+        __m128 p0_d = _mm_mul_ps( d_0_1, p0 );
+        __m128 p1_d = _mm_mul_ps( d_0_1, p1 );
+        __m128 p2_d = _mm_mul_ps( d_2_3, p2 );
+        __m128 p3_d = _mm_mul_ps( d_2_3, p3 );
+
+        // Acummulate
+        acc_0 = _mm_add_ps(p0_d, acc_0);
+        acc_1 = _mm_add_ps(p1_d, acc_1);
+        acc_2 = _mm_add_ps(p2_d, acc_2);
+        acc_3 = _mm_add_ps(p3_d, acc_3);
+    }
+
+    sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+#elif defined(__riscv_v_intrinsic)
+    size_t vl = __riscv_vsetvl_e8m1(qk/2);
+
+    for (; ib < nb; ++ib) {
+        // load elements
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
+
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
+
+        // mask and store lower part of x, and then upper part
+        vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
+        vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
+
+        vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
+        vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
+
+        // subtract offset
+        vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 8, vl);
+        vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 8, vl);
+
+        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
+        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
+
+        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
+
+        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
+        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+    }
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+    const vector signed char v8 = vec_splats((signed char)0x8);
+
+    vector float vsumf0 = vec_splats(0.0f);
+
+#pragma GCC unroll 8
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
+
+        vector signed char q4x0 = vec_and(qxs, lowMask);
+        vector signed char q4x1 = vec_sr(qxs, v4);
+
+        q4x0 = vec_sub(q4x0, v8);
+        q4x1 = vec_sub(q4x1, v8);
+
+        vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
+        vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
+
+        vector signed int vsumi0 = v0;
+
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi0 = vec_sum4s(qv1, vsumi0);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = __lasx_xvreplfr2vr_s( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+
+        // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
+        const __m256i off = __lasx_xvreplgr2vr_b( 8 );
+        qx = __lasx_xvsub_b( qx, off );
+
+        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        /* Multiply q with scale and accumulate */
+        acc = __lasx_xvfmadd_s( d, q, acc );
+    }
+
+    sumf = hsum_float_8(acc);
+#elif defined(__loongarch_sx)
+    // set constants
+    const __m128i low_mask = __lsx_vreplgr2vr_b(0xF);
+    const __m128i off = __lsx_vreplgr2vr_b(8);
+
+    // Initialize accumulator with zeros
+    __m128 acc_0 = __lsx_vldi(0);
+    __m128 acc_1 = __lsx_vldi(0);
+    __m128 acc_2 = __lsx_vldi(0);
+    __m128 acc_3 = __lsx_vldi(0);
+
+    for (; ib + 1 < nb; ib += 2) {
+
+        // Compute combined scale for the block 0 and 1
+        const __m128 d_0_1 = __lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+
+        const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);
+
+        __m128i bx_0 = __lsx_vand_v(low_mask, tmp_0_1);
+        __m128i by_0 = __lsx_vld((const __m128i *)y[ib].qs, 0);
+        bx_0 = __lsx_vsub_b(bx_0, off);
+        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
+
+        __m128i bx_1 = __lsx_vand_v(low_mask, __lsx_vsrli_d(tmp_0_1, 4));
+        __m128i by_1 = __lsx_vld((const __m128i *)(y[ib].qs + 16), 0);
+        bx_1 = __lsx_vsub_b(bx_1, off);
+        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
+
+        //_mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
+        //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
+
+        // Compute combined scale for the block 2 and 3
+        const __m128 d_2_3 = __lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
+
+        const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);
+
+        __m128i bx_2 = __lsx_vand_v(low_mask, tmp_2_3);
+        __m128i by_2 = __lsx_vld((const __m128i *)y[ib + 1].qs, 0);
+        bx_2 = __lsx_vsub_b(bx_2, off);
+        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
+
+        __m128i bx_3 = __lsx_vand_v(low_mask, __lsx_vsrli_d(tmp_2_3, 4));
+        __m128i by_3 = __lsx_vld((const __m128i *)(y[ib + 1].qs + 16), 0);
+        bx_3 = __lsx_vsub_b(bx_3, off);
+        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
+
+        // Convert int32_t to float
+        __m128 p0 = __lsx_vffint_s_w(i32_0);
+        __m128 p1 = __lsx_vffint_s_w(i32_1);
+        __m128 p2 = __lsx_vffint_s_w(i32_2);
+        __m128 p3 = __lsx_vffint_s_w(i32_3);
+
+        // Apply the scale
+        __m128 p0_d = __lsx_vfmul_s( d_0_1, p0 );
+        __m128 p1_d = __lsx_vfmul_s( d_0_1, p1 );
+        __m128 p2_d = __lsx_vfmul_s( d_2_3, p2 );
+        __m128 p3_d = __lsx_vfmul_s( d_2_3, p3 );
+
+        // Acummulate
+        acc_0 = __lsx_vfadd_s(p0_d, acc_0);
+        acc_1 = __lsx_vfadd_s(p1_d, acc_1);
+        acc_2 = __lsx_vfadd_s(p2_d, acc_2);
+        acc_3 = __lsx_vfadd_s(p3_d, acc_3);
+    }
+
+    sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    const int qk = QK8_1;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_1 * restrict x = vx;
+    const block_q8_1 * restrict y = vy;
+
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q4_1 * restrict vx0 = vx;
+        const block_q4_1 * restrict vx1 = (const block_q4_1 *) ((const uint8_t*)vx + bx);
+        const block_q8_1 * restrict vy0 = vy;
+        const block_q8_1 * restrict vy1 = (const block_q8_1 *) ((const uint8_t*)vy + by);
+
+        float32x4_t sumv0 = vdupq_n_f32(0.0f);
+        float32x4_t summs0 = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; i++) {
+            const block_q4_1 * restrict b_x0 = &vx0[i];
+            const block_q4_1 * restrict b_x1 = &vx1[i];
+            const block_q8_1 * restrict b_y0 = &vy0[i];
+            const block_q8_1 * restrict b_y1 = &vy1[i];
+
+            float32_t summs_t[4] = {
+                GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
+                GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
+                GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
+                GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)
+            };
+            summs0 = vaddq_f32(summs0, vld1q_f32(summs_t));
+
+            const uint8x16_t m4b = vdupq_n_u8(0x0F);
+
+            const uint8x16_t v0_0 = vld1q_u8(b_x0->qs);
+            const uint8x16_t v0_1 = vld1q_u8(b_x1->qs);
+
+            // 4-bit -> 8-bit
+            const int8x16_t x0_l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+            const int8x16_t x0_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+            const int8x16_t x1_l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+            const int8x16_t x1_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            // mmla into int32x4_t
+            float32_t _scale[4] = {
+                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+            };
+            float32x4_t scale = vld1q_f32(_scale);
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+
+        float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2);
+        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
+
+        sumv2 = vaddq_f32(sumv2, summs0);
+
+        vst1_f32(s,      vget_low_f32 (sumv2));
+        vst1_f32(s + bs, vget_high_f32(sumv2));
+
+        return;
+    }
+#endif
+
+    int ib = 0;
+    float sumf = 0;
+
+    // TODO: add WASM SIMD
+#if defined(__ARM_NEON)
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    float summs = 0;
+
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q4_1 * restrict x0 = &x[ib + 0];
+        const block_q4_1 * restrict x1 = &x[ib + 1];
+        const block_q8_1 * restrict y0 = &y[ib + 0];
+        const block_q8_1 * restrict y1 = &y[ib + 1];
+
+        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
+
+        const uint8x16_t m4b = vdupq_n_u8(0x0F);
+
+        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
+        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
+
+        // 4-bit -> 8-bit
+        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+        // load y
+        const int8x16_t v1_0l = vld1q_s8(y0->qs);
+        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
+        const int8x16_t v1_1l = vld1q_s8(y1->qs);
+        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
+
+        // dot product into int32x4_t
+        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
+        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
+#elif defined(__AVX2__) || defined(__AVX__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
+        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
+
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+
+        const __m256 d0v = _mm256_set1_ps( d0 );
+        const __m256 d1v = _mm256_set1_ps( d1 );
+
+        // Compute combined scales
+        const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
+
+        // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
+        const __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i qy = _mm256_loadu_si256( (const __m256i *)y[ib].qs );
+
+        const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
+
+        // Accumulate d0*d1*x*y
+#if defined(__AVX2__)
+        acc = _mm256_fmadd_ps( d0d1, xy, acc );
+#else
+        acc = _mm256_add_ps( _mm256_mul_ps( d0d1, xy ), acc );
+#endif
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+#elif defined(__riscv_v_intrinsic)
+    size_t vl = __riscv_vsetvl_e8m1(qk/2);
+
+    for (; ib < nb; ++ib) {
+        // load elements
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
+
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
+
+        // mask and store lower part of x, and then upper part
+        vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
+        vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
+
+        vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
+        vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
+
+        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
+        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
+
+        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
+
+        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
+        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+
+#pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
+        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f};
+        vsumf0 = vec_madd(vxmin, vys, vsumf0);
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
+
+        vector unsigned char q4x0 = (vector unsigned char)vec_and(qxs, lowMask);
+        vector unsigned char q4x1 = (vector unsigned char)vec_sr(qxs, v4);
+
+        vector signed int vsumi0 = v0;
+
+        vsumi0 = vec_msum(q8y0, q4x0, vsumi0);
+        vsumi0 = vec_msum(q8y1, q4x1, vsumi0);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    float summs = 0;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
+        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
+
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+
+        const __m256 d0v = __lasx_xvreplfr2vr_s( d0 );
+        const __m256 d1v = __lasx_xvreplfr2vr_s( d1 );
+
+        // Compute combined scales
+        const __m256 d0d1 = __lasx_xvfmul_s( d0v, d1v );
+
+        // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
+        const __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i qy = __lasx_xvld( (const __m256i *)y[ib].qs, 0);
+
+        const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
+
+        // Accumulate d0*d1*x*y
+        acc = __lasx_xvfmadd_s( d0d1, xy, acc );
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F);
+            const int v1 = (x[ib].qs[j] >>   4);
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * restrict x = vx;
+    const block_q8_0 * restrict y = vy;
+
+#if defined(__ARM_NEON)
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    uint32_t qh0;
+    uint32_t qh1;
+
+    uint64_t tmp0[4];
+    uint64_t tmp1[4];
+
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q5_0 * restrict x0 = &x[ib];
+        const block_q5_0 * restrict x1 = &x[ib + 1];
+        const block_q8_0 * restrict y0 = &y[ib];
+        const block_q8_0 * restrict y1 = &y[ib + 1];
+
+        const uint8x16_t m4b = vdupq_n_u8(0x0F);
+
+        // extract the 5th bit via lookup table ((!b) << 4)
+        memcpy(&qh0, x0->qh, sizeof(qh0));
+        memcpy(&qh1, x1->qh, sizeof(qh1));
+
+        tmp0[0] = table_b2b_1[(qh0 >>  0) & 0xFF];
+        tmp0[1] = table_b2b_1[(qh0 >>  8) & 0xFF];
+        tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF];
+        tmp0[3] = table_b2b_1[(qh0 >> 24)       ];
+
+        tmp1[0] = table_b2b_1[(qh1 >>  0) & 0xFF];
+        tmp1[1] = table_b2b_1[(qh1 >>  8) & 0xFF];
+        tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF];
+        tmp1[3] = table_b2b_1[(qh1 >> 24)       ];
+
+        const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
+        const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
+        const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
+        const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));
+
+        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
+        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
+
+        // 4-bit -> 8-bit
+        int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+        int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+        int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+        int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
+        const int8x16_t v0_0lf = vsubq_s8(v0_0l, qhl0);
+        const int8x16_t v0_0hf = vsubq_s8(v0_0h, qhh0);
+        const int8x16_t v0_1lf = vsubq_s8(v0_1l, qhl1);
+        const int8x16_t v0_1hf = vsubq_s8(v0_1h, qhh1);
+
+        // load y
+        const int8x16_t v1_0l = vld1q_s8(y0->qs);
+        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
+        const int8x16_t v1_1l = vld1q_s8(y1->qs);
+        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+#elif defined(__wasm_simd128__)
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    uint32_t qh;
+    uint64_t tmp[4];
+
+    // TODO: check if unrolling this is better
+    for (; ib < nb; ++ib) {
+        const block_q5_0 * restrict x0 = &x[ib];
+        const block_q8_0 * restrict y0 = &y[ib];
+
+        const v128_t m4b  = wasm_i8x16_splat(0x0F);
+
+        // extract the 5th bit
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        tmp[0] = table_b2b_1[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_1[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_1[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_1[(qh >> 24)       ];
+
+        const v128_t qhl = wasm_v128_load(tmp + 0);
+        const v128_t qhh = wasm_v128_load(tmp + 2);
+
+        const v128_t v0 = wasm_v128_load(x0->qs);
+
+        // 4-bit -> 8-bit
+        const v128_t v0l = wasm_v128_and (v0, m4b);
+        const v128_t v0h = wasm_u8x16_shr(v0, 4);
+
+        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
+        const v128_t v0lf = wasm_i8x16_sub(v0l, qhl);
+        const v128_t v0hf = wasm_i8x16_sub(v0h, qhh);
+
+        // load y
+        const v128_t v1l = wasm_v128_load(y0->qs);
+        const v128_t v1h = wasm_v128_load(y0->qs + 16);
+
+        // int8x16 -> int16x8
+        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
+        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
+        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
+        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
+
+        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
+        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
+        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
+        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
+
+        // dot product
+        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
+                        wasm_i32x4_add(
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
+                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
+                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
+                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
+    }
+
+    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
+#elif defined(__AVX2__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0));
+        qx = _mm256_or_si256(qx, bxhi);
+
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        /* Multiply q with scale and accumulate */
+        acc = _mm256_fmadd_ps(d, q, acc);
+    }
+
+    sumf = hsum_float_8(acc);
+#elif defined(__AVX__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+    __m128i mask = _mm_set1_epi8((char)0xF0);
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+
+        __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        __m128i bxhil = _mm256_castsi256_si128(bxhi);
+        __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
+        bxhil = _mm_andnot_si128(bxhil, mask);
+        bxhih = _mm_andnot_si128(bxhih, mask);
+        __m128i bxl = _mm256_castsi256_si128(bx_0);
+        __m128i bxh = _mm256_extractf128_si256(bx_0, 1);
+        bxl = _mm_or_si128(bxl, bxhil);
+        bxh = _mm_or_si128(bxh, bxhih);
+        bx_0 = MM256_SET_M128I(bxh, bxl);
+
+        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_i8_pairs_float(bx_0, by_0);
+
+        /* Multiply q with scale and accumulate */
+        acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc);
+    }
+
+    sumf = hsum_float_8(acc);
+#elif defined(__riscv_v_intrinsic)
+    uint32_t qh;
+
+    size_t vl = __riscv_vsetvl_e8m1(qk/2);
+
+    // These temporary registers are for masking and shift operations
+    vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
+    vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl);
+
+    vuint32m2_t vt_3 = __riscv_vsll_vx_u32m2(vt_2, 16, vl);
+    vuint32m2_t vt_4 = __riscv_vadd_vx_u32m2(vt_1, 12, vl);
+
+    for (; ib < nb; ++ib) {
+        memcpy(&qh, x[ib].qh, sizeof(uint32_t));
+
+        // ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+        vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(vt_2, qh, vl);
+        vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(xha_0, vt_1, vl);
+        vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl);
+
+        // ((qh & (1u << (j + 16))) >> (j + 12));
+        vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(vt_3, qh, vl);
+        vuint32m2_t xhl_1 = __riscv_vsrl_vv_u32m2(xha_1, vt_4, vl);
+
+        // narrowing
+        vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xhl_0, vl);
+        vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl);
+
+        vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xhl_1, vl);
+        vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl);
+
+        // load
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
+
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
+
+        vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
+        vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
+
+        vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl);
+        vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl);
+
+        vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
+        vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
+
+        vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 16, vl);
+        vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 16, vl);
+
+        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
+        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
+
+        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
+
+        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
+        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector unsigned char v4 = vec_splats((unsigned char)4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+
+#pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed long long aux64x2_0 = {(uint64_t)(table_b2b_1[x[ib].qh[0]]), (uint64_t)(table_b2b_1[x[ib].qh[1]])};
+        vector signed long long aux64x2_1 = {(uint64_t)(table_b2b_1[x[ib].qh[2]]), (uint64_t)(table_b2b_1[x[ib].qh[3]])};
+
+        vector signed char qh0 = (vector signed char)aux64x2_0;
+        vector signed char qh1 = (vector signed char)aux64x2_1;
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+
+        vector signed char q5x0 = vec_sub(vec_and (qxs, lowMask), qh0);
+        vector signed char q5x1 = vec_sub(vec_sr(qxs, v4), qh1);
+
+        vector signed char q8y0 = vec_xl(  0, y[ib].qs);
+        vector signed char q8y1 = vec_xl( 16, y[ib].qs);
+
+        vector signed short qv0 = vec_add(vec_mule(q5x0, q8y0), vec_mulo(q5x0, q8y0));
+        vector signed short qv1 = vec_add(vec_mule(q5x1, q8y1), vec_mulo(q5x1, q8y1));
+
+        qv0 = vec_add(qv0, qv1);
+
+        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0));
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); //FIXME
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        bxhi = __lasx_xvandn_v(bxhi, __lasx_xvreplgr2vr_b((char)0xF0));
+        qx = __lasx_xvor_v(qx, bxhi);
+
+        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        /* Multiply q with scale and accumulate */
+        acc = __lasx_xvfmadd_s(d, q, acc);
+    }
+
+    sumf = hsum_float_8(acc);
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
+
+            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    const int qk = QK8_1;
+    const int nb = n / qk;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * restrict x = vx;
+    const block_q8_1 * restrict y = vy;
+
+#if defined(__ARM_NEON)
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    float summs0 = 0.0f;
+    float summs1 = 0.0f;
+
+    uint32_t qh0;
+    uint32_t qh1;
+
+    uint64_t tmp0[4];
+    uint64_t tmp1[4];
+
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q5_1 * restrict x0 = &x[ib];
+        const block_q5_1 * restrict x1 = &x[ib + 1];
+        const block_q8_1 * restrict y0 = &y[ib];
+        const block_q8_1 * restrict y1 = &y[ib + 1];
+
+        const uint8x16_t m4b = vdupq_n_u8(0x0F);
+
+        summs0 += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
+        summs1 += GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
+
+        // extract the 5th bit via lookup table ((b) << 4)
+        memcpy(&qh0, x0->qh, sizeof(qh0));
+        memcpy(&qh1, x1->qh, sizeof(qh1));
+
+        tmp0[0] = table_b2b_0[(qh0 >>  0) & 0xFF];
+        tmp0[1] = table_b2b_0[(qh0 >>  8) & 0xFF];
+        tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF];
+        tmp0[3] = table_b2b_0[(qh0 >> 24)       ];
+
+        tmp1[0] = table_b2b_0[(qh1 >>  0) & 0xFF];
+        tmp1[1] = table_b2b_0[(qh1 >>  8) & 0xFF];
+        tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF];
+        tmp1[3] = table_b2b_0[(qh1 >> 24)       ];
+
+        const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
+        const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
+        const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
+        const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));
+
+        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
+        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
+
+        // 4-bit -> 8-bit
+        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+        // add high bit
+        const int8x16_t v0_0lf = vorrq_s8(v0_0l, qhl0);
+        const int8x16_t v0_0hf = vorrq_s8(v0_0h, qhh0);
+        const int8x16_t v0_1lf = vorrq_s8(v0_1l, qhl1);
+        const int8x16_t v0_1hf = vorrq_s8(v0_1h, qhh1);
+
+        // load y
+        const int8x16_t v1_0l = vld1q_s8(y0->qs);
+        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
+        const int8x16_t v1_1l = vld1q_s8(y1->qs);
+        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
+#elif defined(__wasm_simd128__)
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    float summs = 0.0f;
+
+    uint32_t qh;
+    uint64_t tmp[4];
+
+    // TODO: check if unrolling this is better
+    for (; ib < nb; ++ib) {
+        const block_q5_1 * restrict x0 = &x[ib];
+        const block_q8_1 * restrict y0 = &y[ib];
+
+        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
+
+        const v128_t m4b = wasm_i8x16_splat(0x0F);
+
+        // extract the 5th bit
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        tmp[0] = table_b2b_0[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_0[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_0[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_0[(qh >> 24)       ];
+
+        const v128_t qhl = wasm_v128_load(tmp + 0);
+        const v128_t qhh = wasm_v128_load(tmp + 2);
+
+        const v128_t v0 = wasm_v128_load(x0->qs);
+
+        // 4-bit -> 8-bit
+        const v128_t v0l = wasm_v128_and (v0, m4b);
+        const v128_t v0h = wasm_u8x16_shr(v0, 4);
+
+        // add high bit
+        const v128_t v0lf = wasm_v128_or(v0l, qhl);
+        const v128_t v0hf = wasm_v128_or(v0h, qhh);
+
+        // load y
+        const v128_t v1l = wasm_v128_load(y0->qs);
+        const v128_t v1h = wasm_v128_load(y0->qs + 16);
+
+        // int8x16 -> int16x8
+        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
+        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
+        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
+        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
+
+        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
+        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
+        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
+        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
+
+        // dot product
+        sumv = wasm_f32x4_add(sumv,
+                wasm_f32x4_mul(wasm_f32x4_convert_i32x4(wasm_i32x4_add(
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
+                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
+                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
+                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
+    }
+
+    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
+#elif defined(__AVX2__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0.0f;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
+
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
+        qx = _mm256_or_si256(qx, bxhi);
+
+        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
+
+        acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+#elif defined(__AVX__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+    __m128i mask = _mm_set1_epi8(0x10);
+
+    float summs = 0.0f;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
+
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+
+        __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        __m128i bxhil = _mm256_castsi256_si128(bxhi);
+        __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
+        bxhil = _mm_and_si128(bxhil, mask);
+        bxhih = _mm_and_si128(bxhih, mask);
+        __m128i bxl = _mm256_castsi256_si128(bx_0);
+        __m128i bxh = _mm256_extractf128_si256(bx_0, 1);
+        bxl = _mm_or_si128(bxl, bxhil);
+        bxh = _mm_or_si128(bxh, bxhih);
+        bx_0 = MM256_SET_M128I(bxh, bxl);
+
+        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
+
+        acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc);
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+#elif defined(__riscv_v_intrinsic)
+    uint32_t qh;
+
+    size_t vl = __riscv_vsetvl_e8m1(qk/2);
+
+    // temporary registers for shift operations
+    vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
+    vuint32m2_t vt_2 = __riscv_vadd_vx_u32m2(vt_1, 12, vl);
+
+    for (; ib < nb; ++ib) {
+        memcpy(&qh, x[ib].qh, sizeof(uint32_t));
+
+        // load qh
+        vuint32m2_t vqh = __riscv_vmv_v_x_u32m2(qh, vl);
+
+        // ((qh >> (j +  0)) << 4) & 0x10;
+        vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(vqh, vt_1, vl);
+        vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl);
+        vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(xhl_0, 0x10, vl);
+
+        // ((qh >> (j + 12))     ) & 0x10;
+        vuint32m2_t xhr_1 = __riscv_vsrl_vv_u32m2(vqh, vt_2, vl);
+        vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(xhr_1, 0x10, vl);
+
+        // narrowing
+        vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xha_0, vl);
+        vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl);
+
+        vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xha_1, vl);
+        vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl);
+
+        // load
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
+
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
+
+        vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
+        vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
+
+        vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl);
+        vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl);
+
+        vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
+        vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
+
+        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
+        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
+
+        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
+
+        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
+        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+
+#pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
+        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f};
+        vsumf0 = vec_madd(vxmin, vys, vsumf0);
+
+        vector unsigned long long aux64x2_0 = {(uint64_t)(table_b2b_0[x[ib].qh[0]]), (uint64_t)(table_b2b_0[x[ib].qh[1]])};
+        vector unsigned long long aux64x2_1 = {(uint64_t)(table_b2b_0[x[ib].qh[2]]), (uint64_t)(table_b2b_0[x[ib].qh[3]])};
+
+        vector signed char qh0 = (vector signed char)aux64x2_0;
+        vector signed char qh1 = (vector signed char)aux64x2_1;
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+
+        vector unsigned char q5x0 = (vector unsigned char)vec_or(vec_and(qxs, lowMask), qh0);
+        vector unsigned char q5x1 = (vector unsigned char)vec_or(vec_sr(qxs, v4), qh1);
+
+        vector signed char q8y0 = vec_xl(  0, y[ib].qs);
+        vector signed char q8y1 = vec_xl( 16, y[ib].qs);
+
+        vector signed int vsumi0 = v0;
+
+        vsumi0 = vec_msum(q8y0, q5x0, vsumi0);
+        vsumi0 = vec_msum(q8y1, q5x1, vsumi0);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    float summs = 0.0f;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const __m256 dx = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d));
+
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        bxhi = __lasx_xvand_v(bxhi, __lasx_xvreplgr2vr_b(0x10));
+        qx = __lasx_xvor_v(qx, bxhi);
+
+        const __m256 dy = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
+
+        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
+
+        acc = __lasx_xvfmadd_s(q, __lasx_xvfmul_s(dx, dy), acc);
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
+            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q8_0 * restrict x = vx;
+    const block_q8_0 * restrict y = vy;
+
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q8_0 * restrict vx0 = vx;
+        const block_q8_0 * restrict vx1 = (const block_q8_0 *) ((const uint8_t*)vx + bx);
+        const block_q8_0 * restrict vy0 = vy;
+        const block_q8_0 * restrict vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by);
+
+        float32x4_t sumv0 = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; i++) {
+            const block_q8_0 * restrict b_x0 = &vx0[i];
+            const block_q8_0 * restrict b_y0 = &vy0[i];
+
+            const block_q8_0 * restrict b_x1 = &vx1[i];
+            const block_q8_0 * restrict b_y1 = &vy1[i];
+
+            const int8x16_t x0_l = vld1q_s8(b_x0->qs);
+            const int8x16_t x0_h = vld1q_s8(b_x0->qs + 16);
+            const int8x16_t x1_l = vld1q_s8(b_x1->qs);
+            const int8x16_t x1_h = vld1q_s8(b_x1->qs + 16);
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            float32_t _scale[4] = {
+                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+            };
+            float32x4_t scale = vld1q_f32(_scale);
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+
+        float32x4_t sumv1 = vextq_f32 (sumv0, sumv0, 2);
+        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
+
+        vst1_f32(s,      vget_low_f32 (sumv2));
+        vst1_f32(s + bs, vget_high_f32(sumv2));
+
+        return;
+    }
+#endif
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__ARM_FEATURE_SVE)
+    svfloat32_t sumv0 = svdup_n_f32(0.0f);
+    svfloat32_t sumv1 = svdup_n_f32(0.0f);
+
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
+
+    //VLA Implemenation for SVE
+    switch (vector_length) {
+        case 128:
+            {
+                // predicate for activating lanes for 16 Int8 elements
+                const svbool_t ph16 = svptrue_pat_b8 (SV_VL16);
+                const svbool_t pl16 = svptrue_pat_b32(SV_VL4);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q8_0 * restrict x0 = &x[ib + 0];
+                    const block_q8_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    // load x
+                    const svint8_t qx0_0 = svld1_s8(ph16, x0->qs);
+                    const svint8_t qx0_1 = svld1_s8(ph16, x0->qs+16);
+                    const svint8_t qx1_0 = svld1_s8(ph16, x1->qs);
+                    const svint8_t qx1_1 = svld1_s8(ph16, x1->qs+16);
+
+                    // load y
+                    const svint8_t qy0_0 = svld1_s8(ph16, y0->qs);
+                    const svint8_t qy0_1 = svld1_s8(ph16, y0->qs+16);
+                    const svint8_t qy1_0 = svld1_s8(ph16, y1->qs);
+                    const svint8_t qy1_1 = svld1_s8(ph16, y1->qs+16);
+
+                    sumv0 = svmla_n_f32_x(pl16, sumv0, svcvt_f32_s32_x(pl16, svadd_x(pl16,
+                                    svdot_s32(svdup_n_s32(0), qx0_0, qy0_0),
+                                    svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(pl16, sumv1, svcvt_f32_s32_x(pl16, svadd_x(pl16,
+                                    svdot_s32(svdup_n_s32(0), qx1_0, qy1_0),
+                                    svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(pl16, svadd_f32_x(pl16, sumv0, sumv1));
+            } break;
+        case 256:
+            {
+                //printf("sve256");
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q8_0 * restrict x0 = &x[ib + 0];
+                    const block_q8_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    // load x
+                    const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs);
+                    const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs);
+
+                    // load y
+                    const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
+                    const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
+
+                    sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+            } break;
+        case 512:
+            {
+                // predicate for activating high 256 bit
+                const svbool_t ph32 = svptrue_pat_b8(SV_VL32);
+                // predicate for activating low 256 bit
+                const svbool_t pl32 = svnot_b_z(svptrue_b8(), ph32);
+
+                // predicate for activating high lanes for 8 float32 elements
+                const svbool_t ph8 = svptrue_pat_b32(SV_VL8);
+                // predicate for activating low lanes for 8 float32 elements
+                const svbool_t pl8 = svnot_b_z(svptrue_b32(), ph8);
+
+                svfloat32_t sumv00 = svdup_n_f32(0.0f);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q8_0 * restrict x0 = &x[ib + 0];
+                    const block_q8_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    //load 32 int8_t in first half of vector and put another 32 int8_t in second vector lower bits
+                    // and add them to make one 64 element vector
+                    // load x
+                    const svint8_t qx_32 = svld1_s8(ph32, x0->qs);
+                          svint8_t qx_64 = svld1_s8(pl32, x0->qs + 2);
+
+                    qx_64 = svadd_s8_x(svptrue_b8(), qx_32, qx_64);
+
+                    // load y
+                    const svint8_t qy_32 = svld1_s8(ph32, y0->qs);
+                          svint8_t qy_64 = svld1_s8(pl32, y0->qs + 2);
+
+                    qy_64 = svadd_s8_x(svptrue_b8(), qy_32, qy_64);
+
+                    // scale creation
+                    const float32_t deq1 = GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d);
+                    const float32_t deq2 = GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d);
+
+                    // duplicate deq1 in first half of vector and deq2 in second half of vector
+                    const svfloat32_t temp = svdup_f32_m(svdup_f32_z(ph8, deq1), pl8, deq2);
+
+                    const svfloat32_t sumvt = svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx_64, qy_64));
+
+                    sumv00 = svmla_f32_m(svptrue_b32(), sumv00, sumvt, temp);
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), sumv00);
+                break;
+            }
+        default:
+            assert(false && "Unsupported vector length");
+            break;
+    }
+#elif defined(__ARM_NEON)
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q8_0 * restrict x0 = &x[ib + 0];
+        const block_q8_0 * restrict x1 = &x[ib + 1];
+        const block_q8_0 * restrict y0 = &y[ib + 0];
+        const block_q8_0 * restrict y1 = &y[ib + 1];
+
+        const int8x16_t x0_0 = vld1q_s8(x0->qs);
+        const int8x16_t x0_1 = vld1q_s8(x0->qs + 16);
+        const int8x16_t x1_0 = vld1q_s8(x1->qs);
+        const int8x16_t x1_1 = vld1q_s8(x1->qs + 16);
+
+        // load y
+        const int8x16_t y0_0 = vld1q_s8(y0->qs);
+        const int8x16_t y0_1 = vld1q_s8(y0->qs + 16);
+        const int8x16_t y1_0 = vld1q_s8(y1->qs);
+        const int8x16_t y1_1 = vld1q_s8(y1->qs + 16);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
+                        ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
+                        ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+#elif defined(__AVX2__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        // Compute combined scale for the block
+        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        __m256i qx = _mm256_loadu_si256((const __m256i *)x[ib].qs);
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        // Multiply q with scale and accumulate
+        acc = _mm256_fmadd_ps( d, q, acc );
+    }
+
+    sumf = hsum_float_8(acc);
+#elif defined(__AVX__)
+    __m256 accum = _mm256_setzero_ps();
+
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i qx_1_0 = _mm_loadu_si128((const __m128i *)x[ib].qs);
+        const __m128i qx_1_1 = _mm_loadu_si128((const __m128i *)x[ib].qs + 1);
+        const __m128i qx_2_0 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+        const __m128i qx_2_1 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs + 1);
+        const __m128i qy_1_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
+        const __m128i qy_1_1 = _mm_loadu_si128((const __m128i *)y[ib].qs + 1);
+        const __m128i qy_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        const __m128i qy_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
+
+        const __m256 p = mul_sum_i8_quad_float(qx_1_0, qx_1_1, qx_2_0, qx_2_1, qy_1_0, qy_1_1, qy_2_0, qy_2_1);
+        const __m256 deltas = quad_fp16_delta_float(x[ib].d, y[ib].d, x[ib + 1].d, y[ib + 1].d);
+        accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
+    }
+
+    sumf = hsum_float_8(accum);
+#elif defined(__riscv_v_intrinsic)
+    size_t vl = __riscv_vsetvl_e8m1(qk);
+
+    for (; ib < nb; ++ib) {
+        // load elements
+        vint8m1_t bx_0 = __riscv_vle8_v_i8m1(x[ib].qs, vl);
+        vint8m1_t by_0 = __riscv_vle8_v_i8m1(y[ib].qs, vl);
+
+        vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx_0, by_0, vl);
+
+        vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, vl);
+        vint32m1_t v_sum = __riscv_vwredsum_vs_i16m2_i32m1(vw_mul, v_zero, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum);
+
+        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+    }
+#elif defined(__POWER9_VECTOR__)
+    const vector signed int v0 = vec_splats((int32_t)0);
+    vector float vsumf0 = vec_splats(0.0f);
+
+#pragma GCC unroll 8
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed char q8x0 = vec_xl( 0, x[ib].qs);
+        vector signed char q8x1 = vec_xl(16, x[ib].qs);
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
+
+        vector signed short qv0 = vec_mule(q8x0, q8y0);
+        vector signed short qv1 = vec_mulo(q8x0, q8y0);
+        vector signed short qv2 = vec_mule(q8x1, q8y1);
+        vector signed short qv3 = vec_mulo(q8x1, q8y1);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi1 = vec_sum4s(qv1, vsumi1);
+        vsumi0 = vec_sum4s(qv2, vsumi0);
+        vsumi1 = vec_sum4s(qv3, vsumi1);
+
+        vsumi0 = vec_add(vsumi0, vsumi1);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        // Compute combined scale for the block
+        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        __m256i qx = __lasx_xvld((const __m256i *)x[ib].qs, 0);
+        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        // Multiply q with scale and accumulate
+        acc = __lasx_xvfmadd_s( d, q, acc );
+    }
+
+    sumf = hsum_float_8(acc);
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi = 0;
+
+        for (int j = 0; j < qk; j++) {
+            sumi += x[ib].qs[j]*y[ib].qs[j];
+        }
+
+        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_tq1_0_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_tq1_0 * restrict x = vx;
+    const block_q8_K  * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+    float sumf = 0.0f;
+
+    uint8_t k_shift[16] = {1, 1, 1, 1, 3, 3, 3, 3, 9, 9, 9, 9, 27, 27, 27, 27};
+
+    const uint8x16_t shift = vld1q_u8(k_shift);
+
+    for (int i = 0; i < nb; ++i) {
+#if defined(__ARM_FEATURE_DOTPROD)
+        int32x4_t sumi0 = vdupq_n_s32(0);
+        int32x4_t sumi1 = vdupq_n_s32(0);
+#else
+        int16x8_t sumi0 = vdupq_n_s16(0);
+        int16x8_t sumi1 = vdupq_n_s16(0);
+#endif
+
+        // first 32 bytes of 5 elements
+        {
+            uint8x16_t qx0 = vld1q_u8(x[i].qs + 0);
+            uint8x16_t qx1 = vld1q_u8(x[i].qs + 16);
+            uint8x16_t qx2 = vmulq_u8(qx0, vdupq_n_u8(3));
+            uint8x16_t qx3 = vmulq_u8(qx1, vdupq_n_u8(3));
+            uint8x16_t qx4 = vmulq_u8(qx0, vdupq_n_u8(9));
+            uint8x16_t qx5 = vmulq_u8(qx1, vdupq_n_u8(9));
+            uint8x16_t qx6 = vmulq_u8(qx0, vdupq_n_u8(27));
+            uint8x16_t qx7 = vmulq_u8(qx1, vdupq_n_u8(27));
+            uint8x16_t qx8 = vmulq_u8(qx0, vdupq_n_u8(81));
+            uint8x16_t qx9 = vmulq_u8(qx1, vdupq_n_u8(81));
+
+            // multiply by 3 and keep the 2 bits above 8 bits
+            int8x16_t sqx0 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx0, vshrq_n_u8(qx0, 1)), 6));
+            int8x16_t sqx1 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx1, vshrq_n_u8(qx1, 1)), 6));
+            int8x16_t sqx2 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx2, vshrq_n_u8(qx2, 1)), 6));
+            int8x16_t sqx3 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx3, vshrq_n_u8(qx3, 1)), 6));
+            int8x16_t sqx4 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx4, vshrq_n_u8(qx4, 1)), 6));
+            int8x16_t sqx5 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx5, vshrq_n_u8(qx5, 1)), 6));
+            int8x16_t sqx6 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx6, vshrq_n_u8(qx6, 1)), 6));
+            int8x16_t sqx7 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx7, vshrq_n_u8(qx7, 1)), 6));
+            int8x16_t sqx8 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx8, vshrq_n_u8(qx8, 1)), 6));
+            int8x16_t sqx9 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx9, vshrq_n_u8(qx9, 1)), 6));
+
+            const int8x16_t qy0 = vld1q_s8(y[i].qs +   0);
+            const int8x16_t qy1 = vld1q_s8(y[i].qs +  16);
+            const int8x16_t qy2 = vld1q_s8(y[i].qs +  32);
+            const int8x16_t qy3 = vld1q_s8(y[i].qs +  48);
+            const int8x16_t qy4 = vld1q_s8(y[i].qs +  64);
+            const int8x16_t qy5 = vld1q_s8(y[i].qs +  80);
+            const int8x16_t qy6 = vld1q_s8(y[i].qs +  96);
+            const int8x16_t qy7 = vld1q_s8(y[i].qs + 112);
+            const int8x16_t qy8 = vld1q_s8(y[i].qs + 128);
+            const int8x16_t qy9 = vld1q_s8(y[i].qs + 144);
+
+#if defined(__ARM_FEATURE_DOTPROD)
+            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
+            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
+            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
+            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
+            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
+            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
+            sumi0 = vdotq_s32(sumi0, sqx6, qy6);
+            sumi1 = vdotq_s32(sumi1, sqx7, qy7);
+            sumi0 = vdotq_s32(sumi0, sqx8, qy8);
+            sumi1 = vdotq_s32(sumi1, sqx9, qy9);
+#else
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx6), vget_low_s8(qy6));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx6), vget_high_s8(qy6));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx7), vget_low_s8(qy7));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx7), vget_high_s8(qy7));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx8), vget_low_s8(qy8));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx8), vget_high_s8(qy8));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx9), vget_low_s8(qy9));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx9), vget_high_s8(qy9));
+#endif
+        }
+
+        // last 16 bytes of 5-element, along with the 4 bytes of 4 elements
+        {
+            uint8x16_t qx0 = vld1q_u8(x[i].qs + 32);
+            uint8x16_t qx1 = vmulq_u8(qx0, vdupq_n_u8(3));
+            uint8x16_t qx2 = vmulq_u8(qx0, vdupq_n_u8(9));
+            uint8x16_t qx3 = vmulq_u8(qx0, vdupq_n_u8(27));
+            uint8x16_t qx4 = vmulq_u8(qx0, vdupq_n_u8(81));
+            uint32_t qh;
+            memcpy(&qh, x[i].qh, sizeof(qh)); // potentially unaligned
+            uint8x16_t qx5 = vreinterpretq_u8_u32(vdupq_n_u32(qh));
+            qx5 = vmulq_u8(qx5, shift);
+
+            // multiply by 3 and keep the 2 bits above 8 bits
+            int8x16_t sqx0 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx0, vshrq_n_u8(qx0, 1)), 6));
+            int8x16_t sqx1 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx1, vshrq_n_u8(qx1, 1)), 6));
+            int8x16_t sqx2 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx2, vshrq_n_u8(qx2, 1)), 6));
+            int8x16_t sqx3 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx3, vshrq_n_u8(qx3, 1)), 6));
+            int8x16_t sqx4 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx4, vshrq_n_u8(qx4, 1)), 6));
+            int8x16_t sqx5 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx5, vshrq_n_u8(qx5, 1)), 6));
+
+            const int8x16_t qy0 = vld1q_s8(y[i].qs + 160);
+            const int8x16_t qy1 = vld1q_s8(y[i].qs + 176);
+            const int8x16_t qy2 = vld1q_s8(y[i].qs + 192);
+            const int8x16_t qy3 = vld1q_s8(y[i].qs + 208);
+            const int8x16_t qy4 = vld1q_s8(y[i].qs + 224);
+            const int8x16_t qy5 = vld1q_s8(y[i].qs + 240);
+
+#if defined(__ARM_FEATURE_DOTPROD)
+            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
+            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
+            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
+            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
+            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
+            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
+#else
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
+#endif
+        }
+
+        const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
+        const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+
+#if defined(__ARM_FEATURE_DOTPROD)
+        sumi0 = vaddq_s32(sumi0, sumi1);
+        sumi0 = vsubq_s32(sumi0, vpaddlq_s16(vaddq_s16(ysum0, ysum1)));
+
+        sumf += d * (float) vaddvq_s32(sumi0);
+#else
+        sumi0 = vaddq_s16(sumi0, sumi1);
+        sumi0 = vsubq_s16(sumi0, vaddq_s16(ysum0, ysum1));
+
+        sumf += d * (float) vaddlvq_s16(sumi0);
+#endif
+    }
+
+    *s = sumf;
+
+#elif defined(__AVX2__)
+    __m256 sumf = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+        // 16-bit sums
+        __m256i sumi0 = _mm256_setzero_si256();
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+
+        // first 32 bytes of 5 elements
+        {
+            __m256i qx0 = _mm256_loadu_si256((const __m256i *) (x[i].qs));
+            // 8-bit multiplies with shifts, masks and adds
+            __m256i qx1 = _mm256_add_epi8(qx0, _mm256_add_epi8(qx0, qx0)); // 1 * 3
+            __m256i qx2 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx0, 3), _mm256_set1_epi8(-8)), qx0); // 1 * 9
+            __m256i qx3 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx1, 3), _mm256_set1_epi8(-8)), qx1); // 3 * 9
+            __m256i qx4 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx2, 3), _mm256_set1_epi8(-8)), qx2); // 9 * 9
+
+            // TODO: can _mm256_mulhi_epu16 be faster even if 16-bits?
+
+            // Cancel the +1 from avg so that it behaves like a halving add
+            qx0 = _mm256_subs_epu8(qx0, _mm256_set1_epi8(1));
+            qx1 = _mm256_subs_epu8(qx1, _mm256_set1_epi8(1));
+            qx2 = _mm256_subs_epu8(qx2, _mm256_set1_epi8(1));
+            qx3 = _mm256_subs_epu8(qx3, _mm256_set1_epi8(1));
+            qx4 = _mm256_subs_epu8(qx4, _mm256_set1_epi8(1));
+            // Multiply by 3 and get the top 2 bits
+            qx0 = _mm256_avg_epu8(qx0, _mm256_avg_epu8(qx0, _mm256_setzero_si256()));
+            qx1 = _mm256_avg_epu8(qx1, _mm256_avg_epu8(qx1, _mm256_setzero_si256()));
+            qx2 = _mm256_avg_epu8(qx2, _mm256_avg_epu8(qx2, _mm256_setzero_si256()));
+            qx3 = _mm256_avg_epu8(qx3, _mm256_avg_epu8(qx3, _mm256_setzero_si256()));
+            qx4 = _mm256_avg_epu8(qx4, _mm256_avg_epu8(qx4, _mm256_setzero_si256()));
+            qx0 = _mm256_and_si256(_mm256_srli_epi16(qx0, 6), _mm256_set1_epi8(3));
+            qx1 = _mm256_and_si256(_mm256_srli_epi16(qx1, 6), _mm256_set1_epi8(3));
+            qx2 = _mm256_and_si256(_mm256_srli_epi16(qx2, 6), _mm256_set1_epi8(3));
+            qx3 = _mm256_and_si256(_mm256_srli_epi16(qx3, 6), _mm256_set1_epi8(3));
+            qx4 = _mm256_and_si256(_mm256_srli_epi16(qx4, 6), _mm256_set1_epi8(3));
+
+            const __m256i qy0 = _mm256_loadu_si256((const __m256i *) (y[i].qs +   0));
+            const __m256i qy1 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  32));
+            const __m256i qy2 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  64));
+            const __m256i qy3 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  96));
+            const __m256i qy4 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 128));
+
+            qx0 = _mm256_maddubs_epi16(qx0, qy0);
+            qx1 = _mm256_maddubs_epi16(qx1, qy1);
+            qx2 = _mm256_maddubs_epi16(qx2, qy2);
+            qx3 = _mm256_maddubs_epi16(qx3, qy3);
+            qx4 = _mm256_maddubs_epi16(qx4, qy4);
+
+            sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(qx0, qx1));
+            sumi1 = _mm256_add_epi16(sumi1, _mm256_add_epi16(qx2, qx3));
+            sumi2 = _mm256_add_epi16(sumi2, qx4);
+        }
+
+        // last 16 bytes of 5-element, along with the 4 bytes of 4 elements
+        {
+            __m128i qx0 = _mm_loadu_si128((const __m128i *) (x[i].qs + 32));
+            uint32_t qh;
+            memcpy(&qh, x[i].qh, sizeof(qh)); // potentially unaligned
+            __m256i qx5_l = _mm256_cvtepu8_epi16(_mm_set1_epi32(qh));
+            __m128i qx1 = _mm_add_epi8(qx0, _mm_add_epi8(qx0, qx0)); // 1 * 3
+            __m128i qx2 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx0, 3), _mm_set1_epi8(-8)), qx0); // 1 * 9
+            __m128i qx3 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx1, 3), _mm_set1_epi8(-8)), qx1); // 3 * 9
+            __m128i qx4 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx2, 3), _mm_set1_epi8(-8)), qx2); // 9 * 9
+            __m256i qx01 = MM256_SET_M128I(qx1, qx0);
+            __m256i qx23 = MM256_SET_M128I(qx3, qx2);
+
+            // avx2 does not have 8-bit multiplies, so 16-bit it is.
+            qx5_l = _mm256_mullo_epi16(qx5_l, _mm256_set_epi16(27, 27, 27, 27, 9, 9, 9, 9, 3, 3, 3, 3, 1, 1, 1, 1));
+            qx5_l = _mm256_and_si256(qx5_l, _mm256_set1_epi16(0xFF));
+            __m128i qx5 = _mm_packus_epi16(_mm256_castsi256_si128(qx5_l), _mm256_extracti128_si256(qx5_l, 1));
+
+            __m256i qx45 = MM256_SET_M128I(qx5, qx4);
+
+            // Cancel the +1 from avg so that it behaves like a halving add
+            qx01 = _mm256_subs_epu8(qx01, _mm256_set1_epi8(1));
+            qx23 = _mm256_subs_epu8(qx23, _mm256_set1_epi8(1));
+            qx45 = _mm256_subs_epu8(qx45, _mm256_set1_epi8(1));
+            // Multiply by 3 and get the top 2 bits
+            qx01 = _mm256_avg_epu8(qx01, _mm256_avg_epu8(qx01, _mm256_setzero_si256()));
+            qx23 = _mm256_avg_epu8(qx23, _mm256_avg_epu8(qx23, _mm256_setzero_si256()));
+            qx45 = _mm256_avg_epu8(qx45, _mm256_avg_epu8(qx45, _mm256_setzero_si256()));
+            qx01 = _mm256_and_si256(_mm256_srli_epi16(qx01, 6), _mm256_set1_epi8(3));
+            qx23 = _mm256_and_si256(_mm256_srli_epi16(qx23, 6), _mm256_set1_epi8(3));
+            qx45 = _mm256_and_si256(_mm256_srli_epi16(qx45, 6), _mm256_set1_epi8(3));
+
+            const __m256i qy01 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 160));
+            const __m256i qy23 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 192));
+            const __m256i qy45 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 224));
+
+            qx01 = _mm256_maddubs_epi16(qx01, qy01);
+            qx23 = _mm256_maddubs_epi16(qx23, qy23);
+            qx45 = _mm256_maddubs_epi16(qx45, qy45);
+
+            sumi0 = _mm256_add_epi16(sumi0, qx01);
+            sumi1 = _mm256_add_epi16(sumi1, qx23);
+            sumi2 = _mm256_add_epi16(sumi2, qx45);
+        }
+
+        const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
+
+        sumi0 = _mm256_sub_epi16(sumi0, ysum);
+        sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(sumi1, sumi2));
+        sumi0 = _mm256_madd_epi16(sumi0, _mm256_set1_epi16(1));
+
+        sumf = _mm256_add_ps(_mm256_mul_ps(_mm256_cvtepi32_ps(sumi0), d), sumf);
+    }
+
+    *s = hsum_float_8(sumf);
+
+#else
+    const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
+
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        int sum = 0;
+
+        for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
+            for (size_t l = 0; l < 5; ++l) {
+                for (size_t m = 0; m < 32; ++m) {
+                    uint8_t q = x[i].qs[j + m] * pow3[l];
+                    uint16_t xi = ((uint16_t) q * 3) >> 8;
+                    sum += (xi - 1) * y[i].qs[j*5 + l*32 + m];
+                }
+            }
+        }
+        for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
+            for (size_t l = 0; l < 5; ++l) {
+                for (size_t m = 0; m < 16; ++m) {
+                    uint8_t q = x[i].qs[j + m] * pow3[l];
+                    uint16_t xi = ((uint16_t) q * 3) >> 8;
+                    sum += (xi - 1) * y[i].qs[j*5 + l*16 + m];
+                }
+            }
+        }
+
+        for (size_t l = 0; l < 4; ++l) {
+            for (size_t j = 0; j < sizeof(x->qh); ++j) {
+                uint8_t q = x[i].qh[j] * pow3[l];
+                uint16_t xi = ((uint16_t) q * 3) >> 8;
+                sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j];
+            }
+        }
+
+        sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d);
+    }
+
+    *s = sumf;
+#endif
+}
+
+void ggml_vec_dot_tq2_0_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_tq2_0 * restrict x = vx;
+    const block_q8_K  * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+    float sumf = 0.0f;
+
+    const uint8x16_t m3 = vdupq_n_u8(3);
+
+    for (int i = 0; i < nb; ++i) {
+#if defined(__ARM_FEATURE_DOTPROD)
+        int32x4_t sumi0 = vdupq_n_s32(0);
+        int32x4_t sumi1 = vdupq_n_s32(0);
+#else
+        int16x8_t sumi0 = vdupq_n_s16(0);
+        int16x8_t sumi1 = vdupq_n_s16(0);
+#endif
+
+        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+            uint8x16_t qx0 = vld1q_u8(x[i].qs + j);
+            uint8x16_t qx1 = vld1q_u8(x[i].qs + j + 16);
+            uint8x16_t qx2 = vshrq_n_u8(qx0, 2);
+            uint8x16_t qx3 = vshrq_n_u8(qx1, 2);
+            uint8x16_t qx4 = vshrq_n_u8(qx0, 4);
+            uint8x16_t qx5 = vshrq_n_u8(qx1, 4);
+            uint8x16_t qx6 = vshrq_n_u8(qx0, 6);
+            uint8x16_t qx7 = vshrq_n_u8(qx1, 6);
+
+            int8x16_t sqx0 = vreinterpretq_s8_u8(vandq_u8(qx0, m3));
+            int8x16_t sqx1 = vreinterpretq_s8_u8(vandq_u8(qx1, m3));
+            int8x16_t sqx2 = vreinterpretq_s8_u8(vandq_u8(qx2, m3));
+            int8x16_t sqx3 = vreinterpretq_s8_u8(vandq_u8(qx3, m3));
+            int8x16_t sqx4 = vreinterpretq_s8_u8(vandq_u8(qx4, m3));
+            int8x16_t sqx5 = vreinterpretq_s8_u8(vandq_u8(qx5, m3));
+            int8x16_t sqx6 = vreinterpretq_s8_u8(vandq_u8(qx6, m3));
+            int8x16_t sqx7 = vreinterpretq_s8_u8(vandq_u8(qx7, m3));
+
+            const int8x16_t qy0 = vld1q_s8(y[i].qs + j*4 +   0);
+            const int8x16_t qy1 = vld1q_s8(y[i].qs + j*4 +  16);
+            const int8x16_t qy2 = vld1q_s8(y[i].qs + j*4 +  32);
+            const int8x16_t qy3 = vld1q_s8(y[i].qs + j*4 +  48);
+            const int8x16_t qy4 = vld1q_s8(y[i].qs + j*4 +  64);
+            const int8x16_t qy5 = vld1q_s8(y[i].qs + j*4 +  80);
+            const int8x16_t qy6 = vld1q_s8(y[i].qs + j*4 +  96);
+            const int8x16_t qy7 = vld1q_s8(y[i].qs + j*4 + 112);
+
+#if defined(__ARM_FEATURE_DOTPROD)
+            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
+            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
+            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
+            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
+            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
+            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
+            sumi0 = vdotq_s32(sumi0, sqx6, qy6);
+            sumi1 = vdotq_s32(sumi1, sqx7, qy7);
+#else
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx6), vget_low_s8(qy6));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx6), vget_high_s8(qy6));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx7), vget_low_s8(qy7));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx7), vget_high_s8(qy7));
+#endif
+        }
+
+        const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
+        const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+
+#if defined(__ARM_FEATURE_DOTPROD)
+        sumi0 = vaddq_s32(sumi0, sumi1);
+        sumi0 = vsubq_s32(sumi0, vpaddlq_s16(vaddq_s16(ysum0, ysum1)));
+
+        sumf += d * (float) vaddvq_s32(sumi0);
+#else
+        sumi0 = vaddq_s16(sumi0, sumi1);
+        sumi0 = vsubq_s16(sumi0, vaddq_s16(ysum0, ysum1));
+
+        sumf += d * (float) vaddlvq_s16(sumi0);
+#endif
+    }
+
+    *s = sumf;
+
+#elif defined(__AVX2__)
+    __m256 sumf = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+        // 16-bit sums, because 256*127 still fits
+        __m256i sumi0 = _mm256_setzero_si256();
+        __m256i sumi1 = _mm256_setzero_si256();
+
+        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+            __m256i qx0 = _mm256_loadu_si256((const __m256i *) (x[i].qs + j));
+            __m256i qx1 = _mm256_srli_epi16(qx0, 2);
+            __m256i qx2 = _mm256_srli_epi16(qx0, 4);
+            __m256i qx3 = _mm256_srli_epi16(qx0, 6);
+
+            // 0, 1, 2 (should not be 3)
+            qx0 = _mm256_and_si256(qx0, _mm256_set1_epi8(3));
+            qx1 = _mm256_and_si256(qx1, _mm256_set1_epi8(3));
+            qx2 = _mm256_and_si256(qx2, _mm256_set1_epi8(3));
+            qx3 = _mm256_and_si256(qx3, _mm256_set1_epi8(3));
+
+            const __m256i qy0 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 +  0));
+            const __m256i qy1 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 32));
+            const __m256i qy2 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 64));
+            const __m256i qy3 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 96));
+
+            qx0 = _mm256_maddubs_epi16(qx0, qy0);
+            qx1 = _mm256_maddubs_epi16(qx1, qy1);
+            qx2 = _mm256_maddubs_epi16(qx2, qy2);
+            qx3 = _mm256_maddubs_epi16(qx3, qy3);
+
+            sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(qx0, qx1));
+            sumi1 = _mm256_add_epi16(sumi1, _mm256_add_epi16(qx2, qx3));
+        }
+
+        const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
+
+        sumi0 = _mm256_add_epi16(sumi0, sumi1);
+        sumi0 = _mm256_sub_epi16(sumi0, ysum);
+        sumi0 = _mm256_madd_epi16(sumi0, _mm256_set1_epi16(1));
+
+        sumf = _mm256_add_ps(_mm256_mul_ps(_mm256_cvtepi32_ps(sumi0), d), sumf);
+    }
+
+    *s = hsum_float_8(sumf);
+
+#else
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        int32_t sumi = 0;
+
+        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+            for (size_t l = 0; l < 4; ++l) {
+                for (size_t k = 0; k < 32; ++k) {
+                    sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1);
+                }
+            }
+        }
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        sumf += (float) sumi * d;
+    }
+
+    *s = sumf;
+#endif
+}
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q2_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#ifdef __ARM_NEON
+    const uint8x16_t m3 = vdupq_n_u8(0x3);
+    const uint8x16_t m4 = vdupq_n_u8(0xF);
+
+    const int32x4_t vzero = vdupq_n_s32(0);
+
+    ggml_int8x16x2_t q2bytes;
+    uint8_t aux[16];
+
+    float sum = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * restrict sc = x[i].scales;
+
+        const uint8x16_t mins_and_scales = vld1q_u8(sc);
+        const uint8x16_t scales = vandq_u8(mins_and_scales, m4);
+        vst1q_u8(aux, scales);
+
+        const uint8x16_t mins = vshrq_n_u8(mins_and_scales, 4);
+        const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
+        const ggml_int16x8x2_t mins16 = {{vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}};
+        const int32x4_t s0 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[0]), vget_low_s16 (q8sums.val[0])),
+                                       vmull_s16(vget_high_s16(mins16.val[0]), vget_high_s16(q8sums.val[0])));
+        const int32x4_t s1 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[1]), vget_low_s16 (q8sums.val[1])),
+                                       vmull_s16(vget_high_s16(mins16.val[1]), vget_high_s16(q8sums.val[1])));
+        sum += dmin * vaddvq_s32(vaddq_s32(s0, s1));
+
+        int isum = 0;
+        int is = 0;
+
+// We use this macro instead of a function call because for some reason
+// the code runs 2-3% slower, even if the function is declared inline
+#define MULTIPLY_ACCUM_WITH_SCALE(index)\
+        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * aux[is+(index)];\
+        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * aux[is+1+(index)];
+
+#define SHIFT_MULTIPLY_ACCUM_WITH_SCALE(shift, index)\
+        q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;\
+        q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[0], (shift)), m3));\
+        q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[1], (shift)), m3));\
+        MULTIPLY_ACCUM_WITH_SCALE((index));
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            const ggml_uint8x16x2_t q2bits = ggml_vld1q_u8_x2(q2); q2 += 32;
+
+            ggml_int8x16x2_t q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
+            q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[0], m3));
+            q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[1], m3));
+
+            MULTIPLY_ACCUM_WITH_SCALE(0);
+
+            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(2, 2);
+            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(4, 4);
+            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(6, 6);
+
+            is += 8;
+        }
+
+        sum += d * isum;
+    }
+
+    *s = sum;
+
+#elif defined __AVX2__
+
+    const __m256i m3 = _mm256_set1_epi8(3);
+    const __m128i m4 = _mm_set1_epi8(0xF);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+        const __m128i scales8 = _mm_and_si128(mins_and_scales, m4);
+        const __m128i mins8 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
+        const __m256i mins = _mm256_cvtepi8_epi16(mins8);
+        const __m256i prod = _mm256_madd_epi16(mins, _mm256_loadu_si256((const __m256i*)y[i].bsums));
+
+        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(prod), acc);
+
+        const __m256i all_scales = _mm256_cvtepi8_epi16(scales8);
+        const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
+        const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
+        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
+
+        __m256i sumi = _mm256_setzero_si256();
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m256i q2bits = _mm256_loadu_si256((const __m256i*)q2); q2 += 32;
+
+            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            const __m256i q2_0 = _mm256_and_si256(q2bits, m3);
+            const __m256i q2_1 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 2), m3);
+            const __m256i q2_2 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 4), m3);
+            const __m256i q2_3 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 6), m3);
+
+            __m256i p0 = _mm256_maddubs_epi16(q2_0, q8_0);
+            __m256i p1 = _mm256_maddubs_epi16(q2_1, q8_1);
+            __m256i p2 = _mm256_maddubs_epi16(q2_2, q8_2);
+            __m256i p3 = _mm256_maddubs_epi16(q2_3, q8_3);
+
+            p0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(0)), p0);
+            p1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(1)), p1);
+            p2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(2)), p2);
+            p3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(3)), p3);
+
+            p0 = _mm256_add_epi32(p0, p1);
+            p2 = _mm256_add_epi32(p2, p3);
+
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p0, p2));
+        }
+
+        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(0x3);
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i m2 = _mm_set1_epi8(0x2);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        // load mins and scales from block_q2_K.scales[QK_K/16]
+        const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+        const __m128i scales16 = _mm_and_si128(mins_and_scales, m4);
+        const __m128i mins16 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
+        const __m128i mins_0 = _mm_cvtepi8_epi16(mins16);
+        const __m128i mins_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(mins16, mins16));
+
+        // summs = y[i].bsums * (x[i].scales >> 4) in 16bits*8*2 to 32bits*4*2
+        const __m128i summs_0 = _mm_madd_epi16(mins_0, _mm_loadu_si128((const __m128i*)&y[i].bsums[0]));
+        const __m128i summs_1 = _mm_madd_epi16(mins_1, _mm_loadu_si128((const __m128i*)&y[i].bsums[8]));
+
+        // sumf += -dmin * summs in 32bits*8
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(MM256_SET_M128I(summs_1, summs_0))), acc);
+
+        const __m128i scales_0 = _mm_cvtepi8_epi16(scales16);
+        const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales16, scales16));
+        const __m128i scales[2] = { scales_0, scales_1 };
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            // load Q8 quants int8*16*8 from block_q8_K.qs[QK_K]
+            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+            // load 2bits*16*8 from block_q2_K.qs[QK_K/4]
+            __m128i q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
+            const __m128i q2_0 = _mm_and_si128(q2bits, m3);
+            const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
+            const __m128i q2_4 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
+            const __m128i q2_6 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
+            q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
+            const __m128i q2_1 = _mm_and_si128(q2bits, m3);
+            const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
+            const __m128i q2_5 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
+            const __m128i q2_7 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
+
+            // isuml = q8[l] * ((q2[l] >> shift) & 3) in 8bits*16*8 to 16bits*8*8
+            __m128i p0 = _mm_maddubs_epi16(q2_0, q8_0);
+            __m128i p1 = _mm_maddubs_epi16(q2_1, q8_1);
+            __m128i p2 = _mm_maddubs_epi16(q2_2, q8_2);
+            __m128i p3 = _mm_maddubs_epi16(q2_3, q8_3);
+            __m128i p4 = _mm_maddubs_epi16(q2_4, q8_4);
+            __m128i p5 = _mm_maddubs_epi16(q2_5, q8_5);
+            __m128i p6 = _mm_maddubs_epi16(q2_6, q8_6);
+            __m128i p7 = _mm_maddubs_epi16(q2_7, q8_7);
+
+            // isum += (x[i].scales[is++] & 0xF) * isuml in 16bits*8*8 to 32bits*4*8
+            __m128i shuffle = _mm_set1_epi16(0x0100);
+            p0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p0);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p1);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p2);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p3);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p4);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p5);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p6);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p7);
+
+            p0 = _mm_add_epi32(p0, p1);
+            p2 = _mm_add_epi32(p2, p3);
+            p4 = _mm_add_epi32(p4, p5);
+            p6 = _mm_add_epi32(p6, p7);
+
+            // isum in 32bits*4*2
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p0, p2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p4, p6));
+        }
+
+        // sumf += dall * isum - dmin * summs in 32bits
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dall), _mm256_cvtepi32_ps(sumi)), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __riscv_v_intrinsic
+
+    float sumf = 0;
+    uint8_t temp_01[32] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                            1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1};
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * q2 = x[i].qs;
+        const  int8_t * q8 = y[i].qs;
+        const uint8_t * sc = x[i].scales;
+
+        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        size_t vl = 16;
+
+        vuint8m1_t scales = __riscv_vle8_v_u8m1(sc, vl);
+        vuint8m1_t aux = __riscv_vand_vx_u8m1(scales, 0x0F, vl);
+
+        vint16m1_t q8sums = __riscv_vle16_v_i16m1(y[i].bsums, vl);
+
+        vuint8mf2_t scales_2 = __riscv_vle8_v_u8mf2(sc, vl);
+        vuint8mf2_t mins8 = __riscv_vsrl_vx_u8mf2(scales_2, 0x4, vl);
+        vint16m1_t mins = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(mins8, vl));
+        vint32m2_t prod = __riscv_vwmul_vv_i32m2(q8sums, mins, vl);
+        vint32m1_t vsums = __riscv_vredsum_vs_i32m2_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
+
+        sumf  += dmin * __riscv_vmv_x_s_i32m1_i32(vsums);
+
+        vl = 32;
+
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+        vuint8m1_t v_b = __riscv_vle8_v_u8m1(temp_01, vl);
+
+        uint8_t is=0;
+        int isum=0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            // load Q2
+            vuint8m1_t q2_x = __riscv_vle8_v_u8m1(q2, vl);
+
+            vuint8m1_t q2_0 = __riscv_vand_vx_u8m1(q2_x, 0x03, vl);
+            vuint8m1_t q2_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x2, vl), 0x03 , vl);
+            vuint8m1_t q2_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x4, vl), 0x03 , vl);
+            vuint8m1_t q2_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x6, vl), 0x03 , vl);
+
+            // duplicate scale elements for product
+            vuint8m1_t sc0 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 0+is, vl), vl);
+            vuint8m1_t sc1 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 2+is, vl), vl);
+            vuint8m1_t sc2 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 4+is, vl), vl);
+            vuint8m1_t sc3 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 6+is, vl), vl);
+
+            vint16m2_t p0 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_0, sc0, vl));
+            vint16m2_t p1 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_1, sc1, vl));
+            vint16m2_t p2 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_2, sc2, vl));
+            vint16m2_t p3 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_3, sc3, vl));
+
+            // load Q8
+            vint8m1_t q8_0 = __riscv_vle8_v_i8m1(q8, vl);
+            vint8m1_t q8_1 = __riscv_vle8_v_i8m1(q8+32, vl);
+            vint8m1_t q8_2 = __riscv_vle8_v_i8m1(q8+64, vl);
+            vint8m1_t q8_3 = __riscv_vle8_v_i8m1(q8+96, vl);
+
+            vint32m4_t s0 = __riscv_vwmul_vv_i32m4(p0, __riscv_vwcvt_x_x_v_i16m2(q8_0, vl), vl);
+            vint32m4_t s1 = __riscv_vwmul_vv_i32m4(p1, __riscv_vwcvt_x_x_v_i16m2(q8_1, vl), vl);
+            vint32m4_t s2 = __riscv_vwmul_vv_i32m4(p2, __riscv_vwcvt_x_x_v_i16m2(q8_2, vl), vl);
+            vint32m4_t s3 = __riscv_vwmul_vv_i32m4(p3, __riscv_vwcvt_x_x_v_i16m2(q8_3, vl), vl);
+
+            vint32m1_t isum0 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s0, s1, vl), vzero, vl);
+            vint32m1_t isum1 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s2, s3, vl), isum0, vl);
+
+            isum += __riscv_vmv_x_s_i32m1_i32(isum1);
+
+            q2+=32;  q8+=128;  is=8;
+
+        }
+
+        sumf += dall * isum;
+
+    }
+
+    *s = sumf;
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0x3);
+    const vector signed char lowScaleMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
+    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vdmin = vec_mul(vxmin, vyd);
+
+        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
+        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
+
+        vector signed char q2xmins = (vector signed char)vec_xl( 0, x[i].scales);
+        vector signed char vscales = vec_and(q2xmins, lowScaleMask);
+
+        q2xmins = vec_sr(q2xmins, v4);
+        vector signed short q2xmins0 = vec_unpackh(q2xmins);
+        vector signed short q2xmins1 = vec_unpackl(q2xmins);
+
+        vector signed int prod0 = vec_mule(q2xmins0, q8ysums0);
+        vector signed int prod1 = vec_mulo(q2xmins0, q8ysums0);
+        vector signed int prod2 = vec_mule(q2xmins1, q8ysums1);
+        vector signed int prod3 = vec_mulo(q2xmins1, q8ysums1);
+
+        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
+        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
+        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
+        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            __builtin_prefetch(q2, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q2);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q2);
+            q2 += 32;
+
+            vector unsigned char q2x00 = (vector unsigned char)vec_and(qxs0, lowMask);
+            vector unsigned char q2x01 = (vector unsigned char)vec_and(vec_sr(qxs0, v2), lowMask);
+            vector unsigned char q2x02 = (vector unsigned char)vec_and(vec_sr(qxs0, v4), lowMask);
+            vector unsigned char q2x03 = (vector unsigned char)vec_and(vec_sr(qxs0, v6), lowMask);
+            vector unsigned char q2x10 = (vector unsigned char)vec_and(qxs1, lowMask);
+            vector unsigned char q2x11 = (vector unsigned char)vec_and(vec_sr(qxs1, v2), lowMask);
+            vector unsigned char q2x12 = (vector unsigned char)vec_and(vec_sr(qxs1, v4), lowMask);
+            vector unsigned char q2x13 = (vector unsigned char)vec_and(vec_sr(qxs1, v6), lowMask);
+
+            vector signed char q8y00 = vec_xl(  0, q8);
+            vector signed char q8y10 = vec_xl( 16, q8);
+            vector signed char q8y01 = vec_xl( 32, q8);
+            vector signed char q8y11 = vec_xl( 48, q8);
+            vector signed char q8y02 = vec_xl( 64, q8);
+            vector signed char q8y12 = vec_xl( 80, q8);
+            vector signed char q8y03 = vec_xl( 96, q8);
+            vector signed char q8y13 = vec_xl(112, q8);
+            q8 += 128;
+
+            vector signed int qv0 = vec_msum(q8y00, q2x00, v0);
+            vector signed int qv1 = vec_msum(q8y01, q2x01, v0);
+            vector signed int qv2 = vec_msum(q8y02, q2x02, v0);
+            vector signed int qv3 = vec_msum(q8y03, q2x03, v0);
+            vector signed int qv4 = vec_msum(q8y10, q2x10, v0);
+            vector signed int qv5 = vec_msum(q8y11, q2x11, v0);
+            vector signed int qv6 = vec_msum(q8y12, q2x12, v0);
+            vector signed int qv7 = vec_msum(q8y13, q2x13, v0);
+
+            vector signed short vscales_07 = vec_unpackh(vscales);
+            vector signed int vscales_03 = vec_unpackh(vscales_07);
+            vector signed int vscales_47 = vec_unpackl(vscales_07);
+            vector signed int vs0 = vec_splat(vscales_03, 0);
+            vector signed int vs1 = vec_splat(vscales_03, 1);
+            vector signed int vs2 = vec_splat(vscales_03, 2);
+            vector signed int vs3 = vec_splat(vscales_03, 3);
+            vector signed int vs4 = vec_splat(vscales_47, 0);
+            vector signed int vs5 = vec_splat(vscales_47, 1);
+            vector signed int vs6 = vec_splat(vscales_47, 2);
+            vector signed int vs7 = vec_splat(vscales_47, 3);
+            vscales = vec_sld(vscales, vscales, 8);
+
+            vsumi0 = vec_add(vec_mul(qv0, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv1, vs2), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv2, vs4), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv3, vs6), vsumi3);
+            vsumi4 = vec_add(vec_mul(qv4, vs1), vsumi4);
+            vsumi5 = vec_add(vec_mul(qv5, vs3), vsumi5);
+            vsumi6 = vec_add(vec_mul(qv6, vs5), vsumi6);
+            vsumi7 = vec_add(vec_mul(qv7, vs7), vsumi7);
+        }
+
+        vsumi0 = vec_add(vsumi0, vsumi4);
+        vsumi1 = vec_add(vsumi1, vsumi5);
+        vsumi2 = vec_add(vsumi2, vsumi6);
+        vsumi3 = vec_add(vsumi3, vsumi7);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined __loongarch_asx
+
+    const __m256i m3 = __lasx_xvreplgr2vr_b(3);
+    const __m128i m4 = __lsx_vreplgr2vr_b(0xF);
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m128i mins_and_scales = __lsx_vld((const __m128i*)x[i].scales, 0);
+        const __m128i scales8 = __lsx_vand_v(mins_and_scales, m4);
+        const __m128i mins8 = __lsx_vand_v(__lsx_vsrli_h(mins_and_scales, 4), m4);
+        const __m256i mins = lasx_ext8_16(mins8);
+        const __m256i prod = lasx_madd_h(mins, __lasx_xvld((const __m256i*)y[i].bsums, 0));
+
+        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(dmin), __lasx_xvffint_s_w(prod), acc);
+
+        const __m256i all_scales = lasx_ext8_16(scales8);
+        const __m128i l_scales = lasx_extracti128(all_scales, 0);
+        const __m128i h_scales = lasx_extracti128(all_scales, 1);
+        const __m256i scales[2] = {lasx_insertf128(l_scales, l_scales), lasx_insertf128(h_scales, h_scales)};
+
+        __m256i sumi = __lasx_xvldi(0);
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m256i q2bits = __lasx_xvld((const __m256i*)q2, 0); q2 += 32;
+
+            const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+
+            const __m256i q2_0 = __lasx_xvand_v(q2bits, m3);
+            const __m256i q2_1 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 2), m3);
+            const __m256i q2_2 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 4), m3);
+            const __m256i q2_3 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 6), m3);
+
+            __m256i p0 = lasx_maddubs_h(q2_0, q8_0);
+            __m256i p1 = lasx_maddubs_h(q2_1, q8_1);
+            __m256i p2 = lasx_maddubs_h(q2_2, q8_2);
+            __m256i p3 = lasx_maddubs_h(q2_3, q8_3);
+
+            p0 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(0)), p0);
+            p1 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(1)), p1);
+            p2 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(2)), p2);
+            p3 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(3)), p3);
+
+            p0 = __lasx_xvadd_w(p0, p1);
+            p2 = __lasx_xvadd_w(p2, p3);
+
+            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p0, p2));
+        }
+
+        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * q2 = x[i].qs;
+        const  int8_t * q8 = y[i].qs;
+        const uint8_t * sc = x[i].scales;
+
+        int summs = 0;
+        for (int j = 0; j < 16; ++j) {
+            summs += y[i].bsums[j] * (sc[j] >> 4);
+        }
+
+        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        int isum = 0;
+        int is = 0;
+        int d;
+        for (int k = 0; k < QK_K/128; ++k) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+                d = sc[is++] & 0xF;
+                int isuml = 0;
+                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                d = sc[is++] & 0xF;
+                isuml = 0;
+                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                shift += 2;
+                q8 += 32;
+            }
+            q2 += 32;
+        }
+        sumf += dall * isum - dmin * summs;
+    }
+    *s = sumf;
+#endif
+}
+
+void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    const block_q3_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#ifdef __ARM_NEON
+
+    uint32_t aux[3];
+    uint32_t utmp[4];
+
+    const uint8x16_t m3b = vdupq_n_u8(0x3);
+    const int32x4_t  vzero = vdupq_n_s32(0);
+
+    const uint8x16_t m0 = vdupq_n_u8(1);
+    const uint8x16_t m1 = vshlq_n_u8(m0, 1);
+    const uint8x16_t m2 = vshlq_n_u8(m0, 2);
+    const uint8x16_t m3 = vshlq_n_u8(m0, 3);
+    const int8_t m32 = 32;
+
+    ggml_int8x16x4_t q3bytes;
+
+    float sum = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict qh = x[i].hmask;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh);
+
+        ggml_uint8x16x4_t q3h;
+
+        int32_t isum = 0;
+
+        // Set up scales
+        memcpy(aux, x[i].scales, 12);
+        utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
+        utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
+        utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
+        utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
+
+        int8_t * scale = (int8_t *)utmp;
+        for (int j = 0; j < 16; ++j) scale[j] -= m32;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const ggml_uint8x16x2_t q3bits = ggml_vld1q_u8_x2(q3); q3 += 32;
+            const ggml_int8x16x4_t q8bytes_1 = ggml_vld1q_s8_x4(q8); q8 += 64;
+            const ggml_int8x16x4_t q8bytes_2 = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            q3h.val[0] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[0]), 2);
+            q3h.val[1] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[1]), 2);
+            q3h.val[2] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[0]), 1);
+            q3h.val[3] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[1]), 1);
+
+            q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[0], m3b)), vreinterpretq_s8_u8(q3h.val[0]));
+            q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[1], m3b)), vreinterpretq_s8_u8(q3h.val[1]));
+            q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 2), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
+            q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 2), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
+
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes_1.val[0])) * scale[0];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes_1.val[1])) * scale[1];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes_1.val[2])) * scale[2];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes_1.val[3])) * scale[3];
+
+            scale += 4;
+
+            q3h.val[0] = vbicq_u8(m2, qhbits.val[0]);
+            q3h.val[1] = vbicq_u8(m2, qhbits.val[1]);
+            q3h.val[2] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[0]), 1);
+            q3h.val[3] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[1]), 1);
+
+            q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 4), m3b)), vreinterpretq_s8_u8(q3h.val[0]));
+            q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 4), m3b)), vreinterpretq_s8_u8(q3h.val[1]));
+            q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 6), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
+            q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 6), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
+
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes_2.val[0])) * scale[0];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes_2.val[1])) * scale[1];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes_2.val[2])) * scale[2];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes_2.val[3])) * scale[3];
+
+            scale += 4;
+
+            if (j == 0) {
+                qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 4);
+                qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 4);
+            }
+
+        }
+        sum += d * isum;
+
+    }
+
+    *s = sum;
+
+#elif defined __AVX2__
+
+    const __m256i m3 = _mm256_set1_epi8(3);
+    const __m256i mone = _mm256_set1_epi8(1);
+    const __m128i m32 = _mm_set1_epi8(32);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    uint32_t aux[3];
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        // Set up scales
+        memcpy(aux, x[i].scales, 12);
+        __m128i scales128 = _mm_set_epi32(
+                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
+                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
+                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
+                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
+        scales128 = _mm_sub_epi8(scales128, m32);
+        const __m256i all_scales = _mm256_cvtepi8_epi16(scales128);
+        const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
+        const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
+        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
+
+        // high bit
+        const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].hmask);
+
+        // integer accumulator
+        __m256i sumi = _mm256_setzero_si256();
+
+        int bit = 0;
+        int is  = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            // load low 2 bits
+            const __m256i q3bits = _mm256_loadu_si256((const __m256i*)q3); q3 += 32;
+
+            // prepare low and high bits
+            const __m256i q3l_0 = _mm256_and_si256(q3bits, m3);
+            const __m256i q3h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+            ++bit;
+
+            const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 2), m3);
+            const __m256i q3h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+            ++bit;
+
+            const __m256i q3l_2 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 4), m3);
+            const __m256i q3h_2 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+            ++bit;
+
+            const __m256i q3l_3 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 6), m3);
+            const __m256i q3h_3 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+            ++bit;
+
+            // load Q8 quants
+            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
+            // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
+            // and 2 if the high bit was set)
+            __m256i q8s_0 = _mm256_maddubs_epi16(q3h_0, q8_0);
+            __m256i q8s_1 = _mm256_maddubs_epi16(q3h_1, q8_1);
+            __m256i q8s_2 = _mm256_maddubs_epi16(q3h_2, q8_2);
+            __m256i q8s_3 = _mm256_maddubs_epi16(q3h_3, q8_3);
+
+            __m256i p16_0 = _mm256_maddubs_epi16(q3l_0, q8_0);
+            __m256i p16_1 = _mm256_maddubs_epi16(q3l_1, q8_1);
+            __m256i p16_2 = _mm256_maddubs_epi16(q3l_2, q8_2);
+            __m256i p16_3 = _mm256_maddubs_epi16(q3l_3, q8_3);
+
+            p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
+            p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
+            p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
+            p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
+
+            // multiply with scales
+            p16_0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 0)), p16_0);
+            p16_1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 1)), p16_1);
+            p16_2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 2)), p16_2);
+            p16_3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 3)), p16_3);
+
+            // accumulate
+            p16_0 = _mm256_add_epi32(p16_0, p16_1);
+            p16_2 = _mm256_add_epi32(p16_2, p16_3);
+            sumi  = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_2));
+
+        }
+
+        // multiply with block scale and accumulate
+        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(3);
+    const __m128i mone = _mm_set1_epi8(1);
+    const __m128i m32 = _mm_set1_epi8(32);
+    const __m128i m2 = _mm_set1_epi8(2);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    const uint32_t *aux;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        // Set up scales
+        aux = (const uint32_t *)x[i].scales;
+        __m128i scales128 = _mm_set_epi32(
+                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
+                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
+                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
+                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
+        scales128 = _mm_sub_epi8(scales128, m32);
+        const __m128i scales_0 = _mm_cvtepi8_epi16(scales128);
+        const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales128, scales128));
+        const __m128i scales[2] = { scales_0, scales_1 };
+
+        // high bit *128*2 from block_q3_K.hmask[QK_K/8]
+        const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].hmask[0]);
+        const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].hmask[16]);
+
+        // integer accumulator
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            // load low 2 bits *64*2 from block_q3_K.qs[QK_K/4]
+            const __m128i q3bits_0 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
+            const __m128i q3bits_1 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
+
+            // prepare low and high bits
+            const int bit = j << 2;
+
+            const __m128i q3l_0 = _mm_and_si128(q3bits_0, m3);
+            const __m128i q3l_1 = _mm_and_si128(q3bits_1, m3);
+            const __m128i q3h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit)), bit), 2);
+            const __m128i q3h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit)), bit), 2);
+
+            const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 2), m3);
+            const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 2), m3);
+            const __m128i q3h_2 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
+            const __m128i q3h_3 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
+
+            const __m128i q3l_4 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 4), m3);
+            const __m128i q3l_5 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 4), m3);
+            const __m128i q3h_4 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
+            const __m128i q3h_5 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
+
+            const __m128i q3l_6 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 6), m3);
+            const __m128i q3l_7 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 6), m3);
+            const __m128i q3h_6 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
+            const __m128i q3h_7 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
+
+            // load Q8 quants from block_q8_K.qs[QK_K]
+            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
+            // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
+            // and 2 if the high bit was set)
+            __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, q8_0);
+            __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, q8_1);
+            __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, q8_2);
+            __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, q8_3);
+            __m128i q8s_4 = _mm_maddubs_epi16(q3h_4, q8_4);
+            __m128i q8s_5 = _mm_maddubs_epi16(q3h_5, q8_5);
+            __m128i q8s_6 = _mm_maddubs_epi16(q3h_6, q8_6);
+            __m128i q8s_7 = _mm_maddubs_epi16(q3h_7, q8_7);
+
+            __m128i p16_0 = _mm_maddubs_epi16(q3l_0, q8_0);
+            __m128i p16_1 = _mm_maddubs_epi16(q3l_1, q8_1);
+            __m128i p16_2 = _mm_maddubs_epi16(q3l_2, q8_2);
+            __m128i p16_3 = _mm_maddubs_epi16(q3l_3, q8_3);
+            __m128i p16_4 = _mm_maddubs_epi16(q3l_4, q8_4);
+            __m128i p16_5 = _mm_maddubs_epi16(q3l_5, q8_5);
+            __m128i p16_6 = _mm_maddubs_epi16(q3l_6, q8_6);
+            __m128i p16_7 = _mm_maddubs_epi16(q3l_7, q8_7);
+
+            p16_0 = _mm_sub_epi16(p16_0, q8s_0);
+            p16_1 = _mm_sub_epi16(p16_1, q8s_1);
+            p16_2 = _mm_sub_epi16(p16_2, q8s_2);
+            p16_3 = _mm_sub_epi16(p16_3, q8s_3);
+            p16_4 = _mm_sub_epi16(p16_4, q8s_4);
+            p16_5 = _mm_sub_epi16(p16_5, q8s_5);
+            p16_6 = _mm_sub_epi16(p16_6, q8s_6);
+            p16_7 = _mm_sub_epi16(p16_7, q8s_7);
+
+            // multiply with scales
+            __m128i shuffle = _mm_set1_epi16(0x0100);
+            p16_0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_0);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_1);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_2);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_3);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_4);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_5);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_6);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_7);
+
+            // accumulate
+            p16_0 = _mm_add_epi32(p16_0, p16_1);
+            p16_2 = _mm_add_epi32(p16_2, p16_3);
+            p16_4 = _mm_add_epi32(p16_4, p16_5);
+            p16_6 = _mm_add_epi32(p16_6, p16_7);
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_4, p16_6));
+
+        }
+
+        // multiply with block scale and accumulate
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __riscv_v_intrinsic
+
+    uint32_t aux[3];
+    uint32_t utmp[4];
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict qh = x[i].hmask;
+        const  int8_t * restrict q8 = y[i].qs;
+
+        memcpy(aux, x[i].scales, 12);
+        utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
+        utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
+        utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
+        utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
+
+        int8_t * scale = (int8_t *)utmp;
+        for (int j = 0; j < 16; ++j) scale[j] -= 32;
+
+
+        size_t vl = 32;
+        uint8_t m =  1;
+
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+        vuint8m1_t vqh = __riscv_vle8_v_u8m1(qh, vl);
+
+        int sum_t = 0;
+
+        for (int j = 0; j < QK_K; j += 128) {
+
+            vl = 32;
+
+            // load Q3
+            vuint8m1_t q3_x = __riscv_vle8_v_u8m1(q3, vl);
+
+            vint8m1_t q3_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q3_x, 0x03, vl));
+            vint8m1_t q3_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x2, vl), 0x03 , vl));
+            vint8m1_t q3_2 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x4, vl), 0x03 , vl));
+            vint8m1_t q3_3 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x6, vl), 0x03 , vl));
+
+            // compute mask for subtraction
+            vuint8m1_t qh_m0 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_0 = __riscv_vmseq_vx_u8m1_b8(qh_m0, 0, vl);
+            vint8m1_t q3_m0 = __riscv_vsub_vx_i8m1_mu(vmask_0, q3_0, q3_0, 0x4, vl);
+            m <<= 1;
+
+            vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_1 = __riscv_vmseq_vx_u8m1_b8(qh_m1, 0, vl);
+            vint8m1_t q3_m1 = __riscv_vsub_vx_i8m1_mu(vmask_1, q3_1, q3_1, 0x4, vl);
+            m <<= 1;
+
+            vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_2 = __riscv_vmseq_vx_u8m1_b8(qh_m2, 0, vl);
+            vint8m1_t q3_m2 = __riscv_vsub_vx_i8m1_mu(vmask_2, q3_2, q3_2, 0x4, vl);
+            m <<= 1;
+
+            vuint8m1_t qh_m3 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_3 = __riscv_vmseq_vx_u8m1_b8(qh_m3, 0, vl);
+            vint8m1_t q3_m3 = __riscv_vsub_vx_i8m1_mu(vmask_3, q3_3, q3_3, 0x4, vl);
+            m <<= 1;
+
+            // load Q8 and take product with Q3
+            vint16m2_t a0 = __riscv_vwmul_vv_i16m2(q3_m0, __riscv_vle8_v_i8m1(q8, vl), vl);
+            vint16m2_t a1 = __riscv_vwmul_vv_i16m2(q3_m1, __riscv_vle8_v_i8m1(q8+32, vl), vl);
+            vint16m2_t a2 = __riscv_vwmul_vv_i16m2(q3_m2, __riscv_vle8_v_i8m1(q8+64, vl), vl);
+            vint16m2_t a3 = __riscv_vwmul_vv_i16m2(q3_m3, __riscv_vle8_v_i8m1(q8+96, vl), vl);
+
+            vl = 16;
+
+            // retrieve lane to multiply with scale
+            vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl);
+            vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl);
+            vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl);
+            vint32m2_t aux1_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 1), (scale[3]), vl);
+            vint32m2_t aux2_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 0), (scale[4]), vl);
+            vint32m2_t aux2_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 1), (scale[5]), vl);
+            vint32m2_t aux3_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 0), (scale[6]), vl);
+            vint32m2_t aux3_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 1), (scale[7]), vl);
+
+            vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux0_0, aux0_1, vl), vzero, vl);
+            vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux1_0, aux1_1, vl), isum0, vl);
+            vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux2_0, aux2_1, vl), isum1, vl);
+            vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux3_0, aux3_1, vl), isum2, vl);
+
+            sum_t +=  __riscv_vmv_x_s_i32m1_i32(isum3);
+
+            q3 += 32;    q8 += 128;   scale += 8;
+
+        }
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+
+        sumf += d*sum_t;
+
+    }
+
+    *s = sumf;
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0x3);
+    const vector signed char lowMask1 = vec_splats((int8_t)0xf);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector signed char v1 = vec_splats((signed char)0x1);
+    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
+    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
+    const vector signed char off = vec_splats((signed char)0x20);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        UNUSED(kmask1);
+        UNUSED(kmask2);
+
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(u0, lowMask1);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = (vector signed char)vec_mergeh((vector signed int)u2, (vector signed int)vec_sr(u2, v2));
+        vector signed char u30 = vec_sl(vec_and(u3, lowMask), v4);
+        vector signed char u31 = vec_and(u3, lowMask2);
+
+        u1 = vec_or(u1, u30);
+        u2 = vec_or(vec_sr(u0, v4), u31);
+
+        vector signed char vscales = (vector signed char)vec_mergeh((vector signed long long)u1, (vector signed long long)u2);
+        vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].hmask);
+        vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].hmask);
+
+        vscales = vec_sub(vscales, off);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            __builtin_prefetch(q3, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q3);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q3);
+            q3 += 32;
+
+            //the low 2 bits
+            vector signed char qxs00 = vec_and(qxs0, lowMask);
+            vector signed char qxs01 = vec_and(vec_sr(qxs0, v2), lowMask);
+            vector signed char qxs02 = vec_and(vec_sr(qxs0, v4), lowMask);
+            vector signed char qxs03 = vec_and(vec_sr(qxs0, v6), lowMask);
+            vector signed char qxs10 = vec_and(qxs1, lowMask);
+            vector signed char qxs11 = vec_and(vec_sr(qxs1, v2), lowMask);
+            vector signed char qxs12 = vec_and(vec_sr(qxs1, v4), lowMask);
+            vector signed char qxs13 = vec_and(vec_sr(qxs1, v6), lowMask);
+
+            //the 3rd bit
+            vector signed char qxh00 = vec_sl(vec_andc(v1, qxhs0), v2);
+            vector signed char qxh01 = vec_sl(vec_andc(v1, vec_sr(qxhs0, (vector unsigned char)v1)), v2);
+            vector signed char qxh02 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v2)), v2);
+            vector signed char qxh03 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v3)), v2);
+            vector signed char qxh10 = vec_sl(vec_andc(v1, qxhs1), v2);
+            vector signed char qxh11 = vec_sl(vec_andc(v1, vec_sr(qxhs1, (vector unsigned char)v1)), v2);
+            vector signed char qxh12 = vec_sl(vec_andc(v1, vec_sr(qxhs1, v2)), v2);
+            vector signed char qxh13 = vec_sl(vec_andc(v1, vec_sr(qxhs1, v3)), v2);
+            qxhs0 = vec_sr(qxhs0, v4);
+            qxhs1 = vec_sr(qxhs1, v4);
+
+            vector signed char q3x00 = vec_sub(qxs00, qxh00);
+            vector signed char q3x01 = vec_sub(qxs01, qxh01);
+            vector signed char q3x02 = vec_sub(qxs02, qxh02);
+            vector signed char q3x03 = vec_sub(qxs03, qxh03);
+            vector signed char q3x10 = vec_sub(qxs10, qxh10);
+            vector signed char q3x11 = vec_sub(qxs11, qxh11);
+            vector signed char q3x12 = vec_sub(qxs12, qxh12);
+            vector signed char q3x13 = vec_sub(qxs13, qxh13);
+
+            vector signed char q8y00 = vec_xl(  0, q8);
+            vector signed char q8y10 = vec_xl( 16, q8);
+            vector signed char q8y01 = vec_xl( 32, q8);
+            vector signed char q8y11 = vec_xl( 48, q8);
+            vector signed char q8y02 = vec_xl( 64, q8);
+            vector signed char q8y12 = vec_xl( 80, q8);
+            vector signed char q8y03 = vec_xl( 96, q8);
+            vector signed char q8y13 = vec_xl(112, q8);
+            q8 += 128;
+
+            vector signed short vscales_h = vec_unpackh(vscales);
+            vector signed short vs0 = vec_splat(vscales_h, 0);
+            vector signed short vs1 = vec_splat(vscales_h, 1);
+            vector signed short vs2 = vec_splat(vscales_h, 2);
+            vector signed short vs3 = vec_splat(vscales_h, 3);
+            vector signed short vs4 = vec_splat(vscales_h, 4);
+            vector signed short vs5 = vec_splat(vscales_h, 5);
+            vector signed short vs6 = vec_splat(vscales_h, 6);
+            vector signed short vs7 = vec_splat(vscales_h, 7);
+            vscales = vec_sld(vscales, vscales, 8);
+
+            vector signed short qv00 = vec_add(vec_mule(q3x00, q8y00), vec_mulo(q3x00, q8y00));
+            vector signed short qv01 = vec_add(vec_mule(q3x01, q8y01), vec_mulo(q3x01, q8y01));
+            vector signed short qv02 = vec_add(vec_mule(q3x02, q8y02), vec_mulo(q3x02, q8y02));
+            vector signed short qv03 = vec_add(vec_mule(q3x03, q8y03), vec_mulo(q3x03, q8y03));
+            vector signed short qv10 = vec_add(vec_mule(q3x10, q8y10), vec_mulo(q3x10, q8y10));
+            vector signed short qv11 = vec_add(vec_mule(q3x11, q8y11), vec_mulo(q3x11, q8y11));
+            vector signed short qv12 = vec_add(vec_mule(q3x12, q8y12), vec_mulo(q3x12, q8y12));
+            vector signed short qv13 = vec_add(vec_mule(q3x13, q8y13), vec_mulo(q3x13, q8y13));
+
+            vsumi0 = vec_msum(qv00, vs0, vsumi0);
+            vsumi1 = vec_msum(qv01, vs2, vsumi1);
+            vsumi2 = vec_msum(qv02, vs4, vsumi2);
+            vsumi3 = vec_msum(qv03, vs6, vsumi3);
+            vsumi4 = vec_msum(qv10, vs1, vsumi4);
+            vsumi5 = vec_msum(qv11, vs3, vsumi5);
+            vsumi6 = vec_msum(qv12, vs5, vsumi6);
+            vsumi7 = vec_msum(qv13, vs7, vsumi7);
+        }
+
+        vsumi0 = vec_add(vsumi0, vsumi4);
+        vsumi1 = vec_add(vsumi1, vsumi5);
+        vsumi2 = vec_add(vsumi2, vsumi6);
+        vsumi3 = vec_add(vsumi3, vsumi7);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined __loongarch_asx
+
+    const __m256i m3 = __lasx_xvreplgr2vr_b(3);
+    const __m256i mone = __lasx_xvreplgr2vr_b(1);
+    const __m128i m32 = __lsx_vreplgr2vr_b(32);
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    uint32_t aux[3];
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const uint8_t * restrict q3 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+        // Set up scales
+        memcpy(aux, x[i].scales, 12);
+        __m128i scales128 = lsx_set_w(
+                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
+                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
+                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
+                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
+        scales128 = __lsx_vsub_b(scales128, m32);
+        const __m256i all_scales = lasx_ext8_16(scales128);
+        const __m128i l_scales = lasx_extracti128(all_scales, 0);
+        const __m128i h_scales = lasx_extracti128(all_scales, 1);
+        const __m256i scales[2] = {lasx_insertf128(l_scales, l_scales), lasx_insertf128(h_scales, h_scales)};
+
+        // high bit
+        const __m256i hbits = __lasx_xvld((const __m256i*)x[i].hmask, 0);
+
+        // integer accumulator
+        __m256i sumi = __lasx_xvldi(0);
+
+        int bit = 0;
+        int is  = 0;
+        __m256i xvbit;
+
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            // load low 2 bits
+            const __m256i q3bits = __lasx_xvld((const __m256i*)q3, 0); q3 += 32;
+
+            xvbit = __lasx_xvreplgr2vr_h(bit);
+            // prepare low and high bits
+            const __m256i q3l_0 = __lasx_xvand_v(q3bits, m3);
+            const __m256i q3h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
+            ++bit;
+
+            xvbit = __lasx_xvreplgr2vr_h(bit);
+            const __m256i q3l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 2), m3);
+            const __m256i q3h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
+            ++bit;
+
+            xvbit = __lasx_xvreplgr2vr_h(bit);
+            const __m256i q3l_2 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 4), m3);
+            const __m256i q3h_2 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
+            ++bit;
+
+            xvbit = __lasx_xvreplgr2vr_h(bit);
+            const __m256i q3l_3 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 6), m3);
+            const __m256i q3h_3 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
+            ++bit;
+
+            // load Q8 quants
+            const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+
+            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use lasx_maddubs_h,
+            // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
+            // and 2 if the high bit was set)
+            __m256i q8s_0 = lasx_maddubs_h(q3h_0, q8_0);
+            __m256i q8s_1 = lasx_maddubs_h(q3h_1, q8_1);
+            __m256i q8s_2 = lasx_maddubs_h(q3h_2, q8_2);
+            __m256i q8s_3 = lasx_maddubs_h(q3h_3, q8_3);
+
+            __m256i p16_0 = lasx_maddubs_h(q3l_0, q8_0);
+            __m256i p16_1 = lasx_maddubs_h(q3l_1, q8_1);
+            __m256i p16_2 = lasx_maddubs_h(q3l_2, q8_2);
+            __m256i p16_3 = lasx_maddubs_h(q3l_3, q8_3);
+
+            p16_0 = __lasx_xvsub_h(p16_0, q8s_0);
+            p16_1 = __lasx_xvsub_h(p16_1, q8s_1);
+            p16_2 = __lasx_xvsub_h(p16_2, q8s_2);
+            p16_3 = __lasx_xvsub_h(p16_3, q8s_3);
+
+            // multiply with scales
+            p16_0 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 0)), p16_0);
+            p16_1 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 1)), p16_1);
+            p16_2 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 2)), p16_2);
+            p16_3 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 3)), p16_3);
+
+            // accumulate
+            p16_0 = __lasx_xvadd_w(p16_0, p16_1);
+            p16_2 = __lasx_xvadd_w(p16_2, p16_3);
+            sumi  = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_2));
+        }
+        // multiply with block scale and accumulate
+        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);//FIXME
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict hm = x[i].hmask;
+        const  int8_t * restrict q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * restrict a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            q3 += 32;
+        }
+        a = aux8;
+
+        memcpy(auxs, x[i].scales, 12);
+        uint32_t tmp = auxs[2];
+        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
+        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
+        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
+        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+
+#endif
+
+}
+
+void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#ifdef __ARM_NEON
+    const uint8x16_t m4b = vdupq_n_u8(0xf);
+    const int32x4_t mzero = vdupq_n_s32(0);
+
+    ggml_int8x16x2_t q4bytes;
+    ggml_int8x16x2_t q8bytes;
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
+
+        memcpy(utmp, x[i].scales, 12);
+
+        uint32x2_t mins8 = { 0 };
+        mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
+        mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);
+
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[0] &= kmask1;
+
+        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins8)));
+        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
+                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
+        sumf -= dmin * vaddvq_s32(prod);
+
+        const uint8_t * scales = (const uint8_t *)utmp;
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        int32_t sumi1 = 0;
+        int32_t sumi2 = 0;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); q4 += 32;
+
+            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
+            q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[0], m4b));
+            q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[1], m4b));
+
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
+            sumi1 += vaddvq_s32(p1) * scales[2*j+0];
+
+            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
+            q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
+            q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
+
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
+
+            sumi2 += vaddvq_s32(p2) * scales[2*j+1];
+        }
+
+        sumf += d * (sumi1 + sumi2);
+
+    }
+
+    *s = sumf;
+
+#elif defined __AVX2__
+
+    const __m256i m4 = _mm256_set1_epi8(0xF);
+
+    __m256 acc = _mm256_setzero_ps();
+    __m128 acc_m = _mm_setzero_ps();
+
+   for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
+
+        const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
+        const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
+        const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
+        acc_m = _mm_fmadd_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod), acc_m);
+
+        const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
+        const __m256i scales = MM256_SET_M128I(sc128, sc128);
+
+        __m256i sumi = _mm256_setzero_si256();
+
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m256i scale_l = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
+            const __m256i scale_h = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
+
+            const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
+            const __m256i q4l = _mm256_and_si256(q4bits, m4);
+            const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(q4bits, 4), m4);
+
+            const __m256i q8l = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            __m256i p16l = _mm256_maddubs_epi16(q4l, q8l);
+            p16l = _mm256_madd_epi16(scale_l, p16l);
+
+            const __m256i q8h = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            __m256i p16h = _mm256_maddubs_epi16(q4h, q8h);
+            p16h = _mm256_madd_epi16(scale_h, p16h);
+            const __m256i sumj = _mm256_add_epi32(p16l, p16h);
+
+            sumi = _mm256_add_epi32(sumi, sumj);
+        }
+
+        __m256 vd = _mm256_set1_ps(d);
+        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
+
+    }
+
+    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
+    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
+
+    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
+
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i m2 = _mm_set1_epi8(0x2);
+
+    __m256 acc = _mm256_setzero_ps();
+    __m128 acc_m = _mm_setzero_ps();
+
+   for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
+        const __m128i scales = _mm_cvtepu8_epi16(utmps);
+        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
+
+        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
+        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
+        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
+        const __m128i prod = _mm_madd_epi16(mins, q8s);
+        acc_m = _mm_add_ps(_mm_mul_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod)), acc_m);
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        __m128i shuffle = _mm_set1_epi16(0x0100);
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m128i scale_l = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            const __m128i scale_h = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+
+            __m128i q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4l_0 = _mm_and_si128(q4bits, m4);
+            const __m128i q4h_0 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
+            q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4l_1 = _mm_and_si128(q4bits, m4);
+            const __m128i q4h_1 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
+
+            const __m128i q8l_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16l = _mm_maddubs_epi16(q4l_0, q8l_0);
+            p16l = _mm_madd_epi16(scale_l, p16l);
+            sumi_0 = _mm_add_epi32(sumi_0, p16l);
+            const __m128i q8l_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            p16l = _mm_maddubs_epi16(q4l_1, q8l_1);
+            p16l = _mm_madd_epi16(scale_l, p16l);
+            sumi_1 = _mm_add_epi32(sumi_1, p16l);
+
+            const __m128i q8h_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16h = _mm_maddubs_epi16(q4h_0, q8h_0);
+            p16h = _mm_madd_epi16(scale_h, p16h);
+            sumi_0 = _mm_add_epi32(sumi_0, p16h);
+            const __m128i q8h_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            p16h = _mm_maddubs_epi16(q4h_1, q8h_1);
+            p16h = _mm_madd_epi16(scale_h, p16h);
+            sumi_1 = _mm_add_epi32(sumi_1, p16h);
+
+        }
+
+        __m256 vd = _mm256_set1_ps(d);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
+
+    }
+
+    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
+    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
+
+    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
+
+#elif defined __riscv_v_intrinsic
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        size_t vl = 8;
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
+        vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
+        vint16mf2_t q8sums   = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        vuint8mf4_t mins8  = __riscv_vle8_v_u8mf4(mins, vl);
+        vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl));
+        vint32m1_t  prod   = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl);
+
+        vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
+        sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        vl = 32;
+
+        int32_t sum_1 = 0;
+        int32_t sum_2 = 0;
+
+        vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1);
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            // load Q4
+            vuint8m1_t q4_x = __riscv_vle8_v_u8m1(q4, vl);
+
+            // load Q8 and multiply it with lower Q4 nibble
+            vint8m1_t  q8_0 = __riscv_vle8_v_i8m1(q8, vl);
+            vint8m1_t  q4_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q4_x, 0x0F, vl));
+            vint16m2_t qv_0 = __riscv_vwmul_vv_i16m2(q4_0, q8_0, vl);
+            vint16m1_t vs_0 = __riscv_vredsum_vs_i16m2_i16m1(qv_0, vzero, vl);
+
+            sum_1 += __riscv_vmv_x_s_i16m1_i16(vs_0) * scales[2*j+0];
+
+            // load Q8 and multiply it with upper Q4 nibble
+            vint8m1_t  q8_1 = __riscv_vle8_v_i8m1(q8+32, vl);
+            vint8m1_t  q4_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q4_x, 0x04, vl));
+            vint16m2_t qv_1 = __riscv_vwmul_vv_i16m2(q4_1, q8_1, vl);
+            vint16m1_t vs_1 = __riscv_vredsum_vs_i16m2_i16m1(qv_1, vzero, vl);
+
+            sum_2 += __riscv_vmv_x_s_i16m1_i16(vs_1) * scales[2*j+1];
+
+            q4 += 32;    q8 += 64;
+
+        }
+
+        sumf += d*(sum_1 + sum_2);
+
+    }
+
+    *s = sumf;
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed char lowMask1 = vec_splats((int8_t)0x3f);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v2 = vec_splats((uint8_t)2);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vdmin = vec_mul(vxmin, vyd);
+
+        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
+        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
+
+        UNUSED(kmask1);
+        UNUSED(kmask2);
+        UNUSED(kmask3);
+        UNUSED(utmp);
+
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(vec_sr(u0, v2), lowMask2);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = vec_sr(u2, v4);
+
+        vector signed char u30 = u1;
+        vector signed char u31 = (vector signed char)vec_mergeh((vector signed int)vec_and(u2, lowMask), (vector signed int)u3);
+
+        u1 = vec_and(u0, lowMask1);
+        u2 = vec_or(u30, u31);
+
+        vector signed char utmps = (vector signed char)vec_mergeh((vector signed int)u1, (vector signed int)u2);
+
+        vector signed short vscales = vec_unpackh(utmps);
+        vector signed short q4xmins = vec_unpackl(utmps);
+        vector signed short q4xmins0 = vec_mergeh(q4xmins, q4xmins);
+        vector signed short q4xmins1 = vec_mergel(q4xmins, q4xmins);
+
+        vector signed int prod0 = vec_mule(q4xmins0, q8ysums0);
+        vector signed int prod1 = vec_mule(q4xmins1, q8ysums1);
+        vector signed int prod2 = vec_mulo(q4xmins0, q8ysums0);
+        vector signed int prod3 = vec_mulo(q4xmins1, q8ysums1);
+
+        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
+        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
+        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
+        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/64; j+=2) {
+            __builtin_prefetch(q4, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q4);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q4);
+            vector signed char qxs2 = (vector signed char)vec_xl(32, q4);
+            vector signed char qxs3 = (vector signed char)vec_xl(48, q4);
+            q4 += 64;
+
+            vector unsigned char q4x00 = (vector unsigned char)vec_and(qxs0, lowMask);
+            vector unsigned char q4x01 = (vector unsigned char)vec_sr(qxs0, v4);
+            vector unsigned char q4x10 = (vector unsigned char)vec_and(qxs1, lowMask);
+            vector unsigned char q4x11 = (vector unsigned char)vec_sr(qxs1, v4);
+            vector unsigned char q4x20 = (vector unsigned char)vec_and(qxs2, lowMask);
+            vector unsigned char q4x21 = (vector unsigned char)vec_sr(qxs2, v4);
+            vector unsigned char q4x30 = (vector unsigned char)vec_and(qxs3, lowMask);
+            vector unsigned char q4x31 = (vector unsigned char)vec_sr(qxs3, v4);
+
+            vector signed char q8y00 = vec_xl(  0, q8);
+            vector signed char q8y10 = vec_xl( 16, q8);
+            vector signed char q8y01 = vec_xl( 32, q8);
+            vector signed char q8y11 = vec_xl( 48, q8);
+            vector signed char q8y20 = vec_xl( 64, q8);
+            vector signed char q8y30 = vec_xl( 80, q8);
+            vector signed char q8y21 = vec_xl( 96, q8);
+            vector signed char q8y31 = vec_xl(112, q8);
+            q8 += 128;
+
+            vector signed int qv00 = vec_msum(q8y00, q4x00, v0);
+            vector signed int qv01 = vec_msum(q8y01, q4x01, v0);
+            vector signed int qv10 = vec_msum(q8y10, q4x10, v0);
+            vector signed int qv11 = vec_msum(q8y11, q4x11, v0);
+            vector signed int qv20 = vec_msum(q8y20, q4x20, v0);
+            vector signed int qv21 = vec_msum(q8y21, q4x21, v0);
+            vector signed int qv30 = vec_msum(q8y30, q4x30, v0);
+            vector signed int qv31 = vec_msum(q8y31, q4x31, v0);
+
+            vector signed int vscales_h = vec_unpackh(vscales);
+            vector signed int vs0 = vec_splat(vscales_h, 0);
+            vector signed int vs1 = vec_splat(vscales_h, 1);
+            vector signed int vs2 = vec_splat(vscales_h, 2);
+            vector signed int vs3 = vec_splat(vscales_h, 3);
+            vscales = vec_sld(vscales, vscales, 8);
+
+            vsumi0 = vec_add(vec_mul(qv00, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv01, vs1), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv20, vs2), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv21, vs3), vsumi3);
+
+            vsumi0 = vec_add(vec_mul(qv10, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv11, vs1), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv30, vs2), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv31, vs3), vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined __loongarch_asx
+    GGML_UNUSED(kmask1);
+    GGML_UNUSED(kmask2);
+    GGML_UNUSED(kmask3);
+
+    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+    __m128 acc_m = (__m128)__lsx_vldi(0);
+
+   for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]));
+
+        const __m256i q8sums = __lasx_xvld((const __m256i*)y[i].bsums, 0);
+        const __m128i q8s = lsx_hadd_h(lasx_extracti128(q8sums, 0), lasx_extracti128(q8sums, 1));
+        const __m128i prod = lsx_madd_h(lasx_extracti128(mins_and_scales, 1), q8s);
+        acc_m = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(dmin), __lsx_vffint_s_w(prod), acc_m);
+
+        const __m128i sc128  = lasx_extracti128(mins_and_scales, 0);
+        const __m256i scales = lasx_insertf128(sc128, sc128);
+
+        __m256i sumi = __lasx_xvldi(0);
+
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m256i scale_l = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+0));
+            const __m256i scale_h = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+1));
+
+            const __m256i q4bits = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
+            const __m256i q4l = __lasx_xvand_v(q4bits, m4);
+            const __m256i q4h = __lasx_xvand_v(__lasx_xvsrli_h(q4bits, 4), m4);
+
+            const __m256i q8l = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            __m256i p16l = lasx_maddubs_h(q4l, q8l);
+            p16l = lasx_madd_h(scale_l, p16l);
+
+            const __m256i q8h = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            __m256i p16h = lasx_maddubs_h(q4h, q8h);
+            p16h = lasx_madd_h(scale_h, p16h);
+            const __m256i sumj = __lasx_xvadd_w(p16l, p16h);
+
+            sumi = __lasx_xvadd_w(sumi, sumj);
+        }
+
+        __m256 vd = __lasx_xvreplfr2vr_s(d);
+        acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc);
+
+    }
+
+    acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vpermi_w((__m128i)acc_m, (__m128i)acc_m, 0xee));
+    __m128i tmp1 = __lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w((__m128i)acc_m, 1), 0);
+    acc_m = __lsx_vfadd_s(acc_m, (__m128)tmp1);
+
+
+    ft_union fi;
+    fi.i = __lsx_vpickve2gr_w(acc_m, 0);
+    *s = hsum_float_8(acc) + fi.f ;
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q4 = x[i].qs;
+        const  int8_t * restrict q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * restrict a = aux8;
+        for (int j = 0; j < QK_K/64; ++j) {
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
+            a += 32;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            a += 32; q4 += 32;
+        }
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy,  size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#ifdef __ARM_NEON
+    const uint8x16_t m4b = vdupq_n_u8(0xf);
+    const uint8x16_t mone = vdupq_n_u8(1);
+    const uint8x16_t mtwo = vdupq_n_u8(2);
+    const int32x4_t mzero = vdupq_n_s32(0);
+
+    ggml_int8x16x4_t q5bytes;
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const uint8x8_t mins8 = vld1_u8((const uint8_t*)utmp + 8);
+        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(mins8));
+        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
+                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
+        int32_t sumi_mins = vaddvq_s32(prod);
+
+        const uint8_t * scales = (const uint8_t *)utmp;
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh);
+
+        ggml_uint8x16x4_t q5h;
+
+        int32_t sumi = 0;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const ggml_uint8x16x2_t q5bits = ggml_vld1q_u8_x2(q5); q5 += 32;
+            const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            q5h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
+            q5h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
+            q5h.val[2] = vshlq_n_u8(vandq_u8(mtwo, qhbits.val[0]), 3);
+            q5h.val[3] = vshlq_n_u8(vandq_u8(mtwo, qhbits.val[1]), 3);
+            qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 2);
+            qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 2);
+
+            q5bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.val[0], m4b), q5h.val[0]));
+            q5bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.val[1], m4b), q5h.val[1]));
+            q5bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[0], 4), q5h.val[2]));
+            q5bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[1], 4), q5h.val[3]));
+
+            sumi += vaddvq_s32(ggml_vdotq_s32(ggml_vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0]), q5bytes.val[1], q8bytes.val[1])) * *scales++;
+            sumi += vaddvq_s32(ggml_vdotq_s32(ggml_vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2]), q5bytes.val[3], q8bytes.val[3])) * *scales++;
+        }
+
+        sumf += d * sumi - dmin * sumi_mins;
+    }
+
+    *s = sumf;
+
+#elif defined __AVX2__
+
+    const __m256i m4 = _mm256_set1_epi8(0xF);
+    const __m128i mzero = _mm_setzero_si128();
+    const __m256i mone  = _mm256_set1_epi8(1);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0.f;
+
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q5 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
+
+        const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
+        const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
+        const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
+        const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
+        summs += dmin * _mm_extract_epi32(hsum, 0);
+
+        const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
+        const __m256i scales = MM256_SET_M128I(sc128, sc128);
+
+        const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].qh);
+        __m256i hmask = mone;
+
+        __m256i sumi = _mm256_setzero_si256();
+
+        int bit = 0;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m256i scale_0 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
+            const __m256i scale_1 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
+
+            const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5); q5 += 32;
+
+            const __m256i q5l_0 = _mm256_and_si256(q5bits, m4);
+            const __m256i q5h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
+            const __m256i q5_0  = _mm256_add_epi8(q5l_0, q5h_0);
+            hmask = _mm256_slli_epi16(hmask, 1);
+
+            const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4);
+            const __m256i q5h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
+            const __m256i q5_1  = _mm256_add_epi8(q5l_1, q5h_1);
+            hmask = _mm256_slli_epi16(hmask, 1);
+
+            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            __m256i p16_0 = _mm256_maddubs_epi16(q5_0, q8_0);
+            __m256i p16_1 = _mm256_maddubs_epi16(q5_1, q8_1);
+
+            p16_0 = _mm256_madd_epi16(scale_0, p16_0);
+            p16_1 = _mm256_madd_epi16(scale_1, p16_1);
+
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
+
+        }
+
+        __m256 vd = _mm256_set1_ps(d);
+        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc) + summs;
+
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i mzero = _mm_setzero_si128();
+    const __m128i mone  = _mm_set1_epi8(1);
+    const __m128i m2 = _mm_set1_epi8(2);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0.f;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
+        const __m128i scales = _mm_cvtepu8_epi16(utmps);
+        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
+
+        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
+        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
+        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
+        const __m128i prod = _mm_madd_epi16(mins, q8s);
+        const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
+        summs += dmin * _mm_extract_epi32(hsum, 0);
+
+        const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].qh[0]);
+        const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].qh[16]);
+        __m128i hmask = mone;
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        int bit = 0;
+
+        __m128i shuffle = _mm_set1_epi16(0x0100);
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m128i scale_0 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            const __m128i scale_1 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+
+            const __m128i q5bits_0 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
+            const __m128i q5bits_1 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
+
+            __m128i q5l_0 = _mm_and_si128(q5bits_0, m4);
+            __m128i q5l_1 = _mm_and_si128(q5bits_1, m4);
+            __m128i q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
+            __m128i q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
+            __m128i q5_0  = _mm_add_epi8(q5l_0, q5h_0);
+            __m128i q5_1  = _mm_add_epi8(q5l_1, q5h_1);
+            hmask = _mm_slli_epi16(hmask, 1);
+
+            __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16_0 = _mm_maddubs_epi16(q5_0, q8_0);
+            __m128i p16_1 = _mm_maddubs_epi16(q5_1, q8_1);
+            p16_0 = _mm_madd_epi16(scale_0, p16_0);
+            p16_1 = _mm_madd_epi16(scale_0, p16_1);
+
+            q5l_0 = _mm_and_si128(_mm_srli_epi16(q5bits_0, 4), m4);
+            q5l_1 = _mm_and_si128(_mm_srli_epi16(q5bits_1, 4), m4);
+            q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
+            q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
+            q5_0  = _mm_add_epi8(q5l_0, q5h_0);
+            q5_1  = _mm_add_epi8(q5l_1, q5h_1);
+            hmask = _mm_slli_epi16(hmask, 1);
+
+            q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16_2 = _mm_maddubs_epi16(q5_0, q8_0);
+            __m128i p16_3 = _mm_maddubs_epi16(q5_1, q8_1);
+            p16_2 = _mm_madd_epi16(scale_1, p16_2);
+            p16_3 = _mm_madd_epi16(scale_1, p16_3);
+
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
+
+        }
+
+        __m256 vd = _mm256_set1_ps(d);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
+
+    }
+
+    *s = hsum_float_8(acc) + summs;
+
+#elif defined __riscv_v_intrinsic
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    float sumf = 0;
+    float sums = 0.0;
+
+    size_t vl;
+
+    for (int i = 0; i < nb; ++i) {
+
+        vl = 8;
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const uint8_t * restrict hm = x[i].qh;
+        const  int8_t * restrict q8 = y[i].qs;
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+
+        vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
+        vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
+        vint16mf2_t q8sums = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        vuint8mf4_t mins8 = __riscv_vle8_v_u8mf4(mins, vl);
+        vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl));
+        vint32m1_t prod = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl);
+
+        vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
+        sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
+
+        vl = 32;
+        int32_t aux32 = 0;
+        int is = 0;
+
+        uint8_t m = 1;
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+        vuint8m1_t vqh = __riscv_vle8_v_u8m1(hm, vl);
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            // load Q5 and Q8
+            vuint8m1_t q5_x = __riscv_vle8_v_u8m1(q5, vl);
+            vint8m1_t  q8_y1 = __riscv_vle8_v_i8m1(q8, vl);
+            vint8m1_t  q8_y2 = __riscv_vle8_v_i8m1(q8+32, vl);
+
+            // compute mask for addition
+            vint8m1_t q5_a = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q5_x, 0x0F, vl));
+            vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_1 = __riscv_vmsne_vx_u8m1_b8(qh_m1, 0, vl);
+            vint8m1_t q5_m1 = __riscv_vadd_vx_i8m1_mu(vmask_1, q5_a, q5_a, 16, vl);
+            m <<= 1;
+
+            vint8m1_t q5_l = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q5_x, 0x04, vl));
+            vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_2 = __riscv_vmsne_vx_u8m1_b8(qh_m2, 0, vl);
+            vint8m1_t q5_m2 = __riscv_vadd_vx_i8m1_mu(vmask_2, q5_l, q5_l, 16, vl);
+            m <<= 1;
+
+            vint16m2_t v0 = __riscv_vwmul_vv_i16m2(q5_m1, q8_y1, vl);
+            vint16m2_t v1 = __riscv_vwmul_vv_i16m2(q5_m2, q8_y2, vl);
+
+            vint32m4_t vs1 = __riscv_vwmul_vx_i32m4(v0, scales[is++], vl);
+            vint32m4_t vs2 = __riscv_vwmul_vx_i32m4(v1, scales[is++], vl);
+
+            vint32m1_t vacc1 = __riscv_vredsum_vs_i32m4_i32m1(vs1, vzero, vl);
+            vint32m1_t vacc2 = __riscv_vredsum_vs_i32m4_i32m1(vs2, vzero, vl);
+
+            aux32 += __riscv_vmv_x_s_i32m1_i32(vacc1) + __riscv_vmv_x_s_i32m1_i32(vacc2);
+            q5 += 32;    q8 += 64;
+
+        }
+
+        vfloat32m1_t vaux = __riscv_vfmul_vf_f32m1(__riscv_vfmv_v_f_f32m1(aux32, 1), d, 1);
+        sums += __riscv_vfmv_f_s_f32m1_f32(vaux);
+
+    }
+
+    *s = sumf+sums;
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed char lowMask1 = vec_splats((int8_t)0x3f);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v1 = vec_splats((unsigned char)0x1);
+    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
+    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vdmin = vec_mul(vxmin, vyd);
+
+        UNUSED(kmask1);
+        UNUSED(kmask2);
+        UNUSED(kmask3);
+        UNUSED(utmp);
+
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(vec_sr(u0, v2), lowMask2);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = vec_sr(u2, v4);
+
+        vector signed char u30 = u1;
+        vector signed char u31 = (vector signed char)vec_mergeh((vector signed int)vec_and(u2, lowMask), (vector signed int)u3);
+
+        u1 = vec_and(u0, lowMask1);
+        u2 = vec_or(u30, u31);
+
+        vector signed char utmps = (vector signed char)vec_mergeh((vector signed int)u1, (vector signed int)u2);
+
+        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
+        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
+
+        vector signed short vscales = vec_unpackh(utmps);
+
+        vector signed short q5xmins = vec_unpackl(utmps);
+        vector signed short q5xmins0 = vec_mergeh(q5xmins, q5xmins);
+        vector signed short q5xmins1 = vec_mergel(q5xmins, q5xmins);
+
+        vector signed int prod0 = vec_mule(q5xmins0, q8ysums0);
+        vector signed int prod1 = vec_mule(q5xmins1, q8ysums1);
+        vector signed int prod2 = vec_mulo(q5xmins0, q8ysums0);
+        vector signed int prod3 = vec_mulo(q5xmins1, q8ysums1);
+
+        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
+        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
+        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
+        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
+
+        vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].qh);
+        vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].qh);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            __builtin_prefetch(q5, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q5);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q5);
+            q5 += 32;
+
+            vector signed char qxs00 = vec_and(qxs0, lowMask);
+            vector signed char qxs01 = vec_sr(qxs0, v4);
+            vector signed char qxs10 = vec_and(qxs1, lowMask);
+            vector signed char qxs11 = vec_sr(qxs1, v4);
+
+            vector signed char q5h00 = vec_sl(vec_and((vector signed char)v1, qxhs0), v4);
+            vector signed char q5h01 = vec_sl(vec_and((vector signed char)v2, qxhs0), v3);
+            vector signed char q5h10 = vec_sl(vec_and((vector signed char)v1, qxhs1), v4);
+            vector signed char q5h11 = vec_sl(vec_and((vector signed char)v2, qxhs1), v3);
+            qxhs0 = vec_sr(qxhs0, v2);
+            qxhs1 = vec_sr(qxhs1, v2);
+
+            vector unsigned char q5x00 = (vector unsigned char)vec_or(q5h00, qxs00);
+            vector unsigned char q5x01 = (vector unsigned char)vec_or(q5h01, qxs01);
+            vector unsigned char q5x10 = (vector unsigned char)vec_or(q5h10, qxs10);
+            vector unsigned char q5x11 = (vector unsigned char)vec_or(q5h11, qxs11);
+
+            vector signed char q8y00 = vec_xl( 0, q8);
+            vector signed char q8y10 = vec_xl(16, q8);
+            vector signed char q8y01 = vec_xl(32, q8);
+            vector signed char q8y11 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed int qv00 = vec_msum(q8y00, q5x00, v0);
+            vector signed int qv01 = vec_msum(q8y01, q5x01, v0);
+            vector signed int qv10 = vec_msum(q8y10, q5x10, v0);
+            vector signed int qv11 = vec_msum(q8y11, q5x11, v0);
+
+            vector signed int vscales_h = vec_unpackh(vscales);
+            vector signed int vs0 = vec_splat(vscales_h, 0);
+            vector signed int vs1 = vec_splat(vscales_h, 1);
+            vscales = vec_sld(vscales, vscales, 12);
+
+            vsumi0 = vec_add(vec_mul(qv00, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv10, vs0), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv01, vs1), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv11, vs1), vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined __loongarch_asx
+    GGML_UNUSED(kmask1);
+    GGML_UNUSED(kmask2);
+    GGML_UNUSED(kmask3);
+
+    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
+    const __m128i mzero = __lsx_vldi(0);
+    const __m256i mone  = __lasx_xvreplgr2vr_b(1);
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    float summs = 0.f;
+
+   for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]));
+
+        const __m256i q8sums = __lasx_xvld((const __m256i*)y[i].bsums, 0);
+        const __m128i q8s = lsx_hadd_h(lasx_extracti128(q8sums, 0), lasx_extracti128(q8sums, 1));
+        const __m128i prod = lsx_madd_h(lasx_extracti128(mins_and_scales, 1), q8s);
+        const __m128i hsum = lsx_hadd_w(lsx_hadd_w(prod, mzero), mzero);
+        summs += dmin * __lsx_vpickve2gr_w(hsum, 0);    //TODO check
+
+        const __m128i sc128  = lasx_extracti128(mins_and_scales, 0);
+        const __m256i scales = lasx_insertf128(sc128, sc128);
+
+        const __m256i hbits = __lasx_xvld((const __m256i*)x[i].qh, 0);
+        __m256i hmask = mone;
+
+        __m256i sumi = __lasx_xvldi(0);
+
+        int bit = 0;
+        __m256i xvbit;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m256i scale_0 = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+0));
+            const __m256i scale_1 = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+1));
+
+            const __m256i q5bits = __lasx_xvld((const __m256i*)q5, 0); q5 += 32;
+
+            xvbit = __lasx_xvreplgr2vr_h(bit++);
+            const __m256i q5l_0 = __lasx_xvand_v(q5bits, m4);
+            const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4);
+            const __m256i q5_0  = __lasx_xvadd_b(q5l_0, q5h_0);
+            hmask = __lasx_xvslli_h(hmask, 1);
+
+            xvbit = __lasx_xvreplgr2vr_h(bit++);
+            const __m256i q5l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q5bits, 4), m4);
+            const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4);
+            const __m256i q5_1  = __lasx_xvadd_b(q5l_1, q5h_1);
+            hmask = __lasx_xvslli_h(hmask, 1);
+
+            const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+
+            __m256i p16_0 = lasx_maddubs_h(q5_0, q8_0);
+            __m256i p16_1 = lasx_maddubs_h(q5_1, q8_1);
+
+            p16_0 = lasx_madd_h(scale_0, p16_0);
+            p16_1 = lasx_madd_h(scale_1, p16_1);
+
+            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1));
+
+        }
+
+        __m256 vd = __lasx_xvreplfr2vr_s(d);
+        acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc) + summs;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q4 = x[i].qs;
+        const uint8_t * restrict hm = x[i].qh;
+        const  int8_t * restrict q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * restrict a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K/64; ++j) {
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            q4 += 32;
+        }
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q6_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#ifdef __ARM_NEON
+    float sum = 0;
+
+    const uint8x16_t m4b = vdupq_n_u8(0xF);
+    const int32x4_t  vzero = vdupq_n_s32(0);
+    //const int8x16_t  m32s = vdupq_n_s8(32);
+
+    const uint8x16_t mone = vdupq_n_u8(3);
+
+    ggml_int8x16x4_t q6bytes;
+    ggml_uint8x16x4_t q6h;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q6 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const int8_t * restrict scale = x[i].scales;
+
+        const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
+        const int8x16_t scales = vld1q_s8(scale);
+        const ggml_int16x8x2_t q6scales = {{vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}};
+
+        const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])),
+                                                   vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))),
+                                         vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[1]), vget_low_s16 (q6scales.val[1])),
+                                                   vmull_s16(vget_high_s16(q8sums.val[1]), vget_high_s16(q6scales.val[1]))));
+        int32_t isum_mins = vaddvq_s32(prod);
+
+        int32_t isum = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); qh += 32;
+            ggml_uint8x16x4_t q6bits = ggml_vld1q_u8_x4(q6); q6 += 64;
+            ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
+            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
+            uint8x16_t shifted = vshrq_n_u8(qhbits.val[0], 2);
+            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+            shifted = vshrq_n_u8(qhbits.val[1], 2);
+            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+
+            //q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0])), m32s);
+            //q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1])), m32s);
+            //q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2])), m32s);
+            //q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3])), m32s);
+            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0]));
+            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1]));
+            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2]));
+            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3]));
+
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
+
+            scale += 4;
+
+            q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            shifted = vshrq_n_u8(qhbits.val[0], 4);
+            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+            shifted = vshrq_n_u8(qhbits.val[1], 4);
+            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+            shifted = vshrq_n_u8(qhbits.val[0], 6);
+            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+            shifted = vshrq_n_u8(qhbits.val[1], 6);
+            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+
+            //q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0])), m32s);
+            //q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1])), m32s);
+            //q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2])), m32s);
+            //q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3])), m32s);
+            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0]));
+            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1]));
+            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2]));
+            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3]));
+
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
+            scale += 4;
+        }
+        //sum += isum * d_all * y[i].d;
+        sum += d_all * y[i].d * (isum - 32 * isum_mins);
+
+    }
+    *s = sum;
+
+#elif defined __AVX2__
+
+    const __m256i m4 = _mm256_set1_epi8(0xF);
+    const __m256i m2 = _mm256_set1_epi8(3);
+    const __m256i m32s = _mm256_set1_epi8(32);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+
+        __m256i sumi = _mm256_setzero_si256();
+
+        int is = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
+            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
+            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
+            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
+            is += 4;
+
+            const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
+            const __m256i q4bits2 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
+            const __m256i q4bitsH = _mm256_loadu_si256((const __m256i*)qh); qh += 32;
+
+            const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, m2), 4);
+            const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 2), m2), 4);
+            const __m256i q4h_2 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 4), m2), 4);
+            const __m256i q4h_3 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 6), m2), 4);
+
+            const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
+            const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(q4bits2, m4), q4h_1);
+            const __m256i q4_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_2);
+            const __m256i q4_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits2, 4), m4), q4h_3);
+
+            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            __m256i q8s_0 = _mm256_maddubs_epi16(m32s, q8_0);
+            __m256i q8s_1 = _mm256_maddubs_epi16(m32s, q8_1);
+            __m256i q8s_2 = _mm256_maddubs_epi16(m32s, q8_2);
+            __m256i q8s_3 = _mm256_maddubs_epi16(m32s, q8_3);
+
+            __m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
+            __m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
+            __m256i p16_2 = _mm256_maddubs_epi16(q4_2, q8_2);
+            __m256i p16_3 = _mm256_maddubs_epi16(q4_3, q8_3);
+
+            p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
+            p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
+            p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
+            p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
+
+            p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
+            p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
+            p16_2 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_2), p16_2);
+            p16_3 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_3), p16_3);
+
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_2, p16_3));
+
+        }
+
+        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(3);
+    const __m128i m15 = _mm_set1_epi8(15);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        // handle the q6_k -32 offset separately using bsums
+        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)y[i].bsums);
+        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)y[i].bsums + 1);
+        const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+        const __m128i scales_16_0 = _mm_cvtepi8_epi16(scales);
+        const __m128i scales_16_1 = _mm_cvtepi8_epi16(_mm_bsrli_si128(scales, 8));
+        const __m128i q8sclsub_0 = _mm_slli_epi32(_mm_madd_epi16(q8sums_0, scales_16_0), 5);
+        const __m128i q8sclsub_1 = _mm_slli_epi32(_mm_madd_epi16(q8sums_1, scales_16_1), 5);
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        int is = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m128i q4bitsH_0 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
+            const __m128i q4bitsH_1 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
+
+            const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, m3), 4);
+            const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, m3), 4);
+            const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(12)), 2);
+            const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(12)), 2);
+            const __m128i q4h_4 = _mm_and_si128(q4bitsH_0, _mm_set1_epi8(48));
+            const __m128i q4h_5 = _mm_and_si128(q4bitsH_1, _mm_set1_epi8(48));
+            const __m128i q4h_6 = _mm_srli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(-64)), 2);
+            const __m128i q4h_7 = _mm_srli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(-64)), 2);
+
+            const __m128i q4bits1_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4bits1_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4bits2_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4bits2_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+
+            const __m128i q4_0 = _mm_or_si128(_mm_and_si128(q4bits1_0, m15), q4h_0);
+            const __m128i q4_1 = _mm_or_si128(_mm_and_si128(q4bits1_1, m15), q4h_1);
+            const __m128i q4_2 = _mm_or_si128(_mm_and_si128(q4bits2_0, m15), q4h_2);
+            const __m128i q4_3 = _mm_or_si128(_mm_and_si128(q4bits2_1, m15), q4h_3);
+            const __m128i q4_4 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_0, 4), m15), q4h_4);
+            const __m128i q4_5 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_1, 4), m15), q4h_5);
+            const __m128i q4_6 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_0, 4), m15), q4h_6);
+            const __m128i q4_7 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_1, 4), m15), q4h_7);
+
+            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+            __m128i p16_0 = _mm_maddubs_epi16(q4_0, q8_0);
+            __m128i p16_1 = _mm_maddubs_epi16(q4_1, q8_1);
+            __m128i p16_2 = _mm_maddubs_epi16(q4_2, q8_2);
+            __m128i p16_3 = _mm_maddubs_epi16(q4_3, q8_3);
+            __m128i p16_4 = _mm_maddubs_epi16(q4_4, q8_4);
+            __m128i p16_5 = _mm_maddubs_epi16(q4_5, q8_5);
+            __m128i p16_6 = _mm_maddubs_epi16(q4_6, q8_6);
+            __m128i p16_7 = _mm_maddubs_epi16(q4_7, q8_7);
+
+            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
+            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
+            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
+            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
+            is += 4;
+
+            p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
+            p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_0, 8)), p16_1);
+            p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
+            p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_1, 8)), p16_3);
+            p16_4 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_2), p16_4);
+            p16_5 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_2, 8)), p16_5);
+            p16_6 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_3), p16_6);
+            p16_7 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_3, 8)), p16_7);
+
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_4, p16_6));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_5, p16_7));
+
+        }
+
+        sumi_0 = _mm_sub_epi32(sumi_0, q8sclsub_0);
+        sumi_1 = _mm_sub_epi32(sumi_1, q8sclsub_1);
+        const __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi)), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __riscv_v_intrinsic
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+
+        const uint8_t * restrict q6 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const  int8_t * restrict q8 = y[i].qs;
+
+        const int8_t * restrict scale = x[i].scales;
+
+        size_t vl;
+
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+
+        int sum_t = 0;
+        int is = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            vl = 32;
+
+            // load qh
+            vuint8m1_t qh_x = __riscv_vle8_v_u8m1(qh, vl);
+
+            // load Q6
+            vuint8m1_t q6_0 = __riscv_vle8_v_u8m1(q6, vl);
+            vuint8m1_t q6_1 = __riscv_vle8_v_u8m1(q6+32, vl);
+
+            vuint8m1_t q6a_0 = __riscv_vand_vx_u8m1(q6_0, 0x0F, vl);
+            vuint8m1_t q6a_1 = __riscv_vand_vx_u8m1(q6_1, 0x0F, vl);
+            vuint8m1_t q6s_0 = __riscv_vsrl_vx_u8m1(q6_0, 0x04, vl);
+            vuint8m1_t q6s_1 = __riscv_vsrl_vx_u8m1(q6_1, 0x04, vl);
+
+            vuint8m1_t qh_0 = __riscv_vand_vx_u8m1(qh_x, 0x03, vl);
+            vuint8m1_t qh_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x2, vl), 0x03 , vl);
+            vuint8m1_t qh_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x4, vl), 0x03 , vl);
+            vuint8m1_t qh_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x6, vl), 0x03 , vl);
+
+            vuint8m1_t qhi_0 = __riscv_vor_vv_u8m1(q6a_0, __riscv_vsll_vx_u8m1(qh_0, 0x04, vl), vl);
+            vuint8m1_t qhi_1 = __riscv_vor_vv_u8m1(q6a_1, __riscv_vsll_vx_u8m1(qh_1, 0x04, vl), vl);
+            vuint8m1_t qhi_2 = __riscv_vor_vv_u8m1(q6s_0, __riscv_vsll_vx_u8m1(qh_2, 0x04, vl), vl);
+            vuint8m1_t qhi_3 = __riscv_vor_vv_u8m1(q6s_1, __riscv_vsll_vx_u8m1(qh_3, 0x04, vl), vl);
+
+            vint8m1_t a_0 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_0), 32, vl);
+            vint8m1_t a_1 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_1), 32, vl);
+            vint8m1_t a_2 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_2), 32, vl);
+            vint8m1_t a_3 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_3), 32, vl);
+
+            // load Q8 and take product
+            vint16m2_t va_q_0 = __riscv_vwmul_vv_i16m2(a_0, __riscv_vle8_v_i8m1(q8, vl), vl);
+            vint16m2_t va_q_1 = __riscv_vwmul_vv_i16m2(a_1, __riscv_vle8_v_i8m1(q8+32, vl), vl);
+            vint16m2_t va_q_2 = __riscv_vwmul_vv_i16m2(a_2, __riscv_vle8_v_i8m1(q8+64, vl), vl);
+            vint16m2_t va_q_3 = __riscv_vwmul_vv_i16m2(a_3, __riscv_vle8_v_i8m1(q8+96, vl), vl);
+
+            vl = 16;
+
+            vint32m2_t vaux_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 0), scale[is+0], vl);
+            vint32m2_t vaux_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 1), scale[is+1], vl);
+            vint32m2_t vaux_2 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 0), scale[is+2], vl);
+            vint32m2_t vaux_3 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 1), scale[is+3], vl);
+            vint32m2_t vaux_4 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 0), scale[is+4], vl);
+            vint32m2_t vaux_5 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 1), scale[is+5], vl);
+            vint32m2_t vaux_6 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 0), scale[is+6], vl);
+            vint32m2_t vaux_7 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 1), scale[is+7], vl);
+
+            vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_0, vaux_1, vl), vzero, vl);
+            vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_2, vaux_3, vl), isum0, vl);
+            vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_4, vaux_5, vl), isum1, vl);
+            vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_6, vaux_7, vl), isum2, vl);
+
+            sum_t += __riscv_vmv_x_s_i32m1_i32(isum3);
+
+            q6 += 64;   qh += 32;   q8 += 128;   is=8;
+
+        }
+
+        sumf += d * sum_t;
+
+    }
+
+    *s = sumf;
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
+    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
+    const vector signed char off = vec_splats((signed char)0x20);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
+
+        const uint8_t * restrict q6 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict qs = x[i].scales;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            __builtin_prefetch(q6, 0, 0);
+            __builtin_prefetch(qh, 0, 0);
+            __builtin_prefetch(q8, 0, 0);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q6);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q6);
+            vector signed char qxs2 = (vector signed char)vec_xl(32, q6);
+            vector signed char qxs3 = (vector signed char)vec_xl(48, q6);
+            q6 += 64;
+
+            vector signed char qxs00 = vec_and(qxs0, lowMask);
+            vector signed char qxs01 = vec_sr(qxs0, v4);
+            vector signed char qxs10 = vec_and(qxs1, lowMask);
+            vector signed char qxs11 = vec_sr(qxs1, v4);
+            vector signed char qxs20 = vec_and(qxs2, lowMask);
+            vector signed char qxs21 = vec_sr(qxs2, v4);
+            vector signed char qxs30 = vec_and(qxs3, lowMask);
+            vector signed char qxs31 = vec_sr(qxs3, v4);
+
+            vector signed char qxhs0 = (vector signed char)vec_xl( 0, qh);
+            vector signed char qxhs1 = (vector signed char)vec_xl(16, qh);
+            qh += 32;
+
+            vector signed char qxh00 = vec_sl(vec_and((vector signed char)v3, qxhs0), v4);
+            vector signed char qxh01 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v4)), v4);
+            vector signed char qxh10 = vec_sl(vec_and((vector signed char)v3, qxhs1), v4);
+            vector signed char qxh11 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v4)), v4);
+            vector signed char qxh20 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v2)), v4);
+            vector signed char qxh21 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v6)), v4);
+            vector signed char qxh30 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v2)), v4);
+            vector signed char qxh31 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v6)), v4);
+
+            vector signed char q6x00 = vec_sub(vec_or(qxh00, qxs00), off);
+            vector signed char q6x01 = vec_sub(vec_or(qxh01, qxs01), off);
+            vector signed char q6x10 = vec_sub(vec_or(qxh10, qxs10), off);
+            vector signed char q6x11 = vec_sub(vec_or(qxh11, qxs11), off);
+            vector signed char q6x20 = vec_sub(vec_or(qxh20, qxs20), off);
+            vector signed char q6x21 = vec_sub(vec_or(qxh21, qxs21), off);
+            vector signed char q6x30 = vec_sub(vec_or(qxh30, qxs30), off);
+            vector signed char q6x31 = vec_sub(vec_or(qxh31, qxs31), off);
+
+            vector signed char q8y00 = vec_xl(  0, q8);
+            vector signed char q8y10 = vec_xl( 16, q8);
+            vector signed char q8y20 = vec_xl( 32, q8);
+            vector signed char q8y30 = vec_xl( 48, q8);
+            vector signed char q8y01 = vec_xl( 64, q8);
+            vector signed char q8y11 = vec_xl( 80, q8);
+            vector signed char q8y21 = vec_xl( 96, q8);
+            vector signed char q8y31 = vec_xl(112, q8);
+            q8 += 128;
+
+            vector signed short qv00 = vec_add(vec_mule(q6x00, q8y00), vec_mulo(q6x00, q8y00));
+            vector signed short qv10 = vec_add(vec_mule(q6x10, q8y10), vec_mulo(q6x10, q8y10));
+            vector signed short qv20 = vec_add(vec_mule(q6x20, q8y20), vec_mulo(q6x20, q8y20));
+            vector signed short qv30 = vec_add(vec_mule(q6x30, q8y30), vec_mulo(q6x30, q8y30));
+            vector signed short qv01 = vec_add(vec_mule(q6x01, q8y01), vec_mulo(q6x01, q8y01));
+            vector signed short qv11 = vec_add(vec_mule(q6x11, q8y11), vec_mulo(q6x11, q8y11));
+            vector signed short qv21 = vec_add(vec_mule(q6x21, q8y21), vec_mulo(q6x21, q8y21));
+            vector signed short qv31 = vec_add(vec_mule(q6x31, q8y31), vec_mulo(q6x31, q8y31));
+
+            vector signed short vscales = vec_unpackh(vec_xl_len(qs, 8));
+            qs += 8;
+
+            vector signed short vs0 = vec_splat(vscales, 0);
+            vector signed short vs1 = vec_splat(vscales, 1);
+            vector signed short vs2 = vec_splat(vscales, 2);
+            vector signed short vs3 = vec_splat(vscales, 3);
+            vector signed short vs4 = vec_splat(vscales, 4);
+            vector signed short vs5 = vec_splat(vscales, 5);
+            vector signed short vs6 = vec_splat(vscales, 6);
+            vector signed short vs7 = vec_splat(vscales, 7);
+
+            vsumi0 = vec_msum(qv00, vs0, vsumi0);
+            vsumi1 = vec_msum(qv01, vs4, vsumi1);
+            vsumi2 = vec_msum(qv10, vs1, vsumi2);
+            vsumi3 = vec_msum(qv11, vs5, vsumi3);
+            vsumi4 = vec_msum(qv20, vs2, vsumi4);
+            vsumi5 = vec_msum(qv21, vs6, vsumi5);
+            vsumi6 = vec_msum(qv30, vs3, vsumi6);
+            vsumi7 = vec_msum(qv31, vs7, vsumi7);
+        }
+
+        vsumi0 = vec_add(vsumi0, vsumi4);
+        vsumi1 = vec_add(vsumi1, vsumi5);
+        vsumi2 = vec_add(vsumi2, vsumi6);
+        vsumi3 = vec_add(vsumi3, vsumi7);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined __loongarch_asx
+
+    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
+    const __m256i m2 = __lasx_xvreplgr2vr_b(3);
+    const __m256i m32s = __lasx_xvreplgr2vr_b(32);
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m128i scales = __lsx_vld((const __m128i*)x[i].scales, 0);
+
+        __m256i sumi = __lasx_xvldi(0);
+
+        int is = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m128i scale_0 = lsx_shuffle_b(scales, get_scale_shuffle(is + 0));
+            const __m128i scale_1 = lsx_shuffle_b(scales, get_scale_shuffle(is + 1));
+            const __m128i scale_2 = lsx_shuffle_b(scales, get_scale_shuffle(is + 2));
+            const __m128i scale_3 = lsx_shuffle_b(scales, get_scale_shuffle(is + 3));
+            is += 4;
+
+            const __m256i q4bits1 = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
+            const __m256i q4bits2 = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
+            const __m256i q4bitsH = __lasx_xvld((const __m256i*)qh, 0); qh += 32;
+
+            const __m256i q4h_0 = __lasx_xvslli_h(__lasx_xvand_v(q4bitsH, m2), 4);
+            const __m256i q4h_1 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 2), m2), 4);
+            const __m256i q4h_2 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 4), m2), 4);
+            const __m256i q4h_3 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 6), m2), 4);
+
+            const __m256i q4_0 = __lasx_xvor_v(__lasx_xvand_v(q4bits1, m4), q4h_0);
+            const __m256i q4_1 = __lasx_xvor_v(__lasx_xvand_v(q4bits2, m4), q4h_1);
+            const __m256i q4_2 = __lasx_xvor_v(__lasx_xvand_v(__lasx_xvsrli_h(q4bits1, 4), m4), q4h_2);
+            const __m256i q4_3 = __lasx_xvor_v(__lasx_xvand_v(__lasx_xvsrli_h(q4bits2, 4), m4), q4h_3);
+
+            const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+
+            __m256i q8s_0 = lasx_maddubs_h(m32s, q8_0);
+            __m256i q8s_1 = lasx_maddubs_h(m32s, q8_1);
+            __m256i q8s_2 = lasx_maddubs_h(m32s, q8_2);
+            __m256i q8s_3 = lasx_maddubs_h(m32s, q8_3);
+
+            __m256i p16_0 = lasx_maddubs_h(q4_0, q8_0);
+            __m256i p16_1 = lasx_maddubs_h(q4_1, q8_1);
+            __m256i p16_2 = lasx_maddubs_h(q4_2, q8_2);
+            __m256i p16_3 = lasx_maddubs_h(q4_3, q8_3);
+
+            p16_0 = __lasx_xvsub_h(p16_0, q8s_0);
+            p16_1 = __lasx_xvsub_h(p16_1, q8s_1);
+            p16_2 = __lasx_xvsub_h(p16_2, q8s_2);
+            p16_3 = __lasx_xvsub_h(p16_3, q8s_3);
+
+            p16_0 = lasx_madd_h(lasx_ext8_16(scale_0), p16_0);
+            p16_1 = lasx_madd_h(lasx_ext8_16(scale_1), p16_1);
+            p16_2 = lasx_madd_h(lasx_ext8_16(scale_2), p16_2);
+            p16_3 = lasx_madd_h(lasx_ext8_16(scale_3), p16_3);
+
+            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1));
+            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_2, p16_3));
+        }
+
+        acc = __lasx_xvfmadd_s((__m256)__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const  int8_t * restrict q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * restrict a = aux8;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
+                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
+                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
+                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
+            }
+            a  += 128;
+            q4 += 64;
+            qh += 32;
+        }
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+#if defined (__AVX__) || defined (__AVX2__) || defined (__ARM_NEON) || defined (__POWER9_VECTOR__) || defined(__loongarch_asx)
+static const int8_t keven_signs_q2xs[1024] = {
+     1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
+     1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
+     1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1, -1,
+     1,  1, -1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1,  1,
+     1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1, -1,
+     1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1,  1,
+     1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1,  1,
+     1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1,  1,  1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1, -1,
+     1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1, -1,
+     1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1,  1,
+     1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1,  1,
+     1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1, -1,
+     1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1,  1,
+     1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1, -1,
+     1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1, -1,
+     1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1,  1,
+     1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1, -1,
+     1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1,  1,
+     1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1,  1,
+     1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1, -1,
+     1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1,  1,
+     1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1, -1,
+     1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1, -1,
+     1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1,  1,
+     1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1,  1,
+     1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1, -1,
+     1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1, -1,
+     1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1,  1,
+     1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1, -1,
+     1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1,  1,
+     1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1,  1,
+     1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
+};
+#endif
+
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xxs * restrict x = vx;
+    const block_q8_K    * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    ggml_int8x16x4_t q2u;
+    ggml_int8x16x4_t q2s;
+    ggml_int8x16x4_t q8b;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        float sumf1 = 0, sumf2 = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+            q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 0])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 1])));
+            q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 2])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 3])));
+            q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 8])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 9])));
+            q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[10])), vld1_s8((const void *)(iq2xxs_grid + aux8[11])));
+            q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >>  7) & 127))));
+            q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 21) & 127))));
+            q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[3] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[3] >>  7) & 127))));
+            q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[3] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[3] >> 21) & 127))));
+            q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]);
+            q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]);
+            q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]);
+            q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]);
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]), q2u.val[1], q8b.val[1]);
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]), q2u.val[3], q8b.val[3]);
+            sumf1 += vaddvq_s32(p1) * (0.5f + (aux32[1] >> 28));
+            sumf2 += vaddvq_s32(p2) * (0.5f + (aux32[3] >> 28));
+        }
+        sumf += d*(sumf1 + sumf2);
+    }
+    *s = 0.25f * sumf;
+
+#elif defined(__AVX2__)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+            const __m256i q2_1 = _mm256_set_epi64x(iq2xxs_grid[aux8[ 3]], iq2xxs_grid[aux8[ 2]], iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
+            const __m256i q2_2 = _mm256_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]], iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
+            const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127],
+                                                   signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
+            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
+            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const uint16_t ls1 = aux32[1] >> 28;
+            const uint16_t ls2 = aux32[3] >> 28;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
+            sumi1 = _mm256_add_epi32(sumi1, p1);
+            sumi2 = _mm256_add_epi32(sumi2, p2);
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2xxs_grid[aux8[3]], iq2xxs_grid[aux8[2]]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]]);
+            const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
+            const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
+            const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127]);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = aux32[1] >> 28;
+            const uint16_t ls2 = aux32[3] >> 28;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__POWER9_VECTOR__)
+    const vector int v0 = vec_splats((int32_t)0);
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t  *  restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/32; j += 2) {
+            __builtin_prefetch(q2, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            uint32_t aux32[4];
+            const uint8_t * aux8 = (const uint8_t *)aux32;
+
+            memcpy(aux32, q2, 4*sizeof(uint32_t));
+            q2 += 8;
+
+            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2xxs_grid + aux8[ 0]), *(const int64_t *)(iq2xxs_grid + aux8[ 1])};
+            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2xxs_grid + aux8[ 2]), *(const int64_t *)(iq2xxs_grid + aux8[ 3])};
+            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2xxs_grid + aux8[ 8]), *(const int64_t *)(iq2xxs_grid + aux8[ 9])};
+            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2xxs_grid + aux8[10]), *(const int64_t *)(iq2xxs_grid + aux8[11])};
+
+            vector signed long long vsigns0 = {*(const int64_t *)(signs64 + ((aux32[1] >>  0) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >>  7) & 127))};
+            vector signed long long vsigns1 = {*(const int64_t *)(signs64 + ((aux32[1] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >> 21) & 127))};
+            vector signed long long vsigns2 = {*(const int64_t *)(signs64 + ((aux32[3] >>  0) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >>  7) & 127))};
+            vector signed long long vsigns3 = {*(const int64_t *)(signs64 + ((aux32[3] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >> 21) & 127))};
+
+            vector signed char q2x0 = (vector signed char)vec_mul((vector signed char)vsigns0, (vector signed char)aux64x2_0);
+            vector signed char q2x1 = (vector signed char)vec_mul((vector signed char)vsigns1, (vector signed char)aux64x2_1);
+            vector signed char q2x2 = (vector signed char)vec_mul((vector signed char)vsigns2, (vector signed char)aux64x2_2);
+            vector signed char q2x3 = (vector signed char)vec_mul((vector signed char)vsigns3, (vector signed char)aux64x2_3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
+
+            const uint16_t ls0 = aux32[1] >> 28;
+            const uint16_t ls1 = aux32[3] >> 28;
+
+            vector signed short vscales01 = vec_splats((int16_t)(2*ls0+1));
+            vector signed short vscales23 = vec_splats((int16_t)(2*ls1+1));
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = 0.125f * vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+
+            const __m256i q2_1 = lasx_set_d(iq2xxs_grid[aux8[ 3]], iq2xxs_grid[aux8[ 2]], iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
+            const __m256i q2_2 = lasx_set_d(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]], iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
+            const __m256i s2_1 = lasx_set_d(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m256i s2_2 = lasx_set_d(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127],
+                                                   signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
+            const __m256i q8s_1 = __lasx_xvsigncov_b(s2_1, q8_1);
+            const __m256i q8s_2 = __lasx_xvsigncov_b(s2_2, q8_2);
+            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
+            const uint16_t ls1 = aux32[1] >> 28;
+            const uint16_t ls2 = aux32[3] >> 28;
+            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
+            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
+            sumi1 = __lasx_xvadd_w(sumi1, p1);
+            sumi2 = __lasx_xvadd_w(sumi2, p2);
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#else
+
+    uint32_t aux32[2];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(aux32, q2, 2*sizeof(uint32_t));
+            q2 += 4;
+            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+#endif
+}
+
+void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xs * restrict x = vx;
+    const block_q8_K   * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    ggml_int8x16x4_t q2u;
+    ggml_int8x16x4_t q2s;
+    ggml_int8x16x4_t q8b;
+
+    int32x4x4_t scales32;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        const uint8x8_t scales8 = vld1_u8(x[i].scales);
+        const uint8x8_t scales_l = vand_u8(scales8, vdup_n_u8(0xf));
+        const uint8x8_t scales_h = vshr_n_u8(scales8, 4);
+        uint8x16_t scales = vcombine_u8(vzip1_u8(scales_l, scales_h), vzip2_u8(scales_l, scales_h));
+        scales = vaddq_u8(vshlq_n_u8(scales, 1), vdupq_n_u8(1));
+        const uint16x8_t scales1 = vmovl_u8(vget_low_u8(scales));
+        const uint16x8_t scales2 = vmovl_u8(vget_high_u8(scales));
+        scales32.val[0] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales1)));
+        scales32.val[1] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales1)));
+        scales32.val[2] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales2)));
+        scales32.val[3] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales2)));
+        int32x4_t sumi = vdupq_n_s32(0);
+        for (int ib64 = 0; ib64 < QK_K/64; ++ib64) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+            q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[0] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[1] & 511))));
+            q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[2] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[3] & 511))));
+            q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[4] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[5] & 511))));
+            q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[6] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[7] & 511))));
+            q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[0] >> 9))), vld1_s8((const void *)(signs64 + (q2[1] >> 9))));
+            q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[2] >> 9))), vld1_s8((const void *)(signs64 + (q2[3] >> 9))));
+            q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[4] >> 9))), vld1_s8((const void *)(signs64 + (q2[5] >> 9))));
+            q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[6] >> 9))), vld1_s8((const void *)(signs64 + (q2[7] >> 9))));
+            q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]);
+            q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]);
+            q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]);
+            q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]);
+            const int32x4_t p1 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]);
+            const int32x4_t p2 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[1], q8b.val[1]);
+            const int32x4_t p3 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]);
+            const int32x4_t p4 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[3], q8b.val[3]);
+            const int32x4_t p = vpaddq_s32(vpaddq_s32(p1, p2), vpaddq_s32(p3, p4));
+            sumi = vmlaq_s32(sumi, p, scales32.val[ib64]);
+            q2 += 8;
+        }
+        sumf += d*vaddvq_s32(sumi);
+    }
+    *s = 0.125f * sumf;
+
+#elif defined(__AVX2__)
+
+    const __m256i mone = _mm256_set1_epi8(1);
+    static const char block_sign_shuffle_mask_1[32] = {
+        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+    };
+    static const char block_sign_shuffle_mask_2[32] = {
+        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
+        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
+    };
+    static const uint8_t bit_selector_mask_bytes[32] = {
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m256i bit_selector_mask = _mm256_loadu_si256((const __m256i*)bit_selector_mask_bytes);
+    const __m256i block_sign_shuffle_1 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_1);
+    const __m256i block_sign_shuffle_2 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_2);
+
+    static const uint8_t k_bit_helper[32] = {
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+    };
+    const __m256i bit_helper = _mm256_loadu_si256((const __m256i*)k_bit_helper);
+    const __m256i m511 = _mm256_set1_epi16(511);
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    uint64_t aux64;
+
+    // somewhat hacky, but gives a significant boost in performance
+    __m256i aux_gindex;
+    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        __m128i stmp = _mm_set1_epi64x(aux64);
+        stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
+        const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
+
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
+
+            const __m256i q2_data = _mm256_loadu_si256((const __m256i*)q2);  q2 += 16;
+            aux_gindex = _mm256_and_si256(q2_data, m511);
+
+            const __m256i partial_sign_bits = _mm256_srli_epi16(q2_data, 9);
+            const __m256i partial_sign_bits_upper = _mm256_srli_epi16(q2_data, 13);
+            const __m256i partial_sign_bits_for_counting = _mm256_xor_si256(partial_sign_bits, partial_sign_bits_upper);
+
+            const __m256i odd_bits = _mm256_shuffle_epi8(bit_helper, partial_sign_bits_for_counting);
+            const __m256i full_sign_bits = _mm256_or_si256(partial_sign_bits, odd_bits);
+
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_3 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_4 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+
+            const __m256i q2_1 = _mm256_set_epi64x(iq2xs_grid[gindex[ 3]], iq2xs_grid[gindex[ 2]],
+                                                   iq2xs_grid[gindex[ 1]], iq2xs_grid[gindex[ 0]]);
+            const __m256i q2_2 = _mm256_set_epi64x(iq2xs_grid[gindex[ 7]], iq2xs_grid[gindex[ 6]],
+                                                   iq2xs_grid[gindex[ 5]], iq2xs_grid[gindex[ 4]]);
+            const __m256i q2_3 = _mm256_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]],
+                                                   iq2xs_grid[gindex[ 9]], iq2xs_grid[gindex[ 8]]);
+            const __m256i q2_4 = _mm256_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]],
+                                                   iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
+
+            const __m128i full_signs_l = _mm256_castsi256_si128(full_sign_bits);
+            const __m128i full_signs_h = _mm256_extractf128_si256(full_sign_bits, 1);
+            const __m256i full_signs_1 = MM256_SET_M128I(full_signs_l, full_signs_l);
+            const __m256i full_signs_2 = MM256_SET_M128I(full_signs_h, full_signs_h);
+
+            __m256i signs;
+            signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_1);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, _mm256_or_si256(signs, mone));
+
+            signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_2);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, _mm256_or_si256(signs, mone));
+
+            signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_1);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_3 = _mm256_sign_epi8(q8_3, _mm256_or_si256(signs, mone));
+
+            signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_2);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_4 = _mm256_sign_epi8(q8_4, _mm256_or_si256(signs, mone));
+
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const __m256i dot3  = _mm256_maddubs_epi16(q2_3, q8s_3);
+            const __m256i dot4  = _mm256_maddubs_epi16(q2_4, q8s_4);
+
+            const __m256i sc1 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0)));
+            const __m256i sc2 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1)));
+            const __m256i sc3 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2)));
+            const __m256i sc4 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3)));
+
+            sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot1, sc1));
+            sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot2, sc2));
+            sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot3, sc3));
+            sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot4, sc4));
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+    const __m128i mone = _mm_set1_epi8(1);
+    static const char block_sign_shuffle_mask_1[32] = {
+        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+    };
+    static const char block_sign_shuffle_mask_2[32] = {
+        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
+        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
+    };
+    static const uint8_t bit_selector_mask_bytes[32] = {
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i bit_selector_mask_0 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes);
+    const __m128i bit_selector_mask_1 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes + 1);
+    const __m128i block_sign_shuffle_1_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1);
+    const __m128i block_sign_shuffle_1_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1 + 1);
+    const __m128i block_sign_shuffle_2_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2);
+    const __m128i block_sign_shuffle_2_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2 + 1);
+
+    static const uint8_t k_bit_helper[32] = {
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+    };
+    const __m128i bit_helper_0 = _mm_loadu_si128((const __m128i*)k_bit_helper);
+    const __m128i bit_helper_1 = _mm_loadu_si128((const __m128i*)k_bit_helper + 1);
+    const __m128i m511 = _mm_set1_epi16(511);
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    uint64_t aux64;
+
+    // somewhat hacky, but gives a significant boost in performance
+    __m256i aux_gindex;
+    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        __m128i stmp = _mm_set1_epi64x(aux64);
+        stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
+        const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
+
+            const __m128i q2_data_0 = _mm_loadu_si128((const __m128i*)q2);
+            const __m128i q2_data_1 = _mm_loadu_si128((const __m128i*)q2 + 1);  q2 += 16;
+            aux_gindex = MM256_SET_M128I(_mm_and_si128(q2_data_1, m511), _mm_and_si128(q2_data_0, m511));
+
+            const __m128i partial_sign_bits_0 = _mm_srli_epi16(q2_data_0, 9);
+            const __m128i partial_sign_bits_1 = _mm_srli_epi16(q2_data_1, 9);
+            const __m128i partial_sign_bits_upper_0 = _mm_srli_epi16(q2_data_0, 13);
+            const __m128i partial_sign_bits_upper_1 = _mm_srli_epi16(q2_data_1, 13);
+            const __m128i partial_sign_bits_for_counting_0 = _mm_xor_si128(partial_sign_bits_0, partial_sign_bits_upper_0);
+            const __m128i partial_sign_bits_for_counting_1 = _mm_xor_si128(partial_sign_bits_1, partial_sign_bits_upper_1);
+
+            const __m128i odd_bits_0 = _mm_shuffle_epi8(bit_helper_0, partial_sign_bits_for_counting_0);
+            const __m128i odd_bits_1 = _mm_shuffle_epi8(bit_helper_1, partial_sign_bits_for_counting_1);
+            const __m128i full_sign_bits_0 = _mm_or_si128(partial_sign_bits_0, odd_bits_0);
+            const __m128i full_sign_bits_1 = _mm_or_si128(partial_sign_bits_1, odd_bits_1);
+
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_3_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_3_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_4_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_4_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2xs_grid[gindex[1]], iq2xs_grid[gindex[0]]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2xs_grid[gindex[3]], iq2xs_grid[gindex[2]]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2xs_grid[gindex[5]], iq2xs_grid[gindex[4]]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2xs_grid[gindex[7]], iq2xs_grid[gindex[6]]);
+            const __m128i q2_3_0 = _mm_set_epi64x(iq2xs_grid[gindex[9]], iq2xs_grid[gindex[8]]);
+            const __m128i q2_3_1 = _mm_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]]);
+            const __m128i q2_4_0 = _mm_set_epi64x(iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
+            const __m128i q2_4_1 = _mm_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]]);
+
+            // AVX2 full_signs_1 is full_sign_bits_0 here
+            // AVX2 full_signs_2 is full_sign_bits_1 here
+            __m128i signs_0, signs_1;
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_3_0 = _mm_sign_epi8(q8_3_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_3_1 = _mm_sign_epi8(q8_3_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_4_0 = _mm_sign_epi8(q8_4_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_4_1 = _mm_sign_epi8(q8_4_1, _mm_or_si128(signs_1, mone));
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const __m128i dot3_0  = _mm_maddubs_epi16(q2_3_0, q8s_3_0);
+            const __m128i dot3_1  = _mm_maddubs_epi16(q2_3_1, q8s_3_1);
+            const __m128i dot4_0  = _mm_maddubs_epi16(q2_4_0, q8s_4_0);
+            const __m128i dot4_1  = _mm_maddubs_epi16(q2_4_1, q8s_4_1);
+
+            __m128i sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0));
+            const __m128i sc1_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc1_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1));
+            const __m128i sc2_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc2_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2));
+            const __m128i sc3_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc3_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3));
+            const __m128i sc4_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc4_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot1_0, sc1_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot1_1, sc1_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot2_0, sc2_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot2_1, sc2_1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot3_0, sc3_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot3_1, sc3_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot4_0, sc4_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot4_1, sc4_1));
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__loongarch_asx)
+
+    const __m256i mone = __lasx_xvreplgr2vr_b(1);
+    static const char block_sign_shuffle_mask_1[32] = {
+        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+    };
+    static const char block_sign_shuffle_mask_2[32] = {
+        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
+        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
+    };
+    static const uint8_t bit_selector_mask_bytes[32] = {
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m256i bit_selector_mask = __lasx_xvld((const __m256i*)bit_selector_mask_bytes, 0);
+    const __m256i block_sign_shuffle_1 = __lasx_xvld((const __m256i*)block_sign_shuffle_mask_1, 0);
+    const __m256i block_sign_shuffle_2 = __lasx_xvld((const __m256i*)block_sign_shuffle_mask_2, 0);
+
+    static const uint8_t k_bit_helper[32] = {
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+    };
+    const __m256i bit_helper = __lasx_xvld((const __m256i*)k_bit_helper, 0);
+    const __m256i m511 = __lasx_xvreplgr2vr_h(511);
+    const __m128i m4 = __lsx_vreplgr2vr_b(0xf);
+    const __m128i m1 = __lsx_vreplgr2vr_b(1);
+
+    uint64_t aux64;
+
+    // somewhat hacky, but gives a significant boost in performance
+    __m256i aux_gindex;
+    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        __m128i stmp = __lsx_vreplgr2vr_d(aux64);
+        stmp = __lsx_vilvl_b( __lsx_vand_v(__lsx_vsrli_h(stmp, 4), m4), __lsx_vand_v(stmp, m4));
+        const __m128i scales = __lsx_vadd_b(__lsx_vslli_h(stmp, 1), m1);
+
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
+
+            const __m256i q2_data = __lasx_xvld((const __m256i*)q2, 0);  q2 += 16;
+            aux_gindex = __lasx_xvand_v(q2_data, m511);
+
+            const __m256i partial_sign_bits = __lasx_xvsrli_h(q2_data, 9);
+            const __m256i partial_sign_bits_upper = __lasx_xvsrli_h(q2_data, 13);
+            const __m256i partial_sign_bits_for_counting = __lasx_xvxor_v(partial_sign_bits, partial_sign_bits_upper);
+
+            const __m256i odd_bits = lasx_shuffle_b(bit_helper, partial_sign_bits_for_counting);
+            const __m256i full_sign_bits = __lasx_xvor_v(partial_sign_bits, odd_bits);
+
+            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_3 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_4 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+
+            const __m256i q2_1 = lasx_set_d(iq2xs_grid[gindex[ 3]], iq2xs_grid[gindex[ 2]],
+                                                   iq2xs_grid[gindex[ 1]], iq2xs_grid[gindex[ 0]]);
+            const __m256i q2_2 = lasx_set_d(iq2xs_grid[gindex[ 7]], iq2xs_grid[gindex[ 6]],
+                                                   iq2xs_grid[gindex[ 5]], iq2xs_grid[gindex[ 4]]);
+            const __m256i q2_3 = lasx_set_d(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]],
+                                                   iq2xs_grid[gindex[ 9]], iq2xs_grid[gindex[ 8]]);
+            const __m256i q2_4 = lasx_set_d(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]],
+                                                   iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
+
+            const __m128i full_signs_l = lasx_extracti128(full_sign_bits, 0);
+            const __m128i full_signs_h = lasx_extracti128(full_sign_bits, 1);
+            const __m256i full_signs_1 = lasx_insertf128(full_signs_l, full_signs_l);
+            const __m256i full_signs_2 = lasx_insertf128(full_signs_h, full_signs_h);
+
+            __m256i signs;
+            signs = lasx_shuffle_b(full_signs_1, block_sign_shuffle_1);
+            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_1 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_1);
+
+            signs = lasx_shuffle_b(full_signs_1, block_sign_shuffle_2);
+            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_2 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_2);
+
+            signs = lasx_shuffle_b(full_signs_2, block_sign_shuffle_1);
+            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_3 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_3);
+
+            signs = lasx_shuffle_b(full_signs_2, block_sign_shuffle_2);
+            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_4 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_4);
+
+            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
+            const __m256i dot3  = lasx_maddubs_h(q2_3, q8s_3);
+            const __m256i dot4  = lasx_maddubs_h(q2_4, q8s_4);
+
+            const __m256i sc1 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+0)));
+            const __m256i sc2 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+1)));
+            const __m256i sc3 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+2)));
+            const __m256i sc4 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+3)));
+
+            sumi1 = __lasx_xvadd_w(sumi1, lasx_madd_h(dot1, sc1));
+            sumi2 = __lasx_xvadd_w(sumi2, lasx_madd_h(dot2, sc2));
+            sumi1 = __lasx_xvadd_w(sumi1, lasx_madd_h(dot3, sc3));
+            sumi2 = __lasx_xvadd_w(sumi2, lasx_madd_h(dot4, sc4));
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+#elif defined(__POWER9_VECTOR__)
+    const vector int v0 = vec_splats((int32_t)0);
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        const uint16_t * restrict q2 = x[i].qs;
+        const uint8_t  * restrict sc = x[i].scales;
+        const int8_t  *  restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            __builtin_prefetch(q2, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2xs_grid + (q2[0] & 511)), *(const int64_t *)(iq2xs_grid + (q2[1] & 511))};
+            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2xs_grid + (q2[2] & 511)), *(const int64_t *)(iq2xs_grid + (q2[3] & 511))};
+            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2xs_grid + (q2[4] & 511)), *(const int64_t *)(iq2xs_grid + (q2[5] & 511))};
+            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2xs_grid + (q2[6] & 511)), *(const int64_t *)(iq2xs_grid + (q2[7] & 511))};
+
+            vector signed long long vsigns0 = {*(const int64_t *)(signs64 + ((q2[0] >> 9))), *(const int64_t *)(signs64 + ((q2[1] >> 9)))};
+            vector signed long long vsigns1 = {*(const int64_t *)(signs64 + ((q2[2] >> 9))), *(const int64_t *)(signs64 + ((q2[3] >> 9)))};
+            vector signed long long vsigns2 = {*(const int64_t *)(signs64 + ((q2[4] >> 9))), *(const int64_t *)(signs64 + ((q2[5] >> 9)))};
+            vector signed long long vsigns3 = {*(const int64_t *)(signs64 + ((q2[6] >> 9))), *(const int64_t *)(signs64 + ((q2[7] >> 9)))};
+            q2 += 8;
+
+            vector signed char q2x0 = (vector signed char)vec_mul((vector signed char)vsigns0, (vector signed char)aux64x2_0);
+            vector signed char q2x1 = (vector signed char)vec_mul((vector signed char)vsigns1, (vector signed char)aux64x2_1);
+            vector signed char q2x2 = (vector signed char)vec_mul((vector signed char)vsigns2, (vector signed char)aux64x2_2);
+            vector signed char q2x3 = (vector signed char)vec_mul((vector signed char)vsigns3, (vector signed char)aux64x2_3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
+            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
+            const uint16_t ls2 = (uint16_t)(sc[1] & 0xf);
+            const uint16_t ls3 = (uint16_t)(sc[1] >>  4);
+            sc += 2;
+
+            vector signed short vscales0 = vec_splats((int16_t)(2*ls0+1));
+            vector signed short vscales1 = vec_splats((int16_t)(2*ls1+1));
+            vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1));
+            vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1));
+
+            vsumi0 = vec_msum(qv0, vscales0, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales1, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales2, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales3, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = 0.125f * vec_extract(vsumf0, 0);
+#else
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const uint8_t  * restrict sc = x[i].scales;
+        const int8_t   * restrict q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
+            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls2;
+            q2 += 4;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+#endif
+}
+
+void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_s * restrict x = vx;
+    const block_q8_K  * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
+
+    const ggml_uint8x16x2_t mask1 = ggml_vld1q_u8_x2(k_mask1);
+    const uint8x16_t        mask2 = vld1q_u8(k_mask2);
+    const uint8x16_t m1 = vdupq_n_u8(1);
+    const int32x4_t vzero = vdupq_n_s32(0);
+
+    uint8x16x2_t vs;
+    ggml_int8x16x4_t q2s;
+    ggml_int8x16x4_t q8b;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const int8_t  * restrict q8 = y[i].qs;
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+            q2s.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[0] | ((qh[ib32+0] << 8) & 0x300)))),
+                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[1] | ((qh[ib32+0] << 6) & 0x300)))));
+            q2s.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[2] | ((qh[ib32+0] << 4) & 0x300)))),
+                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[3] | ((qh[ib32+0] << 2) & 0x300)))));
+            q2s.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[4] | ((qh[ib32+1] << 8) & 0x300)))),
+                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[5] | ((qh[ib32+1] << 6) & 0x300)))));
+            q2s.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[6] | ((qh[ib32+1] << 4) & 0x300)))),
+                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[7] | ((qh[ib32+1] << 2) & 0x300)))));
+            qs += 8;
+
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16)));
+            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
+            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
+            vs.val[0] = vceqq_u8(vs.val[0], mask2);
+            vs.val[1] = vceqq_u8(vs.val[1], mask2);
+
+            q2s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[0], m1)), q2s.val[0]);
+            q2s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[1], m1)), q2s.val[1]);
+
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16)));
+            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
+            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
+            vs.val[0] = vceqq_u8(vs.val[0], mask2);
+            vs.val[1] = vceqq_u8(vs.val[1], mask2);
+
+            signs += 4;
+
+            q2s.val[2] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[0], m1)), q2s.val[2]);
+            q2s.val[3] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[1], m1)), q2s.val[3]);
+
+            const int32x4_t p1 = ggml_vdotq_s32(vzero, q2s.val[0], q8b.val[0]);
+            const int32x4_t p2 = ggml_vdotq_s32(vzero, q2s.val[1], q8b.val[1]);
+            const int32x4_t p3 = ggml_vdotq_s32(vzero, q2s.val[2], q8b.val[2]);
+            const int32x4_t p4 = ggml_vdotq_s32(vzero, q2s.val[3], q8b.val[3]);
+
+            sumi1 += vaddvq_s32(p1) * (1 + 2*(x[i].scales[ib32+0] & 0xf));
+            sumi2 += vaddvq_s32(p2) * (1 + 2*(x[i].scales[ib32+0] >>  4));
+            sumi1 += vaddvq_s32(p3) * (1 + 2*(x[i].scales[ib32+1] & 0xf));
+            sumi2 += vaddvq_s32(p4) * (1 + 2*(x[i].scales[ib32+1] >>  4));
+        }
+        sumf += d*(sumi1 + sumi2);
+    }
+
+    *s = 0.125f * sumf;
+
+#elif defined(__AVX2__)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1);
+    const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2);
+
+    uint64_t aux64;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const int8_t  * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1);
+        const __m256i scales16 = _mm256_cvtepi8_epi16(scales8); // 0 2 4 6 8 10 12 14 1 3 5 7 9 11 13 15
+
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q2_1 = _mm256_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
+                                                   iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)],
+                                                   iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
+                                                   iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
+            const __m256i q2_2 = _mm256_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
+                                                   iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)],
+                                                   iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
+                                                   iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
+            qs += 8;
+
+            __m256i aux256 = _mm256_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
+            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+            const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2);
+            const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1);
+
+            aux256 = _mm256_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
+            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+            const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2);
+            const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2);
+
+            signs += 4;
+
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1); // blocks 2*ib32+0, 2*ib32+1
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2); // blocks 2*ib32+2, 2*ib32+3
+
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_shuffle_epi8(scales16, get_scale_shuffle_k4(ib32+0)));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_shuffle_epi8(scales16, get_scale_shuffle_k4(ib32+1)));
+            sumi1 = _mm256_add_epi32(sumi1, p1);
+            sumi2 = _mm256_add_epi32(sumi2, p2);
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
+    const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
+    const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
+    const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
+
+    uint64_t aux64;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const int8_t  * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1);
+        const __m128i scales16_0 = _mm_cvtepi8_epi16(scales8);
+        const __m128i scales16_1 = _mm_cvtepi8_epi16(_mm_srli_si128(scales8, 8));
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
+                                                  iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
+                                                  iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
+                                                  iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
+                                                  iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)]);
+            qs += 8;
+
+            __m128i aux128_0 = _mm_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
+            __m128i aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
+            const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
+
+            aux128_0 = _mm_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
+            aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
+            const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
+
+            signs += 4;
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 0)));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 1)));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 0)));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 1)));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__POWER9_VECTOR__)
+    static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                        0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+    };
+
+    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
+
+    const vector int v0 = vec_splats((int32_t)0);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const vector unsigned char mask0 = vec_xl( 0, k_mask1);
+    const vector unsigned char mask1 = vec_xl(16, k_mask1);
+    const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        const uint8_t *  restrict q2 = x[i].qs;
+        const uint8_t *  restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const uint8_t *  restrict sc = x[i].scales;
+        const int8_t  *  restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/32; j += 2) {
+            __builtin_prefetch(q2, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2s_grid + (q2[0] | ((qh[0] << 8) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[1] | ((qh[0] << 6) & 0x300)))};
+            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2s_grid + (q2[2] | ((qh[0] << 4) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[3] | ((qh[0] << 2) & 0x300)))};
+            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2s_grid + (q2[4] | ((qh[1] << 8) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[5] | ((qh[1] << 6) & 0x300)))};
+            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2s_grid + (q2[6] | ((qh[1] << 4) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[7] | ((qh[1] << 2) & 0x300)))};
+            q2 += 8;
+            qh += 2;
+
+            vector signed char vsigns01 = (vector signed char)vec_splats(*(const uint32_t *)&signs[0]);
+            vector signed char vsigns23 = (vector signed char)vec_splats(*(const uint32_t *)&signs[2]);
+            signs += 4;
+
+            vector signed char vsigns0 = vec_perm(vsigns01, vsigns01, mask0);
+            vector signed char vsigns1 = vec_perm(vsigns01, vsigns01, mask1);
+            vector signed char vsigns2 = vec_perm(vsigns23, vsigns23, mask0);
+            vector signed char vsigns3 = vec_perm(vsigns23, vsigns23, mask1);
+
+            vsigns0 = (vector signed char)vec_cmpeq(vec_and(vsigns0, mask2), mask2);
+            vsigns1 = (vector signed char)vec_cmpeq(vec_and(vsigns1, mask2), mask2);
+            vsigns2 = (vector signed char)vec_cmpeq(vec_and(vsigns2, mask2), mask2);
+            vsigns3 = (vector signed char)vec_cmpeq(vec_and(vsigns3, mask2), mask2);
+
+            vector signed char q2x0 = vec_sub(vec_xor(vsigns0, (vector signed char)aux64x2_0), vsigns0);
+            vector signed char q2x1 = vec_sub(vec_xor(vsigns1, (vector signed char)aux64x2_1), vsigns1);
+            vector signed char q2x2 = vec_sub(vec_xor(vsigns2, (vector signed char)aux64x2_2), vsigns2);
+            vector signed char q2x3 = vec_sub(vec_xor(vsigns3, (vector signed char)aux64x2_3), vsigns3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
+            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
+            const uint16_t ls2 = (uint16_t)(sc[1] & 0xf);
+            const uint16_t ls3 = (uint16_t)(sc[1] >>  4);
+            sc += 2;
+
+            vector signed short vscales0 = vec_splats((int16_t)(2*ls0+1));
+            vector signed short vscales1 = vec_splats((int16_t)(2*ls1+1));
+            vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1));
+            vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1));
+
+            vsumi0 = vec_msum(qv0, vscales0, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales1, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales2, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales3, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = 0.125f * vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+
+    const __m128i m4 = __lsx_vreplgr2vr_b(0xf);
+    const __m128i m1 = __lsx_vreplgr2vr_b(1);
+
+    const __m256i mask1 = __lasx_xvld((const __m256i*)k_mask1, 0);
+    const __m256i mask2 = __lasx_xvld((const __m256i*)k_mask2, 0);
+    uint64_t aux64;
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const int8_t  * restrict q8 = y[i].qs;
+
+        __m128i tmp1;
+        memcpy(&aux64, x[i].scales, 8);
+        tmp1 = __lsx_vinsgr2vr_d(tmp1, aux64, 0);
+        tmp1 = __lsx_vinsgr2vr_d(tmp1, aux64 >> 4, 1);
+        const __m128i scales8 = __lsx_vadd_b(__lsx_vslli_h(__lsx_vand_v(tmp1, m4), 1), m1);
+        const __m256i scales16 = lasx_ext8_16(scales8); // 0 2 4 6 8 10 12 14 1 3 5 7 9 11 13 15
+
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q2_1 = lasx_set_d(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
+                                                   iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)],
+                                                   iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
+                                                   iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
+            const __m256i q2_2 = lasx_set_d(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
+                                                   iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)],
+                                                   iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
+                                                   iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
+            qs += 8;
+
+            __m256i aux256 = __lasx_xvreplgr2vr_w(signs[0] | ((uint32_t) signs[1] << 16));
+            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
+            const __m256i s2_1 = __lasx_xvseq_b(aux256, mask2);
+            const __m256i q8s_1 = __lasx_xvsub_b(__lasx_xvxor_v(s2_1, q8_1), s2_1);
+
+            aux256 = __lasx_xvreplgr2vr_w(signs[2] | ((uint32_t) signs[3] << 16));
+            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
+            const __m256i s2_2 = __lasx_xvseq_b(aux256, mask2);
+            const __m256i q8s_2 = __lasx_xvsub_b(__lasx_xvxor_v(s2_2, q8_2), s2_2);
+
+            signs += 4;
+
+            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1); // blocks 2*ib32+0, 2*ib32+1
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2); // blocks 2*ib32+2, 2*ib32+3
+
+            const __m256i p1 = lasx_madd_h(dot1, lasx_shuffle_b(scales16, get_scale_shuffle_k4(ib32+0)));
+            const __m256i p2 = lasx_madd_h(dot2, lasx_shuffle_b(scales16, get_scale_shuffle_k4(ib32+1)));
+            sumi1 = __lasx_xvadd_w(sumi1, p1);
+            sumi2 = __lasx_xvadd_w(sumi2, p2);
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#else
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const int8_t  * q8 = y[i].qs;
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * qh = x[i].qh;
+        const uint8_t * signs = qs + QK_K/8;
+
+        int bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
+            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
+            int sumi1 = 0, sumi2 = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += ls1 * sumi1 + ls2 * sumi2;
+            qs += 4;
+            signs += 4;
+        }
+
+        sumf += d * bsum;
+    }
+
+    *s = 0.125f * sumf;
+
+#endif
+
+}
+
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_xxs * restrict x = vx;
+    const block_q8_K    * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    ggml_int8x16x4_t q3s;
+    ggml_int8x16x4_t q8b;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t   * restrict q8 = y[i].qs;
+        float sumf1 = 0, sumf2 = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+            memcpy(aux32, gas, 2*sizeof(uint32_t)); gas += 2*sizeof(uint32_t);
+            const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]);
+            const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]);
+            const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]);
+            const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]);
+            q3 += 16;
+            q3s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >>  7) & 127))));
+            q3s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >> 21) & 127))));
+            q3s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >>  7) & 127))));
+            q3s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 21) & 127))));
+            q3s.val[0] = vmulq_s8(q3s.val[0], vreinterpretq_s8_u32(aux32x4_0));
+            q3s.val[1] = vmulq_s8(q3s.val[1], vreinterpretq_s8_u32(aux32x4_1));
+            q3s.val[2] = vmulq_s8(q3s.val[2], vreinterpretq_s8_u32(aux32x4_2));
+            q3s.val[3] = vmulq_s8(q3s.val[3], vreinterpretq_s8_u32(aux32x4_3));
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]);
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]);
+            sumf1 += vaddvq_s32(p1) * (0.5f + (aux32[0] >> 28));
+            sumf2 += vaddvq_s32(p2) * (0.5f + (aux32[1] >> 28));
+        }
+        sumf += d*(sumf1 + sumf2);
+    }
+    *s = 0.5f * sumf;
+
+#elif defined(__AVX2__)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q2_1 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+                                                  iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            q3 += 8;
+            const __m256i q2_2 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+                                                  iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            q3 += 8;
+            memcpy(aux32, gas, 8); gas += 8;
+            const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127],
+                                                   signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
+            const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
+            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const uint16_t ls1 = aux32[0] >> 28;
+            const uint16_t ls2 = aux32[1] >> 28;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
+            sumi1 = _mm256_add_epi32(sumi1, p1);
+            sumi2 = _mm256_add_epi32(sumi2, p2);
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.25f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q2_1_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            const __m128i q2_1_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
+            q3 += 8;
+            const __m128i q2_2_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            const __m128i q2_2_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
+            q3 += 8;
+            memcpy(aux32, gas, 8); gas += 8;
+            const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
+            const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127]);
+            const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = aux32[0] >> 28;
+            const uint16_t ls2 = aux32[1] >> 28;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.25f * hsum_float_8(accumf);
+
+#elif defined(__POWER9_VECTOR__)
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    const vector int v0 = vec_splats((int32_t)0);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint32_t * restrict signs = (const uint32_t *)(x[i].qs + QK_K/4);
+        const int8_t  * restrict q8 = y[i].qs;
+
+#pragma GCC unroll 1
+        for (int j = 0; j < QK_K/32; j += 2) {
+            __builtin_prefetch(q3, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector unsigned int aux32x4_0 = {iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]};
+            vector unsigned int aux32x4_1 = {iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]};
+            vector unsigned int aux32x4_2 = {iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]};
+            vector unsigned int aux32x4_3 = {iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]};
+            q3 += 16;
+
+            vector unsigned long long aux64x2_0 = {(uint64_t)(signs64[(signs[0] >>  0) & 127]), (uint64_t)(signs64[(signs[0] >>  7) & 127])};
+            vector unsigned long long aux64x2_1 = {(uint64_t)(signs64[(signs[0] >> 14) & 127]), (uint64_t)(signs64[(signs[0] >> 21) & 127])};
+            vector unsigned long long aux64x2_2 = {(uint64_t)(signs64[(signs[1] >>  0) & 127]), (uint64_t)(signs64[(signs[1] >>  7) & 127])};
+            vector unsigned long long aux64x2_3 = {(uint64_t)(signs64[(signs[1] >> 14) & 127]), (uint64_t)(signs64[(signs[1] >> 21) & 127])};
+
+            vector signed char q3x0 = vec_mul((vector signed char)aux64x2_0, (vector signed char)aux32x4_0);
+            vector signed char q3x1 = vec_mul((vector signed char)aux64x2_1, (vector signed char)aux32x4_1);
+            vector signed char q3x2 = vec_mul((vector signed char)aux64x2_2, (vector signed char)aux32x4_2);
+            vector signed char q3x3 = vec_mul((vector signed char)aux64x2_3, (vector signed char)aux32x4_3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q3x0, q8y0), vec_mulo(q3x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q3x1, q8y1), vec_mulo(q3x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q3x2, q8y2), vec_mulo(q3x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q3x3, q8y3), vec_mulo(q3x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(signs[0] >> 28);
+            const uint16_t ls1 = (uint16_t)(signs[1] >> 28);
+            signs += 2;
+
+            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
+            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = 0.25f * vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q2_1 = lasx_set_w(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+                                                iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            q3 += 8;
+            const __m256i q2_2 = lasx_set_w(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+                                                iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            q3 += 8;
+            memcpy(aux32, gas, 8); gas += 8;
+
+            const __m256i s2_1 = lasx_set_d(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127],
+                                                   signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
+            const __m256i s2_2 = lasx_set_d(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m256i q8s_1 = __lasx_xvsigncov_b(s2_1, q8_1);
+            const __m256i q8s_2 = __lasx_xvsigncov_b(s2_2, q8_2);
+            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
+            const uint16_t ls1 = aux32[0] >> 28;
+            const uint16_t ls2 = aux32[1] >> 28;
+
+            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
+            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
+            sumi1 = __lasx_xvadd_w(sumi1, p1);
+            sumi2 = __lasx_xvadd_w(sumi2, p2);
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+    }
+
+    *s = 0.25f * hsum_float_8(accumf);
+
+#else
+
+    uint32_t aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t  * restrict q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
+            const uint32_t ls = 2*(aux32 >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
+                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            q3 += 8;
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.25f * sumf;
+#endif
+}
+
+void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_s * restrict x = vx;
+    const block_q8_K  * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    typedef union {
+        uint16x8_t vec_index;
+        uint16_t   index[8];
+    } vec_index_t;
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
+
+    static const int16_t k_shift[8] = {8, 7, 6, 5, 4, 3, 2, 1};
+
+    const ggml_uint8x16x2_t mask1 = ggml_vld1q_u8_x2(k_mask1);
+    const uint8x16_t        mask2 = vld1q_u8(k_mask2);
+
+    const int16x8_t  hshift = vld1q_s16(k_shift);
+    const uint16x8_t m256   = vdupq_n_u16(256);
+    const uint8x16_t m1     = vdupq_n_u8(1);
+
+    uint8x16x2_t vs;
+    ggml_int8x16x4_t q3s;
+    ggml_int8x16x4_t q8b;
+    vec_index_t idx;
+
+    uint32_t scales32[2];
+    const uint8_t * scales8 = (const uint8_t *)scales32;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
+        const int8_t   * restrict q8 = y[i].qs;
+
+        memcpy(scales32, x[i].scales, 4);
+        scales32[1] = (((scales32[0] >> 4) & 0x0f0f0f0f) << 1) | 0x01010101;
+        scales32[0] = ((scales32[0] & 0x0f0f0f0f) << 1) | 0x01010101;
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            const uint8x16_t idx_l = vld1q_u8(qs); qs += 16;
+            idx.vec_index = vorrq_u16(vmovl_u8(vget_low_u8 (idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+0]), hshift), m256));
+            const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]],
+                                                        iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]);
+            const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]],
+                                                        iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]);
+            idx.vec_index = vorrq_u16(vmovl_u8(vget_high_u8(idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+1]), hshift), m256));
+            const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]],
+                                                        iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]);
+            const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]],
+                                                        iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]);
+
+
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16)));
+            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
+            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
+            vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1);
+            vs.val[1] = vorrq_u8(vceqq_u8(vs.val[1], mask2), m1);
+
+            q3s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_0));
+            q3s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_1));
+
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16)));
+            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
+            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
+            vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1);
+            vs.val[1] = vorrq_u8(vceqq_u8(vs.val[1], mask2), m1);
+
+            signs += 4;
+
+            q3s.val[2] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_2));
+            q3s.val[3] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_3));
+
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]);
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]);
+
+            sumi1 += vaddvq_s32(p1) * scales8[ib32/2+0];
+            sumi2 += vaddvq_s32(p2) * scales8[ib32/2+4];
+        }
+        sumf += d*(sumi1 + sumi2);
+    }
+    *s = sumf;
+
+#elif defined(__AVX2__)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1);
+    const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2);
+
+    const __m256i idx_shift = _mm256_set_epi32(1, 2, 3, 4, 5, 6, 7, 8);
+    const __m256i idx_mask  = _mm256_set1_epi32(256);
+
+    typedef union {
+        __m256i  vec[2];
+        uint32_t index[16];
+    } index_t;
+
+    index_t idx;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i idx_l = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)qs)); qs += 16;
+            idx.vec[0] = _mm256_set1_epi32(qh[ib32+0]);
+            idx.vec[1] = _mm256_set1_epi32(qh[ib32+1]);
+            idx.vec[0] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[0], idx_shift), idx_mask);
+            idx.vec[1] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[1], idx_shift), idx_mask);
+            idx.vec[0] = _mm256_or_si256(idx.vec[0], _mm256_cvtepi16_epi32(_mm256_castsi256_si128(idx_l)));
+            idx.vec[1] = _mm256_or_si256(idx.vec[1], _mm256_cvtepi16_epi32(_mm256_extractf128_si256(idx_l, 1)));
+
+            // At leat on my CPU (Ryzen 7950X), using _mm256_i32gather_epi32 is slower than _mm256_set_epi32. Strange.
+            //const __m256i q2_1 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[0], 4);
+            //const __m256i q2_2 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[1], 4);
+            const __m256i q2_1 = _mm256_set_epi32(
+                    iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]],
+                    iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]
+            );
+            const __m256i q2_2 = _mm256_set_epi32(
+                    iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]],
+                    iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[ 9]], iq3s_grid[idx.index[ 8]]
+            );
+
+            __m256i aux256 = _mm256_set1_epi32(signs[0] | (signs[1] << 16));
+            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+            const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2);
+            const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1);
+
+            aux256 = _mm256_set1_epi32(signs[2] | (signs[3] << 16));
+            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+            const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2);
+            const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2);
+
+            signs += 4;
+
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
+            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
+            sumi1 = _mm256_add_epi32(sumi1, p1);
+            sumi2 = _mm256_add_epi32(sumi2, p2);
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
+    const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
+    const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
+    const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
+
+    const __m128i idx_mul_0 = _mm_set_epi32(32, 64, 128, 256);
+    const __m128i idx_mul_1 = _mm_set_epi32(2, 4, 8, 16);
+    const __m128i idx_mask  = _mm_set1_epi32(256);
+
+    typedef union {
+        __m128i  vec[4];
+        uint32_t index[16];
+    } index_t;
+
+    index_t idx;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i qs_tmp = _mm_loadu_si128((const __m128i *)qs);
+            const __m128i idx_l_0 = _mm_cvtepu8_epi16(qs_tmp);
+            const __m128i idx_l_1 = _mm_cvtepu8_epi16(_mm_srli_si128(qs_tmp, 8)); qs += 16;
+            idx.vec[0] = _mm_set1_epi32(qh[ib32+0]);
+            idx.vec[1] = idx.vec[0];
+            idx.vec[2] = _mm_set1_epi32(qh[ib32+1]);
+            idx.vec[3] = idx.vec[2];
+
+            idx.vec[0] = _mm_and_si128(_mm_mullo_epi32(idx.vec[0], idx_mul_0), idx_mask);
+            idx.vec[1] = _mm_and_si128(_mm_mullo_epi32(idx.vec[1], idx_mul_1), idx_mask);
+            idx.vec[2] = _mm_and_si128(_mm_mullo_epi32(idx.vec[2], idx_mul_0), idx_mask);
+            idx.vec[3] = _mm_and_si128(_mm_mullo_epi32(idx.vec[3], idx_mul_1), idx_mask);
+
+            idx.vec[0] = _mm_or_si128(idx.vec[0], _mm_cvtepi16_epi32(idx_l_0));
+            idx.vec[1] = _mm_or_si128(idx.vec[1], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_0, 8)));
+            idx.vec[2] = _mm_or_si128(idx.vec[2], _mm_cvtepi16_epi32(idx_l_1));
+            idx.vec[3] = _mm_or_si128(idx.vec[3], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_1, 8)));
+
+            const __m128i q2_1_0 = _mm_set_epi32(iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]);
+            const __m128i q2_1_1 = _mm_set_epi32(iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]]);
+            const __m128i q2_2_0 = _mm_set_epi32(iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[9]], iq3s_grid[idx.index[8]]);
+            const __m128i q2_2_1 = _mm_set_epi32(iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]]);
+
+            __m128i aux128_0 = _mm_set1_epi32(signs[0] | (signs[1] << 16));
+            __m128i aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
+            const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
+
+            aux128_0 = _mm_set1_epi32(signs[2] | (signs[3] << 16));
+            aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
+            const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
+
+            signs += 4;
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
+            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = hsum_float_8(accumf);
+
+#elif defined(__POWER9_VECTOR__)
+    static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                        0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+    };
+
+    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
+
+    const vector int v0 = vec_splats((int32_t)0);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const vector unsigned char mask0 = vec_xl( 0, k_mask1);
+    const vector unsigned char mask1 = vec_xl(16, k_mask1);
+    const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        const uint8_t *  restrict q3 = x[i].qs;
+        const uint8_t *  restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)(x[i].signs);
+        const uint8_t *  restrict sc = x[i].scales;
+        const int8_t  *  restrict q8 = y[i].qs;
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        for (int j = 0; j < QK_K/32; j += 2) {
+            __builtin_prefetch(q3, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector unsigned int aux32x4_0 = {iq3s_grid[q3[ 0] | ((qh[0] << 8) & 256)], iq3s_grid[q3[ 1] | ((qh[0] << 7) & 256)],
+                                             iq3s_grid[q3[ 2] | ((qh[0] << 6) & 256)], iq3s_grid[q3[ 3] | ((qh[0] << 5) & 256)]};
+            vector unsigned int aux32x4_1 = {iq3s_grid[q3[ 4] | ((qh[0] << 4) & 256)], iq3s_grid[q3[ 5] | ((qh[0] << 3) & 256)],
+                                             iq3s_grid[q3[ 6] | ((qh[0] << 2) & 256)], iq3s_grid[q3[ 7] | ((qh[0] << 1) & 256)]};
+            vector unsigned int aux32x4_2 = {iq3s_grid[q3[ 8] | ((qh[1] << 8) & 256)], iq3s_grid[q3[ 9] | ((qh[1] << 7) & 256)],
+                                             iq3s_grid[q3[10] | ((qh[1] << 6) & 256)], iq3s_grid[q3[11] | ((qh[1] << 5) & 256)]};
+            vector unsigned int aux32x4_3 = {iq3s_grid[q3[12] | ((qh[1] << 4) & 256)], iq3s_grid[q3[13] | ((qh[1] << 3) & 256)],
+                                             iq3s_grid[q3[14] | ((qh[1] << 2) & 256)], iq3s_grid[q3[15] | ((qh[1] << 1) & 256)]};
+            q3 += 16;
+            qh += 2;
+
+            vector signed char vsigns01 = (vector signed char)vec_splats(*(const uint32_t *)&signs[0]);
+            vector signed char vsigns02 = (vector signed char)vec_splats(*(const uint32_t *)&signs[2]);
+            signs += 4;
+
+            vector signed char vsigns0 = vec_perm(vsigns01, vsigns01, mask0);
+            vector signed char vsigns1 = vec_perm(vsigns01, vsigns01, mask1);
+            vector signed char vsigns2 = vec_perm(vsigns02, vsigns02, mask0);
+            vector signed char vsigns3 = vec_perm(vsigns02, vsigns02, mask1);
+
+            vsigns0 = (vector signed char)vec_cmpeq(vec_and(vsigns0, mask2), mask2);
+            vsigns1 = (vector signed char)vec_cmpeq(vec_and(vsigns1, mask2), mask2);
+            vsigns2 = (vector signed char)vec_cmpeq(vec_and(vsigns2, mask2), mask2);
+            vsigns3 = (vector signed char)vec_cmpeq(vec_and(vsigns3, mask2), mask2);
+
+            vector signed char q3x0 = vec_sub(vec_xor(vsigns0, (vector signed char)aux32x4_0), vsigns0);
+            vector signed char q3x1 = vec_sub(vec_xor(vsigns1, (vector signed char)aux32x4_1), vsigns1);
+            vector signed char q3x2 = vec_sub(vec_xor(vsigns2, (vector signed char)aux32x4_2), vsigns2);
+            vector signed char q3x3 = vec_sub(vec_xor(vsigns3, (vector signed char)aux32x4_3), vsigns3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q3x0, q8y0), vec_mulo(q3x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q3x1, q8y1), vec_mulo(q3x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q3x2, q8y2), vec_mulo(q3x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q3x3, q8y3), vec_mulo(q3x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
+            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
+            sc ++;
+
+            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
+            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m256i mask1 = __lasx_xvld((const __m256i*)k_mask1, 0);
+    const __m256i mask2 = __lasx_xvld((const __m256i*)k_mask2, 0);
+
+    __m256i idx_shift = lasx_set_w(1, 2, 3, 4, 5, 6, 7, 8);
+    const __m256i idx_mask  = __lasx_xvreplgr2vr_w(256);
+
+    typedef union {
+        __m256i  vec[2];
+        uint32_t index[16];
+    } index_t;
+
+    index_t idx;
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i idx_l = lasx_extu8_16(__lsx_vld(qs, 0)); qs += 16;
+            idx.vec[0] = __lasx_xvreplgr2vr_w(qh[ib32+0]);
+            idx.vec[1] = __lasx_xvreplgr2vr_w(qh[ib32+1]);
+            idx.vec[0] = __lasx_xvand_v(__lasx_xvsll_w(idx.vec[0], idx_shift), idx_mask);
+            idx.vec[1] = __lasx_xvand_v(__lasx_xvsll_w(idx.vec[1], idx_shift), idx_mask);
+            idx.vec[0] = __lasx_xvor_v(idx.vec[0], lasx_ext16_32(lasx_extracti128(idx_l, 0)));
+            idx.vec[1] = __lasx_xvor_v(idx.vec[1], lasx_ext16_32(lasx_extracti128(idx_l, 1)));
+
+            // At leat on my CPU (Ryzen 7950X), using _mm256_i32gather_epi32 is slower than _mm256_set_epi32. Strange.
+            //const __m256i q2_1 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[0], 4);
+            //const __m256i q2_2 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[1], 4);
+            const __m256i q2_1 = lasx_set_w(
+                    iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]],
+                    iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]
+            );
+            const __m256i q2_2 = lasx_set_w(
+                    iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]],
+                    iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[ 9]], iq3s_grid[idx.index[ 8]]
+            );
+
+            __m256i aux256 = __lasx_xvreplgr2vr_w(signs[0] | (signs[1] << 16));
+            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
+            const __m256i s2_1 = __lasx_xvseq_b(aux256, mask2);
+            const __m256i q8s_1 = __lasx_xvsub_b(__lasx_xvxor_v(s2_1, q8_1), s2_1);
+
+            aux256 = __lasx_xvreplgr2vr_w(signs[2] | (signs[3] << 16));
+            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
+            const __m256i s2_2 = __lasx_xvseq_b(aux256, mask2);
+            const __m256i q8s_2 = __lasx_xvsub_b(__lasx_xvxor_v(s2_2, q8_2), s2_2);
+
+            signs += 4;
+
+            const __m256i dot1 = lasx_maddubs_h(q2_1, q8s_1);
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
+            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
+            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
+            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
+            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
+            sumi1 = __lasx_xvadd_w(sumi1, p1);
+            sumi2 = __lasx_xvadd_w(sumi2, p2);
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+    }
+
+    *s = hsum_float_8(accumf);
+
+#else
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint8_t * restrict signs = x[i].signs;
+        const int8_t  * restrict q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
+            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls2;
+        }
+        sumf += d * bsum;
+    }
+    *s = sumf;
+#endif
+}
+
+#if defined(__AVX2__)
+static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
+    const __m256i ax = _mm256_sign_epi8(x, x);
+    const __m256i sy = _mm256_sign_epi8(y, x);
+    return _mm256_maddubs_epi16(ax, sy);
+}
+#elif defined(__loongarch_asx)
+static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
+    const __m256i ax = __lasx_xvsigncov_b(x, x);
+    const __m256i sy = __lasx_xvsigncov_b(x, y);
+    __m256i tmp1, tmp2, tmp3;
+    tmp1 = __lasx_xvmulwev_h_bu_b(ax, sy);
+    tmp2 = __lasx_xvmulwod_h_bu_b(ax, sy);
+    tmp3 = __lasx_xvadd_h(tmp1, tmp2);
+    return __lasx_xvsat_h(tmp3, 15);
+}
+#endif
+
+void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_s * restrict x = vx;
+    const block_q8_K  * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __ARM_NEON
+
+    ggml_int8x16x4_t q1b;
+    ggml_int8x16x4_t q8b;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        int sumi1 = 0, sumi2 = 0, sumi3 = 0;
+
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+
+            q1b.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[0] | ((qh[ib+0] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[1] | ((qh[ib+0] << 5) & 0x700)))));
+            q1b.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[2] | ((qh[ib+0] << 2) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[3] | ((qh[ib+0] >> 1) & 0x700)))));
+            q1b.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[4] | ((qh[ib+1] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[5] | ((qh[ib+1] << 5) & 0x700)))));
+            q1b.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[6] | ((qh[ib+1] << 2) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[7] | ((qh[ib+1] >> 1) & 0x700)))));
+            qs += 8;
+
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q1b.val[0], q8b.val[0]), q1b.val[1], q8b.val[1]);
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q1b.val[2], q8b.val[2]), q1b.val[3], q8b.val[3]);
+
+            const int ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+            sumi1 += vaddvq_s32(p1) * ls1;
+            sumi2 += vaddvq_s32(p2) * ls2;
+            sumi3 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * ls1 * (qh[ib+0] & 0x8000 ? -1 : 1)
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * ls2 * (qh[ib+1] & 0x8000 ? -1 : 1);
+
+        }
+
+        sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3);
+    }
+
+    *s = sumf;
+
+#elif defined __AVX2__
+
+    __m256 accum = _mm256_setzero_ps();
+    float accum1 = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        __m256i sumi = _mm256_setzero_si256();
+        int sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)],
+                                                    iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]);
+            const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)],
+                                                    iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]);
+            qs += 8;
+            const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
+            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
+            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(ls1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(ls2));
+
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p1, p2));
+            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
+        }
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        accum = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi), accum);
+        accum1 += d * sumi1;
+
+    }
+
+    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
+
+#elif defined __AVX__
+    __m256 accum = _mm256_setzero_ps();
+    float accum1 = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        int sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q1b_1_0 = _mm_set_epi64x(iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]);
+            const __m128i q1b_1_1 = _mm_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)]);
+            const __m128i q1b_2_0 = _mm_set_epi64x(iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]);
+            const __m128i q1b_2_1 = _mm_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)]);
+            qs += 8;
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
+            const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
+            const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
+            const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
+            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(ls1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(ls1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(ls2));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(ls2));
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
+            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
+        }
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum);
+        accum1 += d * sumi1;
+
+    }
+
+    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
+
+#elif defined(__POWER9_VECTOR__)
+    const vector unsigned char v0 = vec_splats((unsigned char)0x0);
+    const vector unsigned short vsign = vec_splats((unsigned short)0x8000);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = vec_splats((int32_t)0);
+        vector signed int vsumi1 = vec_splats((int32_t)0);
+        vector signed int vsumi2 = vec_splats((int32_t)0);
+        vector signed int vsumi3 = vec_splats((int32_t)0);
+        vector signed int vsumi8 = vec_splats((int32_t)0);
+
+        const uint8_t  * restrict q1 = x[i].qs;
+        const uint16_t * restrict qh = x[i].qh;
+        const int8_t   * restrict q8 = y[i].qs;
+        const int16_t  * restrict qs = y[i].bsums;
+
+        for (int j = 0; j < QK_K/32; j += 2) {
+            __builtin_prefetch(q1, 0, 1);
+            __builtin_prefetch(qh, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed long long aux64x2_0 = {*(const int64_t *)(iq1s_grid + (q1[0] | ((qh[0] << 8) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[1] | ((qh[0] << 5) & 0x700)))};
+            vector signed long long aux64x2_1 = {*(const int64_t *)(iq1s_grid + (q1[2] | ((qh[0] << 2) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[3] | ((qh[0] >> 1) & 0x700)))};
+            vector signed long long aux64x2_2 = {*(const int64_t *)(iq1s_grid + (q1[4] | ((qh[1] << 8) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[5] | ((qh[1] << 5) & 0x700)))};
+            vector signed long long aux64x2_3 = {*(const int64_t *)(iq1s_grid + (q1[6] | ((qh[1] << 2) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[7] | ((qh[1] >> 1) & 0x700)))};
+            q1 += 8;
+
+            vector signed char q1x0 = (vector signed char)aux64x2_0;
+            vector signed char q1x1 = (vector signed char)aux64x2_1;
+            vector signed char q1x2 = (vector signed char)aux64x2_2;
+            vector signed char q1x3 = (vector signed char)aux64x2_3;
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q1x0, q8y0), vec_mulo(q1x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q1x1, q8y1), vec_mulo(q1x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q1x2, q8y2), vec_mulo(q1x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q1x3, q8y3), vec_mulo(q1x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)((qh[0] >> 12) & 7);
+            const uint16_t ls1 = (uint16_t)((qh[1] >> 12) & 7);
+
+            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
+            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
+            vector signed short vscales = vec_sld(vscales23, vscales01, 8);
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+
+            vector signed short q8ysums = vec_xl_len(qs, 8);
+            qs += 4;
+            q8ysums = vec_mergeh(q8ysums, (vector signed short)v0);
+
+            vector signed short qxh = (vector signed short)vec_sld(vec_splats(qh[1]), vec_splats(qh[0]), 8);
+            qh += 2;
+            vector __bool short vsel = vec_cmpge(qxh, (vector signed short)v0);
+
+            vector signed short q8ysum = vec_sel((vector signed short)vec_xor((vector unsigned short)q8ysums, vsign), q8ysums, vsel);
+
+            vsumi8 = vec_add(vec_mule(q8ysum, vscales), vsumi8);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi8, 0), vec_mul(vd, vec_splats(IQ1S_DELTA)), vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+
+    __m256 accum = (__m256)__lasx_xvldi(0);
+    float accum1 = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        __m256i sumi = __lasx_xvldi(0);
+        int sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            __m256i q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)], 0);
+            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], 1);
+            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)], 2);
+            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], 3);
+
+            __m256i q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)], 0);
+            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], 1);
+            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)], 2);
+            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], 3);
+
+            qs += 8;
+            const __m256i q8b_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8b_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+
+            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
+            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
+            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+
+            __m256i tmp1, tmp5, tmp6;
+            tmp1 = __lasx_xvreplgr2vr_h(ls1);
+            tmp5 = __lasx_xvmulwev_w_h(dot1, tmp1);
+            tmp6 = __lasx_xvmulwod_w_h(dot1, tmp1);
+            const __m256i p1 = __lasx_xvadd_w(tmp5, tmp6);
+
+            tmp1 = __lasx_xvreplgr2vr_h(ls2);
+            tmp5 = __lasx_xvmulwev_w_h(dot2, tmp1);
+            tmp6 = __lasx_xvmulwod_w_h(dot2, tmp1);
+            const __m256i p2 = __lasx_xvadd_w(tmp5, tmp6);
+
+            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p1, p2));
+            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
+        }
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), accum);
+        accum1 += d * sumi1;
+    }
+
+    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
+
+#else
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        int sumi = 0, sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
+            const int delta = qh[ib] & 0x8000 ? -1 : 1;
+            int lsum = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
+                for (int j = 0; j < 8; ++j) {
+                    lsum += q8[j] * grid[j];
+                }
+                q8 += 8;
+            }
+            sumi  += ls * lsum;
+            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
+            qs += 4;
+        }
+
+        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+    }
+
+    *s = sumf;
+
+#endif
+}
+
+void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_m * restrict x = vx;
+    const block_q8_K  * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+    iq1m_scale_t scale;
+
+#if defined __ARM_NEON
+    const int32x4_t mask  = vdupq_n_s32(0x7);
+    const int32x4_t mone  = vdupq_n_s32(1);
+    const int32x4_t mzero = vdupq_n_s32(0);
+
+    ggml_int8x16x4_t deltas;
+    deltas.val[0] = vcombine_s8(vdup_n_s8(+1), vdup_n_s8(+1));
+    deltas.val[1] = vcombine_s8(vdup_n_s8(-1), vdup_n_s8(+1));
+    deltas.val[2] = vcombine_s8(vdup_n_s8(+1), vdup_n_s8(-1));
+    deltas.val[3] = vcombine_s8(vdup_n_s8(-1), vdup_n_s8(-1));
+
+    ggml_int8x16x4_t q1b;
+    ggml_int8x16x4_t q8b;
+
+    uint32_t aux32;
+    const uint8_t * aux8 = (const uint8_t *)&aux32;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+        int32x4_t sumi1 = mzero;
+        int32x4_t sumi2 = mzero;
+
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+
+            q1b.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[0] | ((qh[0] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[1] | ((qh[0] << 4) & 0x700)))));
+            q1b.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[2] | ((qh[1] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[3] | ((qh[1] << 4) & 0x700)))));
+            q1b.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[4] | ((qh[2] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[5] | ((qh[2] << 4) & 0x700)))));
+            q1b.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[6] | ((qh[3] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[7] | ((qh[3] << 4) & 0x700)))));
+
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            const int32x4_t p1 = vpaddq_s32(ggml_vdotq_s32(mzero, q1b.val[0], q8b.val[0]), ggml_vdotq_s32(mzero, q1b.val[1], q8b.val[1]));
+            const int32x4_t p2 = vpaddq_s32(ggml_vdotq_s32(mzero, q1b.val[2], q8b.val[2]), ggml_vdotq_s32(mzero, q1b.val[3], q8b.val[3]));
+            const int32x4_t p12 = vpaddq_s32(p1, p2);
+
+            const uint32_t * qh32 = (const uint32_t *)qh; // we are 4-byte aligned, so we can do that
+            aux32 = ((qh32[0] >> 3) & 0x01010101) | ((qh32[0] >> 6) & 0x02020202);
+
+            const int32x4_t p3 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[0]], q8b.val[0]), ggml_vdotq_s32(mzero, deltas.val[aux8[1]], q8b.val[1]));
+            const int32x4_t p4 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[2]], q8b.val[2]), ggml_vdotq_s32(mzero, deltas.val[aux8[3]], q8b.val[3]));
+            const int32x4_t p34 = vpaddq_s32(p3, p4);
+
+            int32x4_t scales_4 = ggml_vld1q_u32(sc[ib/2] >> 0, sc[ib/2] >> 3, sc[ib/2] >> 6, sc[ib/2] >> 9);
+
+            scales_4 = vaddq_s32(vshlq_n_s32(vandq_s32(scales_4, mask), 1), mone);
+
+            sumi1 = vmlaq_s32(sumi1, scales_4, p12);
+            sumi2 = vmlaq_s32(sumi2, scales_4, p34);
+
+            qs += 8; qh += 4;
+
+        }
+
+        sumf += y[i].d * GGML_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2));
+    }
+
+    *s = sumf;
+
+#elif defined __AVX2__
+
+    const __m256i mask = _mm256_set1_epi16(0x7);
+    const __m256i mone = _mm256_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m256i q1b_1 = _mm256_set_epi64x(
+                    iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)],
+                    iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)]
+            );
+            const __m256i q1b_2 = _mm256_set_epi64x(
+                    iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)],
+                    iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)]
+            );
+            const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
+            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
+
+            const __m256i delta1 = _mm256_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m256i delta2 = _mm256_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+
+            const __m256i dot3 = mul_add_epi8(delta1, q8b_1);
+            const __m256i dot4 = mul_add_epi8(delta2, q8b_2);
+
+            __m256i scale1 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 3), _mm_set1_epi16(sc[ib/2] >> 0));
+            __m256i scale2 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 9), _mm_set1_epi16(sc[ib/2] >> 6));
+
+            scale1 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale1, mask), 1), mone);
+            scale2 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale2, mask), 1), mone);
+            const __m256i p1 = _mm256_madd_epi16(dot1, scale1);
+            const __m256i p2 = _mm256_madd_epi16(dot2, scale2);
+            const __m256i p3 = _mm256_madd_epi16(dot3, scale1);
+            const __m256i p4 = _mm256_madd_epi16(dot4, scale2);
+
+            sumi1 = _mm256_add_epi32(sumi1, _mm256_add_epi32(p1, p2));
+            sumi2 = _mm256_add_epi32(sumi2, _mm256_add_epi32(p3, p4));
+
+            qs += 8; qh += 4;
+        }
+
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
+
+        accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1);
+        accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2);
+    }
+
+    *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
+
+#elif defined __AVX__
+    const __m128i mask = _mm_set1_epi16(0x7);
+    const __m128i mone = _mm_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q1b_1_0 = _mm_set_epi64x(
+                    iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)]);
+            const __m128i q1b_1_1 = _mm_set_epi64x(
+                    iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)]);
+            const __m128i q1b_2_0 = _mm_set_epi64x(
+                    iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)]);
+            const __m128i q1b_2_1 = _mm_set_epi64x(
+                    iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)]);
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
+            const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
+            const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
+            const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
+
+            const __m128i delta1_0 = _mm_set_epi64x(qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta1_1 = _mm_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta2_0 = _mm_set_epi64x(qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta2_1 = _mm_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+
+            const __m128i dot3_0 = mul_add_epi8_sse(delta1_0, q8b_1_0);
+            const __m128i dot3_1 = mul_add_epi8_sse(delta1_1, q8b_1_1);
+            const __m128i dot4_0 = mul_add_epi8_sse(delta2_0, q8b_2_0);
+            const __m128i dot4_1 = mul_add_epi8_sse(delta2_1, q8b_2_1);
+
+            __m128i scale1_0 = _mm_set1_epi16(sc[ib/2] >> 0);
+            __m128i scale1_1 = _mm_set1_epi16(sc[ib/2] >> 3);
+            __m128i scale2_0 = _mm_set1_epi16(sc[ib/2] >> 6);
+            __m128i scale2_1 = _mm_set1_epi16(sc[ib/2] >> 9);
+
+            scale1_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_0, mask), 1), mone);
+            scale1_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_1, mask), 1), mone);
+            scale2_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_0, mask), 1), mone);
+            scale2_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_1, mask), 1), mone);
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, scale1_0);
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, scale1_1);
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, scale2_0);
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, scale2_1);
+            const __m128i p3_0 = _mm_madd_epi16(dot3_0, scale1_0);
+            const __m128i p3_1 = _mm_madd_epi16(dot3_1, scale1_1);
+            const __m128i p4_0 = _mm_madd_epi16(dot4_0, scale2_0);
+            const __m128i p4_1 = _mm_madd_epi16(dot4_1, scale2_1);
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_add_epi32(p3_0, p4_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_add_epi32(p3_1, p4_1));
+
+            qs += 8; qh += 4;
+        }
+
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
+
+        accum1 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum1);
+        accum2 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi2_1, sumi2_0))), accum2);
+    }
+
+    *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
+
+#else
+
+    int sum1[2], sum2[2], delta[4];
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            delta[0] = qh[0] & 0x08 ? -1 : 1;
+            delta[1] = qh[0] & 0x80 ? -1 : 1;
+            delta[2] = qh[1] & 0x08 ? -1 : 1;
+            delta[3] = qh[1] & 0x80 ? -1 : 1;
+            sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
+                int lsum1 = 0, lsum2 = 0;
+                for (int j = 0; j < 8; ++j) {
+                    lsum1 += q8[j] * grid[j];
+                    lsum2 += q8[j];
+                }
+                q8 += 8;
+                sum1[l/2] += lsum1;
+                sum2[l/2] += lsum2*delta[l];
+            }
+
+            const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
+            const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
+
+            sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
+            sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
+            qs += 4;
+            qh += 2;
+        }
+
+        sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+    }
+
+    *s = sumf;
+
+#endif
+}
+
+void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK4_NL == 0);
+    static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
+
+    const block_iq4_nl * restrict x = vx;
+    const block_q8_0   * restrict y = vy;
+
+    const int nb = n / QK4_NL;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined __ARM_NEON
+    const int8x16_t values = vld1q_s8(kvalues_iq4nl);
+    const uint8x16_t m4b = vdupq_n_u8(0x0f);
+    uint8x16x2_t q4bits;
+    int8x16x4_t q4b;
+    int8x16x4_t q8b;
+    int32x4_t prod_1, prod_2;
+
+    for (; ib + 1 < nb; ib += 2) {
+
+        q4bits.val[0] = vld1q_u8(x[ib + 0].qs);
+        q4bits.val[1] = vld1q_u8(x[ib + 1].qs);
+        q8b.val[0]    = vld1q_s8(y[ib + 0].qs);
+        q8b.val[1]    = vld1q_s8(y[ib + 0].qs + 16);
+        q8b.val[2]    = vld1q_s8(y[ib + 1].qs);
+        q8b.val[3]    = vld1q_s8(y[ib + 1].qs + 16);
+
+        q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[0], m4b));
+        q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
+        q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[1], m4b));
+        q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
+
+        prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
+        prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
+
+        sumf +=
+            GGML_FP16_TO_FP32(x[ib+0].d) * GGML_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
+            GGML_FP16_TO_FP32(x[ib+1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
+    }
+
+#elif defined __AVX2__
+
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+    const __m256i mone = _mm256_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
+        const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[ib + 0].qs);
+        const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[ib + 1].qs);
+        const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
+                                              _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
+        const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
+                                              _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
+        const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
+        const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
+        const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
+        const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
+        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+                _mm256_cvtepi32_ps(p_1), accum1);
+        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+                _mm256_cvtepi32_ps(p_2), accum2);
+    }
+
+    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
+
+#elif defined __AVX__
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+
+    __m256 accum = _mm256_setzero_ps();
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
+        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
+        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
+
+        const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
+        const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
+        const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
+        const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
+
+        const __m256 p = mul_sum_i8_quad_float(q4b_1_0, q4b_1_1, q4b_2_0, q4b_2_1, q8b_1_0, q8b_1_1, q8b_2_0, q8b_2_1);
+        const __m256 deltas = quad_fp16_delta_float(x[ib].d, y[ib].d, x[ib + 1].d, y[ib + 1].d);
+        accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
+    }
+
+    sumf = hsum_float_8(accum);
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+
+    const vector signed char values = vec_xl( 0, kvalues_iq4nl);
+
+#pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+        vector signed char q4x0 = vec_and(qxs, lowMask);
+        vector signed char q4x1 = vec_sr(qxs, v4);
+
+        q4x0 = vec_perm(values, values, (vector unsigned char)q4x0);
+        q4x1 = vec_perm(values, values, (vector unsigned char)q4x1);
+
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
+
+        vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
+        vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi1 = vec_sum4s(qv1, vsumi1);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#elif defined (__loongarch_asx)
+
+    const __m128i values128 = __lsx_vld((const __m128i*)kvalues_iq4nl, 0);
+    const __m128i m4b  = __lsx_vreplgr2vr_b(0x0f);
+    const __m256i mone = __lasx_xvreplgr2vr_h(1);
+
+    __m256 accum1 = (__m256)__lasx_xvldi(0);
+    __m256 accum2 = (__m256)__lasx_xvldi(0);
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = __lsx_vld((const __m128i*)x[ib + 0].qs, 0);
+        const __m128i q4bits_2 = __lsx_vld((const __m128i*)x[ib + 1].qs, 0);
+        const __m256i q8b_1 = __lasx_xvld((const __m256i *)y[ib + 0].qs, 0);
+        const __m256i q8b_2 = __lasx_xvld((const __m256i *)y[ib + 1].qs, 0);
+        const __m256i q4b_1 = lasx_insertf128(lsx_shuffle_b(values128, __lsx_vand_v(__lsx_vsrli_h(q4bits_1, 4), m4b)),
+                                              lsx_shuffle_b(values128, __lsx_vand_v(q4bits_1, m4b)));
+        const __m256i q4b_2 = lasx_insertf128(lsx_shuffle_b(values128, __lsx_vand_v(__lsx_vsrli_h(q4bits_2, 4), m4b)),
+                                              lsx_shuffle_b(values128, __lsx_vand_v(q4bits_2, m4b)));
+        const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
+        const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
+        const __m256i p_1 = lasx_madd_h(p16_1, mone);
+        const __m256i p_2 = lasx_madd_h(p16_2, mone);
+        accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+                __lasx_xvffint_s_w(p_1), accum1);
+        accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+                __lasx_xvffint_s_w(p_2), accum2);
+    }
+
+    sumf = hsum_float_8(__lasx_xvfadd_s(accum1, accum2));
+
+#endif
+    for (; ib < nb; ++ib) {
+        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        int sumi1 = 0, sumi2 = 0;
+        for (int j = 0; j < QK4_NL/2; ++j) {
+            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
+            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
+        }
+        sumf += d * (sumi1 + sumi2);
+    }
+    *s = sumf;
+}
+
+void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_K == 0);
+
+    const block_iq4_xs * restrict x = vx;
+    const block_q8_K   * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __ARM_NEON
+    const int8x16_t values = vld1q_s8(kvalues_iq4nl);
+    const uint8x16_t m4b = vdupq_n_u8(0x0f);
+    ggml_uint8x16x2_t q4bits;
+    ggml_int8x16x4_t q4b;
+    ggml_int8x16x4_t q8b;
+    int32x4_t prod_1, prod_2;
+
+    float sumf = 0;
+
+    for (int ibl = 0; ibl < nb; ++ibl) {
+
+        const int8_t  * q8 = y[ibl].qs;
+        const uint8_t * q4 = x[ibl].qs;
+        uint16_t h = x[ibl].scales_h;
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib = 0; ib < QK_K/64; ++ib) {
+
+            q4bits = ggml_vld1q_u8_x2(q4); q4 += 32;
+            q8b    = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[0], m4b));
+            q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
+            q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[1], m4b));
+            q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
+
+            prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
+            prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
+
+            int ls1 = ((x[ibl].scales_l[ib] & 0xf) | ((h << 4) & 0x30)) - 32;
+            int ls2 = ((x[ibl].scales_l[ib] >>  4) | ((h << 2) & 0x30)) - 32;
+            h >>= 4;
+            sumi1 += vaddvq_s32(prod_1) * ls1;
+            sumi2 += vaddvq_s32(prod_2) * ls2;
+
+        }
+
+        sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
+    }
+
+    *s = sumf;
+
+#elif defined __AVX2__
+
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+
+    __m256 accum = _mm256_setzero_ps();
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        uint16_t sh = x[ibl].scales_h;
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)qs);  qs += 16;
+            const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)qs);  qs += 16;
+            const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
+                                                  _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
+            const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
+                                                  _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
+            const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
+            const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
+            const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;
+            const int16_t ls2 = ((x[ibl].scales_l[ib/2] >>  4) | ((sh << 2) & 0x30)) - 32;
+            sh >>= 4;
+            const __m256i p_1 = _mm256_madd_epi16(p16_1, _mm256_set1_epi16(ls1));
+            const __m256i p_2 = _mm256_madd_epi16(p16_2, _mm256_set1_epi16(ls2));
+            sumi1 = _mm256_add_epi32(p_1, sumi1);
+            sumi2 = _mm256_add_epi32(p_2, sumi2);
+        }
+        accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+                _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accum);
+    }
+
+    *s = hsum_float_8(accum);
+
+#elif defined __AVX__
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+
+    __m256 accum = _mm256_setzero_ps();
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        uint16_t sh = x[ibl].scales_h;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
+            const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
+            const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
+            const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
+            const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
+            const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+            const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+            const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+            const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+            const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;
+            const int16_t ls2 = ((x[ibl].scales_l[ib/2] >>  4) | ((sh << 2) & 0x30)) - 32;
+            sh >>= 4;
+            const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, _mm_set1_epi16(ls1));
+            const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, _mm_set1_epi16(ls1));
+            const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, _mm_set1_epi16(ls2));
+            const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, _mm_set1_epi16(ls2));
+            sumi1_0 = _mm_add_epi32(p_1_0, sumi1_0);
+            sumi1_1 = _mm_add_epi32(p_1_1, sumi1_1);
+            sumi2_0 = _mm_add_epi32(p_2_0, sumi2_0);
+            sumi2_1 = _mm_add_epi32(p_2_1, sumi2_1);
+        }
+        __m128i sumi12_0 = _mm_add_epi32(sumi1_0, sumi2_0);
+        __m128i sumi12_1 = _mm_add_epi32(sumi1_1, sumi2_1);
+        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(sumi12_1, sumi12_0))), accum);
+    }
+
+    *s = hsum_float_8(accum);
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const vector signed char values = vec_xl( 0, kvalues_iq4nl);
+
+    for (int ibl = 0; ibl < nb; ++ibl) {
+
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ibl].d));
+        vector float vyd = vec_splats(y[ibl].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        uint16_t h = x[ibl].scales_h;
+
+        const uint8_t * restrict q4 = x[ibl].qs;
+        const uint8_t * restrict sc = x[ibl].scales_l;
+        const int8_t  * restrict q8 = y[ibl].qs;
+
+        for (int ib = 0; ib < QK_K/64; ib ++ ) {
+            __builtin_prefetch(q4, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q4);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q4);
+            q4 += 32;
+
+            vector signed char q4x00 = (vector signed char)vec_and(qxs0, lowMask);
+            vector signed char q4x01 = (vector signed char)vec_sr(qxs0, v4);
+            vector signed char q4x10 = (vector signed char)vec_and(qxs1, lowMask);
+            vector signed char q4x11 = (vector signed char)vec_sr(qxs1, v4);
+
+            q4x00 = vec_perm(values, values, (vector unsigned char)q4x00);
+            q4x01 = vec_perm(values, values, (vector unsigned char)q4x01);
+            q4x10 = vec_perm(values, values, (vector unsigned char)q4x10);
+            q4x11 = vec_perm(values, values, (vector unsigned char)q4x11);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q4x00, q8y0), vec_mulo(q4x00, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q4x01, q8y1), vec_mulo(q4x01, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q4x10, q8y2), vec_mulo(q4x10, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q4x11, q8y3), vec_mulo(q4x11, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(((sc[0] & 0xf) | ((h << 4) & 0x30)) - 32);
+            const uint16_t ls1 = (uint16_t)(((sc[0] >>  4) | ((h << 2) & 0x30)) - 32);
+            h >>= 4;
+            sc ++;
+
+            vector signed short vscales01 = vec_splats((int16_t)ls0);
+            vector signed short vscales23 = vec_splats((int16_t)ls1);
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+
+    const __m128i values128 = __lsx_vld((const __m128i*)kvalues_iq4nl, 0);
+    const __m128i m4b  = __lsx_vreplgr2vr_b(0x0f);
+
+    __m256 accum = (__m256)__lasx_xvldi(0);
+    __m256i tmp1;
+    __m128i tmp0, tmp2, tmp3, tmp4, mask_8f, mask;
+
+    mask_8f = __lsx_vreplgr2vr_b(0x8f);
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        uint16_t sh = x[ibl].scales_h;
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        __m128i zero = __lsx_vldi(0);
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q4bits_1 = __lsx_vld((const __m128i*)qs, 0);  qs += 16;
+            const __m128i q4bits_2 = __lsx_vld((const __m128i*)qs, 0);  qs += 16;
+            const __m256i q8b_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8b_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            tmp2 = __lsx_vand_v(__lsx_vand_v(__lsx_vsrli_h(q4bits_1, 4), m4b), mask_8f);
+            tmp0 = __lsx_vori_b(tmp2, 0x10);
+            mask = __lsx_vsle_b(zero, tmp2);
+            tmp3 = __lsx_vand_v(tmp0, mask);
+            tmp3 = __lsx_vshuf_b(values128, zero, tmp3);
+
+            tmp2 = __lsx_vand_v(__lsx_vand_v(q4bits_1, m4b), mask_8f);
+            tmp0 = __lsx_vori_b(tmp2, 0x10);
+            mask = __lsx_vsle_b(zero, tmp2);
+            tmp4 = __lsx_vand_v(tmp0, mask);
+            tmp4 = __lsx_vshuf_b(values128, zero, tmp4);
+
+            const __m256i q4b_1 = lasx_insertf128(tmp3, tmp4);
+
+            tmp2 = __lsx_vand_v(__lsx_vand_v(__lsx_vsrli_h(q4bits_2, 4), m4b), mask_8f);
+            tmp0 = __lsx_vori_b(tmp2, 0x10);
+            mask = __lsx_vsle_b(zero, tmp2);
+            tmp3 = __lsx_vand_v(tmp0, mask);
+            tmp3 = __lsx_vshuf_b(values128, zero, tmp3);
+
+            tmp2 = __lsx_vand_v(__lsx_vand_v(q4bits_2, m4b), mask_8f);
+            tmp0 = __lsx_vori_b(tmp2, 0x10);
+            mask = __lsx_vsle_b(zero, tmp2);
+            tmp4 = __lsx_vand_v(tmp0, mask);
+            tmp4 = __lsx_vshuf_b(values128, zero, tmp4);
+
+            const __m256i q4b_2 = lasx_insertf128(tmp3, tmp4);
+
+            const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
+            const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
+            const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;
+            const int16_t ls2 = ((x[ibl].scales_l[ib/2] >>  4) | ((sh << 2) & 0x30)) - 32;
+            sh >>= 4;
+            __m256i tmp5, tmp6;
+            tmp1 = __lasx_xvreplgr2vr_h(ls1);
+            tmp5 = __lasx_xvmulwev_w_h(p16_1, tmp1);
+            tmp6 = __lasx_xvmulwod_w_h(p16_1, tmp1);
+            const __m256i p_1 = __lasx_xvadd_w(tmp5, tmp6);
+            tmp1 = __lasx_xvreplgr2vr_h(ls2);
+            tmp5 = __lasx_xvmulwev_w_h(p16_2, tmp1);
+            tmp6 = __lasx_xvmulwod_w_h(p16_2, tmp1);
+            const __m256i p_2 = __lasx_xvadd_w(tmp5, tmp6);
+            sumi1 = __lasx_xvadd_w(p_1, sumi1);
+            sumi2 = __lasx_xvadd_w(p_2, sumi2);
+        }
+        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+                __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accum);
+    }
+
+    *s = hsum_float_8(accum);
+
+#else
+    float sumf = 0;
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        uint16_t h = x[ibl].scales_h;
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
+            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
+            h >>= 4;
+            const float d1 = d4d8*(ls1 - 32);
+            const float d2 = d4d8*(ls2 - 32);
+            int sumi1 = 0, sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d1 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+            sumi1 = sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d2 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+        }
+    }
+    *s = sumf;
+#endif
+}
+
+// ============================ 4-bit non-linear quants
+
+void quantize_row_iq4_nl(const float * restrict x, void * restrict y, int64_t k) {
+    assert(k % QK4_NL == 0);
+    quantize_row_iq4_nl_ref(x, y, k);
+}
+
+void quantize_row_iq4_xs(const float * restrict x, void * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    quantize_iq4_xs(x, y, 1, k, NULL);
+}
diff --git a/ggml/src/ggml-cpu/ggml-cpu-quants.h b/ggml/src/ggml-cpu/ggml-cpu-quants.h
new file mode 100644
index 00000000000..e33d9d473ea
--- /dev/null
+++ b/ggml/src/ggml-cpu/ggml-cpu-quants.h
@@ -0,0 +1,63 @@
+#pragma once
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML CPU internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+// Dot product
+void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/ggml/src/ggml-cpu/ggml-cpu-traits.cpp b/ggml/src/ggml-cpu/ggml-cpu-traits.cpp
new file mode 100644
index 00000000000..62a0712dabb
--- /dev/null
+++ b/ggml/src/ggml-cpu/ggml-cpu-traits.cpp
@@ -0,0 +1,36 @@
+#include "ggml-cpu-traits.h"
+
+#include "ggml-backend-impl.h"
+#include "ggml-backend.h"
+
+namespace ggml::cpu {
+tensor_traits::~tensor_traits() {}
+
+extra_buffer_type::~extra_buffer_type() {}
+}  // namespace ggml::cpu
+
+bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra->context) {
+            auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
+            auto tensor_traits = buf_extra->get_tensor_traits(op);
+            if (tensor_traits && tensor_traits->compute_forward(params, op)) {
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
+bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra->context) {
+            auto buf_extra     = (ggml::cpu::extra_buffer_type *) extra->context;
+            auto tensor_traits = buf_extra->get_tensor_traits(op);
+            if (tensor_traits && tensor_traits->work_size(n_threads, op, *size)) {
+                return true;
+            }
+        }
+    }
+    return false;
+}
diff --git a/ggml/src/ggml-cpu/ggml-cpu-traits.h b/ggml/src/ggml-cpu/ggml-cpu-traits.h
new file mode 100644
index 00000000000..99a6186b1d6
--- /dev/null
+++ b/ggml/src/ggml-cpu/ggml-cpu-traits.h
@@ -0,0 +1,38 @@
+#pragma once
+#include "ggml-backend-impl.h"
+#include "ggml-cpu-impl.h"
+#include "ggml.h"
+
+#ifdef __cplusplus
+#    include <vector>
+extern "C" {
+#endif
+
+// return true if op part of extra "accelerator"
+bool ggml_cpu_extra_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op);
+bool ggml_cpu_extra_work_size(int n_threads, const struct ggml_tensor * op, size_t * size);
+
+#ifdef __cplusplus
+}
+
+namespace ggml::cpu {
+// register in tensor->extra
+class tensor_traits {
+  public:
+    virtual ~tensor_traits();
+    virtual bool work_size(int n_threads, const struct ggml_tensor * op, size_t & size)        = 0;
+    virtual bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) = 0;
+};
+
+class extra_buffer_type {
+  public:
+    virtual ~extra_buffer_type();
+    virtual bool            supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) = 0;
+    virtual tensor_traits * get_tensor_traits(const struct ggml_tensor * op)                   = 0;
+};
+}  // namespace ggml::cpu
+
+// implemented in ggml-cpu.cpp.
+std::vector<ggml_backend_buffer_type_t> & ggml_backend_cpu_get_extra_buffers_type();
+
+#endif
diff --git a/ggml/src/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
similarity index 93%
rename from ggml/src/ggml-cpu.c
rename to ggml/src/ggml-cpu/ggml-cpu.c
index de1de18ecea..67e67a08968 100644
--- a/ggml/src/ggml-cpu.c
+++ b/ggml/src/ggml-cpu/ggml-cpu.c
@@ -1,13 +1,16 @@
 #define _CRT_SECURE_NO_DEPRECATE // Disables "unsafe" warnings on Windows
 #define _USE_MATH_DEFINES // For M_PI on MSVC
 
-#include "ggml-aarch64.h"
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
+#include "ggml-cpu-traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
 #include "ggml-quants.h"
+#include "ggml-cpu-quants.h"
+#include "ggml-threading.h"
+#include "amx/amx.h"
 #include "ggml.h"
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
@@ -42,7 +45,7 @@
 #endif
 
 #ifdef GGML_USE_LLAMAFILE
-#include <llamafile/sgemm.h>
+#include "llamafile/sgemm.h"
 #endif
 
 #if defined(_MSC_VER)
@@ -104,16 +107,14 @@ static ggml_fp16_t ggml_table_gelu_f16[1 << 16];
 // precomputed quick gelu table for f16 (128 KB)
 static ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
 
-// precomputed f32 table for f16 (256 KB) (ggml-impl.h)
-float ggml_table_f32_f16[1 << 16];
-
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
     int has_neon;
+    int has_dotprod;
     int has_i8mm;
     int has_sve;
     int sve_cnt;
-} ggml_arm_arch_features = {-1, -1, -1, 0};
+} ggml_arm_arch_features = {-1, -1, -1, -1, 0};
 #endif
 
 
@@ -125,8 +126,7 @@ struct ggml_arm_arch_features_type {
 #endif
 #include <windows.h>
 
-
-#if !defined(__clang__)
+#if defined(_MSC_VER) && !defined(__clang__)
 #define GGML_CACHE_ALIGN __declspec(align(GGML_CACHE_LINE))
 
 typedef volatile LONG atomic_int;
@@ -223,10 +223,6 @@ typedef void * thread_ret_t;
 
 typedef pthread_t ggml_thread_t;
 
-#ifdef GGML_USE_CPU_HBM
-#include <hbwmalloc.h>
-#endif
-
 #if defined(__APPLE__)
 #include <unistd.h>
 #include <mach/mach.h>
@@ -261,11 +257,13 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .nrows                    = 1,
     },
     [GGML_TYPE_F16] = {
+        .from_float               = (ggml_from_float_t) ggml_fp32_to_fp16_row,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f16,
         .vec_dot_type             = GGML_TYPE_F16,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q4_0] = {
+        .from_float               = quantize_row_q4_0,
         .vec_dot                  = ggml_vec_dot_q4_0_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
 #if defined (__ARM_FEATURE_MATMUL_INT8)
@@ -275,6 +273,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
 #endif
     },
     [GGML_TYPE_Q4_1] = {
+        .from_float               = quantize_row_q4_1,
         .vec_dot                  = ggml_vec_dot_q4_1_q8_1,
         .vec_dot_type             = GGML_TYPE_Q8_1,
 #if defined (__ARM_FEATURE_MATMUL_INT8)
@@ -283,28 +282,20 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .nrows                    = 1,
 #endif
     },
-    [4] = { // GGML_TYPE_Q4_2
-        .vec_dot                  = NULL,
-        .vec_dot_type             = GGML_TYPE_COUNT,
-        .nrows                    = 1,
-    },
-    [5] = { // GGML_TYPE_Q4_3
-        .vec_dot                  = NULL,
-        .vec_dot_type             = GGML_TYPE_COUNT,
-        .nrows                    = 1,
-    },
     [GGML_TYPE_Q5_0] = {
+        .from_float               = quantize_row_q5_0,
         .vec_dot                  = ggml_vec_dot_q5_0_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q5_1] = {
+        .from_float               = quantize_row_q5_1,
         .vec_dot                  = ggml_vec_dot_q5_1_q8_1,
         .vec_dot_type             = GGML_TYPE_Q8_1,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q8_0] = {
-        .from_float_to_mat        = quantize_mat_q8_0,
+        .from_float               = quantize_row_q8_0,
         .vec_dot                  = ggml_vec_dot_q8_0_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
 #if defined (__ARM_FEATURE_MATMUL_INT8)
@@ -314,114 +305,112 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
 #endif
     },
     [GGML_TYPE_Q8_1] = {
+        .from_float               = quantize_row_q8_1,
         .vec_dot_type             = GGML_TYPE_Q8_1,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q2_K] = {
+        .from_float               = quantize_row_q2_K,
         .vec_dot                  = ggml_vec_dot_q2_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q3_K] = {
+        .from_float               = quantize_row_q3_K,
         .vec_dot                  = ggml_vec_dot_q3_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q4_K] = {
+        .from_float               = quantize_row_q4_K,
         .vec_dot                  = ggml_vec_dot_q4_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q5_K] = {
+        .from_float               = quantize_row_q5_K,
         .vec_dot                  = ggml_vec_dot_q5_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_Q6_K] = {
+        .from_float               = quantize_row_q6_K,
         .vec_dot                  = ggml_vec_dot_q6_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ2_XXS] = {
+        .from_float               = NULL,
         .vec_dot                  = ggml_vec_dot_iq2_xxs_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ2_XS] = {
+        .from_float               = NULL,
         .vec_dot                  = ggml_vec_dot_iq2_xs_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ3_XXS] = {
+        // NOTE: from_float for iq3 and iq2_s was removed because these quants require initialization in ggml_quantize_init
+        //.from_float               = quantize_row_iq3_xxs,
         .vec_dot                  = ggml_vec_dot_iq3_xxs_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ3_S] = {
+        //.from_float               = quantize_row_iq3_s,
         .vec_dot                  = ggml_vec_dot_iq3_s_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ2_S] = {
+        //.from_float               = quantize_row_iq2_s,
         .vec_dot                  = ggml_vec_dot_iq2_s_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ1_S] = {
+        .from_float               = NULL,
         .vec_dot                  = ggml_vec_dot_iq1_s_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ1_M] = {
+        .from_float               = NULL,
         .vec_dot                  = ggml_vec_dot_iq1_m_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ4_NL] = {
+        .from_float               = quantize_row_iq4_nl,
         .vec_dot                  = ggml_vec_dot_iq4_nl_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
         .nrows                    = 1,
     },
     [GGML_TYPE_IQ4_XS] = {
+        .from_float               = quantize_row_iq4_xs,
         .vec_dot                  = ggml_vec_dot_iq4_xs_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
+    [GGML_TYPE_Q8_K] = {
+        .from_float               = quantize_row_q8_K,
+    },
     [GGML_TYPE_BF16] = {
+        .from_float               = (ggml_from_float_t) ggml_fp32_to_bf16_row,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_bf16,
         .vec_dot_type             = GGML_TYPE_BF16,
         .nrows                    = 1,
     },
-    [GGML_TYPE_Q4_0_4_4] = {
-        .vec_dot                  = NULL,
-        .vec_dot_type             = GGML_TYPE_Q8_0,
-        .nrows                    = 1,
-        .ncols                    = 4,
-        .gemv                     = ggml_gemv_q4_0_4x4_q8_0,
-        .gemm                     = ggml_gemm_q4_0_4x4_q8_0,
-    },
-    [GGML_TYPE_Q4_0_4_8] = {
-        .vec_dot                  = NULL,
-        .vec_dot_type             = GGML_TYPE_Q8_0,
-        .nrows                    = 1,
-        .ncols                    = 4,
-        .gemv                     = ggml_gemv_q4_0_4x8_q8_0,
-        .gemm                     = ggml_gemm_q4_0_4x8_q8_0,
-    },
-    [GGML_TYPE_Q4_0_8_8] = {
-        .vec_dot                  = NULL,
-        .vec_dot_type             = GGML_TYPE_Q8_0,
-        .nrows                    = 1,
-        .ncols                    = 8,
-        .gemv                     = ggml_gemv_q4_0_8x8_q8_0,
-        .gemm                     = ggml_gemm_q4_0_8x8_q8_0,
-    },
     [GGML_TYPE_TQ1_0] = {
+        .from_float               = quantize_row_tq1_0,
         .vec_dot                  = ggml_vec_dot_tq1_0_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
     },
     [GGML_TYPE_TQ2_0] = {
+        .from_float               = quantize_row_tq2_0,
         .vec_dot                  = ggml_vec_dot_tq2_0_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
         .nrows                    = 1,
@@ -465,21 +454,21 @@ const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type
 #define GGML_F32x4_ADD          vaddq_f32
 #define GGML_F32x4_MUL          vmulq_f32
 #define GGML_F32x4_REDUCE_ONE(x) vaddvq_f32(x)
-#define GGML_F32x4_REDUCE(res, x)                  \
-{                                                  \
-    int offset = GGML_F32_ARR >> 1;                \
-    for (int i = 0; i < offset; ++i) {             \
-        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
-    }                                              \
-    offset >>= 1;                                  \
-    for (int i = 0; i < offset; ++i) {             \
-        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
-    }                                              \
-    offset >>= 1;                                  \
-    for (int i = 0; i < offset; ++i) {             \
-        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
-    }                                              \
-    (res) = GGML_F32x4_REDUCE_ONE((x)[0]);         \
+#define GGML_F32x4_REDUCE(res, x)                       \
+{                                                       \
+    int offset = GGML_F32_ARR >> 1;                     \
+    for (int i = 0; i < offset; ++i) {                  \
+        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]);      \
+    }                                                   \
+    offset >>= 1;                                       \
+    for (int i = 0; i < offset; ++i) {                  \
+        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]);      \
+    }                                                   \
+    offset >>= 1;                                       \
+    for (int i = 0; i < offset; ++i) {                  \
+        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]);      \
+    }                                                   \
+    (res) = (ggml_float) GGML_F32x4_REDUCE_ONE((x)[0]); \
 }
 
 #define GGML_F32_VEC        GGML_F32x4
@@ -594,7 +583,7 @@ do {                                                                  \
     for (int i = 0; i < offset; ++i) {                                \
         x[i] = _mm512_add_ps(x[i], x[offset+i]);                      \
     }                                                                 \
-    res = _mm512_reduce_add_ps(x[0]);                                 \
+    res = (ggml_float) _mm512_reduce_add_ps(x[0]);                    \
 } while (0)
 
 // TODO: is this optimal ?
@@ -644,7 +633,7 @@ do {                                                              \
     for (int i = 0; i < offset; ++i) {                            \
         x[i] = _mm512_add_ps(x[i], x[offset+i]);                  \
     }                                                             \
-    res = _mm512_reduce_add_ps(x[0]);                             \
+    res = (ggml_float) _mm512_reduce_add_ps(x[0]);                \
 } while (0)
 
 #define GGML_F16_VEC                GGML_F32Cx16
@@ -655,8 +644,8 @@ do {                                                              \
 #define GGML_F16_VEC_FMA            GGML_F32Cx16_FMA
 #define GGML_F16_VEC_ADD            GGML_F32Cx16_ADD
 #define GGML_F16_VEC_MUL            GGML_F32Cx16_MUL
-#define GGML_F16_VEC_REDUCE         GGML_F32Cx16_REDUCE
 
+#define GGML_F16_VEC_REDUCE         GGML_F32Cx16_REDUCE
 #elif defined(__AVX__)
 
 #define GGML_SIMD
@@ -725,7 +714,7 @@ do {                                                              \
 #define GGML_F32Cx8_LOAD(x)     _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x)))
 #define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0))
 #else
-static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
+static inline __m256 __avx_f32cx8_load(const ggml_fp16_t * x) {
     float tmp[8];
 
     for (int i = 0; i < 8; i++) {
@@ -1148,28 +1137,28 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
 #define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a)
 #define GGML_F32x4_ADD     __lsx_vfadd_s
 #define GGML_F32x4_MUL     __lsx_vfmul_s
-#define GGML_F32x4_REDUCE(res, x)                                 \
-{                                                                 \
-    int offset = GGML_F32_ARR >> 1;                               \
-    for (int i = 0; i < offset; ++i) {                            \
-        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                     \
-    }                                                             \
-    offset >>= 1;                                                 \
-    for (int i = 0; i < offset; ++i) {                            \
-        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                     \
-    }                                                             \
-    offset >>= 1;                                                 \
-    for (int i = 0; i < offset; ++i) {                            \
-        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                     \
-    }                                                             \
-    __m128i tmp = __lsx_vsrli_d((__m128i)x[0], 32); \
-    tmp = (__m128i)__lsx_vfadd_s((__m128)tmp, x[0]); \
-    tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \
-    const __m128 t0 = __lsx_vshuf4i_w(tmp, 0x88); \
-    tmp = __lsx_vsrli_d((__m128i)t0, 32); \
-    tmp = (__m128i)__lsx_vfadd_s((__m128)tmp, t0); \
-    tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \
-    res = (ggml_float) __lsx_vpickve2gr_w(__lsx_vshuf4i_w(tmp, 0x88), 0);        \
+#define GGML_F32x4_REDUCE(res, x)                                                     \
+{                                                                                     \
+    int offset = GGML_F32_ARR >> 1;                                                   \
+    for (int i = 0; i < offset; ++i) {                                                \
+        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
+    }                                                                                 \
+    offset >>= 1;                                                                     \
+    for (int i = 0; i < offset; ++i) {                                                \
+        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
+    }                                                                                 \
+    offset >>= 1;                                                                     \
+    for (int i = 0; i < offset; ++i) {                                                \
+        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
+    }                                                                                 \
+    __m128i tmp     = __lsx_vsrli_d((__m128i) x[0], 32);                              \
+    tmp             = (__m128i) __lsx_vfadd_s((__m128) tmp, x[0]);                    \
+    tmp             = __lsx_vpickev_w(__lsx_vldi(0), tmp);                            \
+    const __m128 t0 = __lsx_vshuf4i_w(tmp, 0x88);                                     \
+    tmp             = __lsx_vsrli_d((__m128i) t0, 32);                                \
+    tmp             = (__m128i) __lsx_vfadd_s((__m128) tmp, t0);                      \
+    tmp             = __lsx_vpickev_w(__lsx_vldi(0), tmp);                            \
+    res             = (ggml_float) __lsx_vpickve2gr_w(__lsx_vshuf4i_w(tmp, 0x88), 0); \
 }
 
 #define GGML_F32_VEC        GGML_F32x4
@@ -1337,31 +1326,18 @@ struct ggml_compute_state {
     int ith;
 };
 
-struct ggml_compute_params {
-    // ith = thread index, nth = number of threads
-    int ith, nth;
-
-    // work buffer for all threads
-    size_t wsize;
-    void * wdata;
-
-    struct ggml_threadpool * threadpool;
-};
-
 //
 // fundamental operations
 //
 
 inline static void ggml_vec_set_i8(const int n, int8_t * x, const int8_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
-
 inline static void ggml_vec_set_i16(const int n, int16_t * x, const int16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
 
-inline static void ggml_vec_set_i32(const int n, int32_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+inline static void ggml_vec_set_i32(const int n, int32_t * x, const int32_t   v) { for (int i = 0; i < n; ++i) x[i] = v;    }
+inline static void ggml_vec_cpy_i32(const int n, int32_t * y, const int32_t * x) { for (int i = 0; i < n; ++i) y[i] = x[i]; }
 
 inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
-
 inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
-
 inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i] + y[i]; }
 inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float   v) { for (int i = 0; i < n; ++i) z[i]  = x[i] + v;    }
 inline static void ggml_vec_acc_f32 (const int n, float * y, const float * x)                  { for (int i = 0; i < n; ++i) y[i] += x[i];        }
@@ -1449,8 +1425,12 @@ static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t
     sumf += (ggml_float)_mm512_reduce_add_ps(c2);
 
 #undef LOAD
-#elif defined(__AVX2__)
+#elif defined(__AVX2__) || defined(__AVX__)
+#if defined(__AVX2__)
 #define LOAD(p) _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16))
+#else
+#define LOAD(p) _mm256_castsi256_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16)), (_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_bsrli_si128(_mm_loadu_si128((const __m128i *)(p)), 8)), 16)), 1))
+#endif
     __m256 c1 = _mm256_setzero_ps();
     __m256 c2 = _mm256_setzero_ps();
     __m256 c3 = _mm256_setzero_ps();
@@ -2250,24 +2230,9 @@ struct ggml_state {
     struct ggml_numa_nodes numa;
 };
 
-// global state
 static struct ggml_state g_state = {0};
-static atomic_flag g_state_critical = ATOMIC_FLAG_INIT;
-
-// TODO: move to threading file
-// critical section via spin lock
-void ggml_critical_section_start(void) {
-    while (atomic_flag_test_and_set(&g_state_critical)) {
-        // spin
-        sched_yield();
-    }
-}
-
-void ggml_critical_section_end(void) {
-    atomic_flag_clear(&g_state_critical);
-}
 
-static void ggml_barrier(struct ggml_threadpool * tp) {
+void ggml_barrier(struct ggml_threadpool * tp) {
     int n_threads = atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed);
     if (n_threads == 1) {
         return;
@@ -2360,7 +2325,7 @@ void ggml_numa_init(enum ggml_numa_strategy numa_flag) {
     // figure out which node we're on
     uint current_cpu;
     int getcpu_ret = 0;
-#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 28) || defined(__COSMOPOLITAN__)
+#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 33) || defined(__COSMOPOLITAN__)
     getcpu_ret = getcpu(&current_cpu, &g_state.numa.current_node);
 #else
     // old glibc doesn't have a wrapper for this call. Fall back on direct syscall
@@ -2421,7 +2386,7 @@ bool ggml_is_numa(void) {
 #endif
 
 #if !defined(HWCAP2_I8MM)
-#define HWCAP2_I8MM 0
+#define HWCAP2_I8MM (1 << 13)
 #endif
 
 static void ggml_init_arm_arch_features(void) {
@@ -2430,6 +2395,7 @@ static void ggml_init_arm_arch_features(void) {
     uint32_t hwcap2 = getauxval(AT_HWCAP2);
 
     ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD);
+    ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
     ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
     ggml_arm_arch_features.has_sve  = !!(hwcap & HWCAP_SVE);
 
@@ -2444,6 +2410,11 @@ static void ggml_init_arm_arch_features(void) {
     }
     ggml_arm_arch_features.has_neon = oldp;
 
+    if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_dotprod = oldp;
+
     if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) {
         oldp = 0;
     }
@@ -2997,8 +2968,8 @@ static void ggml_compute_forward_dup_f16(
                         id += ne00 * (ne01 - ir1);
                     }
                 }
-            } else if (ggml_get_type_traits(dst->type)->from_float) {
-                ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
+            } else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
+                ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
                 float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
 
                 size_t id = 0;
@@ -3278,8 +3249,8 @@ static void ggml_compute_forward_dup_bf16(
                         id += ne00 * (ne01 - ir1);
                     }
                 }
-            } else if (ggml_get_type_traits(dst->type)->from_float) {
-                ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
+            } else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
+                ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
                 float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
 
                 size_t id = 0;
@@ -3594,8 +3565,8 @@ static void ggml_compute_forward_dup_f32(
                         id += rs * (ne01 - ir1);
                     }
                 }
-            } else if (ggml_get_type_traits(dst->type)->from_float) {
-                ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dst->type)->from_float;
+            } else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
+                ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
 
                 size_t id = 0;
                 size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type));
@@ -4377,7 +4348,7 @@ static void ggml_compute_forward_add_q_f32(
     const enum ggml_type type = src0->type;
     const enum ggml_type dtype = dst->type;
     ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float;
-    ggml_from_float_t const quantize_row_q = ggml_get_type_traits(dtype)->from_float;
+    ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dtype)->from_float;
 
     // we don't support permuted src0 or src1
     GGML_ASSERT(nb00 == ggml_type_size(type));
@@ -4496,9 +4467,6 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-        case GGML_TYPE_Q4_0_8_8:
             {
                 ggml_compute_forward_add_q_f32(params, dst);
             } break;
@@ -4679,7 +4647,7 @@ static void ggml_compute_forward_add1_q_f32(
 
     const enum ggml_type type = src0->type;
     ggml_to_float_t const dequantize_row_q = ggml_get_type_traits(type)->to_float;
-    ggml_from_float_t const quantize_row_q = ggml_get_type_traits(type)->from_float;
+    ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(type)->from_float;
 
     // we don't support permuted src0
     GGML_ASSERT(nb00 == ggml_type_size(type));
@@ -4876,9 +4844,6 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-        case GGML_TYPE_Q4_0_8_8:
             {
                 ggml_compute_forward_add1_q_f32(params, dst);
             } break;
@@ -5006,9 +4971,6 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-        case GGML_TYPE_Q4_0_8_8:
         default:
             {
                 GGML_ABORT("fatal error");
@@ -7325,6 +7287,7 @@ static void ggml_compute_forward_group_norm(
 static void ggml_compute_forward_mul_mat_one_chunk(
     const struct ggml_compute_params * params,
     struct ggml_tensor * dst,
+    const enum ggml_type type,
     const int64_t num_rows_per_vec_dot,
     const int64_t ir0_start,
     const int64_t ir0_end,
@@ -7336,8 +7299,6 @@ static void ggml_compute_forward_mul_mat_one_chunk(
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    const enum ggml_type type = src0->type;
-
     const bool src1_cont = ggml_is_contiguous(src1);
 
     ggml_vec_dot_t const vec_dot      = type_traits_cpu[type].vec_dot;
@@ -7425,16 +7386,9 @@ static void ggml_compute_forward_mul_mat(
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const enum ggml_type type = src0->type;
-
-    enum ggml_type           const vec_dot_type         = type_traits_cpu[type].vec_dot_type;
-    ggml_from_float_t        const from_float           = ggml_get_type_traits(vec_dot_type)->from_float;
-    ggml_from_float_to_mat_t const from_float_to_mat    = type_traits_cpu[vec_dot_type].from_float_to_mat;
-    int64_t                  const vec_dot_num_rows     = type_traits_cpu[type].nrows;
-    int64_t                  const matmul_num_cols      = type_traits_cpu[type].ncols;
-    int64_t                  const blck_size_interleave = ggml_get_type_traits(type)->blck_size_interleave;
-    ggml_gemv_t              const gemv                 = type_traits_cpu[type].gemv;
-    ggml_gemm_t              const gemm                 = type_traits_cpu[type].gemm;
+    enum ggml_type           const vec_dot_type         = type_traits_cpu[src0->type].vec_dot_type;
+    ggml_from_float_t        const from_float           = type_traits_cpu[vec_dot_type].from_float;
+    int64_t                  const vec_dot_num_rows     = type_traits_cpu[src0->type].nrows;
 
     GGML_ASSERT(ne0 == ne01);
     GGML_ASSERT(ne1 == ne11);
@@ -7442,7 +7396,7 @@ static void ggml_compute_forward_mul_mat(
     GGML_ASSERT(ne3 == ne13);
 
     // we don't support permuted src0 or src1
-    GGML_ASSERT(nb00 == ggml_type_size(type));
+    GGML_ASSERT(nb00 == ggml_type_size(src0->type));
     GGML_ASSERT(nb10 == ggml_type_size(src1->type));
 
     // dst cannot be transposed or permuted
@@ -7454,6 +7408,7 @@ static void ggml_compute_forward_mul_mat(
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
+    // TODO: extract to "extra_op"
 #if GGML_USE_LLAMAFILE
     // broadcast factors
     const int64_t r2 = ne12 / ne02;
@@ -7493,19 +7448,10 @@ UseGgmlGemm1:;
 
         for (int64_t i13 = 0; i13 < ne13; ++i13) {
             for (int64_t i12 = 0; i12 < ne12; ++i12) {
-                int64_t i11_processed = 0;
-                if ((ggml_n_dims(src1) == 2) && from_float_to_mat && gemm) {
-                    for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
-                        from_float_to_mat((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
-                                          (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
-                                          4, ne10, blck_size_interleave);
-                    }
-                    i11_processed = ne11 - ne11 % 4;
-                }
-                for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) {
+                for (int64_t i11 = ith; i11 < ne11; i11 += nth) {
                     from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
-                           (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
-                           ne10);
+                               (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
+                                ne10);
                 }
             }
         }
@@ -7548,14 +7494,6 @@ UseGgmlGemm2:;
     // This is the size of the rest of the dimensions of the result
     const int64_t nr1 = ne1 * ne2 * ne3;
 
-    // dot kernels can handle 1 row and col at a time, but mmla kernels can process 2 rows and cols
-    int64_t num_rows_per_vec_dot = vec_dot_num_rows;
-    // TODO: currently the mmla kernels support only even numbered rows/cols.
-    // this check can be removed once they are extended to support odd numbered rows/cols too
-    if ((nr0 % 2 != 0) || (ne11 % 2 != 0)) {
-        num_rows_per_vec_dot = 1;
-    }
-
     // Now select a reasonable chunk size.
     int chunk_size = 16;
 
@@ -7583,28 +7521,6 @@ UseGgmlGemm2:;
     const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
     const int64_t dr1 = (nr1 + nchunk1 - 1) / nchunk1;
 
-    if ((ggml_n_dims(src0) == 2) && gemv) {
-        const void * src1_wdata      = (src1->type == vec_dot_type) ? src1->data : params->wdata;
-        const size_t src1_col_stride = ggml_is_contiguous(src1) || src1->type != vec_dot_type ? ggml_row_size(vec_dot_type, ne10) : nb11;
-        int64_t src0_start = (ith * ne01) / nth;
-        int64_t src0_end   = ((ith + 1) * ne01) / nth;
-        src0_start = (src0_start % matmul_num_cols) ? src0_start + matmul_num_cols - (src0_start % matmul_num_cols): src0_start;
-        src0_end   = (src0_end   % matmul_num_cols) ? src0_end   + matmul_num_cols - (src0_end   % matmul_num_cols): src0_end;
-        if (src0_start >= src0_end) return;
-
-        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
-        if (gemm && (ne11 > 3)) {
-            gemm(ne00, (float *)((char *) dst->data) + src0_start, ne01, (const char *) src0->data + src0_start * nb01,
-                 (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
-        }
-        for (int iter = gemm ? ne11 - ne11 % 4 : 0; iter < ne11; iter++) {
-            gemv(ne00, (float *)((char *) dst->data + (iter * nb1)) + src0_start, ne01,
-                 (const char *) src0->data + src0_start * nb01, (const char *) src1_wdata + (src1_col_stride * iter), 1,
-                 src0_end - src0_start);
-        }
-        return;
-    }
-
     // The first chunk comes from our thread_id, the rest will get auto-assigned.
     int current_chunk = ith;
 
@@ -7618,7 +7534,16 @@ UseGgmlGemm2:;
         const int64_t ir1_start = dr1 * ith1;
         const int64_t ir1_end = MIN(ir1_start + dr1, nr1);
 
-        ggml_compute_forward_mul_mat_one_chunk(params, dst, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
+        // dot kernels can handle 1 row and col at a time, but mmla kernels can process 2 rows and cols
+        int64_t num_rows_per_vec_dot = vec_dot_num_rows;
+
+        // these checks are needed to avoid crossing dim1 boundaries
+        // can be optimized, but the logic would become more complicated, so keeping it like this for simplicity
+        if ((nr0 % 2 != 0) || (ne11 % 2 != 0) || ((ir0_end - ir0_start) % 2 != 0) || ((ir1_end - ir1_start) % 2 != 0)) {
+            num_rows_per_vec_dot = 1;
+        }
+
+        ggml_compute_forward_mul_mat_one_chunk(params, dst, src0->type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
 
         if (nth >= nchunk0 * nchunk1) {
             break;
@@ -7649,9 +7574,7 @@ static void ggml_compute_forward_mul_mat_id(
 
     ggml_vec_dot_t    const vec_dot         = type_traits_cpu[type].vec_dot;
     enum ggml_type    const vec_dot_type    = type_traits_cpu[type].vec_dot_type;
-    ggml_from_float_t const from_float      = ggml_get_type_traits(vec_dot_type)->from_float;
-    int64_t           const matmul_num_cols = type_traits_cpu[type].ncols;
-    ggml_gemv_t       const gemv            = type_traits_cpu[type].gemv;
+    ggml_from_float_t const from_float      = type_traits_cpu[vec_dot_type].from_float;
 
     // we don't support permuted src0 or src1
     GGML_ASSERT(nb00 == ggml_type_size(type));
@@ -7737,34 +7660,6 @@ static void ggml_compute_forward_mul_mat_id(
         const int64_t nr0 = ne01; // src0 rows
         const int64_t nr1 = cne1; // src1 rows
 
-        if (((ggml_n_dims(src0) - 1) == 2) && gemv) {
-            int64_t src0_cur_start = (ith * ne01) / nth;
-            int64_t src0_cur_end   = ((ith + 1) * ne01) / nth;
-            src0_cur_start = (src0_cur_start % matmul_num_cols) ? src0_cur_start + matmul_num_cols - (src0_cur_start % matmul_num_cols): src0_cur_start;
-            src0_cur_end   = (src0_cur_end % matmul_num_cols) ? src0_cur_end + matmul_num_cols - (src0_cur_end % matmul_num_cols): src0_cur_end;
-            if (src0_cur_start >= src0_cur_end) return;
-
-            for (int ir1 = 0; ir1 < nr1; ir1++) {
-                struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1);
-                const int id       = row_mapping.i1; // selected expert index
-
-                const int64_t  i11 = id % ne11;
-                const int64_t  i12 = row_mapping.i2; // row index in src1
-
-                const int64_t  i1 = id;  // selected expert index
-                const int64_t  i2 = i12; // row
-
-                const char * src1_col = (const char *) wdata +
-                    (src1_cont || src1->type != vec_dot_type
-                    ? (i11        + i12 * ne11) * row_size
-                    : (i11 * nb11 + i12 * nb12));
-
-                gemv(ne00, (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01,
-                     (const char *) src0_cur + src0_cur_start * nb01, src1_col, 1, src0_cur_end - src0_cur_start);
-            }
-            continue;
-        }
-
         // distribute the thread work across the inner or outer loop based on which one is larger
 
         const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
@@ -8072,9 +7967,6 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-        case GGML_TYPE_Q4_0_8_8:
             {
                 ggml_compute_forward_out_prod_q_f32(params, dst);
             } break;
@@ -8227,6 +8119,77 @@ static void ggml_compute_forward_set_f32(
     }
 }
 
+static void ggml_compute_forward_set_i32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
+
+    // view src0 and dst with these strides and data offset inbytes during set
+    // nb0 is implicitly element_size because src0 and dst are contiguous
+    size_t nb1     = ((int32_t *) dst->op_params)[0];
+    size_t nb2     = ((int32_t *) dst->op_params)[1];
+    size_t nb3     = ((int32_t *) dst->op_params)[2];
+    size_t offset  = ((int32_t *) dst->op_params)[3];
+    bool   inplace = (bool) ((int32_t *) dst->op_params)[4];
+
+    if (!inplace) {
+        if (params->ith == 0) {
+            // memcpy needs to be synchronized across threads to avoid race conditions.
+            // => do it in INIT phase
+            memcpy(
+                ((char *)  dst->data),
+                ((char *) src0->data),
+                ggml_nbytes(dst));
+        }
+        ggml_barrier(params->threadpool);
+    }
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nr = ggml_nrows(src1);
+    const int nc = src1->ne[0];
+
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb)
+
+    // src0 and dst as viewed during set
+    const size_t nb0 = ggml_element_size(src0);
+
+    const int im0 = (ne10 == 0 ? 0 : ne10-1);
+    const int im1 = (ne11 == 0 ? 0 : ne11-1);
+    const int im2 = (ne12 == 0 ? 0 : ne12-1);
+    const int im3 = (ne13 == 0 ? 0 : ne13-1);
+
+    GGML_ASSERT(offset + im0*nb0  + im1*nb1  + im2*nb2  + im3*nb3  <= ggml_nbytes(dst));
+
+    GGML_ASSERT(nb10 == sizeof(int32_t));
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int ir = ir0; ir < ir1; ++ir) {
+        // src0 and dst are viewed with shape of src1 and offset
+        // => same indices
+        const int i3 = ir/(ne12*ne11);
+        const int i2 = (ir - i3*ne12*ne11)/ne11;
+        const int i1 = (ir - i3*ne12*ne11 - i2*ne11);
+
+        ggml_vec_cpy_i32(nc,
+                (int32_t *) ((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + offset),
+                (int32_t *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11));
+    }
+}
+
 static void ggml_compute_forward_set(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -8238,6 +8201,10 @@ static void ggml_compute_forward_set(
             {
                 ggml_compute_forward_set_f32(params, dst);
             } break;
+        case GGML_TYPE_I32:
+            {
+                ggml_compute_forward_set_i32(params, dst);
+            } break;
         case GGML_TYPE_F16:
         case GGML_TYPE_BF16:
         case GGML_TYPE_Q4_0:
@@ -8262,9 +8229,6 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-        case GGML_TYPE_Q4_0_8_8:
         default:
             {
                 GGML_ABORT("fatal error");
@@ -8526,9 +8490,6 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-        case GGML_TYPE_Q4_0_8_8:
             {
                 ggml_compute_forward_get_rows_q(params, dst);
             } break;
@@ -9118,9 +9079,6 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_Q8_K:
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-        case GGML_TYPE_Q4_0_8_8:
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
         case GGML_TYPE_I32:
@@ -9159,12 +9117,6 @@ static void rope_yarn(
     *sin_theta = sinf(theta) * mscale;
 }
 
-// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
-// `corr_dim(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
-static float ggml_rope_yarn_corr_dim(int n_dims, int n_ctx_orig, float n_rot, float base) {
-    return n_dims * logf(n_ctx_orig / (n_rot * 2 * (float)M_PI)) / (2 * logf(base));
-}
-
 static void ggml_rope_cache_init(
      float theta_base, float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
      float * cache, float sin_sign, float theta_scale) {
@@ -9181,14 +9133,62 @@ static void ggml_rope_cache_init(
     }
 }
 
-void ggml_rope_yarn_corr_dims(
-    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]
-) {
-    // start and end correction dims
-    float start = floorf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_fast, freq_base));
-    float end   =  ceilf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_slow, freq_base));
-    dims[0] = MAX(0, start);
-    dims[1] = MIN(n_dims - 1, end);
+static void ggml_mrope_cache_init(
+     float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool indep_sects,
+     float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
+     float * cache, float sin_sign, float theta_scale) {
+    // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
+    float theta_t = theta_base_t;
+    float theta_h = theta_base_h;
+    float theta_w = theta_base_w;
+    float theta_e = theta_base_e;  // extra position id for vision encoder
+    int sect_dims = sections[0] + sections[1] + sections[2] + sections[3];
+    int sec_w = sections[1] + sections[0];
+    int sec_e = sections[2] + sec_w;
+    GGML_ASSERT(sect_dims <= ne0);
+
+    for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
+        const float ff = freq_factors ? freq_factors[i0/2] : 1.0f;
+
+        int sector = (i0 / 2) % sect_dims;
+        if (indep_sects) {
+            // compute theta independently for each dim sections
+            // (i.e. reset corresponding theta when `i0` go from one section to another)
+            if (sector == 0) {
+                theta_t = theta_base_t;
+            }
+            else if (sector == sections[0]) {
+                theta_h = theta_base_h;;
+            }
+            else if (sector == sec_w) {
+                theta_w = theta_base_w;
+            }
+            else if (sector == sec_e) {
+                theta_e = theta_base_e;
+            }
+        }
+
+        float theta = theta_t;
+        if (sector >= sections[0] && sector < sec_w) {
+            theta = theta_h;
+        }
+        else if (sector >= sec_w && sector < sec_w + sections[2]) {
+            theta = theta_w;
+        }
+        else if (sector >= sec_w + sections[2]) {
+            theta = theta_e;
+        }
+
+        rope_yarn(
+            theta/ff, freq_scale, corr_dims, i0, ext_factor, mscale, &cache[i0 + 0], &cache[i0 + 1]
+        );
+        cache[i0 + 1] *= sin_sign;
+
+        theta_t *= theta_scale;
+        theta_w *= theta_scale;
+        theta_h *= theta_scale;
+        theta_e *= theta_scale;
+    }
 }
 
 static void ggml_compute_forward_rope_f32(
@@ -9201,6 +9201,7 @@ static void ggml_compute_forward_rope_f32(
     const struct ggml_tensor * src2 = dst->src[2];
 
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+    int sections[4];
 
     //const int n_past     = ((int32_t *) dst->op_params)[0];
     const int n_dims     = ((int32_t *) dst->op_params)[1];
@@ -9214,6 +9215,7 @@ static void ggml_compute_forward_rope_f32(
     memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
@@ -9246,6 +9248,16 @@ static void ggml_compute_forward_rope_f32(
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, multimodal rotary position embedding
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+
+    if (is_mrope) {
+        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
+    }
+
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne0/2);
+    }
 
     const float * freq_factors = NULL;
     if (src2 != NULL) {
@@ -9261,18 +9273,63 @@ static void ggml_compute_forward_rope_f32(
 
     const int32_t * pos = (const int32_t *) src1->data;
 
-    for (int64_t i3 = 0; i3 < ne3; i3++) {
-        for (int64_t i2 = 0; i2 < ne2; i2++) {
-            const int64_t p = pos[i2];
+    for (int64_t i3 = 0; i3 < ne3; i3++) { // batch
+        for (int64_t i2 = 0; i2 < ne2; i2++) { // seq-len
 
             float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
-            ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            if (!is_mrope) {
+                const int64_t p = pos[i2];
+                ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
+            else {
+                const int64_t p_t = pos[i2];
+                const int64_t p_h = pos[i2 + ne2];
+                const int64_t p_w = pos[i2 + ne2 * 2];
+                const int64_t p_e = pos[i2 + ne2 * 3];
+                ggml_mrope_cache_init(
+                    p_t, p_h, p_w, p_e, sections, is_vision,
+                    freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
 
-            for (int64_t i1 = 0; i1 < ne1; i1++) {
+            for (int64_t i1 = 0; i1 < ne1; i1++) { // attn-heads
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
 
-                if (!is_neox) {
+                if (is_neox || is_mrope) {
+                    if (is_vision){
+                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                            const int64_t ic = i0/2;
+
+                            const float cos_theta = cache[i0 + 0];
+                            const float sin_theta = cache[i0 + 1];
+
+                            const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+                            float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+                            const float x0 = src[0];
+                            const float x1 = src[n_dims];
+
+                            dst_data[0]      = x0*cos_theta - x1*sin_theta;
+                            dst_data[n_dims] = x0*sin_theta + x1*cos_theta;
+                        }
+                    } else {
+                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                            const int64_t ic = i0/2;
+
+                            const float cos_theta = cache[i0 + 0];
+                            const float sin_theta = cache[i0 + 1];
+
+                            const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+                            float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+                            const float x0 = src[0];
+                            const float x1 = src[n_dims/2];
+
+                            dst_data[0]        = x0*cos_theta - x1*sin_theta;
+                            dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
+                        }
+                    }
+                } else {
                     for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
                         const float cos_theta = cache[i0 + 0];
                         const float sin_theta = cache[i0 + 1];
@@ -9286,8 +9343,10 @@ static void ggml_compute_forward_rope_f32(
                         dst_data[0] = x0*cos_theta - x1*sin_theta;
                         dst_data[1] = x0*sin_theta + x1*cos_theta;
                     }
-                } else {
-                    for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                }
+
+                if (is_vision) {
+                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
                         const int64_t ic = i0/2;
 
                         const float cos_theta = cache[i0 + 0];
@@ -9297,19 +9356,20 @@ static void ggml_compute_forward_rope_f32(
                         float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
 
                         const float x0 = src[0];
-                        const float x1 = src[n_dims/2];
+                        const float x1 = src[n_dims];
 
-                        dst_data[0]        = x0*cos_theta - x1*sin_theta;
-                        dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
+                        dst_data[0]      = x0*cos_theta - x1*sin_theta;
+                        dst_data[n_dims] = x0*sin_theta + x1*cos_theta;
                     }
-                }
-
-                for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
-                    const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                    float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+                } else {
+                    // fill the remain channels with data from src tensor
+                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
+                        const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                        float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
-                    dst_data[0] = src[0];
-                    dst_data[1] = src[1];
+                        dst_data[0] = src[0];
+                        dst_data[1] = src[1];
+                    }
                 }
             }
         }
@@ -9327,6 +9387,7 @@ static void ggml_compute_forward_rope_f16(
     const struct ggml_tensor * src2 = dst->src[2];
 
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+    int sections[4];
 
     //const int n_past     = ((int32_t *) dst->op_params)[0];
     const int n_dims     = ((int32_t *) dst->op_params)[1];
@@ -9339,6 +9400,8 @@ static void ggml_compute_forward_rope_f16(
     memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);
+
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
@@ -9371,6 +9434,16 @@ static void ggml_compute_forward_rope_f16(
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+
+    if (is_mrope) {
+        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
+    }
+
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne0/2);
+    }
 
     const float * freq_factors = NULL;
     if (src2 != NULL) {
@@ -9388,16 +9461,61 @@ static void ggml_compute_forward_rope_f16(
 
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
-            const int64_t p = pos[i2];
 
             float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
-            ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            if (!is_mrope) {
+                const int64_t p = pos[i2];
+                ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
+            else {
+                const int64_t p_t = pos[i2];
+                const int64_t p_h = pos[i2 + ne2];
+                const int64_t p_w = pos[i2 + ne2 * 2];
+                const int64_t p_e = pos[i2 + ne2 * 3];
+                ggml_mrope_cache_init(
+                    p_t, p_h, p_w, p_e, sections, is_vision,
+                    freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
 
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
 
-                if (!is_neox) {
+                if (is_neox || is_mrope) {
+                    if (is_vision) {
+                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                            const int64_t ic = i0/2;
+
+                            const float cos_theta = cache[i0 + 0];
+                            const float sin_theta = cache[i0 + 1];
+
+                            const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+                            ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+                            const float x0 = GGML_FP16_TO_FP32(src[0]);
+                            const float x1 = GGML_FP16_TO_FP32(src[n_dims]);
+
+                            dst_data[0]      = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                            dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                        }
+                    } else {
+                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                            const int64_t ic = i0/2;
+
+                            const float cos_theta = cache[i0 + 0];
+                            const float sin_theta = cache[i0 + 1];
+
+                            const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+                            ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+                            const float x0 = GGML_FP16_TO_FP32(src[0]);
+                            const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]);
+
+                            dst_data[0]        = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                            dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                        }
+                    }
+                } else {
                     for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
                         const float cos_theta = cache[i0 + 0];
                         const float sin_theta = cache[i0 + 1];
@@ -9411,8 +9529,10 @@ static void ggml_compute_forward_rope_f16(
                         dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
                         dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                     }
-                } else {
-                    for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                }
+
+                if (is_vision) {
+                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
                         const int64_t ic = i0/2;
 
                         const float cos_theta = cache[i0 + 0];
@@ -9422,19 +9542,19 @@ static void ggml_compute_forward_rope_f16(
                         ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
 
                         const float x0 = GGML_FP16_TO_FP32(src[0]);
-                        const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]);
+                        const float x1 = GGML_FP16_TO_FP32(src[n_dims]);
 
-                        dst_data[0]        = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                        dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                        dst_data[0]      = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                        dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                     }
-                }
-
-                for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
-                    const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                    ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+                } else {
+                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
+                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                        ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
-                    dst_data[0] = src[0];
-                    dst_data[1] = src[1];
+                        dst_data[0] = src[0];
+                        dst_data[1] = src[1];
+                    }
                 }
             }
         }
@@ -10433,6 +10553,40 @@ static void ggml_compute_forward_pad(
     }
 }
 
+// ggml_compute_forward_pad_reflect_1d
+
+static void ggml_compute_forward_pad_reflect_1d(
+        const struct ggml_compute_params * params,
+              struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int32_t * opts = (const int32_t *) dst->op_params;
+    const int p0 = opts[0];
+    const int p1 = opts[1];
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    for (int64_t i3 = 0; i3 < ne3; i3++) {
+        for (int64_t i2 = 0; i2 < ne2; i2++) {
+            for (int64_t i1 = ith; i1 < ne1; i1 += nth) {
+                float * left  = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 +         p0*nb0);
+                float * right = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (ne0-p1-1)*nb0);
+
+                ggml_vec_cpy_f32(ne00, left, (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
+
+                for (int i0 = 1; i0 <= p0; i0++) { left[-i0] = left[i0];   }
+                for (int i0 = 1; i0 <= p1; i0++) { right[i0] = right[-i0]; }
+            }
+        }
+    }
+}
 
 // ggml_compute_forward_arange
 
@@ -10668,7 +10822,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
     enum ggml_type    const k_vec_dot_type = type_traits_cpu[k->type].vec_dot_type;
-    ggml_from_float_t const q_to_vec_dot   = ggml_get_type_traits(k_vec_dot_type)->from_float;
+    ggml_from_float_t const q_to_vec_dot   = type_traits_cpu[k_vec_dot_type].from_float;
     ggml_vec_dot_t    const kq_vec_dot     = type_traits_cpu[k->type].vec_dot;
     ggml_to_float_t   const v_to_float     = ggml_get_type_traits(v->type)->to_float;
 
@@ -12220,11 +12374,16 @@ static void ggml_compute_forward_opt_step_adamw_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
 
-    const struct ggml_tensor * src0        = dst->src[0];
-    const struct ggml_tensor * src0_grad   = dst->src[1];
-    const struct ggml_tensor * src0_grad_m = dst->src[2];
-    const struct ggml_tensor * src0_grad_v = dst->src[3];
+    const struct ggml_tensor * src0         = dst->src[0];
+    const struct ggml_tensor * src0_grad    = dst->src[1];
+    const struct ggml_tensor * src0_grad_m  = dst->src[2];
+    const struct ggml_tensor * src0_grad_v  = dst->src[3];
+    const struct ggml_tensor * adamw_params = dst->src[4];
+
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m));
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v));
+    GGML_ASSERT(ggml_nelements(adamw_params) == 7);
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -12241,16 +12400,14 @@ static void ggml_compute_forward_opt_step_adamw_f32(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
-    /* const float   gnorm = 1.0f; */
-    int64_t       iter;   memcpy(&iter, &dst->op_params[0], sizeof(int64_t));
-    const float   alpha = ggml_get_op_params_f32(dst, 2);
-    const float   beta1 = ggml_get_op_params_f32(dst, 3);
-    const float   beta2 = ggml_get_op_params_f32(dst, 4);
-    const float   eps   = ggml_get_op_params_f32(dst, 5);
-    const float   wd    = ggml_get_op_params_f32(dst, 6);
-
-    const float beta1h  = alpha/(1.0f - powf(beta1, iter));
-    const float beta2h  =  1.0f/(1.0f - powf(beta2, iter));
+    const float * adamw_params_ptr = ggml_get_data_f32(adamw_params);
+    const float alpha  = adamw_params_ptr[0];
+    const float beta1  = adamw_params_ptr[1];
+    const float beta2  = adamw_params_ptr[2];
+    const float eps    = adamw_params_ptr[3];
+    const float wd     = adamw_params_ptr[4];
+    const float beta1h = adamw_params_ptr[5];
+    const float beta2h = adamw_params_ptr[6];
 
     for (int ir = ir0; ir < ir1; ++ir) {
         const int64_t i03 = ir/(ne02*ne01);
@@ -12274,17 +12431,9 @@ static void ggml_compute_forward_opt_step_adamw_f32(
             // The weight decay is applied independently of the Adam momenta m and v.
             // This is NOT equivalent to l2 regularization that adds w[i00]*w[i00] to the loss.
             // See: https://arxiv.org/pdf/1711.05101v3.pdf
-            w[i00] = w[i00]*(1.0f - alpha*wd) - mh/vh;
+            w[i00] = w[i00]*(1.0f - alpha*wd) - alpha*mh/vh;
         }
     }
-
-    ggml_barrier(params->threadpool);
-    if (ith != 0) {
-        return;
-    }
-
-    iter++;
-    memcpy(&dst->op_params[0], &iter, sizeof(int64_t));
 }
 
 static void ggml_compute_forward_opt_step_adamw(
@@ -12313,6 +12462,9 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
         return;
     }
 
+    // extra_buffer op?
+    if (ggml_cpu_extra_compute_forward(params, tensor)) return;
+
     switch (tensor->op) {
         case GGML_OP_DUP:
             {
@@ -12534,6 +12686,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_pad(params, tensor);
             } break;
+        case GGML_OP_PAD_REFLECT_1D:
+            {
+                ggml_compute_forward_pad_reflect_1d(params, tensor);
+            } break;
         case GGML_OP_ARANGE:
             {
                 ggml_compute_forward_arange(params, tensor);
@@ -12876,6 +13032,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
+        case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
@@ -12965,7 +13122,7 @@ static thread_ret_t ggml_graph_compute_secondary_thread(void* data);
 #include "windows.h"
 
 // TODO: support > 64 CPUs
-bool ggml_thread_apply_affinity(bool * mask) {
+static bool ggml_thread_apply_affinity(bool * mask) {
     HANDLE    h = GetCurrentThread();
     uint64_t  bitmask = 0ULL;
 
@@ -13255,140 +13412,142 @@ struct ggml_cplan ggml_graph_plan(
 
         size_t cur = 0;
 
-        switch (node->op) {
-            case GGML_OP_CPY:
-            case GGML_OP_DUP:
-                {
-                    if (ggml_is_quantized(node->type) ||
-                        // F16 -> BF16 and BF16 -> F16 copies go through intermediate F32
-                        (node->src[0]->type == GGML_TYPE_F16  && node->src[1] && node->src[1]->type == GGML_TYPE_BF16) ||
-                        (node->src[0]->type == GGML_TYPE_BF16 && node->src[1] && node->src[1]->type == GGML_TYPE_F16)) {
+        if (!ggml_cpu_extra_work_size(n_threads, node, &cur)) {
+
+            switch (node->op) {
+                case GGML_OP_CPY:
+                case GGML_OP_DUP:
+                    {
+                        if (ggml_is_quantized(node->type) ||
+                            // F16 -> BF16 and BF16 -> F16 copies go through intermediate F32
+                            (node->src[0]->type == GGML_TYPE_F16  && node->src[1] && node->src[1]->type == GGML_TYPE_BF16) ||
+                            (node->src[0]->type == GGML_TYPE_BF16 && node->src[1] && node->src[1]->type == GGML_TYPE_F16)) {
+                            cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
+                        }
+                    } break;
+                case GGML_OP_ADD:
+                case GGML_OP_ADD1:
+                    {
+                        if (ggml_is_quantized(node->src[0]->type)) {
+                            cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
+                        }
+                    } break;
+                case GGML_OP_ACC:
+                    {
+                        if (ggml_is_quantized(node->src[0]->type)) {
+                            cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks;
+                        }
+                    } break;
+                case GGML_OP_COUNT_EQUAL:
+                    {
+                        cur = ggml_type_size(node->type)*n_tasks;
+                    } break;
+                case GGML_OP_MUL_MAT:
+                    {
+                        const enum ggml_type vec_dot_type = type_traits_cpu[node->src[0]->type].vec_dot_type;
+
+                        if (node->src[1]->type != vec_dot_type) {
+                            cur = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1]));
+                        }
+                    } break;
+                case GGML_OP_MUL_MAT_ID:
+                    {
+                        cur = 0;
+                        const struct ggml_tensor * src0 = node->src[0];
+                        const struct ggml_tensor * src1 = node->src[1];
+                        const enum ggml_type vec_dot_type = type_traits_cpu[src0->type].vec_dot_type;
+                        if (src1->type != vec_dot_type) {
+                            cur += ggml_row_size(vec_dot_type, ggml_nelements(src1));
+                        }
+                        const int n_as = src0->ne[2];
+                        cur += GGML_PAD(cur, sizeof(int64_t));       // align
+                        cur += n_as * sizeof(int64_t);               // matrix_row_counts
+                        cur += n_as * src1->ne[2] * sizeof(int64_t); // matrix_rows
+                    } break;
+                case GGML_OP_OUT_PROD:
+                    {
+                        if (ggml_is_quantized(node->src[0]->type)) {
+                            cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
+                        }
+                    } break;
+                case GGML_OP_SOFT_MAX:
+                case GGML_OP_ROPE:
+                    {
                         cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
-                    }
-                } break;
-            case GGML_OP_ADD:
-            case GGML_OP_ADD1:
-                {
-                    if (ggml_is_quantized(node->src[0]->type)) {
-                        cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
-                    }
-                } break;
-            case GGML_OP_ACC:
-                {
-                    if (ggml_is_quantized(node->src[0]->type)) {
-                        cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks;
-                    }
-                } break;
-            case GGML_OP_COUNT_EQUAL:
-                {
-                    cur = ggml_type_size(node->type)*n_tasks;
-                } break;
-            case GGML_OP_MUL_MAT:
-                {
-                    const enum ggml_type vec_dot_type = type_traits_cpu[node->src[0]->type].vec_dot_type;
+                    } break;
+                case GGML_OP_CONV_TRANSPOSE_1D:
+                    {
+                        GGML_ASSERT(node->src[0]->ne[3] == 1);
+                        GGML_ASSERT(node->src[1]->ne[2] == 1);
+                        GGML_ASSERT(node->src[1]->ne[3] == 1);
+
+                        const int64_t ne00 = node->src[0]->ne[0];  // K
+                        const int64_t ne01 = node->src[0]->ne[1];  // Cout
+                        const int64_t ne02 = node->src[0]->ne[2];  // Cin
+                        const int64_t ne10 = node->src[1]->ne[0];  // L
+                        const int64_t ne11 = node->src[1]->ne[1];  // Cin
+
+                        if ((node->src[0]->type == GGML_TYPE_F16 ||
+                             node->src[0]->type == GGML_TYPE_BF16) &&
+                            node->src[1]->type == GGML_TYPE_F32) {
+                            cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02;
+                            cur += sizeof(ggml_fp16_t)*ne10*ne11;
+                        } else if (node->src[0]->type == GGML_TYPE_F32 &&
+                                   node->src[1]->type == GGML_TYPE_F32) {
+                            cur += sizeof(float)*ne00*ne01*ne02;
+                            cur += sizeof(float)*ne10*ne11;
+                        } else {
+                            GGML_ABORT("fatal error");
+                        }
+                    } break;
+                case GGML_OP_CONV_TRANSPOSE_2D:
+                    {
+                        const int64_t ne00 = node->src[0]->ne[0]; // W
+                        const int64_t ne01 = node->src[0]->ne[1]; // H
+                        const int64_t ne02 = node->src[0]->ne[2]; // Channels Out
+                        const int64_t ne03 = node->src[0]->ne[3]; // Channels In
 
-                    if (node->src[1]->type != vec_dot_type) {
-                        cur = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1]));
-                    }
-                } break;
-            case GGML_OP_MUL_MAT_ID:
-                {
-                    cur = 0;
-                    const struct ggml_tensor * src0 = node->src[0];
-                    const struct ggml_tensor * src1 = node->src[1];
-                    const enum ggml_type vec_dot_type = type_traits_cpu[src0->type].vec_dot_type;
-                    if (src1->type != vec_dot_type) {
-                        cur += ggml_row_size(vec_dot_type, ggml_nelements(src1));
-                    }
-                    const int n_as = src0->ne[2];
-                    cur += GGML_PAD(cur, sizeof(int64_t));       // align
-                    cur += n_as * sizeof(int64_t);               // matrix_row_counts
-                    cur += n_as * src1->ne[2] * sizeof(int64_t); // matrix_rows
-                } break;
-            case GGML_OP_OUT_PROD:
-                {
-                    if (ggml_is_quantized(node->src[0]->type)) {
-                        cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
-                    }
-                } break;
-            case GGML_OP_SOFT_MAX:
-            case GGML_OP_ROPE:
-                {
-                    cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
-                } break;
-            case GGML_OP_CONV_TRANSPOSE_1D:
-                {
-                    GGML_ASSERT(node->src[0]->ne[3] == 1);
-                    GGML_ASSERT(node->src[1]->ne[2] == 1);
-                    GGML_ASSERT(node->src[1]->ne[3] == 1);
-
-                    const int64_t ne00 = node->src[0]->ne[0];  // K
-                    const int64_t ne01 = node->src[0]->ne[1];  // Cout
-                    const int64_t ne02 = node->src[0]->ne[2];  // Cin
-
-                    const int64_t ne10 = node->src[1]->ne[0];  // L
-                    const int64_t ne11 = node->src[1]->ne[1];  // Cin
-
-                    if ((node->src[0]->type == GGML_TYPE_F16 ||
-                         node->src[0]->type == GGML_TYPE_BF16) &&
-                        node->src[1]->type == GGML_TYPE_F32) {
-                        cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02;
-                        cur += sizeof(ggml_fp16_t)*ne10*ne11;
-                    } else if (node->src[0]->type == GGML_TYPE_F32 &&
-                               node->src[1]->type == GGML_TYPE_F32) {
-                        cur += sizeof(float)*ne00*ne01*ne02;
-                        cur += sizeof(float)*ne10*ne11;
-                    } else {
-                        GGML_ABORT("fatal error");
-                    }
-                } break;
-            case GGML_OP_CONV_TRANSPOSE_2D:
-                {
-                    const int64_t ne00 = node->src[0]->ne[0]; // W
-                    const int64_t ne01 = node->src[0]->ne[1]; // H
-                    const int64_t ne02 = node->src[0]->ne[2]; // Channels Out
-                    const int64_t ne03 = node->src[0]->ne[3]; // Channels In
-
-                    const int64_t ne10 = node->src[1]->ne[0]; // W
-                    const int64_t ne11 = node->src[1]->ne[1]; // H
-                    const int64_t ne12 = node->src[1]->ne[2]; // Channels In
-
-                    cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03;
-                    cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12;
-                } break;
-            case GGML_OP_FLASH_ATTN_EXT:
-                {
-                    const int64_t ne00 = node->src[0]->ne[0]; // D
+                        const int64_t ne10 = node->src[1]->ne[0]; // W
+                        const int64_t ne11 = node->src[1]->ne[1]; // H
+                        const int64_t ne12 = node->src[1]->ne[2]; // Channels In
 
-                    cur = 3*sizeof(float)*ne00*n_tasks; // 3x head size/thread
-                } break;
-            case GGML_OP_FLASH_ATTN_BACK:
-                {
-                    const int64_t    D = node->src[0]->ne[0];
-                    const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL);
-                    const int64_t mxDn = MAX(D, ne11) * 2; // *2 because of S and SM in ggml_compute_forward_flash_attn_back
-                    if (node->src[1]->type == GGML_TYPE_F32) {
-                        cur  = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
-                    } else if (node->src[1]->type == GGML_TYPE_F16) {
-                        cur  = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
-                    } else if (node->src[1]->type == GGML_TYPE_BF16) {
-                        cur  = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
-                    }
-                } break;
+                        cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03;
+                        cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12;
+                    } break;
+                case GGML_OP_FLASH_ATTN_EXT:
+                    {
+                        const int64_t ne00 = node->src[0]->ne[0]; // D
 
-            case GGML_OP_CROSS_ENTROPY_LOSS:
-                {
-                    cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks);
-                } break;
-            case GGML_OP_COUNT:
-                {
-                    GGML_ABORT("fatal error");
-                }
-            default:
-                break;
+                        cur = 3*sizeof(float)*ne00*n_tasks; // 3x head size/thread
+                    } break;
+                case GGML_OP_FLASH_ATTN_BACK:
+                    {
+                        const int64_t    D = node->src[0]->ne[0];
+                        const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL);
+                        const int64_t mxDn = MAX(D, ne11) * 2; // *2 because of S and SM in ggml_compute_forward_flash_attn_back
+                        if (node->src[1]->type == GGML_TYPE_F32) {
+                            cur  = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
+                            cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
+                        } else if (node->src[1]->type == GGML_TYPE_F16) {
+                            cur  = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
+                            cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
+                        } else if (node->src[1]->type == GGML_TYPE_BF16) {
+                            cur  = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
+                            cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
+                        }
+                    } break;
+
+                case GGML_OP_CROSS_ENTROPY_LOSS:
+                    {
+                        cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks);
+                    } break;
+                case GGML_OP_COUNT:
+                    {
+                        GGML_ABORT("fatal error");
+                    }
+                default:
+                    break;
+            }
         }
 
         work_size = MAX(work_size, cur);
@@ -13587,29 +13746,6 @@ static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool, int
 
 #endif // GGML_USE_OPENMP
 
-void ggml_threadpool_params_init(struct ggml_threadpool_params * p, int n_threads) {
-    p->n_threads  = n_threads;
-    p->prio       = 0;     // default priority (usually means normal or inherited)
-    p->poll       = 50;    // hybrid-polling enabled
-    p->strict_cpu = false; // no strict placement (all threads share same cpumask)
-    p->paused     = false; // threads are ready to go
-    memset(p->cpumask, 0, GGML_MAX_N_THREADS); // all-zero means use the default affinity (usually inherited)
-}
-
-struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads) {
-    struct ggml_threadpool_params p;
-    ggml_threadpool_params_init(&p, n_threads);
-    return p;
-}
-
-bool ggml_threadpool_params_match(const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1) {
-    if (p0->n_threads      != p1->n_threads  )    return false;
-    if (p0->prio           != p1->prio       )    return false;
-    if (p0->poll           != p1->poll       )    return false;
-    if (p0->strict_cpu     != p1->strict_cpu )    return false;
-    return memcmp(p0->cpumask, p1->cpumask, GGML_MAX_N_THREADS) == 0;
-}
-
 static struct ggml_threadpool * ggml_threadpool_new_impl(
     struct ggml_threadpool_params * tpp,
                struct ggml_cgraph * cgraph,
@@ -13759,16 +13895,169 @@ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct g
     return ggml_graph_compute(cgraph, &cplan);
 }
 
+
+int ggml_cpu_has_avx(void) {
+#if defined(__AVX__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_avx_vnni(void) {
+#if defined(__AVXVNNI__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_avx2(void) {
+#if defined(__AVX2__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_avx512(void) {
+#if defined(__AVX512F__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_avx512_vbmi(void) {
+#if defined(__AVX512VBMI__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_avx512_vnni(void) {
+#if defined(__AVX512VNNI__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_avx512_bf16(void) {
+#if defined(__AVX512BF16__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_amx_int8(void) {
+#if defined(__AMX_INT8__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_fma(void) {
+#if defined(__FMA__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_arm_fma(void) {
+#if defined(__ARM_FEATURE_FMA)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_riscv_v(void) {
+#if defined(__riscv_v_intrinsic)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_f16c(void) {
+#if defined(__F16C__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_fp16_va(void) {
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_wasm_simd(void) {
+#if defined(__wasm_simd128__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_llamafile(void) {
+#if defined(GGML_USE_LLAMAFILE)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_sse3(void) {
+#if defined(__SSE3__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_ssse3(void) {
+#if defined(__SSSE3__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_vsx(void) {
+#if defined(__POWER9_VECTOR__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_neon(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_NEON)
     return ggml_arm_arch_features.has_neon;
 #else
     return 0;
 #endif
 }
 
+int ggml_cpu_has_dotprod(void) {
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_DOTPROD)
+    return ggml_arm_arch_features.has_dotprod;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_sve(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
     return ggml_arm_arch_features.has_sve;
 #else
     return 0;
@@ -13776,7 +14065,7 @@ int ggml_cpu_has_sve(void) {
 }
 
 int ggml_cpu_has_matmul_int8(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8)
     return ggml_arm_arch_features.has_i8mm;
 #else
     return 0;
@@ -13784,7 +14073,7 @@ int ggml_cpu_has_matmul_int8(void) {
 }
 
 int ggml_cpu_get_sve_cnt(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
     return ggml_arm_arch_features.sve_cnt;
 #else
     return 0;
diff --git a/ggml/src/ggml-cpu/ggml-cpu.cpp b/ggml/src/ggml-cpu/ggml-cpu.cpp
new file mode 100644
index 00000000000..0b6419f83f0
--- /dev/null
+++ b/ggml/src/ggml-cpu/ggml-cpu.cpp
@@ -0,0 +1,622 @@
+#include "ggml-backend.h"
+#include "ggml-backend-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-cpu-aarch64.h"
+#include "ggml-cpu-traits.h"
+#include "ggml-impl.h"
+#include "amx/amx.h"
+
+#include <cctype>
+#include <string>
+#include <vector>
+
+#ifdef GGML_USE_CPU_HBM
+#include "ggml-cpu-hbm.h"
+#endif
+
+#if defined(__APPLE__)
+#include <sys/types.h>
+#include <sys/sysctl.h>
+#endif
+
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+    #define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
+// ggml-backend interface
+
+std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type() {
+    static std::vector<ggml_backend_buffer_type_t> bufts = []() {
+        std::vector<ggml_backend_buffer_type_t> bufts;
+
+#if defined(__AMX_INT8__) && defined(__AVX512VNNI__)
+        if (ggml_backend_amx_buffer_type()) {
+            bufts.push_back(ggml_backend_amx_buffer_type());
+        }
+#endif
+
+#ifdef GGML_USE_CPU_AARCH64
+        if (ggml_backend_cpu_aarch64_buffer_type()) {
+            bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
+        }
+#endif
+
+        bufts.push_back(NULL);
+
+        return bufts;
+    }();
+
+    return bufts;
+}
+
+static ggml_backend_buffer_type_t * ggml_backend_cpu_device_get_extra_buffers_type(ggml_backend_dev_t device) {
+    return ggml_backend_cpu_get_extra_buffers_type().data();
+
+    GGML_UNUSED(device);
+}
+
+static bool ggml_backend_cpu_is_extra_buffer_type(ggml_backend_buffer_type_t buft) {
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra && extra == buft) return true;
+    }
+    return false;
+}
+
+// CPU backend - backend (stream)
+
+struct ggml_backend_cpu_context {
+    int                 n_threads;
+    ggml_threadpool_t   threadpool;
+
+    uint8_t *           work_data;
+    size_t              work_size;
+
+    ggml_abort_callback abort_callback;
+    void *              abort_callback_data;
+};
+
+static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
+    return "CPU";
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_cpu_free(ggml_backend_t backend) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+    delete[] cpu_ctx->work_data;
+    delete cpu_ctx;
+    delete backend;
+}
+
+struct ggml_backend_plan_cpu {
+    struct ggml_cplan cplan;
+    struct ggml_cgraph cgraph;
+};
+
+static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;
+
+    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
+    cpu_plan->cgraph = *cgraph; // FIXME: deep copy
+
+    if (cpu_plan->cplan.work_size > 0) {
+        cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
+        if (cpu_plan->cplan.work_data == NULL) {
+            delete cpu_plan;
+            return NULL;
+        }
+    }
+
+    cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
+    cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+
+    return cpu_plan;
+}
+
+static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    delete[] cpu_plan->cplan.work_data;
+    delete cpu_plan;
+
+    GGML_UNUSED(backend);
+}
+
+static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
+
+    GGML_UNUSED(backend);
+}
+
+static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
+
+    if (cpu_ctx->work_size < cplan.work_size) {
+        delete[] cpu_ctx->work_data;
+        cpu_ctx->work_data = new uint8_t[cplan.work_size];
+        if (cpu_ctx->work_data == NULL) {
+            cpu_ctx->work_size = 0;
+            return GGML_STATUS_ALLOC_FAILED;
+        }
+        cpu_ctx->work_size = cplan.work_size;
+    }
+    cplan.work_data = (uint8_t *)cpu_ctx->work_data;
+
+    cplan.abort_callback      = cpu_ctx->abort_callback;
+    cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+
+    return ggml_graph_compute(cgraph, &cplan);
+}
+
+static const struct ggml_backend_i ggml_backend_cpu_i = {
+    /* .get_name                = */ ggml_backend_cpu_get_name,
+    /* .free                    = */ ggml_backend_cpu_free,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,
+    /* .graph_plan_free         = */ ggml_backend_cpu_graph_plan_free,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
+    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_cpu_guid(void) {
+    static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_cpu_init(void) {
+    // initialize CPU backend now to avoid slowing the first graph computation
+    ggml_cpu_init();
+
+    struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
+    if (ctx == NULL) {
+        return NULL;
+    }
+
+    ctx->n_threads           = GGML_DEFAULT_N_THREADS;
+    ctx->threadpool          = NULL;
+    ctx->work_data           = NULL;
+    ctx->work_size           = 0;
+    ctx->abort_callback      = NULL;
+    ctx->abort_callback_data = NULL;
+
+    ggml_backend_t cpu_backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_cpu_guid(),
+        /* .interface = */ ggml_backend_cpu_i,
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context   = */ ctx,
+    };
+
+    if (cpu_backend == NULL) {
+        delete ctx;
+        return NULL;
+    }
+
+    return cpu_backend;
+}
+
+bool ggml_backend_is_cpu(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
+}
+
+void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->n_threads = n_threads;
+}
+
+void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+
+    if (ctx->threadpool && ctx->threadpool != threadpool) {
+        // already had a different threadpool, pause/suspend it before switching
+        ggml_threadpool_pause(ctx->threadpool);
+    }
+    ctx->threadpool = threadpool;
+}
+
+void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->abort_callback = abort_callback;
+    ctx->abort_callback_data = abort_callback_data;
+}
+
+// CPU backend - device
+
+struct ggml_backend_cpu_device_context {
+    std::string description = "CPU";
+
+    ggml_backend_cpu_device_context() {
+#ifdef __APPLE__
+        size_t len = 0;
+        if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) {
+            description.resize(len);
+            sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT
+        }
+#elif defined(__linux__)
+        FILE * f = fopen("/proc/cpuinfo", "r");
+        if (f) {
+            char buf[1024];
+            while (fgets(buf, sizeof(buf), f)) {
+                if (strncmp(buf, "model name", 10) == 0) {
+                    char * p = strchr(buf, ':');
+                    if (p) {
+                        p++;
+                        while (std::isspace(*p)) {
+                            p++;
+                        }
+                        while (std::isspace(p[strlen(p) - 1])) {
+                            p[strlen(p) - 1] = '\0';
+                        }
+                        description = p;
+                        break;
+                    }
+                }
+            }
+            fclose(f);
+        }
+#elif defined(_WIN32)
+        HKEY hKey;
+        if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
+                        TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
+                        0,
+                        KEY_READ,
+                        &hKey) == ERROR_SUCCESS) {
+            DWORD cpu_brand_size = 0;
+            if (RegQueryValueExA(hKey,
+                                TEXT("ProcessorNameString"),
+                                NULL,
+                                NULL,
+                                NULL,
+                                &cpu_brand_size) == ERROR_SUCCESS) {
+                description.resize(cpu_brand_size);
+                if (RegQueryValueExA(hKey,
+                                    TEXT("ProcessorNameString"),
+                                    NULL,
+                                    NULL,
+                                    (LPBYTE)&description[0], // NOLINT
+                                    &cpu_brand_size) == ERROR_SUCCESS) {
+                    if (description.find('\0') != std::string::npos) {
+                        description.resize(description.find('\0'));
+                    }
+                }
+            }
+            RegCloseKey(hKey);
+        }
+#endif
+    }
+};
+
+static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
+    return "CPU";
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
+    struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context;
+
+    return ctx->description.c_str();
+}
+
+static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    // TODO
+    *free = 0;
+    *total = 0;
+
+    GGML_UNUSED(dev);
+}
+
+static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_CPU;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_cpu_device_get_name(dev);
+    props->description = ggml_backend_cpu_device_get_description(dev);
+    props->type        = ggml_backend_cpu_device_get_type(dev);
+    ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ true,
+        /* .events                = */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_cpu_device_init_backend(ggml_backend_dev_t dev, const char * params) {
+    return ggml_backend_cpu_init();
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_backend_dev_t dev) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(dev);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    return ggml_backend_cpu_buffer_from_ptr(ptr, size);
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(max_tensor_size);
+}
+
+static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * src1 = op->src[1];
+
+    if (op->op == GGML_OP_NONE || op->op == GGML_OP_RESHAPE || op->op == GGML_OP_VIEW || op->op == GGML_OP_PERMUTE || op->op == GGML_OP_TRANSPOSE) {
+        return true;
+    }
+
+    // extra_buffer_op?
+    for (auto extra : ggml_backend_cpu_get_extra_buffers_type()) {
+        if (extra) {
+            auto buf_extra = (ggml::cpu::extra_buffer_type*) extra->context;
+            if (buf_extra && buf_extra->supports_op(dev, op)) {
+                return true;
+            }
+        }
+    }
+
+    // the other case need host buffer.
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        if (op->src[i] && op->src[i]->buffer && !ggml_backend_buft_is_host(op->src[i]->buffer->buft)) {
+            return false;
+        }
+    }
+
+    switch (op->op) {
+        case GGML_OP_CPY:
+            return
+                op->type != GGML_TYPE_IQ3_XXS &&
+                op->type != GGML_TYPE_IQ3_S   &&
+                op->type != GGML_TYPE_IQ2_XXS &&
+                op->type != GGML_TYPE_IQ2_XS  &&
+                op->type != GGML_TYPE_IQ2_S   &&
+                op->type != GGML_TYPE_IQ1_S   &&
+                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
+        case GGML_OP_MUL_MAT:
+            return src1->type == GGML_TYPE_F32 || src1->type == ggml_get_type_traits_cpu(src0->type)->vec_dot_type;
+        case GGML_OP_ROPE_BACK:
+            return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
+        case GGML_OP_IM2COL_BACK:
+            return src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32;
+        case GGML_OP_OUT_PROD:
+            return (src0->type == GGML_TYPE_F32 || ggml_is_quantized(src0->type)) && src1->type == GGML_TYPE_F32;
+        default:
+            return true;
+    }
+}
+
+static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    return ggml_backend_buft_is_host(buft) || ggml_backend_cpu_is_extra_buffer_type(buft);
+    GGML_UNUSED(dev);
+}
+
+static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
+    /* .get_name             = */ ggml_backend_cpu_device_get_name,
+    /* .get_description      = */ ggml_backend_cpu_device_get_description,
+    /* .get_memory           = */ ggml_backend_cpu_device_get_memory,
+    /* .get_type             = */ ggml_backend_cpu_device_get_type,
+    /* .get_props            = */ ggml_backend_cpu_device_get_props,
+    /* .init_backend         = */ ggml_backend_cpu_device_init_backend,
+    /* .get_buffer_type      = */ ggml_backend_cpu_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,
+    /* .supports_op          = */ ggml_backend_cpu_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_cpu_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+// CPU backend - backend (reg)
+
+static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
+    return "CPU";
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
+    return 1;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+
+    static ggml_backend_cpu_device_context ctx;
+    static ggml_backend_device ggml_backend_cpu_device = {
+        /* .iface   = */ ggml_backend_cpu_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ &ctx,
+    };
+
+    return &ggml_backend_cpu_device;
+}
+
+// This is intended to replace the the ggml_cpu_has_* functions when loading the CPU backend dynamically,
+// and additionally to allow other backends to expose their own list of features that applications can query using the same API
+static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t reg) {
+    static std::vector<ggml_backend_feature> features = []() {
+        ggml_cpu_init();
+
+        std::vector<ggml_backend_feature> features;
+        if (ggml_cpu_has_sse3()) {
+            features.push_back({ "SSE3", "1" });
+        }
+        if (ggml_cpu_has_ssse3()) {
+            features.push_back({ "SSSE3", "1" });
+        }
+        if (ggml_cpu_has_avx()) {
+            features.push_back({ "AVX", "1" });
+        }
+        if (ggml_cpu_has_avx_vnni()) {
+            features.push_back({ "AVX_VNNI", "1" });
+        }
+        if (ggml_cpu_has_avx2()) {
+            features.push_back({ "AVX2", "1" });
+        }
+        if (ggml_cpu_has_f16c()) {
+            features.push_back({ "F16C", "1" });
+        }
+        if (ggml_cpu_has_fma()) {
+            features.push_back({ "FMA", "1" });
+        }
+        if (ggml_cpu_has_avx512()) {
+            features.push_back({ "AVX512", "1" });
+        }
+        if (ggml_cpu_has_avx512_vbmi()) {
+            features.push_back({ "AVX512_VBMI", "1" });
+        }
+        if (ggml_cpu_has_avx512_vnni()) {
+            features.push_back({ "AVX512_VNNI", "1" });
+        }
+        if (ggml_cpu_has_avx512_bf16()) {
+            features.push_back({ "AVX512_BF16", "1" });
+        }
+        if (ggml_cpu_has_amx_int8()) {
+            features.push_back({ "AMX_INT8", "1" });
+        }
+        if (ggml_cpu_has_neon()) {
+            features.push_back({ "NEON", "1" });
+        }
+        if (ggml_cpu_has_arm_fma()) {
+            features.push_back({ "ARM_FMA", "1" });
+        }
+        if (ggml_cpu_has_fp16_va()) {
+            features.push_back({ "FP16_VA", "1" });
+        }
+        if (ggml_cpu_has_matmul_int8()) {
+            features.push_back({ "MATMUL_INT8", "1" });
+        }
+        if (ggml_cpu_has_sve()) {
+            features.push_back({ "SVE", "1" });
+        }
+        if (ggml_cpu_get_sve_cnt() > 0) {
+            static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt());
+            features.push_back({ "SVE_CNT", sve_cnt.c_str() });
+        }
+        if (ggml_cpu_has_riscv_v()) {
+            features.push_back({ "RISCV_V", "1" });
+        }
+        if (ggml_cpu_has_vsx()) {
+            features.push_back({ "VSX", "1" });
+        }
+        if (ggml_cpu_has_wasm_simd()) {
+            features.push_back({ "WASM_SIMD", "1" });
+        }
+        if (ggml_cpu_has_llamafile()) {
+            features.push_back({ "LLAMAFILE", "1" });
+        }
+    #ifdef GGML_USE_ACCELERATE
+        features.push_back({ "ACCELERATE", "1" });
+    #endif
+    #ifdef GGML_USE_CPU_HBM
+        features.push_back({ "CPU_HBM", "1" });
+    #endif
+    #ifdef GGML_USE_OPENMP
+        features.push_back({ "OPENMP", "1" });
+    #endif
+    #ifdef GGML_USE_CPU_AARCH64
+        features.push_back({ "AARCH64_REPACK", "1" });
+    #endif
+
+        features.push_back({ nullptr, nullptr });
+
+        return features;
+    }();
+
+    return features.data();
+
+    GGML_UNUSED(reg);
+}
+
+static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
+        ggml_backend_set_n_threads_t fct = ggml_backend_cpu_set_n_threads;
+        return (void *)fct;
+    }
+    if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
+        ggml_backend_dev_get_extra_bufts_t fct = ggml_backend_cpu_device_get_extra_buffers_type;
+        return (void *)fct;
+    }
+    if (strcmp(name, "ggml_backend_get_features") == 0) {
+        return (void *)ggml_backend_cpu_get_features;
+    }
+    if (strcmp(name, "ggml_backend_set_abort_callback") == 0) {
+        return (void *)ggml_backend_cpu_set_abort_callback;
+    }
+    if (strcmp(name, "ggml_backend_cpu_numa_init") == 0) {
+        return (void *)ggml_numa_init;
+    }
+    if (strcmp(name, "ggml_backend_cpu_is_numa") == 0) {
+        return (void *)ggml_is_numa;
+    }
+
+    // threadpool - TODO:  move to ggml-base
+    if (strcmp(name, "ggml_threadpool_new") == 0) {
+        return (void *)ggml_threadpool_new;
+    }
+    if (strcmp(name, "ggml_threadpool_free") == 0) {
+        return (void *)ggml_threadpool_free;
+    }
+    if (strcmp(name, "ggml_backend_cpu_set_threadpool") == 0) {
+        return (void *)ggml_backend_cpu_set_threadpool;
+    }
+
+    return NULL;
+
+    GGML_UNUSED(reg);
+}
+
+static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
+    /* .get_name         = */ ggml_backend_cpu_reg_get_name,
+    /* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_cpu_reg_get_device,
+    /* .get_proc_address = */ ggml_backend_cpu_get_proc_address,
+};
+
+ggml_backend_reg_t ggml_backend_cpu_reg(void) {
+    // init CPU feature detection
+    ggml_cpu_init();
+
+    static struct ggml_backend_reg ggml_backend_cpu_reg = {
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_cpu_reg_i,
+        /* .context     = */ NULL,
+    };
+
+    return &ggml_backend_cpu_reg;
+}
+
+GGML_BACKEND_DL_IMPL(ggml_backend_cpu_reg)
diff --git a/ggml/src/ggml-cpu/llamafile/sgemm.cpp b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
new file mode 100644
index 00000000000..da4146ec4f6
--- /dev/null
+++ b/ggml/src/ggml-cpu/llamafile/sgemm.cpp
@@ -0,0 +1,1884 @@
+// Copyright 2024 Mozilla Foundation
+//
+// Permission is hereby granted, free of charge, to any person obtaining
+// a copy of this software and associated documentation files (the
+// "Software"), to deal in the Software without restriction, including
+// without limitation the rights to use, copy, modify, merge, publish,
+// distribute, sublicense, and/or sell copies of the Software, and to
+// permit persons to whom the Software is furnished to do so, subject to
+// the following conditions:
+//
+// The above copyright notice and this permission notice shall be
+// included in all copies or substantial portions of the Software.
+//
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+// EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+// MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+// NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+// BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+// ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+// CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+// SOFTWARE.
+
+//
+//                   _   _          ___ _      _   ___
+//                  | |_(_)_ _ _  _| _ ) |    /_\ / __|
+//                  |  _| | ' \ || | _ \ |__ / _ \\__ \.
+//                   \__|_|_||_\_, |___/____/_/ \_\___/
+//                             |__/
+//
+//                    BASIC LINEAR ALGEBRA SUBPROGRAMS
+//
+//
+// This file implements multithreaded CPU matrix multiplication for the
+// common contiguous use case C = Aᵀ * B. These kernels are designed to
+// have excellent performance[1] for matrices that fit in the CPU cache
+// without imposing any overhead such as cache filling or malloc calls.
+//
+// This implementation does not guarantee any upper bound with rounding
+// errors, which grow along with k. Our goal's to maximally exploit the
+// hardware for performance, and then use whatever resources remain for
+// improving numerical accuracy.
+//
+// [1] J. Tunney, ‘LLaMA Now Goes Faster on CPUs’, Mar. 2024. [Online].
+//     Available: https://justine.lol/matmul/. [Accessed: 29-Mar-2024].
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Wpedantic"
+#pragma GCC diagnostic ignored "-Wignored-attributes"
+#endif
+
+#include "sgemm.h"
+#include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
+#include "ggml-quants.h"
+
+#ifdef _MSC_VER
+#define NOINLINE __declspec(noinline)
+#else
+#define NOINLINE __attribute__((__noinline__))
+#endif
+
+#if defined(__ARM_NEON) || defined(__AVX512F__)
+#define VECTOR_REGISTERS 32
+#else
+#define VECTOR_REGISTERS 16
+#endif
+
+#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
+
+namespace {
+
+inline float unhalf(ggml_fp16_t d) {
+    return GGML_FP16_TO_FP32(d);
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// VECTORIZED ARITHMETIC OPERATIONS
+
+#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+inline __m128 add(__m128 x, __m128 y) { return _mm_add_ps(x, y); }
+inline __m128 sub(__m128 x, __m128 y) { return _mm_sub_ps(x, y); }
+inline __m128 mul(__m128 x, __m128 y) { return _mm_mul_ps(x, y); }
+#endif  // __SSE__
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+inline __m256 add(__m256 x, __m256 y) { return _mm256_add_ps(x, y); }
+inline __m256 sub(__m256 x, __m256 y) { return _mm256_sub_ps(x, y); }
+inline __m256 mul(__m256 x, __m256 y) { return _mm256_mul_ps(x, y); }
+#endif // __AVX__
+
+#if defined(__AVX512F__)
+inline __m512 add(__m512 x, __m512 y) { return _mm512_add_ps(x, y); }
+inline __m512 sub(__m512 x, __m512 y) { return _mm512_sub_ps(x, y); }
+inline __m512 mul(__m512 x, __m512 y) { return _mm512_mul_ps(x, y); }
+#endif // __AVX512F__
+
+#if defined(__ARM_NEON)
+inline float32x4_t add(float32x4_t x, float32x4_t y) { return vaddq_f32(x, y); }
+inline float32x4_t sub(float32x4_t x, float32x4_t y) { return vsubq_f32(x, y); }
+inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vmulq_f32(x, y); }
+#endif // __ARM_NEON
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+inline float16x8_t add(float16x8_t x, float16x8_t y) { return vaddq_f16(x, y); }
+inline float16x8_t sub(float16x8_t x, float16x8_t y) { return vsubq_f16(x, y); }
+inline float16x8_t mul(float16x8_t x, float16x8_t y) { return vmulq_f16(x, y); }
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+#if defined(__MMA__)
+typedef vector unsigned char vec_t;
+typedef __vector_quad acc_t;
+#endif
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// VECTORIZED FUSED MULTIPLY ADD
+
+/**
+ * Computes a * b + c.
+ */
+template <typename T, typename U>
+inline U madd(T a, T b, U c) {
+    return add(mul(a, b), c);
+}
+
+#if defined(__FMA__)
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+template <>
+inline __m256 madd(__m256 a, __m256 b, __m256 c) {
+    return _mm256_fmadd_ps(a, b, c);
+}
+#endif
+#if defined(__AVX512F__)
+template <>
+inline __m512 madd(__m512 a, __m512 b, __m512 c) {
+    return _mm512_fmadd_ps(a, b, c);
+}
+#endif
+#endif
+
+#if defined(__ARM_FEATURE_FMA)
+template <>
+inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
+    return vfmaq_f32(c, b, a);
+}
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
+template <>
+inline float16x8_t madd(float16x8_t a, float16x8_t b, float16x8_t c) {
+    return vfmaq_f16(c, b, a);
+}
+#endif
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// VECTORIZED HORIZONTAL SUM
+
+#if defined(__ARM_NEON)
+inline float hsum(float32x4_t x) {
+    return vaddvq_f32(x);
+}
+#endif // __ARM_NEON
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
+inline float hsum(float16x8_t x) {
+    return vaddvq_f32(vaddq_f32(vcvt_f32_f16(vget_low_f16(x)),
+                                vcvt_f32_f16(vget_high_f16(x))));
+}
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+inline float hsum(__m128 x) {
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+    x = _mm_add_ps(x, _mm_movehl_ps(x, x));
+    x = _mm_add_ss(x, _mm_movehdup_ps(x));
+#else
+    __m128 t;
+    t = _mm_shuffle_ps(x, x, _MM_SHUFFLE(2, 3, 0, 1));
+    x = _mm_add_ps(x, t);
+    t = _mm_movehl_ps(t, x);
+    x = _mm_add_ss(x, t);
+#endif
+    return _mm_cvtss_f32(x);
+}
+#endif
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+inline float hsum(__m256 x) {
+    return hsum(_mm_add_ps(_mm256_extractf128_ps(x, 1),
+                           _mm256_castps256_ps128(x)));
+}
+#endif // __AVX__
+
+#if defined(__AVX512F__)
+inline float hsum(__m512 x) {
+    return _mm512_reduce_add_ps(x);
+}
+#endif // __AVX512F__
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// VECTORIZED MEMORY LOADING
+
+template <typename T, typename U> T load(const U *);
+
+#if defined(__ARM_NEON)
+template <> inline float32x4_t load(const float *p) {
+    return vld1q_f32(p);
+}
+#if !defined(_MSC_VER)
+template <> inline float16x8_t load(const ggml_fp16_t *p) {
+    return vld1q_f16((const float16_t *)p);
+}
+template <> inline float32x4_t load(const ggml_fp16_t *p) {
+    return vcvt_f32_f16(vld1_f16((const float16_t *)p));
+}
+#endif // _MSC_VER
+#endif // __ARM_NEON
+
+#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+template <> inline __m128 load(const float *p) {
+    return _mm_loadu_ps(p);
+}
+#endif  // __SSE__
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+template <> inline __m256 load(const float *p) {
+    return _mm256_loadu_ps(p);
+}
+#endif // __AVX__
+
+#if defined(__F16C__)
+template <> inline __m256 load(const ggml_fp16_t *p) {
+    return _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)p));
+}
+#endif // __F16C__
+
+#if defined(__AVX512F__)
+template <> inline __m512 load(const float *p) {
+    return _mm512_loadu_ps(p);
+}
+template <> inline __m512 load(const ggml_fp16_t *p) {
+    return _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)p));
+}
+#endif // __AVX512F__
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// CONSTANTS
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
+static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// FLOATING POINT MATRIX MULTIPLICATION
+
+template <int KN, typename D, typename V, typename TA, typename TB, typename TC>
+class tinyBLAS {
+  public:
+    tinyBLAS(int64_t k,
+             const TA *A, int64_t lda,
+             const TB *B, int64_t ldb,
+             TC *C, int64_t ldc,
+             int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+    }
+
+    void matmul(int64_t m, int64_t n) {
+        mnpack(0, m, 0, n);
+    }
+
+  private:
+    NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t mc, nc, mp, np;
+        switch ((MIN(m - m0, 5) << 4) | MIN(n - n0, 5)) {
+#if VECTOR_REGISTERS == 32
+        case 0x55:
+            mc = 5;
+            nc = 5;
+            gemm<5, 5>(m0, m, n0, n);
+            break;
+        case 0x45:
+            mc = 4;
+            nc = 5;
+            gemm<4, 5>(m0, m, n0, n);
+            break;
+        case 0x54:
+            mc = 5;
+            nc = 4;
+            gemm<5, 4>(m0, m, n0, n);
+            break;
+        case 0x44:
+            mc = 4;
+            nc = 4;
+            gemm<4, 4>(m0, m, n0, n);
+            break;
+        case 0x53:
+            mc = 5;
+            nc = 3;
+            gemm<5, 3>(m0, m, n0, n);
+            break;
+        case 0x35:
+            mc = 3;
+            nc = 5;
+            gemm<3, 5>(m0, m, n0, n);
+            break;
+        case 0x43:
+            mc = 4;
+            nc = 3;
+            gemm<4, 3>(m0, m, n0, n);
+            break;
+#else
+        case 0x55:
+        case 0x54:
+        case 0x53:
+        case 0x45:
+        case 0x44:
+        case 0x43:
+            mc = 4;
+            nc = 3;
+            gemm<4, 3>(m0, m, n0, n);
+            break;
+        case 0x35:
+#endif
+        case 0x34:
+            mc = 3;
+            nc = 4;
+            gemm<3, 4>(m0, m, n0, n);
+            break;
+        case 0x52:
+            mc = 5;
+            nc = 2;
+            gemm<5, 2>(m0, m, n0, n);
+            break;
+        case 0x33:
+            mc = 3;
+            nc = 3;
+            gemm<3, 3>(m0, m, n0, n);
+            break;
+        case 0x25:
+            mc = 2;
+            nc = 5;
+            gemm<2, 5>(m0, m, n0, n);
+            break;
+        case 0x42:
+            mc = 4;
+            nc = 2;
+            gemm<4, 2>(m0, m, n0, n);
+            break;
+        case 0x24:
+            mc = 2;
+            nc = 4;
+            gemm<2, 4>(m0, m, n0, n);
+            break;
+        case 0x32:
+            mc = 3;
+            nc = 2;
+            gemm<3, 2>(m0, m, n0, n);
+            break;
+        case 0x23:
+            mc = 2;
+            nc = 3;
+            gemm<2, 3>(m0, m, n0, n);
+            break;
+        case 0x51:
+            mc = 5;
+            nc = 1;
+            gemm<5, 1>(m0, m, n0, n);
+            break;
+        case 0x41:
+            mc = 4;
+            nc = 1;
+            gemm<4, 1>(m0, m, n0, n);
+            break;
+        case 0x22:
+            mc = 2;
+            nc = 2;
+            gemm<2, 2>(m0, m, n0, n);
+            break;
+        case 0x15:
+            mc = 1;
+            nc = 5;
+            gemm<1, 5>(m0, m, n0, n);
+            break;
+        case 0x14:
+            mc = 1;
+            nc = 4;
+            gemm<1, 4>(m0, m, n0, n);
+            break;
+        case 0x31:
+            mc = 3;
+            nc = 1;
+            gemm<3, 1>(m0, m, n0, n);
+            break;
+        case 0x13:
+            mc = 1;
+            nc = 3;
+            gemm<1, 3>(m0, m, n0, n);
+            break;
+        case 0x21:
+            mc = 2;
+            nc = 1;
+            gemm<2, 1>(m0, m, n0, n);
+            break;
+        case 0x12:
+            mc = 1;
+            nc = 2;
+            gemm<1, 2>(m0, m, n0, n);
+            break;
+        case 0x11:
+            mc = 1;
+            nc = 1;
+            gemm<1, 1>(m0, m, n0, n);
+            break;
+        default:
+            return;
+        }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
+    }
+
+    template <int RM, int RN>
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            D Cv[RN][RM] = {};
+            for (int64_t l = 0; l < k; l += KN)
+                for (int64_t j = 0; j < RN; ++j)
+                    for (int64_t i = 0; i < RM; ++i)
+                        Cv[j][i] = madd(load<V>(A + lda * (ii + i) + l),
+                                        load<V>(B + ldb * (jj + j) + l),
+                                        Cv[j][i]);
+            for (int64_t j = 0; j < RN; ++j)
+                for (int64_t i = 0; i < RM; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+
+    const TA *const A;
+    const TB *const B;
+    TC *const C;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
+
+//////////////////////////////////////////////////////////////////////////////////////////
+// QUANT ZERO MATRIX MULTIPLICATION
+
+#if defined(__ARM_FEATURE_DOTPROD)
+template <typename TA>
+class tinyBLAS_Q0_ARM {
+  public:
+    tinyBLAS_Q0_ARM(int64_t k,
+                    const TA *A, int64_t lda,
+                    const block_q8_0 *B, int64_t ldb,
+                    float *C, int64_t ldc,
+                    int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+    }
+
+    void matmul(int64_t m, int64_t n) {
+        mnpack(0, m, 0, n);
+    }
+
+  private:
+    NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t mc, nc, mp, np;
+        switch ((MIN(m - m0, 3) << 4) | MIN(n - n0, 3ll)) {
+        case 0x33:
+            mc = 3;
+            nc = 3;
+            gemm<3, 3>(m0, m, n0, n);
+            break;
+        case 0x32:
+            mc = 3;
+            nc = 2;
+            gemm<3, 2>(m0, m, n0, n);
+            break;
+        case 0x23:
+            mc = 2;
+            nc = 3;
+            gemm<2, 3>(m0, m, n0, n);
+            break;
+        case 0x22:
+            mc = 2;
+            nc = 2;
+            gemm<2, 2>(m0, m, n0, n);
+            break;
+        case 0x31:
+            mc = 3;
+            nc = 1;
+            gemm<3, 1>(m0, m, n0, n);
+            break;
+        case 0x13:
+            mc = 1;
+            nc = 3;
+            gemm<1, 3>(m0, m, n0, n);
+            break;
+        case 0x21:
+            mc = 2;
+            nc = 1;
+            gemm<2, 1>(m0, m, n0, n);
+            break;
+        case 0x12:
+            mc = 1;
+            nc = 2;
+            gemm<1, 2>(m0, m, n0, n);
+            break;
+        case 0x11:
+            mc = 1;
+            nc = 1;
+            gemm<1, 1>(m0, m, n0, n);
+            break;
+        default:
+            return;
+        }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
+    }
+
+    template <int RM, int RN>
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            float32x4_t Cv[RN][RM] = {};
+            for (int64_t l = 0; l < k; ++l)
+                for (int64_t j = 0; j < RN; ++j)
+                    for (int64_t i = 0; i < RM; ++i)
+                        Cv[j][i] = vmlaq_n_f32(Cv[j][i],
+                                               vcvtq_f32_s32(vdotq_s32(
+                                                   vdotq_s32(vdupq_n_s32(0),
+                                                             load_lo(A + lda * (ii + i) + l),
+                                                             load_lo(B + ldb * (jj + j) + l)),
+                                                   load_hi(A + lda * (ii + i) + l),
+                                                   load_hi(B + ldb * (jj + j) + l))),
+                                               unhalf(A[lda * (ii + i) + l].d) *
+                                               unhalf(B[ldb * (jj + j) + l].d));
+            for (int64_t j = 0; j < RN; ++j)
+                for (int64_t i = 0; i < RM; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+
+    inline int8x16_t load_lo(const block_q8_0 *b) {
+        return vld1q_s8(b->qs);
+    }
+
+    inline int8x16_t load_hi(const block_q8_0 *b) {
+        return vld1q_s8(b->qs + 16);
+    }
+
+    inline int8x16_t load_lo(const block_q4_0 *b) {
+        return vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vld1q_u8(b->qs),
+                                                     vdupq_n_u8(0x0f))),
+                        vdupq_n_s8(0x8));
+    }
+
+    inline int8x16_t load_hi(const block_q4_0 *b) {
+        return vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(vld1q_u8(b->qs), 4)),
+                        vdupq_n_s8(0x8));
+    }
+
+    const TA *const A;
+    const block_q8_0 *const B;
+    float *const C;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
+#endif // __ARM_FEATURE_DOTPROD
+
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+template <typename TA, typename TB, typename TC>
+class tinyBLAS_Q0_AVX {
+  public:
+    tinyBLAS_Q0_AVX(int64_t k,
+                    const TA *A, int64_t lda,
+                    const TB *B, int64_t ldb,
+                    TC *C, int64_t ldc,
+                    int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+    }
+
+    void matmul(int64_t m, int64_t n) {
+        mnpack(0, m, 0, n);
+    }
+
+  private:
+    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t mc, nc, mp, np;
+        switch ((MIN(m - m0, 4) << 4) | MIN(n - n0, 4)) {
+#if VECTOR_REGISTERS == 32
+        case 0x44:
+            mc = 4;
+            nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<4>(m0, m, n0, n);
+#else
+            gemm<4, 4>(m0, m, n0, n);
+#endif
+            break;
+        case 0x43:
+            mc = 4;
+            nc = 3;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<3>(m0, m, n0, n);
+#else
+            gemm<4, 3>(m0, m, n0, n);
+#endif
+            break;
+        case 0x34:
+            mc = 3;
+            nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<3>(m0, m, n0, n);
+#else
+            gemm<3, 4>(m0, m, n0, n);
+#endif
+            break;
+        case 0x33:
+            mc = 3;
+            nc = 3;
+            gemm<3, 3>(m0, m, n0, n);
+            break;
+        case 0x42:
+            mc = 4;
+            nc = 2;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<2>(m0, m, n0, n);
+#else
+            gemm<4, 2>(m0, m, n0, n);
+#endif
+            break;
+        case 0x24:
+            mc = 2;
+            nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<2>(m0, m, n0, n);
+#else
+            gemm<2, 4>(m0, m, n0, n);
+#endif
+            break;
+#else
+        case 0x44:
+        case 0x43:
+        case 0x42:
+            mc = 4;
+            nc = 2;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<2>(m0, m, n0, n);
+#else
+            gemm<4, 2>(m0, m, n0, n);
+#endif
+            break;
+        case 0x34:
+        case 0x24:
+            mc = 2;
+            nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<2>(m0, m, n0, n);
+#else
+            gemm<2, 4>(m0, m, n0, n);
+#endif
+            break;
+        case 0x33:
+#endif
+        case 0x32:
+            mc = 3;
+            nc = 2;
+            gemm<3, 2>(m0, m, n0, n);
+            break;
+        case 0x23:
+            mc = 2;
+            nc = 3;
+            gemm<2, 3>(m0, m, n0, n);
+            break;
+        case 0x41:
+            mc = 4;
+            nc = 1;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemm4xN<1>(m0, m, n0, n);
+#else
+            gemm<4, 1>(m0, m, n0, n);
+#endif
+            break;
+        case 0x22:
+            mc = 2;
+            nc = 2;
+            gemm<2, 2>(m0, m, n0, n);
+            break;
+        case 0x14:
+            mc = 1;
+            nc = 4;
+#if defined(__AVX2__) && defined(__F16C__)
+            gemmMx4<1>(m0, m, n0, n);
+#else
+            gemm<1, 4>(m0, m, n0, n);
+#endif
+            break;
+        case 0x31:
+            mc = 3;
+            nc = 1;
+            gemm<3, 1>(m0, m, n0, n);
+            break;
+        case 0x13:
+            mc = 1;
+            nc = 3;
+            gemm<1, 3>(m0, m, n0, n);
+            break;
+        case 0x21:
+            mc = 2;
+            nc = 1;
+            gemm<2, 1>(m0, m, n0, n);
+            break;
+        case 0x12:
+            mc = 1;
+            nc = 2;
+            gemm<1, 2>(m0, m, n0, n);
+            break;
+        case 0x11:
+            mc = 1;
+            nc = 1;
+            gemm<1, 1>(m0, m, n0, n);
+            break;
+        default:
+            return;
+        }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
+    }
+
+#if defined(__AVX2__) && defined(__F16C__)
+// Templated functions for gemm of dimensions 4xN
+    template <int RN>
+    NOINLINE void gemm4xN(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / 4;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * 4;
+            int64_t jj = n0 + job % xtiles * RN;
+            __m256 Cv[RN][4] = {};
+            for (int64_t l = 0; l < k; ++l) {
+                uint64_t a_delta = ((uint64_t)A[lda * (ii + 3) + l].d << 48) | ((uint64_t)A[lda * (ii + 2) + l].d << 32) | ((uint64_t)A[lda * (ii + 1) + l].d << 16) | (A[lda * (ii + 0) + l].d);
+                // Convert delta values for four blocks to float values
+                __m128 da = _mm_cvtph_ps(_mm_set_epi64x(0, a_delta));
+                __m256i avec0 = load(A + lda * (ii + 0) + l);
+                __m256i avec1 = load(A + lda * (ii + 1) + l);
+                __m256i avec2 = load(A + lda * (ii + 2) + l);
+                __m256i avec3 = load(A + lda * (ii + 3) + l);
+                for (int64_t j = 0; j < RN; ++j) {
+                        __m128 db = _mm_set1_ps(unhalf(B[ldb * (jj + j) + l].d));
+                        // Computation of product of delta values for four blocks and replicate it across 256 bit lane
+                        __m256 dvec =  _mm256_castps128_ps256(_mm_mul_ps(da, db));
+                        dvec = _mm256_permute2f128_ps(dvec ,dvec, 0);
+                        // Computation of dot product and multiplication with appropriate delta value products
+                        Cv[j][0] = madd(_mm256_shuffle_ps(dvec, dvec, 0),
+                                    updot(_mm256_sign_epi8(avec0, avec0),
+                                          _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec0)),
+                                    Cv[j][0]);
+                        Cv[j][1] = madd(_mm256_shuffle_ps(dvec, dvec, 85),
+                                    updot(_mm256_sign_epi8(avec1, avec1),
+                                            _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec1)),
+                                    Cv[j][1]);
+                        Cv[j][2] = madd(_mm256_shuffle_ps(dvec, dvec, 170),
+                                    updot(_mm256_sign_epi8(avec2, avec2),
+                                            _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec2)),
+                                    Cv[j][2]);
+                        Cv[j][3] = madd(_mm256_shuffle_ps(dvec, dvec, 255),
+                                    updot(_mm256_sign_epi8(avec3, avec3),
+                                            _mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec3)),
+                                    Cv[j][3]);
+                }
+            }
+
+            for (int64_t j = 0; j < RN; ++j)
+                for (int64_t i = 0; i < 4; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+
+    // Templated functions for gemm of dimensions Mx4
+    template <int RM>
+    NOINLINE void gemmMx4(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / 4;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * 4;
+            __m256 Cv[4][RM] = {};
+            for (int64_t l = 0; l < k; ++l) {
+                uint64_t b_delta = ((uint64_t)B[ldb * (jj + 3) + l].d << 48) | ((uint64_t)B[ldb * (jj + 2) + l].d << 32) | ((uint64_t)B[ldb * (jj + 1) + l].d << 16) | (B[ldb * (jj + 0) + l].d);
+                // Convert delta values for four blocks to float values
+                __m128 db = _mm_cvtph_ps(_mm_set_epi64x(0, b_delta));
+                __m256i bvec0 = load(B + ldb * (jj + 0) + l);
+                __m256i bvec1 = load(B + ldb * (jj + 1) + l);
+                __m256i bvec2 = load(B + ldb * (jj + 2) + l);
+                __m256i bvec3 = load(B + ldb * (jj + 3) + l);
+                for (int64_t i = 0; i < RM; ++i) {
+                    __m128 da = _mm_set1_ps(unhalf((A[lda * (ii + i) + l].d)));
+                    // Computation of product of delta values for four blocks and replicate it across 256 bit lane
+                    __m256 dvec =  _mm256_castps128_ps256(_mm_mul_ps(da, db));
+                    dvec = _mm256_permute2f128_ps(dvec ,dvec, 0);
+                    // Computation of dot product and multiplication with appropriate delta value products
+                    Cv[0][i] = madd(_mm256_shuffle_ps(dvec, dvec, 0),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec0, load(A + lda * (ii + i) + l))),
+                                    Cv[0][i]);
+                    Cv[1][i] = madd(_mm256_shuffle_ps(dvec, dvec, 85),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec1, load(A + lda * (ii + i) + l))),
+                                    Cv[1][i]);
+                    Cv[2][i] = madd(_mm256_shuffle_ps(dvec, dvec, 170),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec2, load(A + lda * (ii + i) + l))),
+                                    Cv[2][i]);
+                    Cv[3][i] = madd(_mm256_shuffle_ps(dvec, dvec, 255),
+                                    updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                            load(A + lda * (ii + i) + l)),
+                                            _mm256_sign_epi8(bvec3, load(A + lda * (ii + i) + l))),
+                                    Cv[3][i]);
+                }
+            }
+            for (int64_t j = 0; j < 4; ++j)
+                for (int64_t i = 0; i < RM; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+#endif
+
+    template <int RM, int RN>
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            __m256 Cv[RN][RM] = {};
+            for (int64_t l = 0; l < k; ++l)
+                for (int64_t j = 0; j < RN; ++j)
+                    for (int64_t i = 0; i < RM; ++i) {
+#if defined(__AVX2__)
+                        __m256 udTmp = updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                              load(A + lda * (ii + i) + l)),
+                                             _mm256_sign_epi8(load(B + ldb * (jj + j) + l),
+                                                              load(A + lda * (ii + i) + l)));
+#else
+                        __m128i ali0 = load0(A + lda * (ii + i) + l);
+                        __m128i ali1 = load1(A + lda * (ii + i) + l);
+                        __m128i blj0 = load0(B + ldb * (jj + j) + l);
+                        __m128i blj1 = load1(B + ldb * (jj + j) + l);
+
+                        __m128i sepAA0 = _mm_sign_epi8(ali0, ali0);
+                        __m128i sepAA1 = _mm_sign_epi8(ali1, ali1);
+                        __m128i sepBA0 = _mm_sign_epi8(blj0, ali0);
+                        __m128i sepBA1 = _mm_sign_epi8(blj1, ali1);
+
+                        // updot
+                        const __m128i oneFill = _mm_set1_epi16(1);
+                        __m128i mad0 = _mm_maddubs_epi16(sepAA0, sepBA0);
+                        __m128i mad1 = _mm_maddubs_epi16(sepAA1, sepBA1);
+                        __m256 udTmp = _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_madd_epi16(oneFill, mad1), _mm_madd_epi16(oneFill, mad0)));
+#endif
+                        Cv[j][i] = madd(_mm256_set1_ps(unhalf(A[lda * (ii + i) + l].d) *
+                                                       unhalf(B[ldb * (jj + j) + l].d)),
+                                                       udTmp,
+                                                       Cv[j][i]);
+                    }
+            for (int64_t j = 0; j < RN; ++j)
+                for (int64_t i = 0; i < RM; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+
+    inline __m256i load(const block_q8_0 *b) {
+        return _mm256_loadu_si256((const __m256i *)b->qs);
+    }
+
+    inline __m128i load0(const block_q8_0 *b) {
+        return _mm_loadu_si128((const __m128i *)b->qs);
+    }
+
+    inline __m128i load1(const block_q8_0 *b) {
+        return _mm_loadu_si128(((const __m128i *)b->qs) + 1);
+    }
+
+    inline __m256i load(const block_q4_0 *b) {
+        return _mm256_sub_epi8(denibble(b->qs), _mm256_set1_epi8(8));
+    }
+
+    inline __m128i load0(const block_q4_0 *b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), x), _mm_set1_epi8(8));
+    }
+
+    inline __m128i load1(const block_q4_0 *b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)), _mm_set1_epi8(8));
+    }
+
+    inline __m256i load(const block_q5_0 *b) {
+        return _mm256_or_si256(denibble(b->qs), bittobyte(b->qh));
+    }
+
+    inline __m128i load0(const block_q5_0* b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        uint32_t x32;
+        memcpy(&x32, b->qh, sizeof(uint32_t));
+        __m128i qxl = _mm_and_si128(_mm_set1_epi8(15), x);
+        __m128i bytesl = _mm_cmpeq_epi8(_mm_set1_epi64x(-1),
+                                        _mm_or_si128(_mm_set1_epi64x(0x7fbfdfeff7fbfdfe),
+                                                     _mm_shuffle_epi8(_mm_set1_epi32(x32),
+                                                                      _mm_set_epi64x(0x0101010101010101, 0x0000000000000000))));
+        bytesl = _mm_andnot_si128(bytesl, _mm_set1_epi8((char)0xF0));
+        return _mm_or_si128(qxl, bytesl);
+    }
+
+    inline __m128i load1(const block_q5_0* b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        uint32_t x32;
+        memcpy(&x32, b->qh, sizeof(uint32_t));
+        __m128i qxh = _mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4));
+        __m128i bytesh = _mm_cmpeq_epi8(_mm_set1_epi64x(-1),
+                                        _mm_or_si128(_mm_set1_epi64x(0x7fbfdfeff7fbfdfe),
+                                                     _mm_shuffle_epi8(_mm_set1_epi32(x32),
+                                                                      _mm_set_epi64x(0x0303030303030303, 0x0202020202020202))));
+        bytesh = _mm_andnot_si128(bytesh, _mm_set1_epi8((char)0xF0));
+        return _mm_or_si128(qxh, bytesh);
+    }
+
+    inline __m256i load(const block_iq4_nl *b) {
+        return MM256_SET_M128I(load1(b), load0(b));
+    }
+
+    inline __m128i load0(const block_iq4_nl *b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        return _mm_shuffle_epi8(iq4nlt, _mm_and_si128(_mm_set1_epi8(15), x));
+    }
+
+    inline __m128i load1(const block_iq4_nl *b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        return _mm_shuffle_epi8(iq4nlt, _mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)));
+    }
+
+    inline __m256 updot(__m256i u, __m256i s) {
+        __m256i res;
+#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+        res = _mm256_dpbusd_epi32(_mm256_setzero_si256(), u, s);
+#else
+        res = _mm256_madd_epi16(_mm256_set1_epi16(1), _mm256_maddubs_epi16(u, s));
+#endif
+        return _mm256_cvtepi32_ps(res);
+    }
+
+    static inline __m256i denibble(const uint8_t *p) {
+        __m128i x = _mm_loadu_si128((const __m128i *)p);
+        return _mm256_and_si256(_mm256_set1_epi8(15),
+                                _mm256_insertf128_si256(_mm256_castsi128_si256(x),
+                                                        _mm_srli_epi16(x, 4), 1));
+    }
+
+    static inline __m256i bittobyte(const uint8_t *p) {
+        uint32_t x32;
+        memcpy(&x32, p, sizeof(uint32_t));
+        __m256i bytes = _mm256_cmpeq_epi8(_mm256_set1_epi64x(-1),
+                                          _mm256_or_si256(_mm256_set1_epi64x(0x7fbfdfeff7fbfdfe),
+                                                          _mm256_shuffle_epi8(_mm256_set1_epi32(x32),
+                                                                              _mm256_set_epi64x(0x0303030303030303, 0x0202020202020202,
+                                                                                                0x0101010101010101, 0x0000000000000000))));
+        return _mm256_andnot_si256(bytes, _mm256_set1_epi8((char)0xF0));
+    }
+
+    const TA *const A;
+    const TB *const B;
+    TC *const C;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
+#endif // __AVX__
+
+//PPC Implementation
+#if defined(__MMA__)
+
+#define SAVE_ACC(ACC, ii, jj) \
+   __builtin_mma_disassemble_acc(vec_C, ACC); \
+   for (int I = 0; I < 4; I++) { \
+      for (int J = 0; J < 4; J++) { \
+         *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&vec_C[I]+J); \
+      } \
+   } \
+
+template <typename TA, typename TB, typename TC>
+class tinyBLAS_PPC {
+  public:
+    tinyBLAS_PPC(int64_t k,
+                const TA *A, int64_t lda,
+                const TB *B, int64_t ldb,
+                TC *C, int64_t ldc,
+                int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+    }
+
+    void matmul(int64_t m, int64_t n) {
+       mnpack(0, m, 0, n);
+    }
+
+  private:
+
+    void (tinyBLAS_PPC::*kernel)(int64_t, int64_t);
+
+    void READ_BLOCK(const float* a, int64_t lda, int rows, int cols, float* vec) {
+        int64_t i, j;
+        float *aoffset = NULL, *boffset = NULL;
+        float *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
+        float *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
+
+        aoffset = const_cast<float*>(a);
+        boffset = vec;
+        j = (rows >> 3);
+        if (j > 0) {
+            do {
+                aoffset1 = aoffset;
+                aoffset2 = aoffset1 + lda;
+                aoffset3 = aoffset2 + lda;
+                aoffset4 = aoffset3 + lda;
+                aoffset5 = aoffset4 + lda;
+                aoffset6 = aoffset5 + lda;
+                aoffset7 = aoffset6 + lda;
+                aoffset8 = aoffset7 + lda;
+                aoffset += 8 * lda;
+                i = (cols >> 3);
+                if (i > 0) {
+                    __vector_pair C1, C2, C3, C4, C5, C6, C7, C8;
+                    vector float c1[2], c2[2], c3[2], c4[2], c5[2], c6[2], c7[2], c8[2];
+                    vector float t1, t2, t3, t4, t5, t6, t7, t8;
+                    do {
+                        C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1);
+                        C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2);
+                        C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3);
+                        C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4);
+                        C5 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset5);
+                        C6 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset6);
+                        C7 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset7);
+                        C8 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset8);
+                        __builtin_vsx_disassemble_pair(c1, &C1);
+                        __builtin_vsx_disassemble_pair(c2, &C2);
+                        __builtin_vsx_disassemble_pair(c3, &C3);
+                        __builtin_vsx_disassemble_pair(c4, &C4);
+                        __builtin_vsx_disassemble_pair(c5, &C5);
+                        __builtin_vsx_disassemble_pair(c6, &C6);
+                        __builtin_vsx_disassemble_pair(c7, &C7);
+                        __builtin_vsx_disassemble_pair(c8, &C8);
+
+                        t1 = vec_mergeh(c1[0], c2[0]);
+                        t2 = vec_mergeh(c3[0], c4[0]);
+                        t3 = vec_mergeh(c5[0], c6[0]);
+                        t4 = vec_mergeh(c7[0], c8[0]);
+                        t5 = vec_xxpermdi(t1, t2, 0);
+                        t6 = vec_xxpermdi(t3, t4, 0);
+                        t7 = vec_xxpermdi(t1, t2, 3);
+                        t8 = vec_xxpermdi(t3, t4, 3);
+                        vec_xst(t5, 0, boffset);
+                        vec_xst(t6, 0, boffset+4);
+                        vec_xst(t7, 0, boffset+8);
+                        vec_xst(t8, 0, boffset+12);
+
+                        t1 = vec_mergel(c1[0], c2[0]);
+                        t2 = vec_mergel(c3[0], c4[0]);
+                        t3 = vec_mergel(c5[0], c6[0]);
+                        t4 = vec_mergel(c7[0], c8[0]);
+                        t5 = vec_xxpermdi(t1, t2, 0);
+                        t6 = vec_xxpermdi(t3, t4, 0);
+                        t7 = vec_xxpermdi(t1, t2, 3);
+                        t8 = vec_xxpermdi(t3, t4, 3);
+                        vec_xst(t5, 0, boffset+16);
+                        vec_xst(t6, 0, boffset+20);
+                        vec_xst(t7, 0, boffset+24);
+                        vec_xst(t8, 0, boffset+28);
+
+                        t1 = vec_mergeh(c1[1], c2[1]);
+                        t2 = vec_mergeh(c3[1], c4[1]);
+                        t3 = vec_mergeh(c5[1], c6[1]);
+                        t4 = vec_mergeh(c7[1], c8[1]);
+                        t5 = vec_xxpermdi(t1, t2, 0);
+                        t6 = vec_xxpermdi(t3, t4, 0);
+                        t7 = vec_xxpermdi(t1, t2, 3);
+                        t8 = vec_xxpermdi(t3, t4, 3);
+                        vec_xst(t5, 0, boffset+32);
+                        vec_xst(t6, 0, boffset+36);
+                        vec_xst(t7, 0, boffset+40);
+                        vec_xst(t8, 0, boffset+44);
+
+                        t1 = vec_mergel(c1[1], c2[1]);
+                        t2 = vec_mergel(c3[1], c4[1]);
+                        t3 = vec_mergel(c5[1], c6[1]);
+                        t4 = vec_mergel(c7[1], c8[1]);
+                        t5 = vec_xxpermdi(t1, t2, 0);
+                        t6 = vec_xxpermdi(t3, t4, 0);
+                        t7 = vec_xxpermdi(t1, t2, 3);
+                        t8 = vec_xxpermdi(t3, t4, 3);
+                        vec_xst(t5, 0, boffset+48);
+                        vec_xst(t6, 0, boffset+52);
+                        vec_xst(t7, 0, boffset+56);
+                        vec_xst(t8, 0, boffset+60);
+
+                        aoffset1 += 8*lda;
+                        aoffset2 += 8*lda;
+                        aoffset3 += 8*lda;
+                        aoffset4 += 8*lda;
+                        boffset += 64;
+                        i--;
+                    } while(i > 0);
+                }
+                if (cols & 4) {
+                    vector float c1, c2, c3, c4, c5, c6, c7, c8;
+                    vector float t1, t2, t3, t4, t5, t6, t7, t8;
+                    c1 = vec_xl(0, aoffset1);
+                    c2 = vec_xl(0, aoffset2);
+                    c3 = vec_xl(0, aoffset3);
+                    c4 = vec_xl(0, aoffset4);
+                    c5 = vec_xl(0, aoffset5);
+                    c6 = vec_xl(0, aoffset6);
+                    c7 = vec_xl(0, aoffset7);
+                    c8 = vec_xl(0, aoffset8);
+
+                    t1 = vec_mergeh(c1, c2);
+                    t2 = vec_mergeh(c3, c4);
+                    t3 = vec_mergeh(c5, c6);
+                    t4 = vec_mergeh(c7, c8);
+                    t5 = vec_xxpermdi(t1, t2, 0);
+                    t6 = vec_xxpermdi(t3, t4, 0);
+                    t7 = vec_xxpermdi(t1, t2, 3);
+                    t8 = vec_xxpermdi(t3, t4, 3);
+                    vec_xst(t5, 0, boffset);
+                    vec_xst(t6, 0, boffset+4);
+                    vec_xst(t7, 0, boffset+8);
+                    vec_xst(t8, 0, boffset+12);
+
+                    t1 = vec_mergel(c1, c2);
+                    t2 = vec_mergel(c3, c4);
+                    t3 = vec_mergel(c5, c6);
+                    t4 = vec_mergel(c7, c8);
+                    t5 = vec_xxpermdi(t1, t2, 0);
+                    t6 = vec_xxpermdi(t3, t4, 0);
+                    t7 = vec_xxpermdi(t1, t2, 3);
+                    t8 = vec_xxpermdi(t3, t4, 3);
+                    vec_xst(t5, 0, boffset+16);
+                    vec_xst(t6, 0, boffset+20);
+                    vec_xst(t7, 0, boffset+24);
+                    vec_xst(t8, 0, boffset+28);
+                }
+            j--;
+            } while(j > 0);
+        }
+
+        if (rows & 4) {
+            aoffset1 = aoffset;
+            aoffset2 = aoffset1 + lda;
+            aoffset3 = aoffset2 + lda;
+            aoffset4 = aoffset3 + lda;
+            aoffset += 4 * lda;
+            i = (cols >> 3);
+            if (i > 0) {
+                __vector_pair C1, C2, C3, C4;
+                vector float c1[2], c2[2], c3[2], c4[2];
+                vector float t1, t2, t3, t4, t5, t6, t7, t8;
+                do {
+                    C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1);
+                    C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2);
+                    C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3);
+                    C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4);
+                    __builtin_vsx_disassemble_pair(c1, &C1);
+                    __builtin_vsx_disassemble_pair(c2, &C2);
+                    __builtin_vsx_disassemble_pair(c3, &C3);
+                    __builtin_vsx_disassemble_pair(c4, &C4);
+
+                    t1 = vec_mergeh(c1[0], c2[0]);
+                    t2 = vec_mergeh(c3[0], c4[0]);
+                    t3 = vec_mergel(c1[0], c2[0]);
+                    t4 = vec_mergel(c3[0], c4[0]);
+                    t5 = vec_xxpermdi(t1, t2, 0);
+                    t6 = vec_xxpermdi(t1, t2, 3);
+                    t7 = vec_xxpermdi(t3, t4, 0);
+                    t8 = vec_xxpermdi(t3, t4, 3);
+                    vec_xst(t5, 0, boffset);
+                    vec_xst(t6, 0, boffset+4);
+                    vec_xst(t7, 0, boffset+8);
+                    vec_xst(t8, 0, boffset+12);
+
+                    t1 = vec_mergeh(c1[1], c2[1]);
+                    t2 = vec_mergeh(c3[1], c4[1]);
+                    t3 = vec_mergel(c1[1], c2[1]);
+                    t4 = vec_mergel(c3[1], c4[1]);
+                    t5 = vec_xxpermdi(t1, t2, 0);
+                    t6 = vec_xxpermdi(t1, t2, 3);
+                    t7 = vec_xxpermdi(t3, t4, 0);
+                    t8 = vec_xxpermdi(t3, t4, 3);
+                    vec_xst(t5, 0, boffset+16);
+                    vec_xst(t6, 0, boffset+20);
+                    vec_xst(t7, 0, boffset+24);
+                    vec_xst(t8, 0, boffset+28);
+
+                    aoffset1 += 8*lda;
+                    aoffset2 += 8*lda;
+                    aoffset3 += 8*lda;
+                    aoffset4 += 8*lda;
+                    boffset += 32;
+                    i--;
+                } while(i > 0);
+            }
+
+            if (cols & 4) {
+                vector float c1, c2, c3, c4;
+                vector float t1, t2, t3, t4;
+                c1 = vec_xl(0, aoffset1);
+                c2 = vec_xl(0, aoffset2);
+                c3 = vec_xl(0, aoffset3);
+                c4 = vec_xl(0, aoffset4);
+
+                t1 = vec_mergeh(c1, c2);
+                t2 = vec_mergeh(c3, c4);
+                t3 = vec_xxpermdi(t1, t2, 0);
+                t4 = vec_xxpermdi(t1, t2, 3);
+                vec_xst(t3, 0, boffset);
+                vec_xst(t4, 0, boffset+4);
+
+                t1 = vec_mergel(c1, c2);
+                t2 = vec_mergel(c3, c4);
+                t3 = vec_xxpermdi(t1, t2, 0);
+                t4 = vec_xxpermdi(t1, t2, 3);
+                vec_xst(t3, 0, boffset+8);
+                vec_xst(t4, 0, boffset+12);
+            }
+        }
+        if (rows & 3) {
+            aoffset1 = aoffset;
+            aoffset2 = aoffset1 + lda;
+            aoffset3 = aoffset2 + lda;
+            if (cols & 4) {
+                vector float c1, c2, c3, c4 = {0};
+                vector float t1, t2, t3, t4;
+                c1 = vec_xl(0, aoffset1);
+                c2 = vec_xl(0, aoffset2);
+                c3 = vec_xl(0, aoffset3);
+
+                t1 = vec_mergeh(c1, c2);
+                t2 = vec_mergeh(c3, c4);
+                t3 = vec_xxpermdi(t1, t2, 0);
+                t4 = vec_xxpermdi(t1, t2, 3);
+                vec_xst(t3, 0, boffset);
+                vec_xst(t4, 0, boffset+4);
+
+                t1 = vec_mergel(c1, c2);
+                t2 = vec_mergel(c3, c4);
+                t3 = vec_xxpermdi(t1, t2, 0);
+                t4 = vec_xxpermdi(t1, t2, 3);
+                vec_xst(t3, 0, boffset+8);
+                vec_xst(t4, 0, boffset+12);
+            }
+        }
+    }
+
+    void KERNEL_4x4(int64_t ii, int64_t jj) {
+        vec_t vec_A[4], vec_B[4], vec_C[4];
+        acc_t acc_0;
+        __builtin_mma_xxsetaccz(&acc_0);
+        for (int l = 0; l < k; l+=4) {
+            READ_BLOCK(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A);
+            READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[2], vec_B[2]);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[3], vec_B[3]);
+        }
+        SAVE_ACC(&acc_0, ii, jj);
+    }
+
+    void KERNEL_4x8(int64_t ii, int64_t jj) {
+        vec_t vec_A[4], vec_B[8], vec_C[4];
+        acc_t acc_0, acc_1;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        for (int64_t l = 0; l < k; l+=4) {
+            READ_BLOCK(A+(ii*lda)+l, lda, 4, 4, (float*)vec_A);
+            READ_BLOCK(B+(jj*ldb)+l, ldb, 8, 4, (float*)vec_B);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], (vec_t)vec_B[0]);
+            __builtin_mma_xvf32gerpp(&acc_1, vec_A[0], (vec_t)vec_B[1]);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], (vec_t)vec_B[2]);
+            __builtin_mma_xvf32gerpp(&acc_1, vec_A[1], (vec_t)vec_B[3]);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[2], (vec_t)vec_B[4]);
+            __builtin_mma_xvf32gerpp(&acc_1, vec_A[2], (vec_t)vec_B[5]);
+            __builtin_mma_xvf32gerpp(&acc_0, vec_A[3], (vec_t)vec_B[6]);
+            __builtin_mma_xvf32gerpp(&acc_1, vec_A[3], (vec_t)vec_B[7]);
+        }
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii, jj+4);
+    }
+
+    void KERNEL_8x4(int64_t ii, int64_t jj) {
+        vec_t vec_A[8], vec_B[4], vec_C[4];
+        acc_t acc_0, acc_1;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        for (int64_t l = 0; l < k; l+=4) {
+            READ_BLOCK(A+(ii*lda)+l, lda, 8, 4, (float*)vec_A);
+            READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
+            __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[0], vec_B[0]);
+            __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[1], vec_B[0]);
+            __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[2], vec_B[1]);
+            __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[3], vec_B[1]);
+            __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[4], vec_B[2]);
+            __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[5], vec_B[2]);
+            __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[6], vec_B[3]);
+            __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[7], vec_B[3]);
+        }
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii+4, jj);
+    }
+
+    void KERNEL_8x8(int64_t ii, int64_t jj) {
+        vec_t vec_A[16], vec_B[16], vec_C[4];
+        acc_t acc_0, acc_1, acc_2, acc_3;
+        __builtin_mma_xxsetaccz(&acc_0);
+        __builtin_mma_xxsetaccz(&acc_1);
+        __builtin_mma_xxsetaccz(&acc_2);
+        __builtin_mma_xxsetaccz(&acc_3);
+        for (int l = 0; l < k; l+=8) {
+            READ_BLOCK(A+(ii*lda)+l, lda, 8, 8, (float*)vec_A);
+            READ_BLOCK(B+(jj*ldb)+l, ldb, 8, 8, (float*)vec_B);
+            for(int x = 0; x < 16; x+=2) {
+                __builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[x], vec_B[x]);
+                __builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[x], vec_B[x+1]);
+                __builtin_mma_xvf32gerpp(&acc_2, (vec_t)vec_A[x+1], vec_B[x]);
+                __builtin_mma_xvf32gerpp(&acc_3, (vec_t)vec_A[x+1], vec_B[x+1]);
+            }
+        }
+        SAVE_ACC(&acc_0, ii, jj);
+        SAVE_ACC(&acc_1, ii, jj+4);
+        SAVE_ACC(&acc_2, ii+4, jj);
+        SAVE_ACC(&acc_3, ii+4, jj+4);
+    }
+
+    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t mc, nc, mp, np;
+        int m_rem = MIN(m - m0, 16);
+        int n_rem = MIN(n - n0, 16);
+        if (m_rem >= 16 && n_rem >= 8) {
+            mc = 8;
+            nc = 8;
+            gemm<8,8>(m0, m, n0, n);
+        } else if(m_rem >= 8 && n_rem >= 16) {
+            mc = 8;
+            nc = 8;
+            gemm<8,8>(m0, m, n0, n);
+        } else if (m_rem >= 8 && n_rem >= 8) {
+            mc = 8;
+            nc = 8;
+            gemm<8,8>(m0, m, n0, n);
+        } else if (m_rem >= 4 && n_rem >= 8) {
+            mc = 4;
+            nc = 8;
+            gemm<4,8>(m0, m, n0, n);
+        } else if (m_rem >= 8 && n_rem >= 4) {
+            mc = 8;
+            nc = 4;
+            gemm<8,4>(m0, m, n0, n);
+        } else if (m_rem >= 4 && n_rem >= 4) {
+            mc = 4;
+            nc = 4;
+            gemm<4,4>(m0, m, n0, n);
+        } else if ((m_rem < 4) && (n_rem > 4)) {
+            nc = 4;
+            switch(m_rem) {
+                case 1:
+                    mc = 1;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 2:
+                    mc = 2;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 3:
+                    mc = 3;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                default:
+                    return;
+            }
+        } else if ((m_rem > 4) && (n_rem < 4)) {
+            mc = 4;
+            switch(n_rem) {
+                case 1:
+                    nc = 1;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 2:
+                    nc = 2;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 3:
+                    nc = 3;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                default:
+                    return;
+            }
+        } else {
+            switch((m_rem << 4) | n_rem) {
+                case 0x43:
+                    mc = 4;
+                    nc = 3;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x42:
+                    mc = 4;
+                    nc = 2;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x41:
+                    mc = 4;
+                    nc = 1;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x34:
+                    mc = 3;
+                    nc = 4;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x33:
+                    mc = 3;
+                    nc = 3;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x32:
+                    mc = 3;
+                    nc = 2;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x31:
+                    mc = 3;
+                    nc = 1;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x24:
+                    mc = 2;
+                    nc = 4;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x23:
+                    mc = 2;
+                    nc = 3;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x22:
+                    mc = 2;
+                    nc = 2;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x21:
+                    mc = 2;
+                    nc = 1;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x14:
+                    mc = 1;
+                    nc = 4;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x13:
+                    mc = 1;
+                    nc = 3;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x12:
+                    mc = 1;
+                    nc = 2;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                case 0x11:
+                    mc = 1;
+                    nc = 1;
+                    gemm_small(m0, m, n0, n, mc, nc);
+                    break;
+                default:
+                    return;
+            }
+        }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
+    }
+
+     void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            vec_t vec_C[4];
+            acc_t acc_0;
+            __builtin_mma_xxsetaccz(&acc_0);
+            vec_t vec_A[4], vec_B[4];
+            for (int l=0; l<k; l+=4) {
+                if (RN >= 4 && RM == 1) {
+                    float* a = const_cast<float*>(A+(ii)*lda+l);
+                    READ_BLOCK(B+(jj*ldb)+l, ldb, 4, 4, (float*)vec_B);
+                    vec_A[0] = (vec_t)vec_xl(0,a);
+                    vec_A[1] = (vec_t)vec_splats(*((float*)&vec_A+1));
+                    vec_A[2] = (vec_t)vec_splats(*((float*)&vec_A+2));
+                    vec_A[3] = (vec_t)vec_splats(*((float*)&vec_A+3));
+                } else {
+                    READ_BLOCK(A+(ii*lda)+l, lda, RM, 4, (float*)vec_A);
+                    READ_BLOCK(B+(jj*ldb)+l, ldb, RN, 4, (float*)vec_B);
+                }
+                __builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
+                __builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
+                __builtin_mma_xvf32gerpp(&acc_0, vec_A[2], vec_B[2]);
+                __builtin_mma_xvf32gerpp(&acc_0, vec_A[3], vec_B[3]);
+            }
+            __builtin_mma_disassemble_acc(vec_C, &acc_0);
+            for (int I = 0; I < RM; I++) {
+                for (int J = 0; J < RN; J++) {
+                    *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&vec_C[I]+J);
+                }
+            }
+       }
+    }
+
+    template <int RM, int RN>
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (RM == 4 && RN == 4) {
+            kernel = &tinyBLAS_PPC::KERNEL_4x4;
+        } else if (RM == 4 && RN == 8) {
+            kernel = &tinyBLAS_PPC::KERNEL_4x8;
+        } else if (RM == 8 && RN == 4) {
+            kernel = &tinyBLAS_PPC::KERNEL_8x4;
+        } else if (RM == 8 && RN == 8) {
+            kernel = &tinyBLAS_PPC::KERNEL_8x8;
+        }
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            (this->*kernel)(ii, jj);
+        }
+    }
+
+    const TA *const A;
+    const TB *const B;
+    TC *C;
+    TA *At;
+    TB *Bt;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
+#endif
+} // namespace
+
+/**
+ * Performs optimized matrix multiplication on CPU.
+ *
+ * This subroutine may compute C = Aᵀ * B with column major ordering.
+ * Despite its name, this isn't a generalized implementation. Work is
+ * only performed when a handwritten kernel is written and available.
+ * Otherwise the caller should fall back to a general matmul routine.
+ *
+ * For example, for single-threaded single-precision GEMM you can say
+ *
+ *     llamafile_sgemm(m, n, k, A, lda, B, ldb, C, ldc,
+ *                     0, 1,
+ *                     GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32);
+ *
+ * @param m is rows in `A` and `C`
+ * @param n is cols in `B` and `C`
+ * @param k is cols in `A` and rows in `B`
+ * @param A is first input matrix (always transposed)
+ * @param lda is row stride of `A`
+ * @param B is second input matrix (never transposed)
+ * @param ldb is row stride of `B`
+ * @param C is input/output array of output matrices
+ * @param ldc is row stride of `C`
+ * @param ith is thread id (must be less than `nth`)
+ * @param nth is number of threads (must be greater than zero)
+ * @param Atype is GGML data type of `A`
+ * @param Btype is GGML data type of `B`
+ * @param Ctype is GGML data type of `C`
+ * @return true if this function was able to service the matmul request
+ */
+bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
+                     int64_t ldc, int ith, int nth, int Atype, int Btype, int Ctype) {
+
+    assert(m >= 0);
+    assert(n >= 0);
+    assert(k >= 0);
+    assert(lda >= k);
+    assert(ldb >= k);
+    assert(ldc >= m);
+    assert(nth > 0);
+    assert(ith < nth);
+
+    // only enable sgemm for prompt processing
+    if (n < 2)
+        return false;
+
+    if (Ctype != GGML_TYPE_F32)
+        return false;
+
+    switch (Atype) {
+
+    case GGML_TYPE_F32: {
+        if (Btype != GGML_TYPE_F32)
+            return false;
+#if defined(__AVX512F__)
+        if (k % 16)
+            return false;
+        tinyBLAS<16, __m512, __m512, float, float, float> tb{
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__AVX__) || defined(__AVX2__)
+        if (k % 8)
+            return false;
+        tinyBLAS<8, __m256, __m256, float, float, float> tb{
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_NEON)
+        if (n < 4)
+            return false;
+        if (k % 4)
+            return false;
+        tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__MMA__)
+        if (k % 8)
+            return false;
+        tinyBLAS_PPC<float, float, float> tb{
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    case GGML_TYPE_F16: {
+#if defined(__AVX512F__)
+        if (k % 16)
+            return false;
+        if (Btype != GGML_TYPE_F32)
+            return false;
+        tinyBLAS<16, __m512, __m512, ggml_fp16_t, float, float> tb{
+            k, (const ggml_fp16_t *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif (defined(__AVX__) || defined(__AVX2__)) && defined(__F16C__)
+        if (k % 8)
+            return false;
+        if (Btype != GGML_TYPE_F32)
+            return false;
+        tinyBLAS<8, __m256, __m256, ggml_fp16_t, float, float> tb{
+            k, (const ggml_fp16_t *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
+        if (n < 8)
+            return false;
+        if (k % 8)
+            return false;
+        if (Btype != GGML_TYPE_F16)
+            return false;
+        tinyBLAS<8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, float> tb{
+            k, (const ggml_fp16_t *)A, lda,
+            (const ggml_fp16_t *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_NEON) && !defined(_MSC_VER)
+        if (k % 4)
+            return false;
+        if (Btype != GGML_TYPE_F32)
+            return false;
+        tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, float, float> tb{
+            k, (const ggml_fp16_t *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    case GGML_TYPE_Q8_0: {
+        if (Btype != GGML_TYPE_Q8_0)
+           return false;
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+        tinyBLAS_Q0_AVX<block_q8_0, block_q8_0, float> tb{
+            k, (const block_q8_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_FEATURE_DOTPROD)
+        tinyBLAS_Q0_ARM<block_q8_0> tb{
+            k, (const block_q8_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    case GGML_TYPE_Q4_0: {
+        if (Btype != GGML_TYPE_Q8_0)
+            return false;
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+        tinyBLAS_Q0_AVX<block_q4_0, block_q8_0, float> tb{
+            k, (const block_q4_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_FEATURE_DOTPROD)
+        tinyBLAS_Q0_ARM<block_q4_0> tb{
+            k, (const block_q4_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    case GGML_TYPE_Q5_0: {
+        if (Btype != GGML_TYPE_Q8_0)
+            return false;
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+        tinyBLAS_Q0_AVX<block_q5_0, block_q8_0, float> tb{
+            k, (const block_q5_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    case GGML_TYPE_IQ4_NL: {
+        if (Btype != GGML_TYPE_Q8_0)
+            return false;
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+        tinyBLAS_Q0_AVX<block_iq4_nl, block_q8_0, float> tb{
+            k, (const block_iq4_nl *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    default:
+        return false;
+    }
+
+    (void)m;
+    (void)n;
+    (void)k;
+    (void)A;
+    (void)lda;
+    (void)B;
+    (void)ldb;
+    (void)C;
+    (void)ldc;
+    (void)ith;
+    (void)nth;
+    (void)Atype;
+    (void)Btype;
+    (void)Ctype;
+}
diff --git a/ggml/src/sgemm.h b/ggml/src/ggml-cpu/llamafile/sgemm.h
similarity index 100%
rename from ggml/src/sgemm.h
rename to ggml/src/ggml-cpu/llamafile/sgemm.h
diff --git a/ggml/src/ggml-cuda/CMakeLists.txt b/ggml/src/ggml-cuda/CMakeLists.txt
new file mode 100644
index 00000000000..14761650f8a
--- /dev/null
+++ b/ggml/src/ggml-cuda/CMakeLists.txt
@@ -0,0 +1,152 @@
+cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
+
+find_package(CUDAToolkit)
+
+if (CUDAToolkit_FOUND)
+    message(STATUS "CUDA Toolkit found")
+
+    if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+        # native == GPUs available at build time
+        # 52     == Maxwell, lowest CUDA 12 standard
+        # 60     == P100, FP16 CUDA intrinsics
+        # 61     == Pascal, __dp4a instruction (per-byte integer dot product)
+        # 70     == V100, FP16 tensor cores
+        # 75     == Turing, int8 tensor cores
+        if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
+            set(CMAKE_CUDA_ARCHITECTURES "native")
+        elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+            set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
+        else()
+            set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
+        endif()
+    endif()
+    message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
+
+    enable_language(CUDA)
+
+    file(GLOB   GGML_HEADERS_CUDA "*.cuh")
+    list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")
+
+    file(GLOB   GGML_SOURCES_CUDA "*.cu")
+    file(GLOB   SRCS "template-instances/fattn-wmma*.cu")
+    list(APPEND GGML_SOURCES_CUDA ${SRCS})
+    file(GLOB   SRCS "template-instances/mmq*.cu")
+    list(APPEND GGML_SOURCES_CUDA ${SRCS})
+
+    if (GGML_CUDA_FA_ALL_QUANTS)
+        file(GLOB   SRCS "template-instances/fattn-vec*.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+    else()
+        file(GLOB   SRCS "template-instances/fattn-vec*q4_0-q4_0.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        file(GLOB   SRCS "template-instances/fattn-vec*q8_0-q8_0.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        file(GLOB   SRCS "template-instances/fattn-vec*f16-f16.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+    endif()
+
+    ggml_add_backend_library(ggml-cuda
+                             ${GGML_HEADERS_CUDA}
+                             ${GGML_SOURCES_CUDA}
+                            )
+
+    add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
+
+    if (GGML_CUDA_GRAPHS)
+        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
+    endif()
+
+    if (GGML_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    if (GGML_CUDA_FORCE_CUBLAS)
+        add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        add_compile_definitions(GGML_CUDA_NO_VMM)
+    endif()
+
+    if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+        add_compile_definitions(GGML_CUDA_F16)
+    endif()
+
+    if (GGML_CUDA_NO_PEER_COPY)
+        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+    endif()
+
+    if (GGML_STATIC)
+        if (WIN32)
+            # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
+            target_link_libraries(ggml-cuda PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+        else ()
+            target_link_libraries(ggml-cuda PRIVATE  CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+        endif()
+    else()
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        # No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
+    else()
+        target_link_libraries(ggml-cuda PRIVATE CUDA::cuda_driver)
+    endif()
+
+    set(CUDA_CXX_FLAGS "")
+
+    set(CUDA_FLAGS -use_fast_math)
+
+    if (GGML_FATAL_WARNINGS)
+        list(APPEND CUDA_FLAGS -Werror all-warnings)
+    endif()
+
+    if (GGML_ALL_WARNINGS AND NOT MSVC)
+        set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
+        if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
+            list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
+        endif()
+
+        execute_process(
+            COMMAND ${NVCC_CMD} -Xcompiler --version
+            OUTPUT_VARIABLE CUDA_CCFULLVER
+            ERROR_QUIET
+        )
+
+        if (NOT CUDA_CCFULLVER MATCHES clang)
+            set(CUDA_CCID "GNU")
+            execute_process(
+                COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
+                OUTPUT_VARIABLE CUDA_CCVER
+                ERROR_QUIET
+            )
+        else()
+            if (CUDA_CCFULLVER MATCHES Apple)
+                set(CUDA_CCID "AppleClang")
+            else()
+                set(CUDA_CCID "Clang")
+            endif()
+            string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
+        endif()
+
+        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
+
+        ggml_get_flags(${CUDA_CCID} ${CUDA_CCVER})
+        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
+    endif()
+
+    if (NOT MSVC)
+        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
+    endif()
+
+    list(JOIN   CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
+
+    if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
+        list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
+    endif()
+
+    target_compile_options(ggml-cuda PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
+else()
+    message(FATAL_ERROR "CUDA Toolkit not found")
+endif()
diff --git a/ggml/src/ggml-cuda/argmax.cu b/ggml/src/ggml-cuda/argmax.cu
index aab04eca7a3..5340eedc089 100644
--- a/ggml/src/ggml-cuda/argmax.cu
+++ b/ggml/src/ggml-cuda/argmax.cu
@@ -1,57 +1,69 @@
-#include "common.cuh"
+#include <algorithm>
+#include <cstdint>
+
 #include "argmax.cuh"
+#include "common.cuh"
 #include "sum.cuh"
 
-#include <cstdint>
+static __global__ void argmax_f32(const float * __restrict__ x, int32_t * __restrict__ dst, const int64_t ncols) {
+    const int64_t row = blockIdx.x;
 
-static __global__ void argmax_f32(
-    const float * x, int32_t * dst, const int64_t ncols, const int64_t nrows) {
+    float maxval = -FLT_MAX;
+    int   argmax = -1;
+    const float * rowx = x + row * ncols;
 
-    int argmax_thread = 0;
-    const int64_t row0 = (int64_t)blockIdx.x*WARP_SIZE;
+    for (int32_t col = threadIdx.x; col < ncols; col += blockDim.x) {
+        const float val = rowx[col];
+        if (val > maxval) {
+            maxval = val;
+            argmax = col;
+        }
+    }
 
 #pragma unroll
-    for (int64_t row1 = 0; row1 < WARP_SIZE; ++row1) {
-        const int64_t row = row0 + row1;
-
-        if (row >= nrows) {
-            break;
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        const float val = __shfl_xor_sync(0xFFFFFFFF, maxval, offset, WARP_SIZE);
+        const int   col = __shfl_xor_sync(0xFFFFFFFF, argmax, offset, WARP_SIZE);
+        if (val > maxval) {
+            maxval = val;
+            argmax = col;
         }
+    }
 
-        float maxval = -FLT_MAX;
-        int   argmax = -1;
-
-        for (int32_t col = threadIdx.x; col < ncols; col += WARP_SIZE) {
-            const float val        = x[row*ncols + col];
-            const int   bigger     = val > maxval;
-            const int   not_bigger = bigger ^ 0x00000001;
-
-            maxval = maxval*not_bigger + val*bigger;
-            argmax = argmax*not_bigger + col*bigger;
+    const int n_warps = blockDim.x / WARP_SIZE;
+    const int lane_id = threadIdx.x % WARP_SIZE;
+    const int warp_id = threadIdx.x / WARP_SIZE;
+    if (n_warps > 1) {
+        constexpr int    max_warps = 1024 / WARP_SIZE;
+        __shared__ float shared_maxval[max_warps];
+        __shared__ int   shared_argmax[max_warps];
+        if (lane_id == 0) {
+            shared_maxval[warp_id] = maxval;
+            shared_argmax[warp_id] = argmax;
         }
 
+        __syncthreads();
+
+        if (warp_id == 0) {
+            if (lane_id < n_warps) {
+                maxval = shared_maxval[lane_id];
+                argmax = shared_argmax[lane_id];
+            }
 #pragma unroll
-        for (int mask = 16; mask > 0; mask >>= 1) {
-            const float val        = __shfl_xor_sync(0xFFFFFFFF, maxval, mask, WARP_SIZE);
-            const int   col        = __shfl_xor_sync(0xFFFFFFFF, argmax, mask, WARP_SIZE);
-            const int   bigger     = val > maxval;
-            const int   not_bigger = bigger ^ 0x00000001;
-
-            maxval = maxval*not_bigger + val*bigger;
-            argmax = argmax*not_bigger + col*bigger;
+            for (int offset = 16; offset > 0; offset >>= 1) {
+                const float val = __shfl_xor_sync(0xFFFFFFFF, maxval, offset, WARP_SIZE);
+                const int   col = __shfl_xor_sync(0xFFFFFFFF, argmax, offset, WARP_SIZE);
+                if (val > maxval) {
+                    maxval = val;
+                    argmax = col;
+                }
+            }
         }
-
-        const int store = row1 == threadIdx.x;
-        argmax_thread += store*argmax;
     }
 
-    const int row = row0 + threadIdx.x;
-
-    if (row >= nrows) {
-        return;
+    if (warp_id == 0 && lane_id == 0) {
+        dst[row] = argmax;
     }
-
-    dst[row] = argmax_thread;
 }
 
 void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -70,10 +82,10 @@ void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     cudaStream_t stream = ctx.stream();
 
-    const int64_t num_blocks = (nrows + WARP_SIZE - 1) / WARP_SIZE;
-
-    const dim3 blocks_dim(WARP_SIZE, 1, 1);
+    const int64_t num_blocks = nrows;
+    const int64_t num_threads = std::min<int64_t>(1024, (ne00 + WARP_SIZE - 1) / WARP_SIZE * WARP_SIZE);
+    const dim3 blocks_dim(num_threads, 1, 1);
     const dim3 blocks_num(num_blocks, 1, 1);
 
-    argmax_f32<<<blocks_num, blocks_dim, 0, stream>>>(src0_d, dst_d, ne00, nrows);
+    argmax_f32<<<blocks_num, blocks_dim, 0, stream>>>(src0_d, dst_d, ne00);
 }
diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index dd203fcded3..2c0a5622672 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -6,7 +6,7 @@
 #include <cstdint>
 #include <memory>
 
-#if defined(GGML_USE_HIPBLAS)
+#if defined(GGML_USE_HIP)
 #define GGML_COMMON_DECL_HIP
 #define GGML_COMMON_IMPL_HIP
 #else
@@ -26,13 +26,13 @@
 #include <string>
 #include <vector>
 
-#if defined(GGML_USE_HIPBLAS)
+#if defined(GGML_USE_HIP)
 #include "vendors/hip.h"
 #elif defined(GGML_USE_MUSA)
 #include "vendors/musa.h"
 #else
 #include "vendors/cuda.h"
-#endif // defined(GGML_USE_HIPBLAS)
+#endif // defined(GGML_USE_HIP)
 
 #define STRINGIZE_IMPL(...) #__VA_ARGS__
 #define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
@@ -41,17 +41,28 @@
 #define CUDART_HMAX   11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
 #define CUDART_HMASK  12000 // CUDA 12.0, min. ver. for half2 -> uint mask comparisons
 
-#define CC_PASCAL     600
-#define MIN_CC_DP4A   610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
-#define CC_VOLTA      700
-#define CC_TURING     750
-#define CC_AMPERE     800
-#define CC_OFFSET_AMD 1000000
-#define CC_RDNA1      (CC_OFFSET_AMD + 1010)
-#define CC_RDNA2      (CC_OFFSET_AMD + 1030)
-#define CC_RDNA3      (CC_OFFSET_AMD + 1100)
-#define CC_QY1        210
-#define CC_QY2        220
+#define GGML_CUDA_CC_PASCAL     600
+#define GGML_CUDA_CC_DP4A       610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
+#define GGML_CUDA_CC_VOLTA      700
+#define GGML_CUDA_CC_TURING     750
+#define GGML_CUDA_CC_AMPERE     800
+#define GGML_CUDA_CC_OFFSET_AMD 1000000
+
+// GCN/CNDA, wave size is 64
+#define GGML_CUDA_CC_GCN4       (GGML_CUDA_CC_OFFSET_AMD + 803)  // Tonga, Fiji, Polaris, minimum for fast fp16
+#define GGML_CUDA_CC_VEGA       (GGML_CUDA_CC_OFFSET_AMD + 900)  // Vega56/64, minimum for fp16 dual issue
+#define GGML_CUDA_CC_VEGA20     (GGML_CUDA_CC_OFFSET_AMD + 906)  // MI50/Radeon VII, minimum for dp4a
+#define GGML_CUDA_CC_CDNA       (GGML_CUDA_CC_OFFSET_AMD + 908)  // MI100, minimum for MFMA, acc registers
+#define GGML_CUDA_CC_CDNA2      (GGML_CUDA_CC_OFFSET_AMD + 910)  // MI210, minimum acc register renameing
+#define GGML_CUDA_CC_CDNA3      (GGML_CUDA_CC_OFFSET_AMD + 942)  // MI300
+
+// RNDA removes MFMA, dp4a, xnack, acc registers, wave size is 32
+#define GGML_CUDA_CC_RDNA1      (GGML_CUDA_CC_OFFSET_AMD + 1010) // RX 5000
+#define GGML_CUDA_CC_RDNA2      (GGML_CUDA_CC_OFFSET_AMD + 1030) // RX 6000, minimum for dp4a
+#define GGML_CUDA_CC_RDNA3      (GGML_CUDA_CC_OFFSET_AMD + 1100) // RX 7000, minimum for WMMA
+
+#define GGML_CUDA_CC_QY1        210
+#define GGML_CUDA_CC_QY2        220
 
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
@@ -97,7 +108,7 @@ void ggml_cuda_error(const char * stmt, const char * func, const char * file, in
 
 #define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str)
 
-#if !defined(GGML_USE_HIPBLAS)
+#if !defined(GGML_USE_HIP)
 static const char * cu_get_error_str(CUresult err) {
     const char * err_str;
     cuGetErrorString(err, &err_str);
@@ -120,50 +131,50 @@ typedef float dfloat; // dequantize float
 typedef float2 dfloat2;
 #endif // GGML_CUDA_F16
 
-#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
+#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 #define FP16_AVAILABLE
-#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
+#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 
 #if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 #define FAST_FP16_AVAILABLE
 #endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 #define FP16_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 #define INT8_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
 
-#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
+#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
 #define FLASH_ATTN_AVAILABLE
-#endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
+#endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= GGML_CUDA_CC_QY1)
 
 static constexpr bool fast_fp16_available(const int cc) {
-    return cc >= CC_PASCAL && cc != 610;
+    return cc >= GGML_CUDA_CC_PASCAL && cc != 610;
 }
 
 static constexpr bool fp16_mma_available(const int cc) {
-    return cc < CC_OFFSET_AMD && cc >= CC_VOLTA;
+    return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_CC_VOLTA;
 }
 
 static constexpr bool int8_mma_available(const int cc) {
-    return cc < CC_OFFSET_AMD && cc >= CC_TURING;
+    return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_CC_TURING;
 }
 
 [[noreturn]]
 static __device__ void no_device_code(
     const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
     printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
            file_name, line, function_name, arch);
     GGML_UNUSED(arch_list);
 #else
     printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
            file_name, line, function_name, arch, arch_list);
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
     __trap();
 
     GGML_UNUSED(no_device_code); // suppress unused function warning
@@ -176,30 +187,30 @@ static __device__ void no_device_code(
 #endif // __CUDA_ARCH__
 
 static __device__ __forceinline__ int warp_reduce_sum(int x) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
     return __reduce_add_sync(0xffffffff, x);
 #else
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        x += __shfl_xor_sync(0xffffffff, x, offset, 32);
     }
     return x;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 }
 
 static __device__ __forceinline__ float warp_reduce_sum(float x) {
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        x += __shfl_xor_sync(0xffffffff, x, offset, 32);
     }
     return x;
 }
 
 static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
-        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        a.x += __shfl_xor_sync(0xffffffff, a.x, offset, 32);
+        a.y += __shfl_xor_sync(0xffffffff, a.y, offset, 32);
     }
     return a;
 }
@@ -207,21 +218,21 @@ static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
 static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 #ifdef FP16_AVAILABLE
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        const half2 a_other = __shfl_xor_sync(0xffffffff, a, mask, 32);
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        const half2 a_other = __shfl_xor_sync(0xffffffff, a, offset, 32);
         reinterpret_cast<half&>(a.x) +=  __low2half(a_other);
         reinterpret_cast<half&>(a.y) += __high2half(a_other);
     }
     return a;
 #else
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, offset, 32));
     }
     return a;
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 
 #else
     NO_DEVICE_CODE;
@@ -231,8 +242,8 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+    for (int offset = 16; offset > 0; offset >>= 1) {
+        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, offset, 32));
     }
     return x;
 }
@@ -240,11 +251,11 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
 static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
 #ifdef FP16_AVAILABLE
 
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
     return __float2half(fmaxf(__half2float(a), __half2float(b)));
 #else
     return __hmax(a, b);
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
 
 #else
    NO_DEVICE_CODE;
@@ -254,7 +265,7 @@ static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b
 }
 
 static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 
 #if CUDART_VERSION >= CUDART_HMAX
     return __hmax2(a, b);
@@ -269,20 +280,20 @@ static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const hal
     GGML_UNUSED(a);
     GGML_UNUSED(b);
     NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 }
 
 static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 #pragma unroll
-   for (int mask = 16; mask > 0; mask >>= 1) {
-       x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+   for (int offset = 16; offset > 0; offset >>= 1) {
+       x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, offset, 32));
    }
    return x;
 #else
    GGML_UNUSED(x);
    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 }
 
 #if CUDART_VERSION < CUDART_HMASK
@@ -294,7 +305,7 @@ static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half
 #endif // CUDART_VERSION < CUDART_HMASK
 
 static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2)
     c = __builtin_amdgcn_sdot4(a, b, c, false);
 #elif defined(RDNA3)
@@ -320,17 +331,17 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #endif
     return c;
 
-#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 
-#if __CUDA_ARCH__ >= MIN_CC_DP4A
+#if __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
     return __dp4a(a, b, c);
-#else // __CUDA_ARCH__ >= MIN_CC_DP4A
+#else // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
     const int8_t * a8 = (const int8_t *) &a;
     const int8_t * b8 = (const int8_t *) &b;
     return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3];
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_DP4A
 
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }
 
 // TODO: move to ggml-common.h
diff --git a/ggml/src/ggml-cuda/concat.cu b/ggml/src/ggml-cuda/concat.cu
index dac10ec36b0..2f42b8a9538 100644
--- a/ggml/src/ggml-cuda/concat.cu
+++ b/ggml/src/ggml-cuda/concat.cu
@@ -94,7 +94,9 @@ static void concat_f32_cuda(const float * x, const float * y, float * dst, int n
 }
 
 // non-contiguous kernel (slow)
-static __global__ void concat_f32_non_cont(
+template <int dim>
+static __global__ void __launch_bounds__(CUDA_CONCAT_BLOCK_SIZE)
+    concat_f32_non_cont(
         const char * src0,
         const char * src1,
               char * dst,
@@ -121,22 +123,28 @@ static __global__ void concat_f32_non_cont(
           uint64_t   nb0,
           uint64_t   nb1,
           uint64_t   nb2,
-          uint64_t   nb3,
-          int32_t   dim) {
+          uint64_t   nb3){
+    static_assert(dim >= 0 && dim <= 3);
+
     const int64_t i3 = blockIdx.z;
     const int64_t i2 = blockIdx.y;
     const int64_t i1 = blockIdx.x;
 
-    int64_t o[4] = {0, 0, 0, 0};
-    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
-
     const float * x;
 
-    for (int i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
+    for (int64_t i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
         if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
             x = (const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
         } else {
-            x = (const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
+            if constexpr (dim == 0) {
+                x = (const float *) (src1 + i3 * nb13 + i2 * nb12 + i1 * nb11 + (i0 - ne00) * nb10);
+            } else if constexpr (dim == 1) {
+                x = (const float *) (src1 + i3 * nb13 + i2 * nb12 + (i1 - ne01) * nb11 + i0 * nb10);
+            } else if constexpr (dim == 2) {
+                x = (const float *) (src1 + i3 * nb13 + (i2 - ne02) * nb12 + i1 * nb11 + i0 * nb10);
+            } else if constexpr (dim == 3) {
+                x = (const float *) (src1 + (i3 - ne03) * nb13 + i2 * nb12 + i1 * nb11 + i0 * nb10);
+            }
         }
 
         float * y = (float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
@@ -182,15 +190,32 @@ void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
         }
     } else {
         dim3 grid_dim(dst->ne[1], dst->ne[2], dst->ne[3]);
-        concat_f32_non_cont<<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
-                (const char *)src0->data,
-                (const char *)src1->data,
-                (      char *)dst->data,
+        auto launch_kernel = [&](auto dim) {
+            concat_f32_non_cont<dim><<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
+                (const char *) src0->data, (const char *) src1->data, (char *) dst->data,
                 src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
                 src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
                 src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
                 src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
-                dst->ne[0],  dst->ne[1],  dst->ne[2],  dst->ne[3],
-                dst->nb[0],  dst->nb[1],  dst->nb[2],  dst->nb[3], dim);
+                dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+                dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3]);
+        };
+        switch (dim) {
+            case 0:
+                launch_kernel(std::integral_constant<int, 0>{});
+                break;
+            case 1:
+                launch_kernel(std::integral_constant<int, 1>{});
+                break;
+            case 2:
+                launch_kernel(std::integral_constant<int, 2>{});
+                break;
+            case 3:
+                launch_kernel(std::integral_constant<int, 3>{});
+                break;
+            default:
+                GGML_ABORT("Invalid dim: %d", dim);
+                break;
+        }
     }
 }
diff --git a/ggml/src/ggml-cuda/convert.cu b/ggml/src/ggml-cuda/convert.cu
index c0a4447075c..3896f956d73 100644
--- a/ggml/src/ggml-cuda/convert.cu
+++ b/ggml/src/ggml-cuda/convert.cu
@@ -26,7 +26,7 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
 
 template <bool need_check>
 static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, half * __restrict__ y, const int64_t k) {
-#if __CUDA_ARCH__ >= CC_PASCAL
+#if __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
     constexpr int nint = CUDA_Q8_0_NE_ALIGN/sizeof(int) + WARP_SIZE;
 
     const int64_t   i0 = CUDA_Q8_0_NE_ALIGN*blockIdx.x;
@@ -64,7 +64,7 @@ static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, h
     GGML_UNUSED(y);
     GGML_UNUSED(k);
     NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= CC_PASCAL
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 }
 
 template<typename dst_t>
@@ -599,7 +599,7 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
         case GGML_TYPE_Q5_1:
             return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
         case GGML_TYPE_Q8_0:
-            if (ggml_cuda_info().devices[ggml_cuda_get_device()].cc >= CC_PASCAL) {
+            if (ggml_cuda_info().devices[ggml_cuda_get_device()].cc >= GGML_CUDA_CC_PASCAL) {
                 return dequantize_block_q8_0_f16_cuda;
             }
             return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
diff --git a/ggml/src/ggml-cuda/dmmv.cu b/ggml/src/ggml-cuda/dmmv.cu
deleted file mode 100644
index 00e21b5d77e..00000000000
--- a/ggml/src/ggml-cuda/dmmv.cu
+++ /dev/null
@@ -1,699 +0,0 @@
-#include "dmmv.cuh"
-#include "dequantize.cuh"
-#include "convert.cuh"
-
-#ifndef K_QUANTS_PER_ITERATION
-#define K_QUANTS_PER_ITERATION 2
-#else
-static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
-#endif
-
-static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q2_K * x = (const block_q2_K *)vx + ib0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...15
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
-
-    const int step = 16/K_QUANTS_PER_ITERATION;
-
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0...15 or 0...7
-
-    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15 or 0...14 in steps of 2
-    const int q_offset = 32*im + l0;
-    const int s_offset = 8*im;
-    const int y_offset = 128*im + l0;
-
-    uint32_t aux[4];
-    const uint8_t * d = (const uint8_t *)aux;
-    const uint8_t * m = (const uint8_t *)(aux + 2);
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + y_offset;
-        const uint8_t * q = x[i].qs + q_offset;
-
-        const float dall = __low2half(x[i].dm);
-        const float dmin = __high2half(x[i].dm);
-
-        const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset);
-        aux[0] = a[0] & 0x0f0f0f0f;
-        aux[1] = a[1] & 0x0f0f0f0f;
-        aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
-        aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
-
-        float sum1 = 0, sum2 = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
-                  + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
-                  + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
-                  + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
-                  + y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
-                  + y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
-                  + y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
-                  +y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
-            sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
-                  + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
-
-        }
-        tmp += dall * sum1 - dmin * sum2;
-
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q3_K * x = (const block_q3_K *)vx + ib0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    const uint16_t kmask1 = 0x0303;
-    const uint16_t kmask2 = 0x0f0f;
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
-
-    const int n  = K_QUANTS_PER_ITERATION;               // iterations in the inner loop
-    const int step = 16/K_QUANTS_PER_ITERATION;
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0....15 or 0...7
-
-    const uint8_t m = 1 << (4*im);
-
-    const int l0 = n*in;                                 // 0...15 or 0...14 in steps of 2
-    const int q_offset =  32*im + l0;
-    const int y_offset = 128*im + l0;
-
-    uint16_t utmp[4];
-    const int8_t * s = (const int8_t *)utmp;
-
-    const uint16_t s_shift = 4*im;
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y  = yy + i * QK_K + y_offset;
-        const uint8_t * q = x[i].qs + q_offset;
-        const uint8_t * h = x[i].hmask + l0;
-
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
-        utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
-        utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
-        utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
-
-        const float d = x[i].d;
-
-        float sum = 0;
-        for (int l = 0; l < n; ++l) {
-            sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
-                 + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
-                 + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
-                 + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
-            sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
-                 + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
-                 + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
-                + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
-        }
-        tmp += d * sum;
-
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q4_K * x = (const block_q4_K *)vx + ib0;
-
-    const uint16_t kmask1 = 0x3f3f;
-    const uint16_t kmask2 = 0x0f0f;
-    const uint16_t kmask3 = 0xc0c0;
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
-
-    const int step = 8/K_QUANTS_PER_ITERATION;           // 8 or 4
-
-    const int il  = tid/step;                            // 0...3
-    const int ir  = tid - step*il;                       // 0...7 or 0...3
-    const int n   = 2 * K_QUANTS_PER_ITERATION;          // 2 or 4
-
-    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const int in = il%2;
-
-    const int l0 = n*(2*ir + in);
-    const int q_offset = 32*im + l0;
-    const int y_offset = 64*im + l0;
-
-    uint16_t aux[4];
-    const uint8_t * sc = (const uint8_t *)aux;
-
-#if K_QUANTS_PER_ITERATION == 2
-    uint32_t q32[4];
-    const uint8_t * q4 = (const uint8_t *)q32;
-#else
-    uint16_t q16[4];
-    const uint8_t * q4 = (const uint8_t *)q16;
-#endif
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y1 = yy + i*QK_K + y_offset;
-        const float   * y2 = y1 + 128;
-
-        const float dall = __low2half(x[i].dm);
-        const float dmin = __high2half(x[i].dm);
-
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux[0] = a[im+0] & kmask1;
-        aux[1] = a[im+2] & kmask1;
-        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
-        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
-
-#if K_QUANTS_PER_ITERATION == 2
-        const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset);
-        const uint32_t * q2 = q1 + 16;
-
-        q32[0] = q1[0] & 0x0f0f0f0f;
-        q32[1] = q1[0] & 0xf0f0f0f0;
-        q32[2] = q2[0] & 0x0f0f0f0f;
-        q32[3] = q2[0] & 0xf0f0f0f0;
-
-        float4 s = {0.f, 0.f, 0.f, 0.f};
-        float smin = 0;
-        for (int l = 0; l < 4; ++l) {
-            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+ 4];
-            s.z += y2[l] * q4[l+8]; s.w += y2[l+32] * q4[l+12];
-            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
-        }
-        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
-#else
-        const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset);
-        const uint16_t * q2 = q1 + 32;
-
-        q16[0] = q1[0] & 0x0f0f;
-        q16[1] = q1[0] & 0xf0f0;
-        q16[2] = q2[0] & 0x0f0f;
-        q16[3] = q2[0] & 0xf0f0;
-
-        float4 s = {0.f, 0.f, 0.f, 0.f};
-        float smin = 0;
-        for (int l = 0; l < 2; ++l) {
-            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2];
-            s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6];
-            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
-        }
-        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
-#endif
-
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (tid == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols) {
-
-    const int row = blockIdx.x;
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q5_K * x = (const block_q5_K *)vx + ib0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    const uint16_t kmask1 = 0x3f3f;
-    const uint16_t kmask2 = 0x0f0f;
-    const uint16_t kmask3 = 0xc0c0;
-
-    const int tid = threadIdx.x/2;  // 0...15
-    const int ix  = threadIdx.x%2;
-
-    const int il  = tid/4;     // 0...3
-    const int ir  = tid - 4*il;// 0...3
-    const int n   = 2;
-
-    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const int in = il%2;
-
-    const int l0 = n*(2*ir + in);
-    const int q_offset = 32*im + l0;
-    const int y_offset = 64*im + l0;
-
-    const uint8_t hm1  = 1 << (2*im);
-    const uint8_t hm2  = hm1 << 4;
-
-    uint16_t aux[4];
-    const uint8_t * sc = (const uint8_t *)aux;
-
-    uint16_t q16[8];
-    const uint8_t * q4 = (const uint8_t *)q16;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2) {
-
-        const uint8_t * ql1 = x[i].qs + q_offset;
-        const uint8_t * qh  = x[i].qh + l0;
-        const float   * y1  = yy + i*QK_K + y_offset;
-        const float   * y2  = y1 + 128;
-
-        const float dall = __low2half(x[i].dm);
-        const float dmin = __high2half(x[i].dm);
-
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux[0] = a[im+0] & kmask1;
-        aux[1] = a[im+2] & kmask1;
-        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
-        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
-
-        float4 sum = {0.f, 0.f, 0.f, 0.f};
-        float smin = 0;
-        const uint16_t * q1 = (const uint16_t *)ql1;
-        const uint16_t * q2 = q1 + 32;
-        q16[0] = q1[0] & 0x0f0f;
-        q16[1] = q1[8] & 0x0f0f;
-        q16[2] = (q1[0] >> 4) & 0x0f0f;
-        q16[3] = (q1[8] >> 4) & 0x0f0f;
-        q16[4] = q2[0] & 0x0f0f;
-        q16[5] = q2[8] & 0x0f0f;
-        q16[6] = (q2[0] >> 4) & 0x0f0f;
-        q16[7] = (q2[8] >> 4) & 0x0f0f;
-        for (int l = 0; l < n; ++l) {
-            sum.x += y1[l+ 0] * (q4[l +0] + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
-                   + y1[l+16] * (q4[l +2] + (qh[l+16] & (hm1 << 0) ? 16 : 0));
-            sum.y += y1[l+32] * (q4[l +4] + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
-                   + y1[l+48] * (q4[l +6] + (qh[l+16] & (hm1 << 1) ? 16 : 0));
-            sum.z += y2[l+ 0] * (q4[l +8] + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
-                   + y2[l+16] * (q4[l+10] + (qh[l+16] & (hm2 << 0) ? 16 : 0));
-            sum.w += y2[l+32] * (q4[l+12] + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
-                   + y2[l+48] * (q4[l+14] + (qh[l+16] & (hm2 << 1) ? 16 : 0));
-            smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
-                  + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
-        }
-        tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q6_K * x = (const block_q6_K *)vx + ib0;
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
-
-    const int step = 16/K_QUANTS_PER_ITERATION;          // 16 or 8
-
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0...15 or 0...7
-
-#if K_QUANTS_PER_ITERATION == 1
-    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15
-    const int is = 0;
-#else
-    const int l0 = 4 * in;                               // 0, 4, 8, ..., 28
-    const int is = in / 4;
-#endif
-    const int ql_offset = 64*im + l0;
-    const int qh_offset = 32*im + l0;
-    const int s_offset  =  8*im + is;
-    const int y_offset = 128*im + l0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y  = yy + i * QK_K + y_offset;
-        const uint8_t * ql = x[i].ql + ql_offset;
-        const uint8_t * qh = x[i].qh + qh_offset;
-        const int8_t  * s  = x[i].scales + s_offset;
-
-        const float d = x[i].d;
-
-#if K_QUANTS_PER_ITERATION == 1
-        float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
-                  + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
-                  + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
-                  + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
-                  + y[64] * s[4] * d * ((int8_t)((ql[ 0]  >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
-                  + y[80] * s[5] * d * ((int8_t)((ql[16]  >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
-                  + y[96] * s[6] * d * ((int8_t)((ql[32]  >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
-                  +y[112] * s[7] * d * ((int8_t)((ql[48]  >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
-        tmp += sum;
-#else
-        float sum = 0;
-        for (int l = 0; l < 4; ++l) {
-            sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
-                 + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
-                 + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
-                 + y[l+96] * s[6] * d * ((int8_t)((ql[l+32]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
-        }
-        tmp += sum;
-#endif
-
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (tid == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __device__ void convert_f16(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
-    const half * x = (const half *) vx;
-    // load 2 halfs into register in a single instruction
-    const half2 x_reg = *((half2 *) &(x[ib + iqs]));
-    // automatic half -> float type cast if dfloat == float
-    v.x = __low2float(x_reg);
-    v.y = __high2float(x_reg);
-}
-
-static constexpr __device__ dequantize_kernel_t get_dequantize_kernel(ggml_type type) {
-    return type == GGML_TYPE_Q4_0 ? dequantize_q4_0 :
-        type == GGML_TYPE_Q4_1 ? dequantize_q4_1 :
-        type == GGML_TYPE_Q5_0 ? dequantize_q5_0 :
-        type == GGML_TYPE_Q5_1 ? dequantize_q5_1 :
-        type == GGML_TYPE_Q8_0 ? dequantize_q8_0 :
-        type == GGML_TYPE_F16 ? convert_f16 :
-        nullptr;
-}
-
-template <ggml_type type>
-static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) {
-    constexpr int qk = ggml_cuda_type_traits<type>::qk; // quantized weights per x block
-    constexpr int qr = ggml_cuda_type_traits<type>::qr; // number of quantized weights per data value in x block
-    constexpr dequantize_kernel_t dequantize_kernel = get_dequantize_kernel(type);
-
-    const int64_t row = (int64_t)blockIdx.x*blockDim.y + threadIdx.y;
-
-    if (row >= nrows) {
-        return;
-    }
-
-    const int tid = threadIdx.x;
-
-    const int iter_stride = 2*GGML_CUDA_DMMV_X;
-    const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter
-    const int y_offset = qr == 1 ? 1 : qk/2;
-
-// partial sum for each thread
-#ifdef GGML_CUDA_F16
-    half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics
-#else
-    float tmp = 0.0f;
-#endif // GGML_CUDA_F16
-
-    for (int i = 0; i < ncols; i += iter_stride) {
-        const int col = i + vals_per_iter*tid;
-        const int64_t ib = ((int64_t)row*ncols + col)/qk; // x block index
-        const int iqs = (col%qk)/qr; // x quant index
-        const int iybs = col - col%qk; // y block start index
-
-// processing >2 values per i iter is faster for fast GPUs
-#pragma unroll
-        for (int j = 0; j < vals_per_iter; j += 2) {
-            // process 2 vals per j iter
-
-            // dequantize
-            // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
-            dfloat2 v;
-            dequantize_kernel(vx, ib, iqs + j/qr, v);
-
-            // matrix multiplication
-            // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
-#ifdef GGML_CUDA_F16
-            if ( y_offset == 1 ) {
-                // load 2 dfloats into register in a single instruction
-                const dfloat2 y_reg = *((dfloat2 *) &(y[iybs + iqs + j/qr]));
-                tmp += __hmul2(v, y_reg);
-            }
-            else {
-                tmp += __hmul2(v, {
-                        y[iybs + iqs + j/qr + 0],
-                        y[iybs + iqs + j/qr + y_offset]
-                    });
-            }
-#else
-            if ( y_offset == 1 ) {
-                // load 2 dfloats into register in a single instruction
-                const dfloat2 y_reg = *((dfloat2 *) &(y[iybs + iqs + j/qr]));
-                tmp += v.x * y_reg.x;
-                tmp += v.y * y_reg.y;
-            }
-            else {
-                tmp += v.x * y[iybs + iqs + j/qr + 0];
-                tmp += v.y * y[iybs + iqs + j/qr + y_offset];
-            }
-#endif // GGML_CUDA_F16
-        }
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (tid == 0) {
-#ifdef GGML_CUDA_F16
-        dst[row] = tmp.x + tmp.y;
-#else
-        dst[row] = tmp;
-#endif // GGML_CUDA_F16
-    }
-}
-
-static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q4_0>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q4_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q5_0>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q5_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_Q8_0>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q2_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2 / K_QUANTS_PER_ITERATION;
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q3_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2 / K_QUANTS_PER_ITERATION;
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q4_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q5_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const dim3 block_dims(32, 1, 1);
-    dequantize_mul_mat_vec_q5_k<<<nrows, block_dims, 0, stream>>>(vx, y, dst, ncols);
-}
-
-static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2 / K_QUANTS_PER_ITERATION;
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q6_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<GGML_TYPE_F16>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-void ggml_cuda_op_dequantize_mul_mat_vec(
-    ggml_backend_cuda_context & ctx,
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
-    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
-    const int64_t src1_padded_row_size, cudaStream_t stream) {
-    GGML_UNUSED(ctx);
-    const int64_t ne00 = src0->ne[0];
-    const int64_t row_diff = row_high - row_low;
-
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
-#ifdef GGML_CUDA_F16
-    ggml_cuda_pool_alloc<half> src1_dfloat_a(ctx.pool());
-    half * src1_dfloat = nullptr; // dfloat == half
-
-    bool src1_convert_f16 =
-        src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
-        src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
-        src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
-
-    if (src1_convert_f16) {
-        src1_dfloat = src1_dfloat_a.alloc(ne00);
-        const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
-        GGML_ASSERT(to_fp16_cuda != nullptr);
-        to_fp16_cuda(src1_ddf_i, src1_dfloat, ne00, stream);
-    }
-#else
-    const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion
-#endif // GGML_CUDA_F16
-
-    switch (src0->type) {
-        case GGML_TYPE_Q4_0:
-            dequantize_mul_mat_vec_q4_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            dequantize_mul_mat_vec_q4_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            dequantize_mul_mat_vec_q5_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            dequantize_mul_mat_vec_q5_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            dequantize_mul_mat_vec_q8_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q2_K:
-            dequantize_mul_mat_vec_q2_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q3_K:
-            dequantize_mul_mat_vec_q3_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q4_K:
-            dequantize_mul_mat_vec_q4_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_K:
-            dequantize_mul_mat_vec_q5_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q6_K:
-            dequantize_mul_mat_vec_q6_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_F16:
-            convert_mul_mat_vec_f16_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        default:
-            GGML_ABORT("fatal error");
-            break;
-    }
-
-    GGML_UNUSED(src1);
-    GGML_UNUSED(dst);
-    GGML_UNUSED(src1_ddq_i);
-    GGML_UNUSED(src1_ncols);
-    GGML_UNUSED(src1_padded_row_size);
-}
-
-bool ggml_cuda_dmmv_type_supported(ggml_type src0_type) {
-    return src0_type == GGML_TYPE_Q4_0 || src0_type == GGML_TYPE_Q4_1 ||
-        src0_type == GGML_TYPE_Q5_0 || src0_type == GGML_TYPE_Q5_1 ||
-        src0_type == GGML_TYPE_Q8_0 || src0_type == GGML_TYPE_Q2_K ||
-        src0_type == GGML_TYPE_Q3_K || src0_type == GGML_TYPE_Q4_K ||
-        src0_type == GGML_TYPE_Q5_K || src0_type == GGML_TYPE_Q6_K ||
-        src0_type == GGML_TYPE_F16;
-}
diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh
index 1fb5c09c3b1..ee9752da6a8 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -517,9 +517,9 @@ constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
 }
 
 template<int D, int parallel_blocks> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_combine_results(
         const float  * __restrict__ VKQ_parts,
         const float2 * __restrict__ VKQ_meta,
diff --git a/ggml/src/ggml-cuda/fattn-tile-f16.cu b/ggml/src/ggml-cuda/fattn-tile-f16.cu
index 5af02c7ecbe..4d314dacb19 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f16.cu
@@ -5,9 +5,9 @@
 #define FATTN_KQ_STRIDE_TILE_F16 64
 
 template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,
diff --git a/ggml/src/ggml-cuda/fattn-tile-f32.cu b/ggml/src/ggml-cuda/fattn-tile-f32.cu
index f402195ce0b..bb33604470f 100644
--- a/ggml/src/ggml-cuda/fattn-tile-f32.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f32.cu
@@ -5,9 +5,9 @@
 #define FATTN_KQ_STRIDE_TILE_F32 32
 
 template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_tile_ext_f32(
         const char * __restrict__ Q,
         const char * __restrict__ K,
diff --git a/ggml/src/ggml-cuda/fattn-vec-f16.cuh b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
index 2ed6509acb8..34a2992c769 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
@@ -2,9 +2,9 @@
 #include "fattn-common.cuh"
 
 template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_vec_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,
@@ -220,7 +220,6 @@ static __global__ void flash_attn_vec_ext_f16(
         for (int j = 0; j < ncols; ++j) {
             half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];
 
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
             if (threadIdx.x == 0) {
                 kqmax_shared[j][threadIdx.y] = kqmax_new_j;
             }
diff --git a/ggml/src/ggml-cuda/fattn-vec-f32.cuh b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
index bf512590253..a28fc8b7fc8 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f32.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f32.cuh
@@ -2,9 +2,9 @@
 #include "fattn-common.cuh"
 
 template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(D, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_vec_ext_f32(
         const char * __restrict__ Q,
         const char * __restrict__ K,
@@ -206,7 +206,6 @@ static __global__ void flash_attn_vec_ext_f32(
         for (int j = 0; j < ncols; ++j) {
             float kqmax_new_j = kqmax_new_arr[j];
 
-            kqmax_new_j = warp_reduce_max(kqmax_new_j);
             if (threadIdx.x == 0) {
                 kqmax_shared[j][threadIdx.y] = kqmax_new_j;
             }
diff --git a/ggml/src/ggml-cuda/fattn-wmma-f16.cuh b/ggml/src/ggml-cuda/fattn-wmma-f16.cuh
index b10d19d9327..860d0e6dc2f 100644
--- a/ggml/src/ggml-cuda/fattn-wmma-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-wmma-f16.cuh
@@ -7,9 +7,9 @@
 
 // D == head size, VKQ_stride == num VKQ rows calculated in parallel:
 template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t, bool use_logit_softcap>
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 __launch_bounds__(nwarps*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void flash_attn_ext_f16(
         const char * __restrict__ Q,
         const char * __restrict__ K,
diff --git a/ggml/src/ggml-cuda/fattn.cu b/ggml/src/ggml-cuda/fattn.cu
index 0e7ebbc5393..0b26b0f8e05 100644
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -304,7 +304,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
     const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);
 
     // On AMD the tile kernels perform poorly, use the vec kernel instead:
-    if (cc >= CC_OFFSET_AMD) {
+    if (cc >= GGML_CUDA_CC_OFFSET_AMD) {
         if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
             ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
         } else {
diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
similarity index 93%
rename from ggml/src/ggml-cuda.cu
rename to ggml/src/ggml-cuda/ggml-cuda.cu
index 357cee660cd..c180adc84b8 100644
--- a/ggml/src/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -16,11 +16,11 @@
 #include "ggml-cuda/cpy.cuh"
 #include "ggml-cuda/cross-entropy-loss.cuh"
 #include "ggml-cuda/diagmask.cuh"
-#include "ggml-cuda/dmmv.cuh"
 #include "ggml-cuda/fattn.cuh"
 #include "ggml-cuda/getrows.cuh"
 #include "ggml-cuda/im2col.cuh"
 #include "ggml-cuda/mmq.cuh"
+#include "ggml-cuda/mmv.cuh"
 #include "ggml-cuda/mmvq.cuh"
 #include "ggml-cuda/norm.cuh"
 #include "ggml-cuda/opt-step-adamw.cuh"
@@ -91,7 +91,7 @@ int ggml_cuda_get_device() {
 
 static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
     ggml_cuda_set_device(device);
-#if defined(GGML_USE_HIPBLAS) && defined(GGML_HIP_UMA)
+#if defined(GGML_USE_HIP) && defined(GGML_HIP_UMA)
     auto res = hipMallocManaged(ptr, size);
     if (res == hipSuccess) {
         // if error we "need" to know why...
@@ -100,7 +100,7 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
     return res;
 #else
 
-#if !defined(GGML_USE_HIPBLAS)
+#if !defined(GGML_USE_HIP)
     cudaError_t err;
     if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
     {
@@ -113,7 +113,7 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
     return err;
 #else
     return cudaMalloc(ptr, size);
-#endif // !defined(GGML_USE_HIPBLAS)
+#endif // !defined(GGML_USE_HIP)
 
 #endif
 }
@@ -151,7 +151,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
     for (int id = 0; id < info.device_count; ++id) {
         int device_vmm = 0;
 
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
         CUdevice device;
         CU_CHECK(cuDeviceGet(&device, id));
         CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));
@@ -163,7 +163,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
             alloc_prop.location.id = id;
             CU_CHECK(cuMemGetAllocationGranularity(&info.devices[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED));
         }
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
         info.devices[id].vmm = !!device_vmm;
 
         cudaDeviceProp prop;
@@ -175,13 +175,13 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
         info.devices[id].nsm   = prop.multiProcessorCount;
         info.devices[id].smpb  = prop.sharedMemPerBlock;
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
         info.devices[id].smpbo = prop.sharedMemPerBlock;
-        info.devices[id].cc = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD;
+        info.devices[id].cc = 100*prop.major + 10*prop.minor + GGML_CUDA_CC_OFFSET_AMD;
 #else
         info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
         info.devices[id].cc = 100*prop.major + 10*prop.minor;
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
     }
 
     for (int id = 0; id < info.device_count; ++id) {
@@ -299,7 +299,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
 };
 
 // pool with virtual memory
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
 struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
     static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
 
@@ -393,14 +393,14 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
         GGML_ASSERT(ptr == (void *) (pool_addr + pool_used));
     }
 };
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
 
 std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#if !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
     if (ggml_cuda_info().devices[device].vmm) {
         return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
     }
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)
     return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
 }
 
@@ -1020,114 +1020,6 @@ typedef void (*ggml_cuda_op_mul_mat_t)(
 
 #define MUL_MAT_SRC1_COL_STRIDE 128
 
-static __global__ void mul_mat_p021_f16_f32(
-    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y) {
-
-    const half * x = (const half *) vx;
-
-    const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
-    const int channel = blockDim.z*blockIdx.z + threadIdx.z;
-    const int channel_x = channel / (nchannels_y / nchannels_x);
-
-    const int nrows_y = ncols_x;
-    const int nrows_dst = nrows_x;
-    const int row_dst = row_x;
-
-    float tmp = 0.0f;
-
-    for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
-        const int col_x = col_x0 + threadIdx.x;
-
-        if (col_x >= ncols_x) {
-            break;
-        }
-
-        // x is transposed and permuted
-        const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x;
-        const float xi = __half2float(x[ix]);
-
-        const int row_y = col_x;
-
-        // y is not transposed but permuted
-        const int iy = channel*nrows_y + row_y;
-
-        tmp += xi * y[iy];
-    }
-
-    // dst is not transposed and not permuted
-    const int idst = channel*nrows_dst + row_dst;
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[idst] = tmp;
-    }
-}
-
-static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
-    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
-    const int row_stride_x, const int channel_stride_x, const int channel_x_divisor) {
-
-    const half * x = (const half *) vx;
-
-    const int row_x     = blockDim.y*blockIdx.y + threadIdx.y;
-    const int channel   = blockDim.z*blockIdx.z + threadIdx.z;
-    const int channel_x = channel / channel_x_divisor;
-
-    const int nrows_y   = ncols_x;
-    const int nrows_dst = nrows_x;
-    const int row_dst   = row_x;
-
-    const int idst = channel*nrows_dst + row_dst;
-
-    float tmp = 0.0f;
-
-    for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
-        const int col_x = col_x0 + threadIdx.x;
-
-        if (col_x >= ncols_x) {
-            break;
-        }
-
-        const int row_y = col_x;
-
-        const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
-        const int iy = channel*nrows_y + row_y;
-
-        const float xi = __half2float(x[ix]);
-
-        tmp += xi * y[iy];
-    }
-
-    // sum up partial sums and write back result
-    tmp = warp_reduce_sum(tmp);
-
-    if (threadIdx.x == 0) {
-        dst[idst] = tmp;
-    }
-}
-
-static void ggml_mul_mat_p021_f16_f32_cuda(
-    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x,
-    const int nchannels_x, const int nchannels_y, cudaStream_t stream) {
-
-    const dim3 block_nums(1, nrows_x, nchannels_y);
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    mul_mat_p021_f16_f32<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols_x, nrows_x, nchannels_x, nchannels_y);
-}
-
-static void ggml_mul_mat_vec_nc_f16_f32_cuda(
-    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int row_stride_x,
-    const int nchannels_x, const int nchannels_y, const int channel_stride_x, cudaStream_t stream) {
-
-    const dim3 block_nums(1, nrows_x, nchannels_y);
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    mul_mat_vec_nc_f16_f32<<<block_nums, block_dims, 0, stream>>>
-        (vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y/nchannels_x);
-}
-
 static cudaError_t ggml_cuda_cpy_tensor_2d(
     void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
 
@@ -1189,7 +1081,7 @@ static void ggml_cuda_op_mul_mat_cublas(
 
     const int compute_capability = ggml_cuda_info().devices[id].cc;
 
-    if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) {
+    if (compute_capability >= GGML_CUDA_CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) {
         // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
         ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
         if (src0->type != GGML_TYPE_F16) {
@@ -1215,6 +1107,11 @@ static void ggml_cuda_op_mul_mat_cublas(
         const half alpha_f16 = 1.0f;
         const half beta_f16 = 0.0f;
 
+        cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
+        if (ggml_cuda_info().devices[ctx.device].cc == GGML_CUDA_CC_CDNA) {
+            cu_compute_type = CUBLAS_COMPUTE_32F;
+        }
+
         CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
         CUBLAS_CHECK(
             cublasGemmEx(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N,
@@ -1222,7 +1119,7 @@ static void ggml_cuda_op_mul_mat_cublas(
                     &alpha_f16, src0_ptr,       CUDA_R_16F, ne00,
                                 src1_ptr,       CUDA_R_16F, ne10,
                     &beta_f16,   dst_f16.get(), CUDA_R_16F, ldc,
-                    CUBLAS_COMPUTE_16F,
+                    cu_compute_type,
                     CUBLAS_GEMM_DEFAULT_TENSOR_OP));
 
         const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
@@ -1325,7 +1222,7 @@ static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
 static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
     void * dst, int dstDevice, size_t dpitch, void * src, int srcDevice, size_t spitch, size_t width, size_t height, cudaStream_t stream) {
 
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
     // cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
     cudaMemcpy3DPeerParms p = {};
     p.dstDevice = dstDevice;
@@ -1339,7 +1236,7 @@ static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
     GGML_UNUSED(dstDevice);
     GGML_UNUSED(srcDevice);
     return cudaMemcpy2DAsync(dst, dpitch, src, spitch, width, height, cudaMemcpyDeviceToDevice, stream);
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
 }
 
 static void ggml_cuda_op_mul_mat(
@@ -1654,58 +1551,6 @@ static void ggml_cuda_op_mul_mat(
     }
 }
 
-static void ggml_cuda_mul_mat_vec_p021(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
-    GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
-    GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-
-    const int64_t ne12 = src1->ne[2];
-
-    cudaStream_t main_stream = ctx.stream();
-
-    void  * src0_ddq = src0->data;
-    float * src1_ddf = (float *) src1->data;
-    float * dst_ddf  = (float *) dst->data;
-
-    ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream);
-}
-
-static void ggml_cuda_mul_mat_vec_nc(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(!ggml_is_transposed(src0));
-    GGML_ASSERT(!ggml_is_transposed(src1));
-    GGML_ASSERT(!ggml_is_permuted(src0));
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-
-    const int64_t nb01 = src0->nb[1];
-    const int64_t nb02 = src0->nb[2];
-
-    const int64_t ne12 = src1->ne[2];
-
-    cudaStream_t main_stream = ctx.stream();
-
-    void  * src0_ddq = src0->data;
-    float * src1_ddf = (float *) src1->data;
-    float * dst_ddf  = (float *) dst->data;
-
-    const int64_t row_stride_x = nb01 / sizeof(half);
-    const int64_t channel_stride_x = nb02 / sizeof(half);
-
-    ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
-}
-
 static __global__ void k_compute_batched_ptrs(
         const half * src0_as_f16, const half * src1_as_f16, char * dst,
         const void ** ptrs_src, void ** ptrs_dst,
@@ -1767,6 +1612,10 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
     cudaDataType_t      cu_data_type    = CUDA_R_16F;
 
+    if (ggml_cuda_info().devices[ctx.device].cc == GGML_CUDA_CC_CDNA) {
+        cu_compute_type = CUBLAS_COMPUTE_32F;
+    }
+
     // dst strides
     size_t nbd2 = dst->nb[2];
     size_t nbd3 = dst->nb[3];
@@ -1879,21 +1728,17 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
 
-    bool use_dequantize_mul_mat_vec = ggml_cuda_dmmv_type_supported(src0->type)
+    bool use_mul_mat_vec   = src0->type == GGML_TYPE_F16
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
-        && src0->ne[0] % (GGML_CUDA_DMMV_X*2) == 0 && src1->ne[1] == 1;
-    bool          use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
+        && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
+    bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
-    bool              use_mul_mat_q =  ggml_is_quantized(src0->type)
+    bool use_mul_mat_q     = ggml_is_quantized(src0->type)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
 
-    // if mmvq is available it's a better choice than dmmv:
-#ifndef GGML_CUDA_FORCE_DMMV
-    use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
-#endif // GGML_CUDA_FORCE_DMMV
-
-    bool any_gpus_with_slow_fp16 = false;
+    bool any_gpus_with_slow_fp16   = false;
+    bool any_gpus_without_fp16_mma = false;
 
     if (split) {
         ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
@@ -1904,14 +1749,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
                 continue;
             }
 
-            const int cc            = ggml_cuda_info().devices[id].cc;
-            use_mul_mat_q           = use_mul_mat_q           && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-            any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
+            const int cc              = ggml_cuda_info().devices[id].cc;
+            use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+            any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_available(cc);
+            any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_available(cc);
         }
     } else {
-        const int cc            = ggml_cuda_info().devices[ctx.device].cc;
-        use_mul_mat_q           = use_mul_mat_q           && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-        any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
+        const int cc              = ggml_cuda_info().devices[ctx.device].cc;
+        use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+        any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_available(cc);
+        any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_available(cc);
     }
 
     // debug helpers
@@ -1922,18 +1769,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
-    if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
-        // FP32 precision KQ single-batch for batch size 1 without FlashAttention
-        ggml_cuda_mul_mat_vec_p021(ctx, src0, src1, dst);
-    } else if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
-        // FP32 precision KQV single-batch for batch size 1 without FlashAttention
-        ggml_cuda_mul_mat_vec_nc(ctx, src0, src1, dst);
+    if (!split && use_mul_mat_vec && dst->ne[3] == 1 && (src0->ne[1] < MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
+        // the custom F16 vector kernel can be used over batched cuBLAS GEMM
+        // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
+        ggml_cuda_mul_mat_vec(ctx, src0, src1, dst);
     } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
                && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
-        // KQ + KQV multi-batch without FlashAttention
+        // general KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
-    } else if (use_dequantize_mul_mat_vec) {
-        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, nullptr);
+    } else if (use_mul_mat_vec) {
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec, nullptr);
     } else if (use_mul_mat_vec_q) {
         ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, quantize_row_q8_1_cuda);
     } else if (use_mul_mat_q) {
@@ -2512,7 +2357,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
     std::vector<void *> ggml_cuda_cpy_fn_ptrs;
 
     if (cuda_ctx->cuda_graph->graph == nullptr) {
-        if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) {
+        if (ggml_cuda_info().devices[cuda_ctx->device].cc < GGML_CUDA_CC_AMPERE) {
             cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
 #ifndef NDEBUG
             GGML_LOG_DEBUG("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
@@ -2978,6 +2823,17 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 struct ggml_tensor * a = op->src[0];
                 struct ggml_tensor * b = op->src[1];
+                // for small weight matrices the active device can end up without any rows, don't use row split in those cases
+                // this avoids some edge cases (and the performance would not be good anyways)
+                if (a->buffer && ggml_backend_buft_is_cuda_split(a->buffer->buft)) {
+                    ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) a->buffer->buft->context;
+                    int64_t row_low;
+                    int64_t row_high;
+                    get_row_split(&row_low, &row_high, a, buft_ctx->tensor_split, dev_ctx->device);
+                    if (row_low == row_high) {
+                        return false;
+                    }
+                }
                 if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) {
                     return false;
                 }
@@ -3172,7 +3028,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 return true;
             }
             const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
-            return cc >= CC_VOLTA && cc < CC_OFFSET_AMD && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
+            return cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
         }
         case GGML_OP_CROSS_ENTROPY_LOSS:
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
@@ -3279,6 +3135,61 @@ static ggml_backend_dev_t ggml_backend_cuda_reg_get_device(ggml_backend_reg_t re
     return ctx->devices[index];
 }
 
+static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t reg) {
+    static std::vector<ggml_backend_feature> features = []() {
+        std::vector<ggml_backend_feature> features;
+    #define _STRINGIFY(...) #__VA_ARGS__
+    #define STRINGIFY(...) _STRINGIFY(__VA_ARGS__)
+
+    #ifdef __CUDA_ARCH_LIST__
+        features.push_back({ "ARCHS", STRINGIFY(__CUDA_ARCH_LIST__) });
+    #endif
+
+    #ifdef GGML_CUDA_FORCE_MMQ
+        features.push_back({ "FORCE_MMQ", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_FORCE_CUBLAS
+        features.push_back({ "FORCE_CUBLAS", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_NO_VMM
+        features.push_back({ "NO_VMM", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_NO_PEER_COPY
+        features.push_back({ "NO_PEER_COPY", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_F16
+        features.push_back({ "F16", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_USE_GRAPHS
+        features.push_back({ "USE_GRAPHS", "1" });
+    #endif
+
+    #ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
+        features.push_back({ "PEER_MAX_BATCH_SIZE", STRINGIFY(GGML_CUDA_PEER_MAX_BATCH_SIZE) });
+    #endif
+
+    #ifdef GGML_CUDA_FA_ALL_QUANTS
+        features.push_back({ "FA_ALL_QUANTS", "1" });
+    #endif
+
+    #undef _STRINGIFY
+    #undef STRINGIFY
+
+        features.push_back({ nullptr, nullptr });
+
+        return features;
+    }();
+
+    return features.data();
+
+    GGML_UNUSED(reg);
+}
+
 static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
     GGML_UNUSED(reg);
     if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
@@ -3290,13 +3201,16 @@ static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, con
     if (strcmp(name, "ggml_backend_unregister_host_buffer") == 0) {
         return (void *)ggml_backend_cuda_unregister_host_buffer;
     }
+    if (strcmp(name, "ggml_backend_get_features") == 0) {
+        return (void *)ggml_backend_cuda_get_features;
+    }
     return nullptr;
 }
 
 static const ggml_backend_reg_i ggml_backend_cuda_reg_interface = {
     /* .get_name          = */ ggml_backend_cuda_reg_get_name,
     /* .get_device_count  = */ ggml_backend_cuda_reg_get_device_count,
-    /* .get_device_get    = */ ggml_backend_cuda_reg_get_device,
+    /* .get_device        = */ ggml_backend_cuda_reg_get_device,
     /* .get_proc_address  = */ ggml_backend_cuda_reg_get_proc_address,
 };
 
@@ -3322,16 +3236,17 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                 dev_ctx->description = prop.name;
 
                 ggml_backend_dev_t dev = new ggml_backend_device {
-                    /* .interface = */ ggml_backend_cuda_device_interface,
-                    /* .reg       = */ &reg,
-                    /* .context   = */ dev_ctx
+                    /* .iface   = */ ggml_backend_cuda_device_interface,
+                    /* .reg     = */ &reg,
+                    /* .context = */ dev_ctx
                 };
                 ctx->devices.push_back(dev);
             }
 
             reg = ggml_backend_reg {
-                /* .interface = */ ggml_backend_cuda_reg_interface,
-                /* .context   = */ ctx
+                /* .api_version = */ GGML_BACKEND_API_VERSION,
+                /* .iface       = */ ggml_backend_cuda_reg_interface,
+                /* .context     = */ ctx
             };
         }
 
@@ -3362,3 +3277,5 @@ ggml_backend_t ggml_backend_cuda_init(int device) {
 
     return cuda_backend;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_cuda_reg)
diff --git a/ggml/src/ggml-cuda/mma.cuh b/ggml/src/ggml-cuda/mma.cuh
index a452a3cc3d1..7d11540afd2 100644
--- a/ggml/src/ggml-cuda/mma.cuh
+++ b/ggml/src/ggml-cuda/mma.cuh
@@ -171,7 +171,7 @@ struct mma_int_C_I16J8 {
 
     __device__ __forceinline__ void mma_K4(const mma_int_A_I16K4 & mma_A, const mma_int_B_J8K4 & mma_B) {
 #ifdef INT8_MMA_AVAILABLE
-#if __CUDA_ARCH__ >= CC_AMPERE
+#if __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
         asm("mma.sync.aligned.m16n8k16.row.col.s32.s8.s8.s32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};"
             : "+r"(x[0]), "+r"(x[1]), "+r"(x[2]), "+r"(x[3])
             : "r"(mma_A.x[0]), "r"(mma_A.x[1]), "r"(mma_B.x[0]));
@@ -183,7 +183,7 @@ struct mma_int_C_I16J8 {
         asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
             : "+r"(x[2]), "+r"(x[3])
             : "r"(mma_A.x[1]), "r"(mma_B.x[0]));
-#endif // __CUDA_ARCH__ >= CC_AMPERE
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #else
         GGML_UNUSED(mma_A);
         GGML_UNUSED(mma_B);
@@ -193,7 +193,7 @@ struct mma_int_C_I16J8 {
 
     __device__ __forceinline__ void mma_K8(const mma_int_A_I16K8 & mma_A, const mma_int_B_J8K8 & mma_B) {
 #ifdef INT8_MMA_AVAILABLE
-#if __CUDA_ARCH__ >= CC_AMPERE
+#if __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
         asm("mma.sync.aligned.m16n8k32.row.col.s32.s8.s8.s32 {%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3};"
             : "+r"(x[0]), "+r"(x[1]), "+r"(x[2]), "+r"(x[3])
             : "r"(mma_A.x[0]), "r"(mma_A.x[1]), "r"(mma_A.x[2]), "r"(mma_A.x[3]), "r"(mma_B.x[0]), "r"(mma_B.x[1]));
@@ -211,7 +211,7 @@ struct mma_int_C_I16J8 {
         asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
             : "+r"(x[2]), "+r"(x[3])
             : "r"(mma_A.x[3]), "r"(mma_B.x[1]));
-#endif // __CUDA_ARCH__ >= CC_AMPERE
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #else
         GGML_UNUSED(mma_A);
         GGML_UNUSED(mma_B);
diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu
index ae5c68ab351..270251df4f1 100644
--- a/ggml/src/ggml-cuda/mmq.cu
+++ b/ggml/src/ggml-cuda/mmq.cu
@@ -27,7 +27,7 @@ void ggml_cuda_op_mul_mat_q(
     // The stream-k decomposition is only faster for recent NVIDIA GPUs.
     // Also its fixup needs to allocate a temporary buffer in the memory pool.
     // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
-    const bool use_stream_k = compute_capability >= CC_VOLTA && compute_capability < CC_OFFSET_AMD && src1_ncols == ne11;
+    const bool use_stream_k = compute_capability >= GGML_CUDA_CC_VOLTA && compute_capability < GGML_CUDA_CC_OFFSET_AMD && src1_ncols == ne11;
     const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
 
     switch (src0->type) {
@@ -136,7 +136,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
         return true;
     }
 
-    if (cc < MIN_CC_DP4A) {
+    if (cc < GGML_CUDA_CC_DP4A) {
         return false;
     }
 
@@ -144,9 +144,9 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
     return true;
 #endif //GGML_CUDA_FORCE_MMQ
 
-    if (cc < CC_OFFSET_AMD) {
-        return cc < CC_VOLTA || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    if (cc < GGML_CUDA_CC_OFFSET_AMD) {
+        return cc < GGML_CUDA_CC_VOLTA || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
     }
 
-    return cc < CC_RDNA3 || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    return (cc < GGML_CUDA_CC_RDNA3 && cc != GGML_CUDA_CC_CDNA && cc != GGML_CUDA_CC_VEGA20) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
 }
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
index 021a25682c8..3cd508a1d4a 100644
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
@@ -89,9 +89,9 @@ struct tile_x_sizes {
 static constexpr int get_mmq_x_max_host(const int cc) {
     return int8_mma_available(cc) ? 128 :
 #ifdef GGML_CUDA_FORCE_MMQ
-        cc >= CC_VOLTA && cc < CC_OFFSET_AMD ? 128                     : 64;
+        cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD ? 128                     : 64;
 #else
-        cc >= CC_VOLTA && cc < CC_OFFSET_AMD ? MMQ_DP4A_MAX_BATCH_SIZE : 64;
+        cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD ? MMQ_DP4A_MAX_BATCH_SIZE : 64;
 #endif // GGML_CUDA_FORCE_MMQ
 }
 
@@ -100,43 +100,43 @@ static constexpr __device__ int get_mmq_x_max_device() {
     return 128;
 #else // INT8_MMA_AVAILABLE
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
     return 128;
-#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#else // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 
-#if __CUDA_ARCH__ >= CC_VOLTA
+#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 #ifdef GGML_CUDA_FORCE_MMQ
     return MMQ_DP4A_MAX_BATCH_SIZE;
 #else // GGML_CUDA_FORCE_MMQ
     return 128;
 #endif // GGML_CUDA_FORCE_MMQ
-#else // __CUDA_ARCH__ >= CC_VOLTA
+#else // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 
     return 64;
-#endif // __CUDA_ARCH__ >= CC_VOLTA
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #endif // INT8_MMA_AVAILABLE
 }
 
 static constexpr int get_mmq_y_host(const int cc) {
-    return cc >= CC_OFFSET_AMD ? (cc == CC_RDNA1 ? 64 : 128) : (cc >= CC_VOLTA ? 128 : 64);
+    return cc >= GGML_CUDA_CC_OFFSET_AMD ? (cc == GGML_CUDA_CC_RDNA1 ? 64 : 128) : (cc >= GGML_CUDA_CC_VOLTA ? 128 : 64);
 }
 
 static constexpr __device__ int get_mmq_y_device() {
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 #if defined(RDNA1)
     return 64;
 #else
     return 128;
 #endif // defined RDNA1
 #else
-#if __CUDA_ARCH__ >= CC_VOLTA
+#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
     return 128;
 #else
     return 64;
-#endif // __CUDA_ARCH__ >= CC_VOLTA
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }
 
 #define MMQ_DP4A_TXS_Q4_0    tile_x_sizes{mmq_y*WARP_SIZE   + mmq_y, mmq_y*WARP_SIZE/QI4_0   + mmq_y/QI4_0,     0}
@@ -2569,17 +2569,17 @@ static __device__ void mul_mat_q_process_tile(
 // The mul_mat_q kernel implements "stream-k" work partitioning as described in https://arxiv.org/abs/2301.03598
 
 template <ggml_type type, int mmq_x, int nwarps, bool need_check>
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
     __launch_bounds__(WARP_SIZE*nwarps, 2)
-#endif // defined(RDNA3) || defined(RDNA2)
+#endif // defined(RDNA3) || defined(RDNA2) || defined(CDNA) || defined(GCN)
 #else
-#if __CUDA_ARCH__ >= CC_VOLTA
+#if __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
     __launch_bounds__(WARP_SIZE*nwarps, 1)
 #else
     __launch_bounds__(WARP_SIZE*nwarps, 2)
-#endif // __CUDA_ARCH__ >= CC_VOLTA
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 static __global__ void mul_mat_q(
     const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
     const int ne00, const int ne01, const int stride01, const int ne10, const int ne11, const int stride11, const int ne0) {
@@ -2594,7 +2594,7 @@ static __global__ void mul_mat_q(
     constexpr int mmq_y = get_mmq_y_device();
 
     // On AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
-#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
+#if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
     {
         constexpr bool fixup = false;
         mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
@@ -2602,7 +2602,7 @@ static __global__ void mul_mat_q(
                 blockIdx.x, blockIdx.y, 0, ne00/qk);
         return;
     }
-#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
+#endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
 
     const     int64_t blocks_per_ne00 = ne00 / qk;
     constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
@@ -2765,14 +2765,14 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
     const int shmem = mmq_get_shmem<type>(mmq_x, mmq_y, cc);
 
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
     static bool shmem_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
     if (!shmem_limit_raised[id]) {
         CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
         CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
         shmem_limit_raised[id] = true;
     }
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 
     const int nty = (args.ne01 + mmq_y - 1) / mmq_y;
     const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;
@@ -2825,7 +2825,7 @@ void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cuda
     const int mmq_x_max = get_mmq_x_max_host(cc);
     const int mmq_y = get_mmq_y_host(cc);
     const int block_num_y = (args.ne01 + mmq_y - 1) / mmq_y;
-    const bool use_stream_k = cc >= CC_VOLTA && cc < CC_OFFSET_AMD;
+    const bool use_stream_k = cc >= GGML_CUDA_CC_VOLTA && cc < GGML_CUDA_CC_OFFSET_AMD;
 
     int mmq_x_best  = 0;
     int nparts_best = INT_MAX;
diff --git a/ggml/src/ggml-cuda/mmv.cu b/ggml/src/ggml-cuda/mmv.cu
new file mode 100644
index 00000000000..a4b4f6bc10d
--- /dev/null
+++ b/ggml/src/ggml-cuda/mmv.cu
@@ -0,0 +1,223 @@
+#include "common.cuh"
+#include "mmv.cuh"
+
+template <typename type_acc, int block_size>
+static __global__ void mul_mat_vec(
+        const half * __restrict__ x, const float * __restrict__ y, float * __restrict__ dst, const int64_t ncols2, const int64_t stride_row,
+        const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst) {
+    const int64_t row     = blockIdx.x;
+    const int64_t channel = blockIdx.z;
+    const int     tid     = threadIdx.x;
+
+    x   += (channel/channel_ratio)*stride_channel_x + row*stride_row;
+    y   +=  channel               *stride_channel_y;
+    dst +=  channel               *stride_channel_dst;
+
+    const half2  * x2 = (const half2  *) x;
+    const float2 * y2 = (const float2 *) y;
+
+    extern __shared__ char data_mmv[];
+    float * buf_iw = (float *) data_mmv;
+
+    if (block_size > WARP_SIZE) {
+        if (tid < WARP_SIZE) {
+            buf_iw[tid] = 0.0f;
+        }
+        __syncthreads();
+    }
+
+    float sumf;
+
+    if (std::is_same<type_acc, float>::value) {
+        sumf = 0.0f;
+
+        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+            const float2 tmpx = __half22float2(x2[col2]);
+            const float2 tmpy = y2[col2];
+            sumf += tmpx.x * tmpy.x;
+            sumf += tmpx.y * tmpy.y;
+        }
+    } else {
+#ifdef FP16_AVAILABLE
+        half2 sumh2 = make_half2(0.0f, 0.0f);
+
+        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+            const float2 tmp = y2[col2];
+            sumh2 += x2[col2] * make_half2(tmp.x, tmp.y);
+        }
+
+        sumf = __low2float(sumh2) + __high2float(sumh2);
+#else
+        NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+    }
+
+    sumf = warp_reduce_sum(sumf);
+
+    if (block_size > WARP_SIZE) {
+        buf_iw[tid/WARP_SIZE] = sumf;
+        __syncthreads();
+        if (tid >= WARP_SIZE) {
+            return;
+        }
+        sumf = buf_iw[tid];
+        sumf = warp_reduce_sum(sumf);
+    }
+
+    if (tid != 0) {
+        return;
+    }
+
+    dst[row] = sumf;
+}
+
+template <typename type_acc>
+static void launch_mul_mat_vec_cuda(
+        const half * x, const float * y, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
+        cudaStream_t stream) {
+    GGML_ASSERT(ncols      % 2 == 0);
+    GGML_ASSERT(stride_row % 2 == 0);
+    GGML_ASSERT(nchannels_y % nchannels_x == 0);
+    const int64_t channel_ratio = nchannels_y / nchannels_x;
+
+    int64_t block_size_best = WARP_SIZE;
+    int64_t niter_best      = (ncols + 2*WARP_SIZE - 1) / (2*WARP_SIZE);
+    for (int64_t block_size = 2*WARP_SIZE; block_size <= 256; block_size += WARP_SIZE) {
+        const int64_t niter = (ncols + 2*block_size - 1) / (2*block_size);
+        if (niter < niter_best) {
+            niter_best      = niter;
+            block_size_best = block_size;
+        }
+    }
+
+    const int smem = WARP_SIZE*sizeof(float);
+    const dim3 block_nums(nrows, 1, nchannels_y);
+    const dim3 block_dims(block_size_best, 1, 1);
+    switch (block_size_best) {
+        case   32: {
+            mul_mat_vec<type_acc,  32><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case   64: {
+            mul_mat_vec<type_acc,  64><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case   96: {
+            mul_mat_vec<type_acc,  96><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  128: {
+            mul_mat_vec<type_acc, 128><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  160: {
+            mul_mat_vec<type_acc, 160><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  192: {
+            mul_mat_vec<type_acc, 192><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  224: {
+            mul_mat_vec<type_acc, 224><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        case  256: {
+            mul_mat_vec<type_acc, 256><<<block_nums, block_dims, smem, stream>>>
+                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst);
+        } break;
+        default: {
+            GGML_ABORT("fatal error");
+        } break;
+    }
+}
+
+static void mul_mat_vec_cuda(
+        const half * x, const float * y, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
+        const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
+        enum ggml_prec prec, cudaStream_t stream) {
+    switch (prec) {
+        case GGML_PREC_DEFAULT: {
+            launch_mul_mat_vec_cuda<half>(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y,
+                stride_channel_x, stride_channel_y, stride_channel_dst, stream);
+        } break;
+        case GGML_PREC_F32: {
+            launch_mul_mat_vec_cuda<float>(x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y,
+                stride_channel_x, stride_channel_y, stride_channel_dst, stream);
+        } break;
+    }
+}
+
+void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+
+    GGML_ASSERT(src1->ne[1] == 1);
+
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
+
+    const half  * src0_d = (const half  *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       *  dst_d = (float       *)  dst->data;
+
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne12 = src1->ne[2];
+    GGML_ASSERT(dst->ne[2] == ne12);
+
+    GGML_ASSERT(src0->ne[3] == 1);
+    GGML_ASSERT(src1->ne[3] == 1);
+    GGML_ASSERT( dst->ne[3] == 1);
+
+    const int64_t stride_row         = src0->nb[1] / ggml_type_size(src0->type);
+    const int64_t channel_stride_x   = src0->nb[2] / ggml_type_size(src0->type);
+    const int64_t channel_stride_y   = src1->nb[2] / ggml_type_size(src1->type);
+    const int64_t channel_stride_dst =  dst->nb[2] / ggml_type_size( dst->type);
+
+    mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, stride_row, ne02, ne12, channel_stride_x, channel_stride_y, channel_stride_dst, prec, ctx.stream());
+}
+
+void ggml_cuda_op_mul_mat_vec(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t row_diff = row_high - row_low;
+
+    GGML_ASSERT(src1_ncols == 1);
+
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
+
+
+    // ggml_cuda_op provides single, contiguous matrices
+    const int64_t stride_row         = ne00;
+    const int64_t nchannels_x        = 1;
+    const int64_t nchannels_y        = 1;
+    const int64_t channel_stride_x   = 0;
+    const int64_t channel_stride_y   = 0;
+    const int64_t channel_stride_dst = 0;
+
+    mul_mat_vec_cuda((const half *) src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
+        nchannels_x, nchannels_y, channel_stride_x, channel_stride_y, channel_stride_dst, prec, stream);
+
+    GGML_UNUSED(ctx);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddq_i);
+    GGML_UNUSED(src1_ncols);
+    GGML_UNUSED(src1_padded_row_size);
+}
diff --git a/ggml/src/ggml-cuda/dmmv.cuh b/ggml/src/ggml-cuda/mmv.cuh
similarity index 55%
rename from ggml/src/ggml-cuda/dmmv.cuh
rename to ggml/src/ggml-cuda/mmv.cuh
index e727eb97f6a..78a1cd4a690 100644
--- a/ggml/src/ggml-cuda/dmmv.cuh
+++ b/ggml/src/ggml-cuda/mmv.cuh
@@ -1,20 +1,12 @@
 #include "common.cuh"
 
-// dmmv = dequantize_mul_mat_vec
+// maximum number of src0 rows with which to use mul_mat_vec over cuBLAS if FP16 tensor cores are available
+#define MMV_MAX_ROWS 512
 
-// TODO: remove this?
-#ifndef GGML_CUDA_DMMV_X
-#define GGML_CUDA_DMMV_X 32
-#endif
+void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
 
-#ifndef GGML_CUDA_MMV_Y
-#define GGML_CUDA_MMV_Y 1
-#endif
-
-void ggml_cuda_op_dequantize_mul_mat_vec(
+void ggml_cuda_op_mul_mat_vec(
     ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
     const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
     const int64_t src1_padded_row_size, cudaStream_t stream);
-
-bool ggml_cuda_dmmv_type_supported(ggml_type src0_type);
diff --git a/ggml/src/ggml-cuda/mmvq.cu b/ggml/src/ggml-cuda/mmvq.cu
index 7dbbc993903..e3b912d875b 100644
--- a/ggml/src/ggml-cuda/mmvq.cu
+++ b/ggml/src/ggml-cuda/mmvq.cu
@@ -48,10 +48,10 @@ static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
 }
 
 template <ggml_type type, int ncols_y>
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 // tell the compiler to use as many registers as it wants, see nwarps definition below
 __launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1)
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void mul_mat_vec_q(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
     const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
@@ -62,13 +62,13 @@ static __global__ void mul_mat_vec_q(
 
     constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
 
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
     constexpr int nwarps              = 1;
     constexpr int rows_per_cuda_block = 1;
 #else
     constexpr int nwarps              = ncols_y <= 4 ? 4 : 2;
     constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
 
     const     int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
     const     int row0 = rows_per_cuda_block*blockIdx.x;
@@ -142,7 +142,7 @@ static void mul_mat_vec_q_cuda(
     int64_t nwarps = 1;
     int64_t rows_per_cuda_block = 1;
 
-    if (ggml_cuda_info().devices[id].cc < CC_RDNA2) { // NVIDIA and AMD older than RDNA2
+    if (ggml_cuda_info().devices[id].cc < GGML_CUDA_CC_CDNA || ggml_cuda_info().devices[id].cc == GGML_CUDA_CC_RDNA1) { // NVIDIA and AMD older than RDNA2 but not CDNA
         switch(ncols_y) {
             case 1:
                 nwarps = 4;
diff --git a/ggml/src/ggml-cuda/opt-step-adamw.cu b/ggml/src/ggml-cuda/opt-step-adamw.cu
index d6f13a9c62d..35154f29966 100644
--- a/ggml/src/ggml-cuda/opt-step-adamw.cu
+++ b/ggml/src/ggml-cuda/opt-step-adamw.cu
@@ -1,11 +1,11 @@
+#include "ggml-impl.h"
 #include "opt-step-adamw.cuh"
 
 #include <cstdint>
 
 static __global__ void opt_step_adamw_f32(
-    float * __restrict__ x, const float * __restrict__ g, float * __restrict__ g_m, float * __restrict__ g_v, const int64_t k,
-    const float alpha, const float beta1, const float beta2, const float eps, const float wd,
-    const float beta1h, const float beta2h) {
+    float * __restrict__ x, const float * __restrict__ g, float * __restrict__ g_m, float * __restrict__ g_v,
+    const float * __restrict__ pars, const int64_t k) {
 
     const int64_t i = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
 
@@ -13,6 +13,14 @@ static __global__ void opt_step_adamw_f32(
         return;
     }
 
+    const float alpha  = pars[0];
+    const float beta1  = pars[1];
+    const float beta2  = pars[2];
+    const float eps    = pars[3];
+    const float wd     = pars[4];
+    const float beta1h = pars[5];
+    const float beta2h = pars[6];
+
     const float gi = g[i];
     const float gmi = g_m[i]*beta1 +    gi*(1.0f - beta1);
     const float gvi = g_v[i]*beta2 + gi*gi*(1.0f - beta2);
@@ -23,58 +31,48 @@ static __global__ void opt_step_adamw_f32(
     const float mh =       gmi*beta1h;
     const float vh = sqrtf(gvi*beta2h) + eps;
 
-    x[i] = x[i]*(1.0f - alpha*wd) - mh/vh;
+    x[i] = x[i]*(1.0f - alpha*wd) - alpha*mh/vh;
 }
 
 static void opt_step_adamw_f32_cuda(
-    float * x, const float * g, float * g_m, float * g_v, const int64_t k,
-    const float alpha, const float beta1, const float beta2, const float eps, const float wd,
-    const float beta1h, const float beta2h, cudaStream_t stream) {
+    float * x, const float * g, float * g_m, float * g_v, const float * pars, const int64_t k, cudaStream_t stream) {
 
     const dim3 block_dims(CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
     const dim3 block_nums((k + CUDA_OPT_STEP_ADAMW_BLOCK_SIZE - 1) / CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
-    opt_step_adamw_f32<<<block_nums, block_dims, 0, stream>>>(x, g, g_m, g_v, k, alpha, beta1, beta2, eps, wd, beta1h, beta2h);
+    opt_step_adamw_f32<<<block_nums, block_dims, 0, stream>>>(x, g, g_m, g_v, pars, k);
 }
 
 void ggml_cuda_opt_step_adamw(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0        = dst->src[0];
-    const ggml_tensor * src0_grad   = dst->src[1];
-    const ggml_tensor * src0_grad_m = dst->src[2];
-    const ggml_tensor * src0_grad_v = dst->src[3];
-
-    GGML_ASSERT(src0->type        == GGML_TYPE_F32);
-    GGML_ASSERT(src0_grad->type   == GGML_TYPE_F32);
-    GGML_ASSERT(src0_grad_m->type == GGML_TYPE_F32);
-    GGML_ASSERT(src0_grad_v->type == GGML_TYPE_F32);
+    const ggml_tensor * src0         = dst->src[0];
+    const ggml_tensor * src0_grad    = dst->src[1];
+    const ggml_tensor * src0_grad_m  = dst->src[2];
+    const ggml_tensor * src0_grad_v  = dst->src[3];
+    const ggml_tensor * adamw_params = dst->src[4];
+
+    GGML_ASSERT(src0->type         == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad->type    == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad_m->type  == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad_v->type  == GGML_TYPE_F32);
+    GGML_ASSERT(adamw_params->type == GGML_TYPE_F32);
     GGML_ASSERT(ggml_is_contiguous(src0));
     GGML_ASSERT(ggml_is_contiguous(src0_grad));
     GGML_ASSERT(ggml_is_contiguous(src0_grad_m));
     GGML_ASSERT(ggml_is_contiguous(src0_grad_v));
+    GGML_ASSERT(ggml_is_contiguous(adamw_params));
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m));
     GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v));
+    GGML_ASSERT(ggml_nelements(adamw_params) == 7);
 
-    float       * src0_d        = (float       *) src0->data;
-    const float * src0_grad_d   = (const float *) src0_grad->data;
-    float       * src0_grad_m_d = (float       *) src0_grad_m->data;
-    float       * src0_grad_v_d = (float       *) src0_grad_v->data;
+    float       * src0_d         = (float       *) src0->data;
+    const float * src0_grad_d    = (const float *) src0_grad->data;
+    float       * src0_grad_m_d  = (float       *) src0_grad_m->data;
+    float       * src0_grad_v_d  = (float       *) src0_grad_v->data;
+    const float * adamw_params_d = (const float *) adamw_params->data;
 
     cudaStream_t stream = ctx.stream();
 
     const int64_t ne = ggml_nelements(src0);
 
-    int64_t iter;  memcpy(&iter,  &dst->op_params[0], sizeof(int64_t));
-    float   alpha; memcpy(&alpha, &dst->op_params[2], sizeof(float));
-    float   beta1; memcpy(&beta1, &dst->op_params[3], sizeof(float));
-    float   beta2; memcpy(&beta2, &dst->op_params[4], sizeof(float));
-    float   eps;   memcpy(&eps,   &dst->op_params[5], sizeof(float));
-    float   wd;    memcpy(&wd,    &dst->op_params[6], sizeof(float));
-
-    const float beta1h  = alpha/(1.0f - powf(beta1, iter));
-    const float beta2h  =  1.0f/(1.0f - powf(beta2, iter));
-
-    opt_step_adamw_f32_cuda(src0_d, src0_grad_d, src0_grad_m_d, src0_grad_v_d, ne, alpha, beta1, beta2, eps, wd, beta1h, beta2h, stream);
-
-    iter++;
-    memcpy(&dst->op_params[0], &iter, sizeof(int64_t));
+    opt_step_adamw_f32_cuda(src0_d, src0_grad_d, src0_grad_m_d, src0_grad_v_d, adamw_params_d, ne, stream);
 }
diff --git a/ggml/src/ggml-cuda/quantize.cu b/ggml/src/ggml-cuda/quantize.cu
index 45408ce8684..1702e4ce2fe 100644
--- a/ggml/src/ggml-cuda/quantize.cu
+++ b/ggml/src/ggml-cuda/quantize.cu
@@ -69,8 +69,8 @@ static __global__ void quantize_mmq_q8_1(
 
     // Exchange max. abs. value between vals_per_scale/4 threads.
 #pragma unroll
-    for (int mask = vals_per_scale/8; mask > 0; mask >>= 1) {
-        amax = fmaxf(amax, __shfl_xor_sync(0xFFFFFFFF, amax, mask, WARP_SIZE));
+    for (int offset = vals_per_scale/8; offset > 0; offset >>= 1) {
+        amax = fmaxf(amax, __shfl_xor_sync(0xFFFFFFFF, amax, offset, WARP_SIZE));
     }
 
     float sum;
@@ -79,8 +79,8 @@ static __global__ void quantize_mmq_q8_1(
 
         // Exchange calculate sum across vals_per_sum/4 threads.
 #pragma unroll
-        for (int mask = vals_per_sum/8; mask > 0; mask >>= 1) {
-            sum += __shfl_xor_sync(0xFFFFFFFF, sum, mask, WARP_SIZE);
+        for (int offset = vals_per_sum/8; offset > 0; offset >>= 1) {
+            sum += __shfl_xor_sync(0xFFFFFFFF, sum, offset, WARP_SIZE);
         }
     }
 
diff --git a/ggml/src/ggml-cuda/rope.cu b/ggml/src/ggml-cuda/rope.cu
index 88f586d689c..2c84778d29c 100644
--- a/ggml/src/ggml-cuda/rope.cu
+++ b/ggml/src/ggml-cuda/rope.cu
@@ -4,6 +4,11 @@ struct rope_corr_dims {
     float v[2];
 };
 
+
+struct mrope_sections {
+    int v[4];
+};
+
 static __device__ float rope_yarn_ramp(const float low, const float high, const int i0) {
     const float y = (i0 / 2 - low) / max(0.001f, high - low);
     return 1.0f - min(1.0f, max(0.0f, y));
@@ -108,6 +113,105 @@ static __global__ void rope_neox(
     dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta;
 }
 
+template<typename T, bool has_ff>
+static __global__ void rope_multi(
+    const T * x, T * dst, int ne0, int ne2, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors, mrope_sections sections) {
+    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i0 >= n_dims) {
+        const int i = row*ne0 + i0;
+
+        dst[i + 0] = x[i + 0];
+        dst[i + 1] = x[i + 1];
+
+        return;
+    }
+
+    const int i  = row*ne0 + i0/2;
+    const int i2 = row/p_delta_rows;
+
+    int sect_dims = sections.v[0] + sections.v[1] + sections.v[2] + sections.v[3];
+    int sec_w = sections.v[1] + sections.v[0];
+    int sector = (i0 / 2) % sect_dims;
+
+    float theta_base = 0.0;
+    if (sector < sections.v[0]) {
+        theta_base = pos[i2]*powf(theta_scale, i0/2.0f);
+    }
+    else if (sector >= sections.v[0] && sector < sec_w) {
+        theta_base = pos[i2 + ne2 * 1]*powf(theta_scale, i0/2.0f);
+    }
+    else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
+        theta_base = pos[i2 + ne2 * 2]*powf(theta_scale, i0/2.0f);
+    }
+    else if (sector >= sec_w + sections.v[2]) {
+        theta_base = pos[i2 + ne2 * 3]*powf(theta_scale, i0/2.0f);
+    }
+
+    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+
+    float cos_theta;
+    float sin_theta;
+
+    rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+
+    const float x0 = x[i + 0];
+    const float x1 = x[i + n_dims/2];
+
+    dst[i + 0]        = x0*cos_theta - x1*sin_theta;
+    dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta;
+}
+
+template<typename T, bool has_ff>
+static __global__ void rope_vision(
+    const T * x, T * dst, int ne0, int ne2, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors, mrope_sections sections) {
+    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
+
+    const int i  = row*ne0 + i0/2;
+    const int i2 = row/p_delta_rows; // i2-th tokens
+
+    int sect_dims = sections.v[0] + sections.v[1];
+    int sec_w = sections.v[1] + sections.v[0];
+    int sector = (i0 / 2) % sect_dims;
+
+    float theta_base = 0.0;
+    if (sector < sections.v[0]) {
+        const int p = sector;
+        theta_base = pos[i2]*powf(theta_scale, p);
+    }
+    else if (sector >= sections.v[0] && sector < sec_w) {
+        const int p = sector - sections.v[0];
+        theta_base = pos[i2 + ne2]*powf(theta_scale, p);
+    }
+
+    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+
+    float cos_theta;
+    float sin_theta;
+
+    rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+
+    const float x0 = x[i + 0];
+    const float x1 = x[i + n_dims];
+
+    dst[i + 0]      = x0*cos_theta - x1*sin_theta;
+    dst[i + n_dims] = x0*sin_theta + x1*cos_theta;
+}
+
 template<typename T>
 static void rope_norm_cuda(
     const T * x, T * dst, int ne0, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
@@ -156,6 +260,56 @@ static void rope_neox_cuda(
     }
 }
 
+template<typename T>
+static void rope_multi_cuda(
+    const T * x, T * dst, int ne0, int ne2, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, mrope_sections sections, cudaStream_t stream) {
+    GGML_ASSERT(ne0 % 2 == 0);
+    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
+    const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
+    const dim3 block_nums(nr, n_blocks_x, 1);
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+
+    if (freq_factors == nullptr) {
+        rope_multi<T, false><<<block_nums, block_dims, 0, stream>>>(
+                x, dst, ne0, ne2, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+                theta_scale, freq_factors, sections
+                );
+    } else {
+        rope_multi<T, true><<<block_nums, block_dims, 0, stream>>>(
+                x, dst, ne0, ne2, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+                theta_scale, freq_factors, sections
+                );
+    }
+}
+
+template<typename T>
+static void rope_vision_cuda(
+    const T * x, T * dst, int ne0, int ne2, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, mrope_sections sections, cudaStream_t stream) {
+    GGML_ASSERT(ne0 % 2 == 0);
+    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
+    const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
+    const dim3 block_nums(nr, n_blocks_x, 1);
+    // break down (head_dim, heads, seq) into (CUDA_ROPE_BLOCK_SIZE, x, heads * seq)
+    // where x ~= ceil(head_dim / CUDA_ROPE_BLOCK_SIZE);
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+
+    if (freq_factors == nullptr) {
+        rope_vision<T, false><<<block_nums, block_dims, 0, stream>>>(
+                x, dst, ne0, ne2, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+                theta_scale, freq_factors, sections
+                );
+    } else {
+        rope_vision<T, true><<<block_nums, block_dims, 0, stream>>>(
+                x, dst, ne0, ne2, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+                theta_scale, freq_factors, sections
+                );
+    }
+}
+
 static void rope_norm_cuda_f16(
     const half * x, half * dst, int ne0, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
     float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
@@ -185,6 +339,38 @@ static void rope_neox_cuda_f32(
     rope_neox_cuda<float>(x, dst, ne0, n_dims, nr, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
 }
 
+static void rope_multi_cuda_f16(
+    const half * x, half * dst, int ne0, int ne2, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, mrope_sections sections, cudaStream_t stream
+) {
+
+    rope_multi_cuda<half>(x, dst, ne0, ne2, n_dims, nr, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+}
+
+static void rope_multi_cuda_f32(
+    const float * x, float * dst, int ne0, int ne2, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, mrope_sections sections, cudaStream_t stream
+) {
+
+    rope_multi_cuda<float>(x, dst, ne0, ne2, n_dims, nr, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+}
+
+static void rope_vision_cuda_f16(
+    const half * x, half * dst, int ne0, int ne2, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, mrope_sections sections, cudaStream_t stream
+) {
+
+    rope_vision_cuda<half>(x, dst, ne0, ne2, n_dims, nr, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+}
+
+static void rope_vision_cuda_f32(
+    const float * x, float * dst, int ne0, int ne2, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, mrope_sections sections, cudaStream_t stream
+) {
+
+    rope_vision_cuda<float>(x, dst, ne0, ne2, n_dims, nr, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, stream);
+}
+
 void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
@@ -201,8 +387,9 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
     GGML_ASSERT(src0->type == dst->type);
 
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
+    const int64_t ne00 = src0->ne[0]; // head dims
+    const int64_t ne01 = src0->ne[1]; // num heads
+    const int64_t ne02 = src0->ne[2]; // num heads
     const int64_t nr = ggml_nrows(src0);
 
     //const int n_past     = ((int32_t *) dst->op_params)[0];
@@ -210,6 +397,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int mode       = ((int32_t *) dst->op_params)[2];
     //const int n_ctx      = ((int32_t *) dst->op_params)[3];
     const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+    mrope_sections sections;
 
     // RoPE alteration for extended context
     float freq_base;
@@ -225,8 +413,19 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections.v,  (int32_t *) dst->op_params + 11, sizeof(int)*4);
 
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+
+    if (is_mrope) {
+        GGML_ASSERT(sections.v[0] > 0 || sections.v[1] > 0 || sections.v[2] > 0);
+    }
+
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne00/2);
+    }
 
     const int32_t * pos = (const int32_t *) src1_d;
 
@@ -253,6 +452,34 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
         } else {
             GGML_ABORT("fatal error");
         }
+    } else if (is_mrope && !is_vision) {
+        if (src0->type == GGML_TYPE_F32) {
+            rope_multi_cuda_f32(
+                (const float *)src0_d, (float *)dst_d, ne00, ne02, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, sections, stream
+            );
+        } else if (src0->type == GGML_TYPE_F16) {
+            rope_multi_cuda_f16(
+                (const half *)src0_d, (half *)dst_d, ne00, ne02, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, sections, stream
+            );
+        } else {
+            GGML_ABORT("fatal error");
+        }
+    } else if (is_vision) {
+        if (src0->type == GGML_TYPE_F32) {
+            rope_vision_cuda_f32(
+                (const float *)src0_d, (float *)dst_d, ne00, ne02, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, sections, stream
+            );
+        } else if (src0->type == GGML_TYPE_F16) {
+            rope_vision_cuda_f16(
+                (const half *)src0_d, (half *)dst_d, ne00, ne02, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, sections, stream
+            );
+        } else {
+            GGML_ABORT("fatal error");
+        }
     } else {
         if (src0->type == GGML_TYPE_F32) {
             rope_norm_cuda_f32(
diff --git a/ggml/src/ggml-cuda/rwkv-wkv.cu b/ggml/src/ggml-cuda/rwkv-wkv.cu
deleted file mode 100644
index 098e92d3521..00000000000
--- a/ggml/src/ggml-cuda/rwkv-wkv.cu
+++ /dev/null
@@ -1,89 +0,0 @@
-#include "common.cuh"
-#include "rwkv-wkv.cuh"
-
-static __global__ void rwkv_wkv_f32(const int B, const int T, const int C, const int H, const float * k, const float * v, const float * r, const float * tf, const float * td, const float * s, float * dst) {
-    const int tid = threadIdx.x;
-    const int bid = blockIdx.x;
-
-    const int head_size = CUDA_WKV_BLOCK_SIZE;
-    const int batch_i = bid / H;
-    const int head_i = bid % H;
-    const int state_size = C * head_size;
-    const int n_seq_tokens = T / B;
-
-    float state[head_size];
-    __shared__ float _k[head_size], _r[head_size], _tf[head_size], _td[head_size];
-
-    #pragma unroll
-    for (int i = 0; i < head_size; i++) {
-        state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid];
-    }
-
-    __syncthreads();
-    _tf[tid] = tf[head_i * head_size + tid];
-    __syncthreads();
-
-    for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; t += C) {
-        __syncthreads();
-        _k[tid] = k[t];
-        _r[tid] = r[t];
-        _td[tid] = td[t];
-        __syncthreads();
-
-        const float _v = v[t];
-        float y = 0;
-        for (int j = 0; j < head_size; j += 4) {
-            const float4& k = (float4&)(_k[j]);
-            const float4& r = (float4&)(_r[j]);
-            const float4& tf = (float4&)(_tf[j]);
-            const float4& td = (float4&)(_td[j]);
-            float4& s = (float4&)(state[j]);
-            float4 kv;
-
-            kv.x = k.x * _v;
-            kv.y = k.y * _v;
-            kv.z = k.z * _v;
-            kv.w = k.w * _v;
-
-            y += r.x * (tf.x * kv.x + s.x);
-            y += r.y * (tf.y * kv.y + s.y);
-            y += r.z * (tf.z * kv.z + s.z);
-            y += r.w * (tf.w * kv.w + s.w);
-
-            s.x = s.x * td.x + kv.x;
-            s.y = s.y * td.y + kv.y;
-            s.z = s.z * td.z + kv.z;
-            s.w = s.w * td.w + kv.w;
-        }
-        dst[t] = y;
-    }
-
-    #pragma unroll
-    for (int i = 0; i < head_size; i++) {
-        dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i];
-    }
-}
-
-void ggml_cuda_op_rwkv_wkv(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const float * k_d  = (const float *)dst->src[0]->data;
-    const float * v_d  = (const float *)dst->src[1]->data;
-    const float * r_d  = (const float *)dst->src[2]->data;
-    const float * tf_d = (const float *)dst->src[3]->data;
-    const float * td_d = (const float *)dst->src[4]->data;
-    const float * s_d  = (const float *)dst->src[5]->data;
-
-    const int64_t B = dst->src[5]->ne[1];
-    const int64_t T = dst->src[0]->ne[3];
-    const int64_t C = dst->ne[0];
-    const int64_t H = dst->src[0]->ne[2];
-
-    float * dst_d = (float *)dst->data;
-
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32);
-    GGML_ASSERT(C % H == 0);
-    GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE);
-
-    rwkv_wkv_f32<<<B * H, C / H, 0, stream>>>(B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d);
-}
diff --git a/ggml/src/ggml-cuda/rwkv-wkv.cuh b/ggml/src/ggml-cuda/rwkv-wkv.cuh
deleted file mode 100644
index 13795247fbe..00000000000
--- a/ggml/src/ggml-cuda/rwkv-wkv.cuh
+++ /dev/null
@@ -1,5 +0,0 @@
-#include "common.cuh"
-
-#define CUDA_WKV_BLOCK_SIZE 64
-
-void ggml_cuda_op_rwkv_wkv(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/ggml/src/ggml-cuda/sum.cu b/ggml/src/ggml-cuda/sum.cu
index 0583e4fe0c4..e0dafc1d204 100644
--- a/ggml/src/ggml-cuda/sum.cu
+++ b/ggml/src/ggml-cuda/sum.cu
@@ -1,10 +1,8 @@
-#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
 #define USE_CUB
-#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
 
 #ifdef USE_CUB
-// On Windows CUB uses libraries with variables called CC_PASCAL which conflict with the define in common.cuh.
-// For this reason CUB must be included BEFORE anything else.
 #include <cub/cub.cuh>
 using namespace cub;
 #endif // USE_CUB
diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
index 1f3c70c2e69..3205534d66f 100644
--- a/ggml/src/ggml-cuda/vendors/hip.h
+++ b/ggml/src/ggml-cuda/vendors/hip.h
@@ -95,6 +95,14 @@
 
 #define __CUDA_ARCH__ 1300
 
+#if defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__)
+#define GCN
+#endif
+
+#if defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx942__)
+#define CDNA
+#endif
+
 #if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
     defined(__gfx1150__) || defined(__gfx1151__)
 #define RDNA3
diff --git a/ggml/src/ggml-hip/CMakeLists.txt b/ggml/src/ggml-hip/CMakeLists.txt
new file mode 100644
index 00000000000..b15fbd24d6b
--- /dev/null
+++ b/ggml/src/ggml-hip/CMakeLists.txt
@@ -0,0 +1,104 @@
+if (NOT EXISTS $ENV{ROCM_PATH})
+    if (NOT EXISTS /opt/rocm)
+        set(ROCM_PATH /usr)
+    else()
+        set(ROCM_PATH /opt/rocm)
+    endif()
+else()
+    set(ROCM_PATH $ENV{ROCM_PATH})
+endif()
+
+list(APPEND CMAKE_PREFIX_PATH  ${ROCM_PATH})
+list(APPEND CMAKE_PREFIX_PATH "${ROCM_PATH}/lib64/cmake")
+
+# CMake on Windows doesn't support the HIP language yet
+if (WIN32)
+    set(CXX_IS_HIPCC TRUE)
+else()
+    string(REGEX MATCH "hipcc(\.bat)?$" CXX_IS_HIPCC "${CMAKE_CXX_COMPILER}")
+endif()
+
+if (CXX_IS_HIPCC)
+    if (LINUX)
+        if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+            message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
+        endif()
+
+        message(WARNING "Setting hipcc as the C++ compiler is legacy behavior."
+                " Prefer setting the HIP compiler directly. See README for details.")
+    endif()
+else()
+    # Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
+    if (AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
+        set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
+    endif()
+    cmake_minimum_required(VERSION 3.21)
+    enable_language(HIP)
+endif()
+
+find_package(hip     REQUIRED)
+find_package(hipblas REQUIRED)
+find_package(rocblas REQUIRED)
+
+message(STATUS "HIP and hipBLAS found")
+
+file(GLOB   GGML_HEADERS_ROCM "../ggml-cuda/*.cuh")
+list(APPEND GGML_HEADERS_ROCM "../../include/ggml-cuda.h")
+
+file(GLOB   GGML_SOURCES_ROCM "../ggml-cuda/*.cu")
+file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-wmma*.cu")
+list(APPEND GGML_SOURCES_ROCM ${SRCS})
+file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")
+list(APPEND GGML_SOURCES_ROCM ${SRCS})
+
+if (GGML_CUDA_FA_ALL_QUANTS)
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+else()
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+endif()
+
+ggml_add_backend_library(ggml-hip
+                         ${GGML_HEADERS_ROCM}
+                         ${GGML_SOURCES_ROCM}
+                        )
+
+# TODO: do not use CUDA definitions for HIP
+target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
+
+add_compile_definitions(GGML_USE_HIP)
+
+if (GGML_HIP_UMA)
+    add_compile_definitions(GGML_HIP_UMA)
+endif()
+
+if (GGML_CUDA_FORCE_MMQ)
+    add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+endif()
+
+if (GGML_CUDA_FORCE_CUBLAS)
+    add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+endif()
+
+if (GGML_CUDA_NO_PEER_COPY)
+    add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+endif()
+
+if (CXX_IS_HIPCC)
+    set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
+    target_link_libraries(ggml-hip PRIVATE hip::device)
+else()
+    set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE HIP)
+endif()
+
+if (GGML_STATIC)
+    message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+endif()
+
+target_link_libraries(ggml-hip PRIVATE ggml-base hip::host roc::rocblas roc::hipblas)
diff --git a/ggml/src/ggml-impl.h b/ggml/src/ggml-impl.h
index af29a26f0e0..f961134edd7 100644
--- a/ggml/src/ggml-impl.h
+++ b/ggml/src/ggml-impl.h
@@ -3,22 +3,40 @@
 // GGML internal header
 
 #include "ggml.h"
-
 #include <assert.h>
+#include <math.h>
 #include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
 #include <stdbool.h>
 #include <stdint.h>
 #include <string.h>
 
+#ifdef __ARM_FEATURE_SVE
+#include <arm_sve.h>
+#endif // __ARM_FEATURE_SVE
+
+#if defined(__ARM_NEON) && !defined(__CUDACC__)
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+#endif
+
+#if defined(__F16C__)
+#include <immintrin.h>
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-#undef MIN
-#undef MAX
+#ifndef MIN
+#    define MIN(a, b) ((a) < (b) ? (a) : (b))
+#endif
 
-#define MIN(a, b) ((a) < (b) ? (a) : (b))
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
+#ifndef MAX
+#    define MAX(a, b) ((a) > (b) ? (a) : (b))
+#endif
 
 // required for mmap as gguf only guarantees 32-byte alignment
 #define TENSOR_ALIGNMENT 32
@@ -28,13 +46,13 @@ extern "C" {
 // if C99 - static_assert is noop
 // ref: https://stackoverflow.com/a/53923785/4039976
 #ifndef __cplusplus
-#ifndef static_assert
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
-#define static_assert(cond, msg) _Static_assert(cond, msg)
-#else
-#define static_assert(cond, msg) struct global_scope_noop_trick
-#endif
-#endif
+    #ifndef static_assert
+        #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+            #define static_assert(cond, msg) _Static_assert(cond, msg)
+        #else
+            #define static_assert(cond, msg) struct global_scope_noop_trick
+        #endif
+    #endif
 #endif
 
 static inline int ggml_up32(int n) {
@@ -56,8 +74,8 @@ static inline int ggml_up(int n, int m) {
 //
 
 GGML_ATTRIBUTE_FORMAT(2, 3)
-void ggml_log_internal        (enum ggml_log_level level, const char * format, ...);
-void ggml_log_callback_default(enum ggml_log_level level, const char * text, void * user_data);
+GGML_API void ggml_log_internal        (enum ggml_log_level level, const char * format, ...);
+GGML_API void ggml_log_callback_default(enum ggml_log_level level, const char * text, void * user_data);
 
 #define GGML_LOG(...)       ggml_log_internal(GGML_LOG_LEVEL_NONE , __VA_ARGS__)
 #define GGML_LOG_INFO(...)  ggml_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
@@ -120,14 +138,12 @@ struct ggml_map_custom1_op_params {
     void             * userdata;
 };
 
-
 struct ggml_map_custom2_op_params {
     ggml_custom2_op_t   fun;
     int                 n_tasks;
     void              * userdata;
 };
 
-
 struct ggml_map_custom3_op_params {
     ggml_custom3_op_t fun;
     int n_tasks;
@@ -182,7 +198,7 @@ void ggml_hash_set_reset(struct ggml_hash_set * hash_set);
 static bool ggml_hash_contains(const struct ggml_hash_set * hash_set, struct ggml_tensor * key);
 
 // returns GGML_HASHSET_FULL if table is full, otherwise the current index of the key or where it should be inserted
-static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key);
+static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, const struct ggml_tensor * key);
 
 // returns GGML_HASHSET_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
 static size_t ggml_hash_insert(struct ggml_hash_set * hash_set, struct ggml_tensor * key);
@@ -196,7 +212,7 @@ static inline size_t ggml_hash(const struct ggml_tensor * p) {
     return (size_t)(uintptr_t)p >> 4;
 }
 
-static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key) {
+static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, const struct ggml_tensor * key) {
     size_t h = ggml_hash(key) % hash_set->size;
 
     // linear probing
@@ -267,29 +283,273 @@ enum ggml_cgraph_eval_order {
 };
 
 struct ggml_cgraph {
-    int size;
-    int n_nodes;
-    int n_leafs;
+    int size;    // maximum number of nodes/leafs/grads/grad_accs
+    int n_nodes; // number of nodes currently in use
+    int n_leafs; // number of leafs currently in use
 
-    struct ggml_tensor ** nodes;
-    struct ggml_tensor ** grads;
-    struct ggml_tensor ** leafs;
+    struct ggml_tensor ** nodes;     // tensors with data that can change if the graph is evaluated
+    struct ggml_tensor ** grads;     // the outputs of these tensors are the gradients of the nodes
+    struct ggml_tensor ** grad_accs; // accumulators for node gradients
+    struct ggml_tensor ** leafs;     // tensors with constant data
 
     struct ggml_hash_set visited_hash_set;
 
     enum ggml_cgraph_eval_order order;
 };
 
+// returns a slice of cgraph with nodes [i0, i1)
+// the slice does not have leafs or gradients
+// if you need the gradients, get them from the original graph
 struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
 
 // Memory allocation
 
-void * ggml_aligned_malloc(size_t size);
-void ggml_aligned_free(void * ptr, size_t size);
+GGML_API void * ggml_aligned_malloc(size_t size);
+GGML_API void ggml_aligned_free(void * ptr, size_t size);
+
+// FP16 to FP32 conversion
+
+#if defined(__ARM_NEON)
+    #if defined(_MSC_VER) || (defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11)
+        typedef uint16_t ggml_fp16_internal_t;
+    #else
+        typedef __fp16 ggml_fp16_internal_t;
+    #endif
+#endif
+
+#if defined(__ARM_NEON) && !defined(_MSC_VER) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11)
+    #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+    #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+    #define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+
+    static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+        ggml_fp16_internal_t tmp;
+        memcpy(&tmp, &h, sizeof(ggml_fp16_t));
+        return (float)tmp;
+    }
+
+    static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+        ggml_fp16_t res;
+        ggml_fp16_internal_t tmp = f;
+        memcpy(&res, &tmp, sizeof(ggml_fp16_t));
+        return res;
+    }
+
+#elif defined(__F16C__)
+
+    #ifdef _MSC_VER
+        #define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+        #define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+    #else
+        #define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+        #define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+    #endif
+
+#elif defined(__POWER9_VECTOR__)
+
+    #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+    #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+    /* the inline asm below is about 12% faster than the lookup method */
+    #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
+    #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+    static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+        register float f;
+        register double d;
+        __asm__(
+            "mtfprd %0,%2\n"
+            "xscvhpdp %0,%0\n"
+            "frsp %1,%0\n" :
+            /* temp */ "=d"(d),
+            /* out */  "=f"(f):
+            /* in */   "r"(h));
+        return f;
+    }
+
+    static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+        register double d;
+        register ggml_fp16_t r;
+        __asm__( /* xscvdphp can work on double or single precision */
+            "xscvdphp %0,%2\n"
+            "mffprd %1,%0\n" :
+            /* temp */ "=d"(d),
+            /* out */  "=r"(r):
+            /* in */   "f"(f));
+        return r;
+    }
+
+#else
+
+    // FP16 <-> FP32
+    // ref: https://github.com/Maratyszcza/FP16
+
+    static inline float fp32_from_bits(uint32_t w) {
+        union {
+            uint32_t as_bits;
+            float as_value;
+        } fp32;
+        fp32.as_bits = w;
+        return fp32.as_value;
+    }
+
+    static inline uint32_t fp32_to_bits(float f) {
+        union {
+            float as_value;
+            uint32_t as_bits;
+        } fp32;
+        fp32.as_value = f;
+        return fp32.as_bits;
+    }
+
+    static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+        const uint32_t w = (uint32_t) h << 16;
+        const uint32_t sign = w & UINT32_C(0x80000000);
+        const uint32_t two_w = w + w;
+
+        const uint32_t exp_offset = UINT32_C(0xE0) << 23;
+    #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
+        const float exp_scale = 0x1.0p-112f;
+    #else
+        const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
+    #endif
+        const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
+
+        const uint32_t magic_mask = UINT32_C(126) << 23;
+        const float magic_bias = 0.5f;
+        const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
+
+        const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
+        const uint32_t result = sign |
+            (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
+        return fp32_from_bits(result);
+    }
+
+    static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
+        const float scale_to_inf = 0x1.0p+112f;
+        const float scale_to_zero = 0x1.0p-110f;
+    #else
+        const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
+        const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
+    #endif
+        float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+
+        const uint32_t w = fp32_to_bits(f);
+        const uint32_t shl1_w = w + w;
+        const uint32_t sign = w & UINT32_C(0x80000000);
+        uint32_t bias = shl1_w & UINT32_C(0xFF000000);
+        if (bias < UINT32_C(0x71000000)) {
+            bias = UINT32_C(0x71000000);
+        }
+
+        base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
+        const uint32_t bits = fp32_to_bits(base);
+        const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+        const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
+        const uint32_t nonsign = exp_bits + mantissa_bits;
+        return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+    }
+
+    #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+    #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml.c, initialized in ggml_init()
+GGML_API float ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_FP16_TO_FP32)
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return ggml_table_f32_f16[s];
+}
+
+#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#endif
+
+#if !defined(GGML_FP32_TO_FP16)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+#endif
+
+/**
+ * Converts brain16 to float32.
+ *
+ * The bfloat16 floating point format has the following structure:
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───┐
+ *     0b0000000000000000 brain16
+ *
+ * Since bf16 has the same number of exponent bits as a 32bit float,
+ * encoding and decoding numbers becomes relatively straightforward.
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───────────────────┐
+ *     0b00000000000000000000000000000000 IEEE binary32
+ *
+ * For comparison, the standard fp16 format has fewer exponent bits.
+ *
+ *       ┌sign
+ *       │
+ *       │  ┌exponent
+ *       │  │
+ *       │  │    ┌mantissa
+ *       │  │    │
+ *       │┌─┴─┐┌─┴──────┐
+ *     0b0000000000000000 IEEE binary16
+ *
+ * @see IEEE 754-2008
+ */
+static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.i = (uint32_t)h.bits << 16;
+    return u.f;
+}
+
+/**
+ * Converts float32 to brain16.
+ *
+ * This is binary identical with Google Brain float conversion.
+ * Floats shall round to nearest even, and NANs shall be quiet.
+ * Subnormals aren't flushed to zero, except perhaps when used.
+ * This code should vectorize nicely if using modern compilers.
+ */
+static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
+    ggml_bf16_t h;
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.f = s;
+    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
+        h.bits = (u.i >> 16) | 64; /* force to quiet */
+        return h;
+    }
+    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
+    return h;
+}
 
-// TODO: move to threading file
-void ggml_critical_section_start(void);
-void ggml_critical_section_end(void);
+#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
+#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
 
 #ifdef __cplusplus
 }
diff --git a/ggml/src/ggml-kompute/CMakeLists.txt b/ggml/src/ggml-kompute/CMakeLists.txt
new file mode 100644
index 00000000000..c9109d5e8ee
--- /dev/null
+++ b/ggml/src/ggml-kompute/CMakeLists.txt
@@ -0,0 +1,166 @@
+
+find_package(Vulkan COMPONENTS glslc REQUIRED)
+find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
+
+if (NOT glslc_executable)
+    message(FATAL_ERROR "glslc not found")
+endif()
+
+ggml_add_backend_library(ggml-kompute
+                         ggml-kompute.cpp
+                         ../../include/ggml-kompute.h
+                        )
+
+target_link_libraries(ggml-kompute PRIVATE ggml-base kompute)
+target_include_directories(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
+
+add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
+
+function(compile_shader)
+    set(options)
+    set(oneValueArgs)
+    set(multiValueArgs SOURCES)
+    cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+    foreach(source ${compile_shader_SOURCES})
+        get_filename_component(filename ${source} NAME)
+        set(spv_file ${filename}.spv)
+        add_custom_command(
+            OUTPUT ${spv_file}
+            DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
+            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
+            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
+            ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
+            COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+            COMMENT "Compiling ${source} to ${spv_file}"
+            )
+
+        get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
+        set(FILE_NAME "shader${RAW_FILE_NAME}")
+        string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
+        string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
+        string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
+        set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
+        message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
+        if(CMAKE_GENERATOR MATCHES "Visual Studio")
+            add_custom_command(
+                OUTPUT ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                DEPENDS ${spv_file} xxd
+                COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
+                )
+        else()
+            add_custom_command(
+                OUTPUT ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                DEPENDS ${spv_file} xxd
+                COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
+                )
+        endif()
+    endforeach()
+endfunction()
+
+if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
+    message(STATUS "Kompute found")
+    set(KOMPUTE_OPT_LOG_LEVEL Error CACHE STRING "Kompute log level")
+    add_subdirectory(kompute)
+
+    # Compile our shaders
+    compile_shader(SOURCES
+        kompute-shaders/op_scale.comp
+        kompute-shaders/op_scale_8.comp
+        kompute-shaders/op_add.comp
+        kompute-shaders/op_addrow.comp
+        kompute-shaders/op_mul.comp
+        kompute-shaders/op_silu.comp
+        kompute-shaders/op_relu.comp
+        kompute-shaders/op_gelu.comp
+        kompute-shaders/op_softmax.comp
+        kompute-shaders/op_norm.comp
+        kompute-shaders/op_rmsnorm.comp
+        kompute-shaders/op_diagmask.comp
+        kompute-shaders/op_mul_mat_mat_f32.comp
+        kompute-shaders/op_mul_mat_f16.comp
+        kompute-shaders/op_mul_mat_q8_0.comp
+        kompute-shaders/op_mul_mat_q4_0.comp
+        kompute-shaders/op_mul_mat_q4_1.comp
+        kompute-shaders/op_mul_mat_q4_k.comp
+        kompute-shaders/op_mul_mat_q6_k.comp
+        kompute-shaders/op_getrows_f32.comp
+        kompute-shaders/op_getrows_f16.comp
+        kompute-shaders/op_getrows_q4_0.comp
+        kompute-shaders/op_getrows_q4_1.comp
+        kompute-shaders/op_getrows_q6_k.comp
+        kompute-shaders/op_rope_norm_f16.comp
+        kompute-shaders/op_rope_norm_f32.comp
+        kompute-shaders/op_rope_neox_f16.comp
+        kompute-shaders/op_rope_neox_f32.comp
+        kompute-shaders/op_cpy_f16_f16.comp
+        kompute-shaders/op_cpy_f16_f32.comp
+        kompute-shaders/op_cpy_f32_f16.comp
+        kompute-shaders/op_cpy_f32_f32.comp
+    )
+
+    # Create a custom target for our generated shaders
+    add_custom_target(generated_shaders DEPENDS
+        shaderop_scale.h
+        shaderop_scale_8.h
+        shaderop_add.h
+        shaderop_addrow.h
+        shaderop_mul.h
+        shaderop_silu.h
+        shaderop_relu.h
+        shaderop_gelu.h
+        shaderop_softmax.h
+        shaderop_norm.h
+        shaderop_rmsnorm.h
+        shaderop_diagmask.h
+        shaderop_mul_mat_mat_f32.h
+        shaderop_mul_mat_f16.h
+        shaderop_mul_mat_q8_0.h
+        shaderop_mul_mat_q4_0.h
+        shaderop_mul_mat_q4_1.h
+        shaderop_mul_mat_q4_k.h
+        shaderop_mul_mat_q6_k.h
+        shaderop_getrows_f32.h
+        shaderop_getrows_f16.h
+        shaderop_getrows_q4_0.h
+        shaderop_getrows_q4_1.h
+        shaderop_getrows_q6_k.h
+        shaderop_rope_norm_f16.h
+        shaderop_rope_norm_f32.h
+        shaderop_rope_neox_f16.h
+        shaderop_rope_neox_f32.h
+        shaderop_cpy_f16_f16.h
+        shaderop_cpy_f16_f32.h
+        shaderop_cpy_f32_f16.h
+        shaderop_cpy_f32_f32.h
+    )
+
+    # Create a custom command that depends on the generated_shaders
+    add_custom_command(
+        OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
+        COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
+        DEPENDS generated_shaders
+        COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
+    )
+
+    # Add the stamp to the main sources to ensure dependency tracking
+    target_sources(ggml-kompute PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+else()
+    message(WARNING "Kompute not found")
+endif()
diff --git a/ggml/src/ggml-kompute.cpp b/ggml/src/ggml-kompute/ggml-kompute.cpp
similarity index 92%
rename from ggml/src/ggml-kompute.cpp
rename to ggml/src/ggml-kompute/ggml-kompute.cpp
index 2fea9e4cc8d..50579227183 100644
--- a/ggml/src/ggml-kompute.cpp
+++ b/ggml/src/ggml-kompute/ggml-kompute.cpp
@@ -28,8 +28,10 @@
 #include "shaderop_getrows_q4_0.h"
 #include "shaderop_getrows_q4_1.h"
 #include "shaderop_getrows_q6_k.h"
-#include "shaderop_rope_f16.h"
-#include "shaderop_rope_f32.h"
+#include "shaderop_rope_norm_f16.h"
+#include "shaderop_rope_norm_f32.h"
+#include "shaderop_rope_neox_f16.h"
+#include "shaderop_rope_neox_f32.h"
 #include "shaderop_cpy_f16_f16.h"
 #include "shaderop_cpy_f16_f32.h"
 #include "shaderop_cpy_f32_f16.h"
@@ -345,7 +347,7 @@ void ggml_vk_allocate_descriptor_pool(struct ggml_kompute_context * ctx, size_t
     std::vector<vk::DescriptorPoolSize> descriptorPoolSizes = {
         vk::DescriptorPoolSize(
           vk::DescriptorType::eStorageBuffer,
-          3 * size // Descriptor count is number of possible tensors to pass into an algorithm
+          4 * size // Descriptor count is number of possible tensors to pass into an algorithm
           )
     };
 
@@ -788,7 +790,8 @@ static void ggml_vk_soft_max(
     const std::shared_ptr<kp::Tensor>& out,
     uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
     int32_t ne00, int32_t ne01, int32_t ne02, uint32_t ne03,
-    float scale
+    float scale, float max_bias, float m0, float m1,
+    uint32_t n_head_log2
 ) {
     const static auto spirv = getSpirvShader(kp::shader_data::op_softmax_comp_spv,
         kp::shader_data::op_softmax_comp_spv_len);
@@ -796,12 +799,14 @@ static void ggml_vk_soft_max(
     struct PushConstants {
         uint32_t inAOff, inBOff, outOff;
         int32_t ne00, ne01, ne02;
-        float scale;
+        float scale, max_bias, m0, m1;
+        uint32_t n_head_log2;
         int32_t mask;
     } pushConsts {
         safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
         ne00, ne01, ne02,
-        scale,
+        scale, max_bias, m0, m1,
+        n_head_log2,
         bool(inB)
     };
 
@@ -911,9 +916,9 @@ static void ggml_vk_mul_mat_f16(
     const std::shared_ptr<kp::Tensor>& out,
     uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
     int32_t ne00, int32_t ne01, int32_t ne02,
-    uint32_t nb00, uint32_t nb01, uint32_t nb02,
+    uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
     int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
-    uint32_t nb10, uint32_t nb11, uint32_t nb12,
+    uint32_t nb10, uint32_t nb11, uint32_t nb12, uint32_t nb13,
     int32_t ne0, int32_t ne1,
     uint32_t r2, uint32_t r3
 ) {
@@ -923,17 +928,17 @@ static void ggml_vk_mul_mat_f16(
     struct PushConstants {
         uint32_t inAOff, inBOff, outOff;
         int32_t ne00, ne01, ne02;
-        uint32_t nb00, nb01, nb02;
+        uint32_t nb00, nb01, nb02, nb03;
         int32_t ne10, ne11, ne12;
-        uint32_t nb10, nb11, nb12;
+        uint32_t nb10, nb11, nb12, nb13;
         int32_t ne0, ne1;
         uint32_t r2, r3;
     } pushConsts {
         safe_divide(inAOff, 2), safe_divide(inBOff, 4), safe_divide(outOff, 4),
         ne00, ne01, ne02,
-        nb00, nb01, nb02,
+        nb00, nb01, nb02, nb03,
         ne10, ne11, ne12,
-        nb10, nb11, nb12,
+        nb10, nb11, nb12, nb13,
         ne0, ne1,
         r2, r3
     };
@@ -1013,6 +1018,8 @@ static void ggml_vk_mul_mat_impl(
     int32_t ne00, int32_t ne01, int32_t ne02,
     int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
     int32_t ne0, int32_t ne1,
+    uint32_t nb01, uint32_t nb02, uint32_t nb03,
+    uint32_t nb11, uint32_t nb12, uint32_t nb13,
     uint32_t r2, uint32_t r3
 ) {
     struct PushConstants {
@@ -1020,19 +1027,23 @@ static void ggml_vk_mul_mat_impl(
         int32_t ne00, ne01, ne02;
         int32_t ne10, ne12;
         int32_t ne0, ne1;
+        uint32_t nb01, nb02, nb03;
+        uint32_t nb11, nb12, nb13;
         uint32_t r2, r3;
     } pushConsts {
         safe_divide(inAOff, block_size), safe_divide(inBOff, 4), safe_divide(outOff, 4),
         ne00, ne01, ne02,
         ne10, ne12,
         ne0, ne1,
+        nb01, nb02, nb03,
+        nb11, nb12, nb13,
         r2, r3
     };
 
     auto name = std::string(__func__) + "_" + suffix;
     std::shared_ptr<kp::Algorithm> s_algo = nullptr;
     if (!komputeManager()->hasAlgorithm(name)) {
-        const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
+        const uint32_t local_x = (ggml_vk_current_device().subgroupSize * 2) / 8;
         s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)}, {local_x}, {pushConsts});
     } else {
         s_algo = komputeManager()->getAlgorithm(name);
@@ -1074,19 +1085,26 @@ static void ggml_vk_mul_mat_q4_k(
     const std::shared_ptr<kp::Tensor>& inB,
     const std::shared_ptr<kp::Tensor>& out,
     uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne10,
-    int32_t ne11, int32_t ne12, int32_t ne13, int32_t ne0,
-    int32_t ne1, int32_t r2, int32_t r3
+    int32_t ne00, int32_t ne01, int32_t ne02,
+    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
+    int32_t ne0, int32_t ne1,
+    uint32_t nb01, uint32_t nb02, uint32_t nb03,
+    uint32_t nb11, uint32_t nb12, uint32_t nb13,
+    uint32_t r2, uint32_t r3
 ) {
     const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_k_comp_spv,
         kp::shader_data::op_mul_mat_q4_k_comp_spv_len);
 
     struct PushConstants {
         uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12, r2, r3;
+        int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12;
+        uint32_t nb01, nb02, nb03, nb11, nb12, nb13;
+        uint32_t r2, r3;
     } pushConsts {
-        0, 0, 0,
-        ne00, ne10, ne0, ne1, ne01, ne02, ne12, r2, r3
+        inAOff, safe_divide(inBOff, 4), safe_divide(outOff, 4),
+        ne00, ne10, ne0, ne1, ne01, ne02, ne12,
+        nb01, nb02, nb03, nb11, nb12, nb13,
+        r2, r3
     };
 
     std::shared_ptr<kp::Algorithm> s_algo = nullptr;
@@ -1108,28 +1126,37 @@ static void ggml_vk_mul_mat_q6_k(
     const std::shared_ptr<kp::Tensor>& inB,
     const std::shared_ptr<kp::Tensor>& out,
     uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
-    int32_t ne00, int32_t ne10, int32_t ne0, int32_t ne1,
-    int32_t ne01, int32_t ne11, int32_t ne12, int32_t ne02
+    int32_t ne00, int32_t ne01, int32_t ne02,
+    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
+    int32_t ne0, int32_t ne1,
+    uint32_t nb01, uint32_t nb02, uint32_t nb03,
+    uint32_t nb11, uint32_t nb12, uint32_t nb13,
+    uint32_t r2, uint32_t r3
 ) {
     const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q6_k_comp_spv,
         kp::shader_data::op_mul_mat_q6_k_comp_spv_len);
 
     struct PushConstants {
         uint32_t inAOff, inBOff, outOff;
-        int32_t ne00, ne10, ne0, ne1, ne01, gqa;
+        int32_t ne00, ne10, ne0, ne1, ne01, ne02, ne12;
+        uint32_t nb01, nb02, nb03, nb11, nb12, nb13;
+        uint32_t r2, r3;
     } pushConsts {
         inAOff, safe_divide(inBOff, 4), safe_divide(outOff, 4),
-        ne00, ne10, ne0, ne1, ne01, ne12/ne02
+        ne00, ne10, ne0, ne1, ne01, ne02, ne12,
+        nb01, nb02, nb03, nb11, nb12, nb13,
+        r2, r3
     };
 
     std::shared_ptr<kp::Algorithm> s_algo = nullptr;
     if (!komputeManager()->hasAlgorithm(__func__)) {
-        const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
-        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)}, {local_x}, {pushConsts});
+        const uint32_t local_x = 2;
+        const uint32_t local_y = ggml_vk_current_device().subgroupSize;
+        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)*unsigned(ne13)}, {local_x, local_y}, {pushConsts});
     } else {
         s_algo = komputeManager()->getAlgorithm(__func__);
         s_algo->setTensors({inA, inB, out});
-        s_algo->setWorkgroup({unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)});
+        s_algo->setWorkgroup({unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)*unsigned(ne13)});
         s_algo->setPushConstants<PushConstants>({pushConsts});
         s_algo->updateDescriptors(s_kompute_context->pool.get());
     }
@@ -1217,10 +1244,11 @@ static void ggml_vk_rope(
     kp::Sequence& seq,
     const std::shared_ptr<kp::Tensor>& inA,
     const std::shared_ptr<kp::Tensor>& inB,
+    const std::shared_ptr<kp::Tensor>& inC,
     const std::shared_ptr<kp::Tensor>& out,
-    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
+    uint32_t inAOff, uint32_t inBOff, uint32_t inCOff, uint32_t outOff,
     ggml_type src0t, int32_t n_dims, int32_t mode, int32_t n_ctx_orig,
-    float freq_base, float freq_scale, float ext_factor, float attn_factor, float beta_fast, float beta_slow,
+    float freq_base, float freq_scale, bool has_freq_factors, float ext_factor, float attn_factor, float beta_fast, float beta_slow,
     int32_t ne01, int32_t ne02, int32_t ne03,
     uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
     int32_t ne0,
@@ -1228,11 +1256,17 @@ static void ggml_vk_rope(
 ) {
     GGML_ASSERT(src0t == GGML_TYPE_F16 || src0t == GGML_TYPE_F32);
 
-    static const auto spirv_f16 = getSpirvShader(
-        kp::shader_data::op_rope_f16_comp_spv, kp::shader_data::op_rope_f16_comp_spv_len
+    static const auto spirv_norm_f16 = getSpirvShader(
+        kp::shader_data::op_rope_norm_f16_comp_spv, kp::shader_data::op_rope_norm_f16_comp_spv_len
+    );
+    static const auto spirv_norm_f32 = getSpirvShader(
+        kp::shader_data::op_rope_norm_f32_comp_spv, kp::shader_data::op_rope_norm_f32_comp_spv_len
     );
-    static const auto spirv_f32 = getSpirvShader(
-        kp::shader_data::op_rope_f32_comp_spv, kp::shader_data::op_rope_f32_comp_spv_len
+    static const auto spirv_neox_f16 = getSpirvShader(
+        kp::shader_data::op_rope_neox_f16_comp_spv, kp::shader_data::op_rope_neox_f16_comp_spv_len
+    );
+    static const auto spirv_neox_f32 = getSpirvShader(
+        kp::shader_data::op_rope_neox_f32_comp_spv, kp::shader_data::op_rope_neox_f32_comp_spv_len
     );
 
     int type_size = src0t == GGML_TYPE_F16 ? 2 : 4;
@@ -1247,32 +1281,40 @@ static void ggml_vk_rope(
     GGML_ASSERT(nb0  % type_size == 0);
 
     struct PushConstants {
-        uint32_t inAOff, inBOff, outOff;
+        uint32_t inAOff, inBOff, inCOff, outOff;
         int32_t n_dims, mode, n_ctx_orig;
-        float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+        float freq_base, freq_scale;
+        bool has_freq_factors;
+        float ext_factor, attn_factor, beta_fast, beta_slow;
         uint32_t nb00, nb01, nb02, nb03;
         int32_t ne0;
         uint32_t nb0, nb1, nb2, nb3;
     } pushConsts {
-        safe_divide(inAOff, type_size), safe_divide(inBOff, 4), safe_divide(outOff, type_size),
+        safe_divide(inAOff, type_size), safe_divide(inBOff, 4), safe_divide(inCOff, type_size), safe_divide(outOff, type_size),
         n_dims, mode, n_ctx_orig,
-        freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow,
+        freq_base, freq_scale,
+        has_freq_factors,
+        ext_factor, attn_factor, beta_fast, beta_slow,
         nb00, nb01, nb02, nb03,
         ne0,
         nb0, nb1, nb2, nb3
     };
 
-    auto name = std::string(__func__) + (src0t == GGML_TYPE_F16 ? "_f16" : "_f32");
+    auto & inC_ = inC ? inC : inA;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_f16 = src0t == GGML_TYPE_F16;
+
+    auto name = std::string(__func__) + (is_neox ? "_neox" : "_norm") + (src0t == GGML_TYPE_F16 ? "_f16" : "_f32");
     std::shared_ptr<kp::Algorithm> s_algo = nullptr;
     if (!komputeManager()->hasAlgorithm(name)) {
+        auto & spirv = is_neox ? is_f16 ? spirv_neox_f16 : spirv_neox_f32 : is_f16 ? spirv_norm_f16 : spirv_norm_f32;
         s_algo = komputeManager()->algorithm<float, PushConstants>(
-            name, s_kompute_context->pool.get(), {inA, inB, out},
-            src0t == GGML_TYPE_F16 ? spirv_f16 : spirv_f32,
+            name, s_kompute_context->pool.get(), {inA, inB, inC_, out}, spirv,
             {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts}
         );
     } else {
         s_algo = komputeManager()->getAlgorithm(name);
-        s_algo->setTensors({inA, inB, out});
+        s_algo->setTensors({inA, inB, inC_, out});
         s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
         s_algo->setPushConstants<PushConstants>({pushConsts});
         s_algo->updateDescriptors(s_kompute_context->pool.get());
@@ -1351,11 +1393,15 @@ static void ggml_vk_cpy_f16_f32(Args&&... args) {
 }
 
 static bool ggml_backend_kompute_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+    int64_t n = ggml_nelements(op);
     switch (op->op) {
         case GGML_OP_UNARY:
+            if (n % 4 != 0) return false;
             switch (ggml_get_unary_op(op)) {
-                case GGML_UNARY_OP_RELU:
                 case GGML_UNARY_OP_GELU:
+                    if (n % 8 != 0) return false;
+                    // fall through
+                case GGML_UNARY_OP_RELU:
                 case GGML_UNARY_OP_SILU:
                     return ggml_is_contiguous(op->src[0]);
                 default:
@@ -1373,8 +1419,18 @@ static bool ggml_backend_kompute_device_supports_op(ggml_backend_dev_t dev, cons
         case GGML_OP_SOFT_MAX:
         case GGML_OP_RMS_NORM:
         case GGML_OP_NORM:
-        case GGML_OP_ROPE:
             return true;
+        case GGML_OP_ROPE:
+            {
+                const int mode = ((const int32_t *) op->op_params)[2];
+                if (mode & GGML_ROPE_TYPE_MROPE) {
+                    return false;
+                }
+                if (mode & GGML_ROPE_TYPE_VISION) {
+                    return false;
+                }
+                return true;
+            }
         case GGML_OP_DUP:
         case GGML_OP_CPY:
         case GGML_OP_CONT:
@@ -1413,8 +1469,8 @@ static bool ggml_backend_kompute_device_supports_op(ggml_backend_dev_t dev, cons
 
             switch (op->src[0]->type) {
                 case GGML_TYPE_F32:
-                case GGML_TYPE_Q6_K:
                     return op->ne[3] == 1;
+                case GGML_TYPE_Q6_K:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_Q8_0:
                 case GGML_TYPE_Q4_0:
@@ -1515,9 +1571,11 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
             const static std::shared_ptr<kp::Tensor> nullTensor = nullptr;
             uint32_t off_src0 = 0;
             uint32_t off_src1 = 0;
+            uint32_t off_src2 = 0;
             uint32_t off_dst  = 0;
             const std::shared_ptr<kp::Tensor>& id_src0 = src0 ? ggml_vk_get_tensor(src0, &off_src0) : nullTensor;
             const std::shared_ptr<kp::Tensor>& id_src1 = src1 ? ggml_vk_get_tensor(src1, &off_src1) : nullTensor;
+            const std::shared_ptr<kp::Tensor>& id_src2 = src2 ? ggml_vk_get_tensor(src2, &off_src2) : nullTensor;
             const std::shared_ptr<kp::Tensor>& id_dst  = dst  ? ggml_vk_get_tensor(dst,  &off_dst)  : nullTensor;
 
             switch (dst->op) {
@@ -1593,11 +1651,16 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
 #pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/5021")
                         GGML_ASSERT(!src1 || src1t == GGML_TYPE_F32);
 
-#pragma message("TODO: add ALiBi support")
-#pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/7192")
-                        GGML_ASSERT(max_bias == 0.0f);
+                        const int64_t nrows_x = ggml_nrows(src0);
+                        const int64_t nrows_y = src0->ne[1];
+
+                        const uint32_t n_head      = nrows_x/nrows_y;
+                        const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
 
-                        ggml_vk_soft_max(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale);
+                        const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+                        const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+                        ggml_vk_soft_max(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale, max_bias, m0, m1, n_head_log2);
                     } break;
                 case GGML_OP_DIAG_MASK_INF:
                     {
@@ -1649,38 +1712,44 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
                             case GGML_TYPE_F16:
                                 ggml_vk_mul_mat_f16(
                                     seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, ne13, nb10, nb11, nb12,
+                                    ne00, ne01, ne02, nb00, nb01, nb02, nb03,
+                                    ne10, ne11, ne12, ne13, nb10, nb11, nb12, nb13,
                                     ne0, ne1, r2, r3
                                 );
                                 break;
                             case GGML_TYPE_Q8_0:
                                 ggml_vk_mul_mat_q8_0(
                                     seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
+                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
+                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
                                 );
                                 break;
                             case GGML_TYPE_Q4_0:
                                 ggml_vk_mul_mat_q4_0(
                                     seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
+                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
+                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
                                 );
                                 break;
                             case GGML_TYPE_Q4_1:
                                 ggml_vk_mul_mat_q4_1(
                                     seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
+                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
+                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
                                 );
                                 break;
                             case GGML_TYPE_Q4_K:
                                 ggml_vk_mul_mat_q4_k(
                                     seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, ne12/ne02, ne13/ne03
+                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
+                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
                                 );
                                 break;
                             case GGML_TYPE_Q6_K:
                                 ggml_vk_mul_mat_q6_k(
                                     seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
-                                    ne00, ne10, ne0, ne1, ne01, ne11, ne12, ne02
+                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1,
+                                    nb01, nb02, nb03, nb11, nb12, nb13, r2, r3
                                 );
                                 break;
                             default: {
@@ -1709,13 +1778,6 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
                     } break;
                 case GGML_OP_ROPE:
                     {
-#pragma message("TODO: implement phi3 frequency factors support")
-#pragma message("      https://github.com/ggerganov/llama.cpp/pull/7225")
-                        GGML_ASSERT(dst->src[2] == nullptr && "phi3 frequency factors not implemented yet");
-
-#pragma message("TODO: update rope NORM mode to match NEOX mode")
-#pragma message("      https://github.com/ggerganov/llama.cpp/pull/7634")
-
                         GGML_ASSERT(ne10 == ne02);
                         GGML_ASSERT(src0t == dstt);
                         // const int n_past = ((int32_t *) dst->op_params)[0];
@@ -1724,6 +1786,8 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
                         // skip 3, n_ctx used in GLM RoPE, unimplemented in Vulkan
                         const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
 
+                        const bool has_freq_factors = dst->src[2] != nullptr;
+
                         float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
                         memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
                         memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
@@ -1732,8 +1796,8 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
                         memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
                         memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
                         ggml_vk_rope(
-                            seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, src0t, n_dims, mode, n_ctx_orig,
-                            freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow,
+                            seq, id_src0, id_src1, id_src2, id_dst, off_src0, off_src1, off_src2, off_dst, src0t, n_dims, mode, n_ctx_orig,
+                            freq_base, freq_scale, has_freq_factors, ext_factor, attn_factor, beta_fast, beta_slow,
                             ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, nb0, nb1, nb2, nb3
                         );
                     } break;
@@ -2176,9 +2240,12 @@ static const struct ggml_backend_reg_i ggml_backend_kompute_reg_i = {
 
 ggml_backend_reg_t ggml_backend_kompute_reg() {
     static ggml_backend_reg reg = {
-        /* .iface   = */ ggml_backend_kompute_reg_i,
-        /* .context = */ nullptr,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_kompute_reg_i,
+        /* .context     = */ nullptr,
     };
 
     return &reg;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_kompute_reg)
diff --git a/ggml/src/kompute-shaders/common.comp b/ggml/src/ggml-kompute/kompute-shaders/common.comp
similarity index 93%
rename from ggml/src/kompute-shaders/common.comp
rename to ggml/src/ggml-kompute/kompute-shaders/common.comp
index 62d62b025ec..dbe4cf804e6 100644
--- a/ggml/src/kompute-shaders/common.comp
+++ b/ggml/src/ggml-kompute/kompute-shaders/common.comp
@@ -3,6 +3,7 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #extension GL_EXT_shader_explicit_arithmetic_types_int8: require
 #extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int64: require
 #extension GL_EXT_control_flow_attributes: enable
 #extension GL_KHR_shader_subgroup_arithmetic : require
 #extension GL_EXT_debug_printf : enable
@@ -15,6 +16,7 @@
 #define TWOPI_F 6.283185307179586f
 
 #define QK_K 256
+#define K_SCALE_SIZE 12
 
 #define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
 #define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
@@ -64,6 +66,14 @@ mat4 dequantize_q4_1(const block_q4_1 xb, uint il) {
     return reg;
 }
 
+#define sizeof_block_q4_k 144
+struct block_q4_k {
+    float16_t d;
+    float16_t dmin;
+    uint8_t scales[K_SCALE_SIZE];
+    uint8_t qs[QK_K/2];
+};
+
 #define sizeof_block_q6_k 210
 struct block_q6_k {
     uint8_t ql[QK_K/2];      // quants, lower 4 bits
diff --git a/ggml/src/kompute-shaders/op_add.comp b/ggml/src/ggml-kompute/kompute-shaders/op_add.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_add.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_add.comp
diff --git a/ggml/src/kompute-shaders/op_addrow.comp b/ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_addrow.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_addrow.comp
diff --git a/ggml/src/kompute-shaders/op_cpy_f16_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_cpy_f16_f16.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f16.comp
diff --git a/ggml/src/kompute-shaders/op_cpy_f16_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_cpy_f16_f32.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_cpy_f16_f32.comp
diff --git a/ggml/src/kompute-shaders/op_cpy_f32_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_cpy_f32_f16.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f16.comp
diff --git a/ggml/src/kompute-shaders/op_cpy_f32_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_cpy_f32_f32.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_cpy_f32_f32.comp
diff --git a/ggml/src/kompute-shaders/op_diagmask.comp b/ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_diagmask.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_diagmask.comp
diff --git a/ggml/src/kompute-shaders/op_gelu.comp b/ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_gelu.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_gelu.comp
diff --git a/ggml/src/kompute-shaders/op_getrows.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_getrows.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_getrows.comp
diff --git a/ggml/src/kompute-shaders/op_getrows_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_getrows_f16.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_getrows_f16.comp
diff --git a/ggml/src/kompute-shaders/op_getrows_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_getrows_f32.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_getrows_f32.comp
diff --git a/ggml/src/kompute-shaders/op_getrows_q4_0.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_getrows_q4_0.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_0.comp
diff --git a/ggml/src/kompute-shaders/op_getrows_q4_1.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_getrows_q4_1.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_getrows_q4_1.comp
diff --git a/ggml/src/kompute-shaders/op_getrows_q6_k.comp b/ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_getrows_q6_k.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_getrows_q6_k.comp
diff --git a/ggml/src/kompute-shaders/op_mul.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_mul.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_mul.comp
diff --git a/ggml/src/kompute-shaders/op_mul_mat_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp
similarity index 91%
rename from ggml/src/kompute-shaders/op_mul_mat_f16.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp
index 8f0a9031f7a..0ab1b2fc20e 100644
--- a/ggml/src/kompute-shaders/op_mul_mat_f16.comp
+++ b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_f16.comp
@@ -20,12 +20,14 @@ layout (push_constant) uniform parameter {
     uint nb00;
     uint nb01;
     uint nb02;
+    uint nb03;
     int ne10;
     int ne11;
     int ne12;
     uint nb10;
     uint nb11;
     uint nb12;
+    uint nb13;
     int ne0;
     int ne1;
     uint r2;
@@ -42,7 +44,7 @@ void main() {
     const uint i12 = im%pcs.ne12;
     const uint i13 = im/pcs.ne12;
 
-    const uint offset0 = r0*pcs.nb01 + (i12/pcs.r2)*pcs.nb02 + (i13/pcs.r3)*pcs.nb02*pcs.ne02;
+    const uint offset0 = r0*pcs.nb01 + (i12/pcs.r2)*pcs.nb02 + (i13/pcs.r3)*pcs.nb03;
 
     const uint x = offset0 / 2 + pcs.inAOff; // Based from inA
 
@@ -52,7 +54,7 @@ void main() {
             break;
         }
 
-        const uint y = (r1*pcs.nb11 + im*pcs.nb12) / 4 + pcs.inBOff; // Based from inB
+        const uint y = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
 
         float sumf = 0;
         for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
diff --git a/ggml/src/kompute-shaders/op_mul_mat_mat_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_mul_mat_mat_f32.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_mat_f32.comp
diff --git a/ggml/src/kompute-shaders/op_mul_mat_q4_0.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_mul_mat_q4_0.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_0.comp
diff --git a/ggml/src/kompute-shaders/op_mul_mat_q4_1.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_mul_mat_q4_1.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_1.comp
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp
new file mode 100644
index 00000000000..a5752a3a006
--- /dev/null
+++ b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q4_k.comp
@@ -0,0 +1,140 @@
+#version 450
+
+#include "common.comp"
+
+#define N_DST 4
+#define SIZE_OF_BLOCK sizeof_block_q4_k
+
+layout(local_size_x = 4) in;
+layout(local_size_y = 8) in;
+layout(local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { block_q4_k inA[]; };
+layout (binding = 1) readonly buffer tensorInB { float inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int ne10;
+    int ne0;
+    int ne1;
+    int ne01;
+    int ne02;
+    int ne12;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    uint nb11;
+    uint nb12;
+    uint nb13;
+    uint r2;
+    uint r3;
+} pcs;
+
+void main() {
+    const uint16_t kmask1 = uint16_t(0x3f3f);
+    const uint16_t kmask2 = uint16_t(0x0f0f);
+    const uint16_t kmask3 = uint16_t(0xc0c0);
+
+    const uint ix = gl_SubgroupInvocationID/8;  // 0...3
+    const uint it = gl_SubgroupInvocationID%8;  // 0...7
+    const uint iq = it/4;     // 0 or 1
+    const uint ir = it%4;     // 0...3
+
+    const uint nb = pcs.ne00/QK_K;
+
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+    const uint im = gl_WorkGroupID.z;
+
+    const uint first_row = r0 * N_DST;
+    const uint ib_row = first_row * nb;
+
+    const uint i12 = im%pcs.ne12;
+    const uint i13 = im/pcs.ne12;
+
+    const uint offset0 = first_row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
+    const uint offset1 =        r1*pcs.nb11 + (i12       )*pcs.nb12 + (i13       )*pcs.nb13;
+
+    const uint xblk = offset0 + pcs.inAOff;
+    const uint y = (offset1 / 4) + pcs.inBOff;
+
+    float yl[16];
+    float yh[16];
+    float sumf[N_DST] = {0.f, 0.f, 0.f, 0.f};
+    float all_sum = 0.f;
+
+    uint y4 = y + ix * QK_K + 64 * iq + 8 * ir;
+
+    for (uint ib = ix; ib < nb; ib += 4) {
+        const uint blk_idx = ib + xblk;
+
+        float sumy[4] = {0.f, 0.f, 0.f, 0.f};
+        for (int i = 0; i < 8; ++i) {
+            yl[i+0] = inB[y4+i+  0]; sumy[0] += yl[i+0];
+            yl[i+8] = inB[y4+i+ 32]; sumy[1] += yl[i+8];
+            yh[i+0] = inB[y4+i+128]; sumy[2] += yh[i+0];
+            yh[i+8] = inB[y4+i+160]; sumy[3] += yh[i+8];
+        }
+
+        for (int row = 0; row < N_DST; row++) {
+            uint row_idx = row * (pcs.nb01 / SIZE_OF_BLOCK);
+
+            uint16_t sc_0 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 0);
+            uint16_t sc_1 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 2);
+            uint16_t sc_2 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 4);
+            uint16_t sc_3 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 6);
+            uint16_t sc_4 = u8BufToU16(inA[blk_idx + row_idx].scales, iq * 2 + 8);
+
+            uint16_t sc16[4];
+            sc16[0] = sc_0 & kmask1;
+            sc16[1] = sc_2 & kmask1;
+            sc16[2] = ((sc_4 >> 0) & kmask2) | ((sc_0 & kmask3) >> 2);
+            sc16[3] = ((sc_4 >> 4) & kmask2) | ((sc_2 & kmask3) >> 2);
+
+            float acc1[4] = {0.f, 0.f, 0.f, 0.f};
+            float acc2[4] = {0.f, 0.f, 0.f, 0.f};
+            for (int i = 0; i < 8; i += 2) {
+                uint16_t q1 = u8BufToU16(inA[blk_idx + row_idx].qs, 32 * iq + 8 * ir + i);
+                uint16_t q2 = u8BufToU16(inA[blk_idx + row_idx].qs, 64 + 32 * iq + 8 * ir + i);
+                acc1[0] += yl[i+0] * (q1 & 0x000F);
+                acc1[1] += yl[i+1] * (q1 & 0x0F00);
+                acc1[2] += yl[i+8] * (q1 & 0x00F0);
+                acc1[3] += yl[i+9] * (q1 & 0xF000);
+                acc2[0] += yh[i+0] * (q2 & 0x000F);
+                acc2[1] += yh[i+1] * (q2 & 0x0F00);
+                acc2[2] += yh[i+8] * (q2 & 0x00F0);
+                acc2[3] += yh[i+9] * (q2 & 0xF000);
+            }
+
+            uint8_t sc8_0 = uint8_t(sc16[0] & 0xFF);
+            uint8_t sc8_1 = uint8_t(sc16[0] >> 8 );
+            uint8_t sc8_2 = uint8_t(sc16[1] & 0xFF);
+            uint8_t sc8_3 = uint8_t(sc16[1] >> 8 );
+            uint8_t sc8_4 = uint8_t(sc16[2] & 0xFF);
+            uint8_t sc8_5 = uint8_t(sc16[2] >> 8 );
+            uint8_t sc8_6 = uint8_t(sc16[3] & 0xFF);
+            uint8_t sc8_7 = uint8_t(sc16[3] >> 8 );
+
+            float dall = float(inA[blk_idx + row_idx].d);
+            float dmin = float(inA[blk_idx + row_idx].dmin);
+            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8_0 +
+                               (acc1[2] + 1.f/256.f * acc1[3]) * sc8_1 * 1.f/16.f +
+                               (acc2[0] + 1.f/256.f * acc2[1]) * sc8_4 +
+                               (acc2[2] + 1.f/256.f * acc2[3]) * sc8_5 * 1.f/16.f) -
+                dmin * (sumy[0] * sc8_2 + sumy[1] * sc8_3 + sumy[2] * sc8_6 + sumy[3] * sc8_7);
+        }
+
+        y4 += 4 * QK_K;
+    }
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = subgroupAdd(sumf[row]);
+        if (subgroupElect()) {
+            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = all_sum;
+        }
+    }
+}
diff --git a/ggml/src/kompute-shaders/op_mul_mat_q6_k.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp
similarity index 86%
rename from ggml/src/kompute-shaders/op_mul_mat_q6_k.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp
index c9baebdf4ba..d331d1a7057 100644
--- a/ggml/src/kompute-shaders/op_mul_mat_q6_k.comp
+++ b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q6_k.comp
@@ -21,7 +21,16 @@ layout (push_constant) uniform parameter {
     int ne0;
     int ne1;
     int ne01;
-    int gqa;
+    int ne02;
+    int ne12;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    uint nb11;
+    uint nb12;
+    uint nb13;
+    uint r2;
+    uint r3;
 } pcs;
 
 void main() {
@@ -34,12 +43,15 @@ void main() {
 
     const uint r0 = gl_WorkGroupID.x;
     const uint r1 = gl_WorkGroupID.y;
-    const uint r2 = gl_WorkGroupID.z;
+    const uint im = gl_WorkGroupID.z;
 
     const uint row = (r0 * gl_NumSubgroups + gl_SubgroupID);
-    const uint offset0 = r2/pcs.gqa*(nb*pcs.ne0);
-    const uint x = row * nb + offset0; // Based from inA without base offset
-    const uint yy = r1*pcs.ne10 + r2*pcs.ne00*pcs.ne1+pcs.inBOff; // Based from inB
+
+    const uint i12 = im%pcs.ne12;
+    const uint i13 = im/pcs.ne12;
+
+    const uint x = row*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
+    const uint yy = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
 
     float sumf = 0;
 
@@ -89,6 +101,6 @@ void main() {
 
     const float tot = subgroupAdd(sumf);
     if (subgroupElect()) {
-        out_[r1*pcs.ne0 + r2*pcs.ne0*pcs.ne1 + row + pcs.outOff] = tot;
+        out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + row + pcs.outOff] = tot;
     }
 }
diff --git a/ggml/src/kompute-shaders/op_mul_mat_q8_0.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_mul_mat_q8_0.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_mul_mat_q8_0.comp
diff --git a/ggml/src/kompute-shaders/op_mul_mv_q_n.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp
similarity index 76%
rename from ggml/src/kompute-shaders/op_mul_mv_q_n.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp
index 440b5ab2c81..a6517cc1f19 100644
--- a/ggml/src/kompute-shaders/op_mul_mv_q_n.comp
+++ b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n.comp
@@ -14,10 +14,15 @@ void main() {
     const uint i12 = im%pcs.ne12;
     const uint i13 = im/pcs.ne12;
 
-    const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
+    // pointers to src0 rows
+    uint ax[N_ROWS];
+    for (int row = 0; row < N_ROWS; ++row) {
+        const uint offset0 = (first_row + row)*(pcs.nb01/SIZE_OF_BLOCK) + (i12/pcs.r2)*(pcs.nb02/SIZE_OF_BLOCK) + (i13/pcs.r3)*(pcs.nb03/SIZE_OF_BLOCK);
+
+        ax[row] = offset0 + pcs.inAOff;
+    }
 
-    const uint x = offset0; // Based from inA without base offset
-    const uint y = r1*uint(pcs.ne10)+im*pcs.ne00*pcs.ne1+pcs.inBOff; // Based from inB
+    const uint y = (r1*pcs.nb11 + i12*pcs.nb12 + i13*pcs.nb13) / 4 + pcs.inBOff;
 
     float sumf[N_ROWS] = {0.0f, 0.0f, 0.0f, 0.0f};
 
@@ -32,8 +37,7 @@ void main() {
 
     for (uint ib = ix; ib < nb; ib += 16) {
         for (int row = 0; row < N_ROWS; row++) {
-            const uint block_index = x + ib + row * nb;
-            sumf[row] += block_q_n_dot_y(block_index, yb, il);
+            sumf[row] += block_q_n_dot_y(ax[row] + ib, yb, il);
         }
 
         yb += BLOCKS_IN_QUANT * 16;
diff --git a/ggml/src/kompute-shaders/op_mul_mv_q_n_pre.comp b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp
similarity index 80%
rename from ggml/src/kompute-shaders/op_mul_mv_q_n_pre.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp
index 7912b09ac69..a9a2f22180f 100644
--- a/ggml/src/kompute-shaders/op_mul_mv_q_n_pre.comp
+++ b/ggml/src/ggml-kompute/kompute-shaders/op_mul_mv_q_n_pre.comp
@@ -1,5 +1,5 @@
 layout(local_size_x_id = 0) in;
-layout(local_size_y = 1) in;
+layout(local_size_y = 8) in;
 layout(local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
@@ -17,6 +17,12 @@ layout (push_constant) uniform parameter {
     int  ne12;
     int  ne0;
     int  ne1;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    uint nb11;
+    uint nb12;
+    uint nb13;
     uint r2;
     uint r3;
 } pcs;
diff --git a/ggml/src/kompute-shaders/op_norm.comp b/ggml/src/ggml-kompute/kompute-shaders/op_norm.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_norm.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_norm.comp
diff --git a/ggml/src/kompute-shaders/op_relu.comp b/ggml/src/ggml-kompute/kompute-shaders/op_relu.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_relu.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_relu.comp
diff --git a/ggml/src/kompute-shaders/op_rmsnorm.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_rmsnorm.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_rmsnorm.comp
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp
new file mode 100644
index 00000000000..63659cbfe55
--- /dev/null
+++ b/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f16.comp
@@ -0,0 +1,52 @@
+#version 450
+
+#include "rope_common.comp"
+
+layout(binding = 0) buffer restrict readonly  tensorInA { float16_t inA[]; };
+layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
+layout(binding = 2) buffer restrict readonly  tensorInC { float     inC[]; };
+layout(binding = 3) buffer restrict writeonly tensorOut { float16_t out_[]; };
+
+void main() {
+    const uint i3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.y;
+    const uint i1 = gl_WorkGroupID.x;
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
+
+    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
+
+    float theta_base = float(inB[pcs.inBOff + i2]);
+    float inv_ndims = -1.f/pcs.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
+        if (i0 < pcs.n_dims) {
+            uint ic = i0/2;
+
+            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
+
+            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + ic*pcs.nb00) / 2) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + ic*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
+
+            const float x0 = float(inA[src]);
+            const float x1 = float(inA[src+pcs.n_dims/2]);
+
+            out_[dst_data]              = float16_t(x0*cos_theta - x1*sin_theta);
+            out_[dst_data+pcs.n_dims/2] = float16_t(x0*sin_theta + x1*cos_theta);
+        } else {
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
+
+            out_[dst_data]   = inA[src];
+            out_[dst_data+1] = inA[src+1];
+        }
+    }
+}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp
new file mode 100644
index 00000000000..4df56204d72
--- /dev/null
+++ b/ggml/src/ggml-kompute/kompute-shaders/op_rope_neox_f32.comp
@@ -0,0 +1,52 @@
+#version 450
+
+#include "rope_common.comp"
+
+layout(binding = 0) buffer restrict readonly  tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
+layout(binding = 2) buffer restrict readonly  tensorInC { float inC[]; };
+layout(binding = 3) buffer restrict writeonly tensorOut { float out_[]; };
+
+void main() {
+    const uint i3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.y;
+    const uint i1 = gl_WorkGroupID.x;
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
+
+    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
+
+    float theta_base = float(inB[pcs.inBOff + i2]);
+    float inv_ndims = -1.f/pcs.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
+        if (i0 < pcs.n_dims) {
+            uint ic = i0/2;
+
+            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
+
+            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + ic*pcs.nb00) / 4) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + ic*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
+
+            const float x0 = inA[src];
+            const float x1 = inA[src+pcs.n_dims/2];
+
+            out_[dst_data]              = x0*cos_theta - x1*sin_theta;
+            out_[dst_data+pcs.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
+
+            out_[dst_data]   = inA[src];
+            out_[dst_data+1] = inA[src+1];
+        }
+    }
+}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp
new file mode 100644
index 00000000000..a3c0eda8bd3
--- /dev/null
+++ b/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f16.comp
@@ -0,0 +1,52 @@
+#version 450
+
+#include "rope_common.comp"
+
+layout(binding = 0) buffer restrict readonly  tensorInA { float16_t inA[]; };
+layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
+layout(binding = 2) buffer restrict readonly  tensorInC { float     inC[]; };
+layout(binding = 3) buffer restrict writeonly tensorOut { float16_t out_[]; };
+
+void main() {
+    const uint i3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.y;
+    const uint i1 = gl_WorkGroupID.x;
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
+
+    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
+
+    float theta_base = float(inB[pcs.inBOff + i2]);
+    float inv_ndims = -1.f/pcs.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
+        if (i0 < pcs.n_dims) {
+            uint ic = i0/2;
+
+            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
+
+            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
+
+            const float x0 = float(inA[src]);
+            const float x1 = float(inA[src+1]);
+
+            out_[dst_data]   = float16_t(x0*cos_theta - x1*sin_theta);
+            out_[dst_data+1] = float16_t(x0*sin_theta + x1*cos_theta);
+        } else {
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
+
+            out_[dst_data]   = inA[src];
+            out_[dst_data+1] = inA[src+1];
+        }
+    }
+}
diff --git a/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp b/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp
new file mode 100644
index 00000000000..b7963ae7253
--- /dev/null
+++ b/ggml/src/ggml-kompute/kompute-shaders/op_rope_norm_f32.comp
@@ -0,0 +1,52 @@
+#version 450
+
+#include "rope_common.comp"
+
+layout(binding = 0) buffer restrict readonly  tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly  tensorInB { int   inB[]; };
+layout(binding = 2) buffer restrict readonly  tensorInC { float inC[]; };
+layout(binding = 3) buffer restrict writeonly tensorOut { float out_[]; };
+
+void main() {
+    const uint i3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.y;
+    const uint i1 = gl_WorkGroupID.x;
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
+
+    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
+
+    float theta_base = float(inB[pcs.inBOff + i2]);
+    float inv_ndims = -1.f/pcs.n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (uint i0 = 2*gl_LocalInvocationIndex; i0 < pcs.ne0; i0 += 2*gl_WorkGroupSize.x) {
+        if (i0 < pcs.n_dims) {
+            uint ic = i0/2;
+
+            float theta = theta_base * pow(pcs.freq_base, inv_ndims*i0);
+
+            const float freq_factor = pcs.has_freq_factors ? inC[pcs.inCOff + ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
+
+            const float x0 = inA[src];
+            const float x1 = inA[src+1];
+
+            out_[dst_data]   = x0*cos_theta - x1*sin_theta;
+            out_[dst_data+1] = x0*sin_theta + x1*cos_theta;
+        } else {
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
+
+            out_[dst_data]   = inA[src];
+            out_[dst_data+1] = inA[src+1];
+        }
+    }
+}
diff --git a/ggml/src/kompute-shaders/op_scale.comp b/ggml/src/ggml-kompute/kompute-shaders/op_scale.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_scale.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_scale.comp
diff --git a/ggml/src/kompute-shaders/op_scale_8.comp b/ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_scale_8.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_scale_8.comp
diff --git a/ggml/src/kompute-shaders/op_silu.comp b/ggml/src/ggml-kompute/kompute-shaders/op_silu.comp
similarity index 100%
rename from ggml/src/kompute-shaders/op_silu.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_silu.comp
diff --git a/ggml/src/kompute-shaders/op_softmax.comp b/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp
similarity index 78%
rename from ggml/src/kompute-shaders/op_softmax.comp
rename to ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp
index 7bc9176caba..4165295bf4b 100644
--- a/ggml/src/kompute-shaders/op_softmax.comp
+++ b/ggml/src/ggml-kompute/kompute-shaders/op_softmax.comp
@@ -18,6 +18,10 @@ layout(push_constant) uniform PushConstants {
     int ne01;
     int ne02;
     float scale;
+    float max_bias;
+    float m0;
+    float m1;
+    uint n_head_log2;
     int mask;
 } pcs;
 
@@ -34,17 +38,29 @@ void main() {
     const uint pmask = i01*pcs.ne00 + pcs.inBOff; // Based from inB
     const uint pdst = extra_off + pcs.outOff; // Based from out_
 
+    float slope = 1.0f;
+
+    // ALiBi
+    if (pcs.max_bias > 0.0f) {
+        int64_t h = i02;
+
+        float base = h < pcs.n_head_log2 ? pcs.m0 : pcs.m1;
+        int64_t exp = h < pcs.n_head_log2 ? h + 1 : 2*(h - pcs.n_head_log2) + 1;
+
+        slope = pow(base, float(exp));
+    }
+
     // parallel max
     float localMax = uintBitsToFloat(0xFF800000);
     for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
-        localMax = max(localMax, inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? inB[pmask + i00] : 0.0f));
+        localMax = max(localMax, inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? slope*inB[pmask + i00] : 0.0f));
     }
     float max_ = subgroupMax(localMax);
 
     // parallel sum
     float localSum = 0.0f;
     for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
-        const float exp_psrc0 = exp(inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? inB[pmask + i00] : 0.0f) - max_);
+        const float exp_psrc0 = exp(inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? slope*inB[pmask + i00] : 0.0f) - max_);
         localSum += exp_psrc0;
         out_[pdst + i00] = exp_psrc0;
     }
diff --git a/ggml/src/kompute-shaders/rope_common.comp b/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp
similarity index 96%
rename from ggml/src/kompute-shaders/rope_common.comp
rename to ggml/src/ggml-kompute/kompute-shaders/rope_common.comp
index 7b9394cb2ff..0fca640dcc2 100644
--- a/ggml/src/kompute-shaders/rope_common.comp
+++ b/ggml/src/ggml-kompute/kompute-shaders/rope_common.comp
@@ -1,17 +1,21 @@
 #include "common.comp"
 
+#define GGML_ROPE_TYPE_NEOX 2
+
 // TODO: use a local size of 32 or more (Metal uses 1024)
 layout(local_size_x = 1) in;
 
 layout (push_constant) uniform parameter {
     uint inAOff;
     uint inBOff;
+    uint inCOff;
     uint outOff;
     int n_dims;
     int mode;
     int n_ctx_orig;
     float freq_base;
     float freq_scale;
+    bool has_freq_factors;
     float ext_factor;
     float attn_factor;
     float beta_fast;
diff --git a/ggml/src/ggml-metal/CMakeLists.txt b/ggml/src/ggml-metal/CMakeLists.txt
new file mode 100644
index 00000000000..1bad2720682
--- /dev/null
+++ b/ggml/src/ggml-metal/CMakeLists.txt
@@ -0,0 +1,105 @@
+find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
+find_library(METAL_FRAMEWORK    Metal      REQUIRED)
+find_library(METALKIT_FRAMEWORK MetalKit   REQUIRED)
+
+message(STATUS "Metal framework found")
+
+ggml_add_backend_library(ggml-metal
+                         ggml-metal.m
+                        )
+
+target_link_libraries(ggml-metal PRIVATE
+                      ${FOUNDATION_LIBRARY}
+                      ${METAL_FRAMEWORK}
+                      ${METALKIT_FRAMEWORK}
+                      )
+
+if (GGML_METAL_NDEBUG)
+    add_compile_definitions(GGML_METAL_NDEBUG)
+endif()
+
+if (GGML_METAL_USE_BF16)
+    add_compile_definitions(GGML_METAL_USE_BF16)
+endif()
+
+# copy metal files to bin directory
+configure_file(../ggml-common.h  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h     COPYONLY)
+configure_file(ggml-metal.metal  ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal  COPYONLY)
+configure_file(ggml-metal-impl.h ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h COPYONLY)
+
+if (GGML_METAL_EMBED_LIBRARY)
+    enable_language(ASM)
+
+    add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
+
+    set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/../ggml-common.h")
+    set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+    set(METALLIB_IMPL   "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal-impl.h")
+
+    file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
+
+    # merge ggml-common.h and ggml-metal.metal into a single file
+    set(METALLIB_EMBED_ASM        "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
+    set(METALLIB_SOURCE_EMBED     "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
+    set(METALLIB_SOURCE_EMBED_TMP "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal.tmp")
+
+    add_custom_command(
+        OUTPUT ${METALLIB_EMBED_ASM}
+        COMMAND echo "Embedding Metal library"
+        COMMAND sed -e '/__embed_ggml-common.h__/r         ${METALLIB_COMMON}' -e '/__embed_ggml-common.h__/d'         < ${METALLIB_SOURCE}           > ${METALLIB_SOURCE_EMBED_TMP}
+        COMMAND sed -e '/\#include \"ggml-metal-impl.h\"/r ${METALLIB_IMPL}'   -e '/\#include \"ggml-metal-impl.h\"/d' < ${METALLIB_SOURCE_EMBED_TMP} > ${METALLIB_SOURCE_EMBED}
+        COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
+        COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
+        COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
+        COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
+        COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
+        COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
+        DEPENDS ../ggml-common.h ggml-metal.metal ggml-metal-impl.h
+        COMMENT "Generate assembly for embedded Metal library"
+    )
+
+    target_sources(ggml-metal PRIVATE ${METALLIB_EMBED_ASM})
+else()
+    if (GGML_METAL_SHADER_DEBUG)
+        # custom command to do the following:
+        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
+        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
+        #
+        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
+        #       disabling fast math is needed in order to pass tests/test-backend-ops
+        # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
+        # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
+        #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
+        set(XC_FLAGS -fno-fast-math -fno-inline -g)
+    else()
+        set(XC_FLAGS -O3)
+    endif()
+
+    # Append macOS metal versioning flags
+    if (GGML_METAL_MACOSX_VERSION_MIN)
+        message(STATUS "Adding  -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
+        list   (APPEND XC_FLAGS -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN})
+    endif()
+
+    if (GGML_METAL_STD)
+        message(STATUS "Adding  -std=${GGML_METAL_STD} flag to metal compilation")
+        list   (APPEND XC_FLAGS -std=${GGML_METAL_STD})
+    endif()
+
+    add_custom_command(
+        OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
+        COMMAND xcrun -sdk macosx metallib                ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air   -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
+        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
+        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
+        DEPENDS ggml-metal.metal ggml-common.h
+        COMMENT "Compiling Metal kernels"
+        )
+
+    # FIXME: only add to the ggml-metal target?
+    add_custom_target(
+        ggml-metal-lib ALL
+        DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        )
+endif() # GGML_METAL_EMBED_LIBRARY
diff --git a/ggml/src/ggml-metal/ggml-metal-impl.h b/ggml/src/ggml-metal/ggml-metal-impl.h
new file mode 100644
index 00000000000..e3dc25f1686
--- /dev/null
+++ b/ggml/src/ggml-metal/ggml-metal-impl.h
@@ -0,0 +1,288 @@
+#ifndef GGML_METAL_IMPL
+#define GGML_METAL_IMPL
+
+// kernel argument structs
+//
+// - element counters (e.g. ne00) typically use int32_t to reduce register usage
+//   however, be careful from int overflows when using those in the kernel implementation
+//
+// - strides (e.g. nb00) use uint64_t
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    int32_t  dim;
+} ggml_metal_kargs_concat;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    uint64_t offs;
+} ggml_metal_kargs_bin;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_repeat;
+
+typedef struct {
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int64_t  ne0;
+    int64_t  ne1;
+    int64_t  ne2;
+    int64_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_cpy;
+
+typedef struct {
+    int64_t  ne10;
+    int64_t  ne11;
+    int64_t  ne12;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    uint64_t offs;
+    bool     inplace;
+} ggml_metal_kargs_set;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    int32_t  n_past;
+    int32_t  n_dims;
+    int32_t  n_ctx_orig;
+    float    freq_base;
+    float    freq_scale;
+    float    ext_factor;
+    float    attn_factor;
+    float    beta_fast;
+    float    beta_slow;
+} ggml_metal_kargs_rope;
+
+typedef struct {
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne11;
+    int32_t  ne_12_2; // assume K and V are same shape
+    int32_t  ne_12_3;
+    uint64_t nb_12_1;
+    uint64_t nb_12_2;
+    uint64_t nb_12_3;
+    uint64_t nb31;
+    int32_t  ne1;
+    int32_t  ne2;
+    float    scale;
+    float    max_bias;
+    float    m0;
+    float    m1;
+    uint16_t n_head_log2;
+    float    logit_softcap;
+} ggml_metal_kargs_flash_attn_ext;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne02;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne12;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne0;
+    int32_t  ne1;
+    int16_t  r2;
+    int16_t  r3;
+} ggml_metal_kargs_mul_mm;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne0;
+    int32_t  ne1;
+    int16_t  r2;
+    int16_t  r3;
+} ggml_metal_kargs_mul_mv;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne0;
+    int32_t  ne1;
+    int16_t  r2;
+    int16_t  r3;
+    int16_t  nsg;
+    int16_t  nxpsg;
+    int16_t  r1ptg;
+} ggml_metal_kargs_mul_mv_ext;
+
+typedef struct {
+    int32_t  nei0;
+    int32_t  nei1;
+    uint64_t nbi1;
+    int32_t  ne00;
+    int32_t  ne02;
+    uint64_t nb01;
+    uint64_t nb02;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    int32_t  ne0;
+    int32_t  ne1;
+} ggml_metal_kargs_mul_mm_id;
+
+typedef struct {
+    int32_t  nei0;
+    int32_t  nei1;
+    uint64_t nbi1;
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    int32_t  ne0;
+    int32_t  ne1;
+    uint64_t nb1;
+} ggml_metal_kargs_mul_mv_id;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne00_4;
+    uint64_t nb01;
+    float    eps;
+} ggml_metal_kargs_norm;
+
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne00_4;
+    uint64_t nb01;
+    float    eps;
+} ggml_metal_kargs_rms_norm;
+
+#endif // GGML_METAL_IMPL
diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal/ggml-metal.m
similarity index 75%
rename from ggml/src/ggml-metal.m
rename to ggml/src/ggml-metal/ggml-metal.m
index 04ec5117f64..28f590f9216 100644
--- a/ggml/src/ggml-metal.m
+++ b/ggml/src/ggml-metal/ggml-metal.m
@@ -2,6 +2,7 @@
 
 #import "ggml-impl.h"
 #import "ggml-backend-impl.h"
+#import "ggml-metal-impl.h"
 
 #import <Foundation/Foundation.h>
 
@@ -125,6 +126,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,
     GGML_METAL_KERNEL_TYPE_SILU,
     GGML_METAL_KERNEL_TYPE_SILU_4,
+    GGML_METAL_KERNEL_TYPE_ELU,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,
     GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,
@@ -173,6 +175,46 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_3,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_4,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_5,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_2,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_3,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_4,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_5,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_2,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_3,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_4,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_5,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_2,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_3,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_4,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_5,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_2,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_3,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_4,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_5,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_2,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_3,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_4,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_5,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_2,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_3,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_4,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_5,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_2,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_3,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_4,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_5,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_2,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_3,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_4,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_5,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_2,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_3,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_4,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_5,
     GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32,
@@ -264,8 +306,11 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_IM2COL_F32,
     GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F16,
     GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F32,
+    GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F32_F32,
+    GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F16_F32,
     GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
     GGML_METAL_KERNEL_TYPE_PAD_F32,
+    GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32,
     GGML_METAL_KERNEL_TYPE_ARANGE_F32,
     GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,
     GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
@@ -327,6 +372,8 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H256,
     GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256,
+    GGML_METAL_KERNEL_TYPE_SET_I32,
+    GGML_METAL_KERNEL_TYPE_SET_F32,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F16,
     GGML_METAL_KERNEL_TYPE_CPY_F32_BF16,
@@ -348,6 +395,7 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     GGML_METAL_KERNEL_TYPE_SUM_ROWS,
     GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
     GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,
+    GGML_METAL_KERNEL_TYPE_ARGMAX,
 
     GGML_METAL_KERNEL_TYPE_COUNT
 };
@@ -462,6 +510,35 @@ @implementation GGMLMetalClass
 #endif
 
         NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
+        if (path_lib == nil) {
+            // Try to find the resource in the directory where the current binary located.
+            NSString * current_binary = [[NSProcessInfo processInfo] arguments][0];
+            NSString * bin_dir = [current_binary stringByDeletingLastPathComponent];
+            NSString * default_metallib_path = [NSString pathWithComponents:@[bin_dir, @"default.metallib"]];
+            if ([[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) {
+                GGML_LOG_INFO("%s: found '%s'\n", __func__, [default_metallib_path UTF8String]);
+                NSDictionary * atts = [[NSFileManager defaultManager] attributesOfItemAtPath:default_metallib_path error:&error];
+                if (atts && atts[NSFileType] == NSFileTypeSymbolicLink) {
+                    // Optionally, if this is a symlink, try to resolve it.
+                    default_metallib_path = [[NSFileManager defaultManager] destinationOfSymbolicLinkAtPath:default_metallib_path error:&error];
+                    if (default_metallib_path && [default_metallib_path length] > 0 && ![[default_metallib_path substringToIndex:1] isEqualToString:@"/"]) {
+                        // It is a relative path, adding the binary directory as directory prefix.
+                        default_metallib_path = [NSString pathWithComponents:@[bin_dir, default_metallib_path]];
+                    }
+                    if (!default_metallib_path || ![[NSFileManager defaultManager] isReadableFileAtPath:default_metallib_path]) {
+                        // Link to the resource could not be resolved.
+                        default_metallib_path = nil;
+                    } else {
+                        GGML_LOG_INFO("%s: symlink resolved '%s'\n", __func__, [default_metallib_path UTF8String]);
+                    }
+                }
+            } else {
+                // The resource couldn't be found in the binary's directory.
+                default_metallib_path = nil;
+            }
+            path_lib = default_metallib_path;
+        }
+
         if (try_metallib && path_lib != nil) {
             // pre-compiled library found
             NSURL * libURL = [NSURL fileURLWithPath:path_lib];
@@ -516,6 +593,10 @@ @implementation GGMLMetalClass
                     [prep setObject:@"1" forKey:@"GGML_METAL_USE_BF16"];
                 }
 
+#if GGML_METAL_EMBED_LIBRARY
+                [prep setObject:@"1" forKey:@"GGML_METAL_EMBED_LIBRARY"];
+#endif
+
                 MTLCompileOptions * options = [MTLCompileOptions new];
                 options.preprocessorMacros = prep;
 
@@ -644,6 +725,7 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,                  gelu_quick_4,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU,                          silu,                           true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4,                        silu_4,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ELU,                           elu,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,                  soft_max_f16,                   has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,                soft_max_f16_4,                 has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,                  soft_max_f32,                   has_simdgroup_reduction);
@@ -692,6 +774,46 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,               mul_mv_q5_0_f32,                has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,               mul_mv_q5_1_f32,                has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,               mul_mv_q8_0_f32,                has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2,       mul_mv_ext_f16_f32_r1_2,        has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_3,       mul_mv_ext_f16_f32_r1_3,        has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_4,       mul_mv_ext_f16_f32_r1_4,        has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_5,       mul_mv_ext_f16_f32_r1_5,        has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_2,      mul_mv_ext_q4_0_f32_r1_2,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_3,      mul_mv_ext_q4_0_f32_r1_3,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_4,      mul_mv_ext_q4_0_f32_r1_4,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_5,      mul_mv_ext_q4_0_f32_r1_5,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_2,      mul_mv_ext_q4_1_f32_r1_2,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_3,      mul_mv_ext_q4_1_f32_r1_3,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_4,      mul_mv_ext_q4_1_f32_r1_4,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_5,      mul_mv_ext_q4_1_f32_r1_5,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_2,      mul_mv_ext_q5_0_f32_r1_2,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_3,      mul_mv_ext_q5_0_f32_r1_3,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_4,      mul_mv_ext_q5_0_f32_r1_4,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_5,      mul_mv_ext_q5_0_f32_r1_5,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_2,      mul_mv_ext_q5_1_f32_r1_2,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_3,      mul_mv_ext_q5_1_f32_r1_3,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_4,      mul_mv_ext_q5_1_f32_r1_4,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_5,      mul_mv_ext_q5_1_f32_r1_5,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_2,      mul_mv_ext_q8_0_f32_r1_2,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_3,      mul_mv_ext_q8_0_f32_r1_3,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_4,      mul_mv_ext_q8_0_f32_r1_4,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_5,      mul_mv_ext_q8_0_f32_r1_5,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_2,      mul_mv_ext_q4_K_f32_r1_2,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_3,      mul_mv_ext_q4_K_f32_r1_3,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_4,      mul_mv_ext_q4_K_f32_r1_4,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_5,      mul_mv_ext_q4_K_f32_r1_5,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_2,      mul_mv_ext_q5_K_f32_r1_2,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_3,      mul_mv_ext_q5_K_f32_r1_3,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_4,      mul_mv_ext_q5_K_f32_r1_4,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_5,      mul_mv_ext_q5_K_f32_r1_5,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_2,      mul_mv_ext_q6_K_f32_r1_2,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_3,      mul_mv_ext_q6_K_f32_r1_3,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_4,      mul_mv_ext_q6_K_f32_r1_4,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_5,      mul_mv_ext_q6_K_f32_r1_5,       has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_2,    mul_mv_ext_iq4_nl_f32_r1_2,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_3,    mul_mv_ext_iq4_nl_f32_r1_3,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_4,    mul_mv_ext_iq4_nl_f32_r1_4,     has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_5,    mul_mv_ext_iq4_nl_f32_r1_5,     has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32,               mul_mv_q2_K_f32,                has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32,               mul_mv_q3_K_f32,                has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32,               mul_mv_q4_K_f32,                has_simdgroup_reduction);
@@ -783,8 +905,11 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32,                    im2col_f32,                     true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F16,                im2col_ext_f16,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_EXT_F32,                im2col_ext_f32,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F32_F32,     conv_transpose_1d_f32_f32,      true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F16_F32,     conv_transpose_1d_f16_f32,      true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,                   upscale_f32,                    true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                       pad_f32,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32,            pad_reflect_1d_f32,             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,        timestep_embedding_f32,         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32,                    arange_f32,                     true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,           argsort_f32_i32_asc,            true);
@@ -846,6 +971,8 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_0_H256,  flash_attn_ext_vec_q5_0_h256,   has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q5_1_H256,  flash_attn_ext_vec_q5_1_h256,   has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_Q8_0_H256,  flash_attn_ext_vec_q8_0_h256,   has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_F32,                       set_f32,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_I32,                       set_i32,                        true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,                   cpy_f32_f32,                    true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16,                   cpy_f32_f16,                    true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_BF16,                  cpy_f32_bf16,                   use_bfloat);
@@ -865,6 +992,7 @@ @implementation GGMLMetalClass
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIN,                           sin,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS,                           cos,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                      sum_rows,                       true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX,                        argmax,                         true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,               pool_2d_avg_f32,                true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,               pool_2d_max_f32,                true);
     }
@@ -963,6 +1091,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                 case GGML_UNARY_OP_GELU:
                 case GGML_UNARY_OP_GELU_QUICK:
                 case GGML_UNARY_OP_SILU:
+                case GGML_UNARY_OP_ELU:
                     return ggml_is_contiguous(op->src[0]);
                 default:
                     return false;
@@ -981,6 +1110,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_REPEAT:
         case GGML_OP_SCALE:
         case GGML_OP_CLAMP:
+        case GGML_OP_CONV_TRANSPOSE_1D:
             return true;
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
@@ -989,12 +1119,24 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
             return ggml_is_contiguous(op->src[0]);
         case GGML_OP_SUM_ROWS:
         case GGML_OP_SOFT_MAX:
-        case GGML_OP_RMS_NORM:
         case GGML_OP_GROUP_NORM:
             return has_simdgroup_reduction;
+        case GGML_OP_RMS_NORM:
+            return has_simdgroup_reduction && (op->ne[0] % 4 == 0);
+        case GGML_OP_ARGMAX:
         case GGML_OP_NORM:
-        case GGML_OP_ROPE:
             return true;
+        case GGML_OP_ROPE:
+            {
+                const int mode = ((const int32_t *) op->op_params)[2];
+                if (mode & GGML_ROPE_TYPE_MROPE) {
+                    return false;
+                }
+                if (mode & GGML_ROPE_TYPE_VISION) {
+                    return false;
+                }
+                return true;
+            }
         case GGML_OP_IM2COL:
             return op->src[0]->type == GGML_TYPE_F16;
         case GGML_OP_POOL_1D:
@@ -1002,6 +1144,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_POOL_2D:
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
+        case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
@@ -1059,6 +1202,16 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
                         return false;
                 };
             }
+        case GGML_OP_SET:
+            {
+                switch (op->src[0]->type) {
+                    case GGML_TYPE_F32:
+                    case GGML_TYPE_I32:
+                        return true;
+                    default:
+                        return false;
+                };
+            }
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_GET_ROWS:
             {
@@ -1189,35 +1342,39 @@ static void ggml_metal_encode_node(
 
                 const int32_t dim = ((const int32_t *) dst->op_params)[0];
 
+                ggml_metal_kargs_concat args = {
+                    /*.ne00 =*/ ne00,
+                    /*.ne01 =*/ ne01,
+                    /*.ne02 =*/ ne02,
+                    /*.ne03 =*/ ne03,
+                    /*.nb00 =*/ nb00,
+                    /*.nb01 =*/ nb01,
+                    /*.nb02 =*/ nb02,
+                    /*.nb03 =*/ nb03,
+                    /*.ne10 =*/ ne10,
+                    /*.ne11 =*/ ne11,
+                    /*.ne12 =*/ ne12,
+                    /*.ne13 =*/ ne13,
+                    /*.nb10 =*/ nb10,
+                    /*.nb11 =*/ nb11,
+                    /*.nb12 =*/ nb12,
+                    /*.nb13 =*/ nb13,
+                    /*.ne0  =*/ ne0,
+                    /*.ne1  =*/ ne1,
+                    /*.ne2  =*/ ne2,
+                    /*.ne3  =*/ ne3,
+                    /*.nb0  =*/ nb0,
+                    /*.nb1  =*/ nb1,
+                    /*.nb2  =*/ nb2,
+                    /*.nb3  =*/ nb3,
+                    /*.dim  =*/ dim,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
-                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
-                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
-                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
-                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
-                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
-                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
-                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
-                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
-                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
-                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
-                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
-                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
-                [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
-                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
-                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
-                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
-                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
-                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
-                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
-                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
-                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
-                [encoder setBytes:&dim  length:sizeof(dim)  atIndex:27];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
 
                 const int nth = MIN(1024, ne0);
 
@@ -1235,8 +1392,6 @@ static void ggml_metal_encode_node(
 
                 bool bcast_row = false;
 
-                int64_t nb = ne00; // used by the "row" kernels
-
                 id<MTLComputePipelineState> pipeline = nil;
 
                 if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
@@ -1245,7 +1400,6 @@ static void ggml_metal_encode_node(
                     // src1 is a row
                     GGML_ASSERT(ne11 == 1);
 
-                    nb = ne00 / 4;
                     switch (dst->op) {
                         case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW].pipeline; break;
                         case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB_ROW].pipeline; break;
@@ -1265,36 +1419,39 @@ static void ggml_metal_encode_node(
                     }
                 }
 
+                ggml_metal_kargs_bin args = {
+                    /*.ne00 =*/ ne00,
+                    /*.ne01 =*/ ne01,
+                    /*.ne02 =*/ ne02,
+                    /*.ne03 =*/ ne03,
+                    /*.nb00 =*/ nb00,
+                    /*.nb01 =*/ nb01,
+                    /*.nb02 =*/ nb02,
+                    /*.nb03 =*/ nb03,
+                    /*.ne10 =*/ ne10,
+                    /*.ne11 =*/ ne11,
+                    /*.ne12 =*/ ne12,
+                    /*.ne13 =*/ ne13,
+                    /*.nb10 =*/ nb10,
+                    /*.nb11 =*/ nb11,
+                    /*.nb12 =*/ nb12,
+                    /*.nb13 =*/ nb13,
+                    /*.ne0  =*/ ne0,
+                    /*.ne1  =*/ ne1,
+                    /*.ne2  =*/ ne2,
+                    /*.ne3  =*/ ne3,
+                    /*.nb0  =*/ nb0,
+                    /*.nb1  =*/ nb1,
+                    /*.nb2  =*/ nb2,
+                    /*.nb3  =*/ nb3,
+                    /*.offs =*/ offs,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
-                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
-                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
-                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
-                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
-                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
-                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
-                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
-                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
-                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
-                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
-                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
-                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
-                [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
-                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
-                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
-                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
-                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
-                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
-                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
-                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
-                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
-                [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
-                [encoder setBytes:&nb   length:sizeof(nb)   atIndex:28];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
 
                 if (bcast_row) {
                     const int64_t n = ggml_nelements(dst)/4;
@@ -1318,25 +1475,29 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("fatal error");
                 }
 
+                ggml_metal_kargs_repeat args = {
+                    /*.ne00 =*/ ne00,
+                    /*.ne01 =*/ ne01,
+                    /*.ne02 =*/ ne02,
+                    /*.ne03 =*/ ne03,
+                    /*.nb00 =*/ nb00,
+                    /*.nb01 =*/ nb01,
+                    /*.nb02 =*/ nb02,
+                    /*.nb03 =*/ nb03,
+                    /*.ne0  =*/ ne0,
+                    /*.ne1  =*/ ne1,
+                    /*.ne2  =*/ ne2,
+                    /*.ne3  =*/ ne3,
+                    /*.nb0  =*/ nb0,
+                    /*.nb1  =*/ nb1,
+                    /*.nb2  =*/ nb2,
+                    /*.nb3  =*/ nb3,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
-                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
-                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
-                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
-                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
-                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
-                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
-                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
-                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
-                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
-                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
-                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
-                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
-                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
-                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
 
                 const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
 
@@ -1365,25 +1526,29 @@ static void ggml_metal_encode_node(
 
                     const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline;
 
+                    ggml_metal_kargs_cpy args = {
+                        /*.ne00 =*/ ne00,
+                        /*.ne01 =*/ ne01,
+                        /*.ne02 =*/ ne02,
+                        /*.ne03 =*/ ne03,
+                        /*.nb00 =*/ nb00,
+                        /*.nb01 =*/ nb01,
+                        /*.nb02 =*/ nb02,
+                        /*.nb03 =*/ nb03,
+                        /*.ne0  =*/ ne0,
+                        /*.ne1  =*/ ne1,
+                        /*.ne2  =*/ ne2,
+                        /*.ne3  =*/ ne3,
+                        /*.nb0  =*/ nb0,
+                        /*.nb1  =*/ nb1,
+                        /*.nb2  =*/ nb2,
+                        /*.nb3  =*/ nb3,
+                    };
+
                     [encoder setComputePipelineState:pipeline];
-                    [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                    [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                    [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                    [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
-                    [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
-                    [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
-                    [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
-                    [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
-                    [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
-                    [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
-                    [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
-                    [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
-                    [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
-                    [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
-                    [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
-                    [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
-                    [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
-                    [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
+                    [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
 
                     const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
 
@@ -1392,35 +1557,39 @@ static void ggml_metal_encode_node(
 
                 const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline;
 
+                ggml_metal_kargs_bin args = {
+                    /*.ne00 =*/ ne00,
+                    /*.ne01 =*/ ne01,
+                    /*.ne02 =*/ ne02,
+                    /*.ne03 =*/ ne03,
+                    /*.nb00 =*/ nb00,
+                    /*.nb01 =*/ pnb1,
+                    /*.nb02 =*/ pnb2,
+                    /*.nb03 =*/ pnb3,
+                    /*.ne10 =*/ ne10,
+                    /*.ne11 =*/ ne11,
+                    /*.ne12 =*/ ne12,
+                    /*.ne13 =*/ ne13,
+                    /*.nb10 =*/ nb10,
+                    /*.nb11 =*/ nb11,
+                    /*.nb12 =*/ nb12,
+                    /*.nb13 =*/ nb13,
+                    /*.ne0  =*/ ne0,
+                    /*.ne1  =*/ ne1,
+                    /*.ne2  =*/ ne2,
+                    /*.ne3  =*/ ne3,
+                    /*.nb0  =*/ nb0,
+                    /*.nb1  =*/ pnb1,
+                    /*.nb2  =*/ pnb2,
+                    /*.nb3  =*/ pnb3,
+                    /*.offs =*/ offs,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
-                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
-                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
-                [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:8];
-                [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:9];
-                [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:10];
-                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
-                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
-                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
-                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
-                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
-                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
-                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
-                [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
-                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
-                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
-                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
-                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
-                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
-                [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:24];
-                [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:25];
-                [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:26];
-                [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
 
                 const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
 
@@ -1461,10 +1630,10 @@ static void ggml_metal_encode_node(
                 memcpy(&max, ((const int32_t *) dst->op_params) + 1, sizeof(float));
 
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
-                [encoder setBytes:&min length:sizeof(min) atIndex:2];
-                [encoder setBytes:&max length:sizeof(max) atIndex:3];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBytes:&min   length:sizeof(min) atIndex:2];
+                [encoder setBytes:&max   length:sizeof(max) atIndex:3];
 
                 const int64_t n = ggml_nelements(dst);
 
@@ -1568,6 +1737,18 @@ static void ggml_metal_encode_node(
 
                     [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                 } break;
+                case GGML_UNARY_OP_ELU:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ELU].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
                 default:
                 {
                     GGML_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op));
@@ -1636,6 +1817,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -1711,6 +1893,8 @@ static void ggml_metal_encode_node(
                 const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
                 const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
+                // TODO: add ggml_metal_kargs struct
+                // TODO: optimize (see https://github.com/ggerganov/llama.cpp/pull/10238/commits/7941b6b9ec29a2866fec6fa6c51612515ca509f6)
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
                 if (id_src1) {
@@ -1727,6 +1911,7 @@ static void ggml_metal_encode_node(
                 [encoder setBytes:&m0          length:sizeof(m0)          atIndex:8];
                 [encoder setBytes:&m1          length:sizeof(m1)          atIndex:9];
                 [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:10];
+
                 [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
                 [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
@@ -1743,6 +1928,7 @@ static void ggml_metal_encode_node(
                     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF].pipeline;
                 }
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -1767,6 +1953,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_CONV_F32].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
                 [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
@@ -1837,6 +2024,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
@@ -1884,331 +2072,490 @@ static void ggml_metal_encode_node(
 
                 // find the break-even point where the matrix-matrix kernel becomes more efficient compared
                 // to the matrix-vector kernel
-                int ne11_mm_min = 1;
+                const int ne11_mm_min = 4;
+
+                // first try to use small-batch mat-mv kernels
+                // these should be efficient for BS [2, ~8]
+                if (src1t == GGML_TYPE_F32 && (ne00%256 == 0) &&
+                    (
+                     (
+                      (
+                       src0t == GGML_TYPE_F16  || // TODO: helper function
+                       src0t == GGML_TYPE_Q4_0 ||
+                       src0t == GGML_TYPE_Q4_1 ||
+                       src0t == GGML_TYPE_Q5_0 ||
+                       src0t == GGML_TYPE_Q5_1 ||
+                       src0t == GGML_TYPE_Q8_0 ||
+                       src0t == GGML_TYPE_IQ4_NL ||
+                       false) && (ne11 >= 2 && ne11 <= 8)
+                     ) ||
+                     (
+                      (
+                       src0t == GGML_TYPE_Q4_K ||
+                       src0t == GGML_TYPE_Q5_K ||
+                       src0t == GGML_TYPE_Q6_K ||
+                       false) && (ne11 >= 4 && ne11 <= 8)
+                     )
+                    )
+                   ) {
+                    // TODO: determine the optimal parameters based on grid utilization
+                    //       I still don't know why we should not always use the maximum available threads:
+                    //
+                    //       nsg = pipeline.maxTotalThreadsPerThreadgroup / 32
+                    //
+                    //       my current hypothesis is that the work grid is not evenly divisible for different nsg
+                    //       values and there can be some tail effects when nsg is high. need to confirm this
+                    //
+                    const int nsg    = 2;                 // num simdgroups per threadgroup
+                    const int nxpsg  = ne11 < 3 ? 16 : 8; // num threads along row per simdgroup
+                    const int nypsg  = 32/nxpsg;          // num threads along col per simdgroup (i.e. a simdgroup processes that many src0 rows at a time)
+                    const int r0ptg  = nypsg*nsg;         // num src0 rows per threadgroup
+                          int r1ptg  = 4;                 // num src1 rows per threadgroup
+
+                    // note: not sure how optimal are those across all different hardware. there might be someting cleverer
+                    switch (ne11) {
+                        case 2:
+                            r1ptg = 2; break;
+                        case 3:
+                        case 6:
+                            r1ptg = 3; break;
+                        case 4:
+                        case 7:
+                        case 8:
+                            r1ptg = 4; break;
+                        case 5:
+                            r1ptg = 5; break;
+                    };
 
-#if 0
-                // the numbers below are measured on M2 Ultra for 7B and 13B models
-                // these numbers do not translate to other devices or model sizes
-                // TODO: need to find a better approach
-                        if ([device.name isEqualToString:@"Apple M2 Ultra"]) {
-                            switch (src0t) {
-                                case GGML_TYPE_F16:  ne11_mm_min = 2;  break;
-                                case GGML_TYPE_Q8_0: ne11_mm_min = 7;  break;
-                                case GGML_TYPE_Q2_K: ne11_mm_min = 15; break;
-                                case GGML_TYPE_Q3_K: ne11_mm_min = 7;  break;
-                                case GGML_TYPE_Q4_0:
-                                case GGML_TYPE_Q4_1: ne11_mm_min = 15; break;
-                                case GGML_TYPE_Q4_K: ne11_mm_min = 11; break;
-                                case GGML_TYPE_Q5_0:                          // not tested yet
-                                case GGML_TYPE_Q5_1: ne11_mm_min = 13; break; // not tested yet
-                                case GGML_TYPE_Q5_K: ne11_mm_min = 7;  break;
-                                case GGML_TYPE_Q6_K: ne11_mm_min = 7;  break;
-                                default:             ne11_mm_min = 1;  break;
-                            }
-                        }
-#endif
+                    id<MTLComputePipelineState> pipeline = nil;
 
-                        // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
-                        // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
-                        if ([device supportsFamily:MTLGPUFamilyApple7] &&
-                                !ggml_is_transposed(src0) &&
-                                !ggml_is_transposed(src1) &&
-                                src1t == GGML_TYPE_F32 &&
-                                ne00 % 32 == 0 && ne00 >= 64 &&
-                                (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) {
-                            //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
-
-                            // some Metal matrix data types require aligned pointers
-                            // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
-                            switch (src0->type) {
-                                case GGML_TYPE_F32:  GGML_ASSERT(nb01 % 16 == 0); break;
-                                case GGML_TYPE_F16:  GGML_ASSERT(nb01 % 8  == 0); break;
-                                case GGML_TYPE_BF16: GGML_ASSERT(nb01 % 8  == 0); break;
-                                default: break;
-                            }
+                    switch (src0->type) {
+                        case GGML_TYPE_F16:
+                            switch (r1ptg) {
+                                case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_2].pipeline; break;
+                                case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_3].pipeline; break;
+                                case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_4].pipeline; break;
+                                case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_F16_F32_R1_5].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            } break;
+                        case GGML_TYPE_Q4_0:
+                            switch (r1ptg) {
+                                case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_2].pipeline; break;
+                                case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_3].pipeline; break;
+                                case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_4].pipeline; break;
+                                case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_0_F32_R1_5].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            } break;
+                        case GGML_TYPE_Q4_1:
+                            switch (r1ptg) {
+                                case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_2].pipeline; break;
+                                case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_3].pipeline; break;
+                                case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_4].pipeline; break;
+                                case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_1_F32_R1_5].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            } break;
+                        case GGML_TYPE_Q5_0:
+                            switch (r1ptg) {
+                                case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_2].pipeline; break;
+                                case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_3].pipeline; break;
+                                case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_4].pipeline; break;
+                                case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_0_F32_R1_5].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            } break;
+                        case GGML_TYPE_Q5_1:
+                            switch (r1ptg) {
+                                case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_2].pipeline; break;
+                                case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_3].pipeline; break;
+                                case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_4].pipeline; break;
+                                case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_1_F32_R1_5].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            } break;
+                        case GGML_TYPE_Q8_0:
+                            switch (r1ptg) {
+                                case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_2].pipeline; break;
+                                case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_3].pipeline; break;
+                                case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_4].pipeline; break;
+                                case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q8_0_F32_R1_5].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            } break;
+                        case GGML_TYPE_Q4_K:
+                            switch (r1ptg) {
+                                case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_2].pipeline; break;
+                                case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_3].pipeline; break;
+                                case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_4].pipeline; break;
+                                case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q4_K_F32_R1_5].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            } break;
+                        case GGML_TYPE_Q5_K:
+                            switch (r1ptg) {
+                                case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_2].pipeline; break;
+                                case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_3].pipeline; break;
+                                case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_4].pipeline; break;
+                                case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q5_K_F32_R1_5].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            } break;
+                        case GGML_TYPE_Q6_K:
+                            switch (r1ptg) {
+                                case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_2].pipeline; break;
+                                case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_3].pipeline; break;
+                                case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_4].pipeline; break;
+                                case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_Q6_K_F32_R1_5].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            } break;
+                        case GGML_TYPE_IQ4_NL:
+                            switch (r1ptg) {
+                                case 2: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_2].pipeline; break;
+                                case 3: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_3].pipeline; break;
+                                case 4: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_4].pipeline; break;
+                                case 5: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_EXT_IQ4_NL_F32_R1_5].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            } break;
+                        default: GGML_ABORT("not implemented");
+                    }
 
-                            id<MTLComputePipelineState> pipeline = nil;
-
-                            switch (src0->type) {
-                                case GGML_TYPE_F32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32    ].pipeline; break;
-                                case GGML_TYPE_F16:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32    ].pipeline; break;
-                                case GGML_TYPE_BF16:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_BF16_F32   ].pipeline; break;
-                                case GGML_TYPE_Q4_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32   ].pipeline; break;
-                                case GGML_TYPE_Q4_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32   ].pipeline; break;
-                                case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32   ].pipeline; break;
-                                case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32   ].pipeline; break;
-                                case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32   ].pipeline; break;
-                                case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32   ].pipeline; break;
-                                case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32   ].pipeline; break;
-                                case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32   ].pipeline; break;
-                                case GGML_TYPE_Q5_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32   ].pipeline; break;
-                                case GGML_TYPE_Q6_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32   ].pipeline; break;
-                                case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
-                                case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
-                                case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
-                                case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ1_M:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
-                                case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
-                                default: GGML_ABORT("MUL MAT-MAT not implemented");
-                            }
+                    ggml_metal_kargs_mul_mv_ext args = {
+                        /*.ne00  =*/ ne00,
+                        /*.ne01  =*/ ne01,
+                        /*.ne02  =*/ ne02,
+                        /*.nb00  =*/ nb00,
+                        /*.nb01  =*/ nb01,
+                        /*.nb02  =*/ nb02,
+                        /*.nb03  =*/ nb03,
+                        /*.ne10  =*/ ne10,
+                        /*.ne11  =*/ ne11,
+                        /*.ne12  =*/ ne12,
+                        /*.nb10  =*/ nb10,
+                        /*.nb11  =*/ nb11,
+                        /*.nb12  =*/ nb12,
+                        /*.nb13  =*/ nb13,
+                        /*.ne0   =*/ ne0,
+                        /*.ne1   =*/ ne1,
+                        /*.r2    =*/ r2,
+                        /*.r3    =*/ r3,
+                        /*.nsg   =*/ nsg,
+                        /*.nxpsg =*/ nxpsg,
+                        /*.r1ptg =*/ r1ptg,
+                    };
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
+
+                    //printf("ne01 = %lld nr0ptg = %d\n", ne01, nr0ptg);
+                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + r0ptg - 1)/r0ptg, (ne11 + r1ptg - 1)/r1ptg, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
+                } else
+                // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
+                // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
+                if ([device supportsFamily:MTLGPUFamilyApple7] &&
+                        !ggml_is_transposed(src0) &&
+                        !ggml_is_transposed(src1) &&
+                        src1t == GGML_TYPE_F32 &&
+                        ne00 % 32 == 0 && ne00 >= 64 &&
+                        (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) {
+                    //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
+
+                    // some Metal matrix data types require aligned pointers
+                    // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
+                    switch (src0->type) {
+                        case GGML_TYPE_F32:  GGML_ASSERT(nb01 % 16 == 0); break;
+                        case GGML_TYPE_F16:  GGML_ASSERT(nb01 % 8  == 0); break;
+                        case GGML_TYPE_BF16: GGML_ASSERT(nb01 % 8  == 0); break;
+                        default: break;
+                    }
+
+                    id<MTLComputePipelineState> pipeline = nil;
+
+                    switch (src0->type) {
+                        case GGML_TYPE_F32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32    ].pipeline; break;
+                        case GGML_TYPE_F16:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32    ].pipeline; break;
+                        case GGML_TYPE_BF16:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_BF16_F32   ].pipeline; break;
+                        case GGML_TYPE_Q4_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32   ].pipeline; break;
+                        case GGML_TYPE_Q4_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32   ].pipeline; break;
+                        case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32   ].pipeline; break;
+                        case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32   ].pipeline; break;
+                        case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32   ].pipeline; break;
+                        case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q5_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q6_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32   ].pipeline; break;
+                        case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
+                        case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
+                        case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
+                        case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ1_M:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
+                        case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
+                        default: GGML_ABORT("MUL MAT-MAT not implemented");
+                    }
+
+                    ggml_metal_kargs_mul_mm args = {
+                        /*.ne00 =*/ ne00,
+                        /*.ne02 =*/ ne02,
+                        /*.nb01 =*/ nb01,
+                        /*.nb02 =*/ nb02,
+                        /*.nb03 =*/ nb03,
+                        /*.ne12 =*/ ne12,
+                        /*.nb10 =*/ nb10,
+                        /*.nb11 =*/ nb11,
+                        /*.nb12 =*/ nb12,
+                        /*.nb13 =*/ nb13,
+                        /*.ne0  =*/ ne0,
+                        /*.ne1  =*/ ne1,
+                        /*.r2   =*/ r2,
+                        /*.r3   =*/ r3,
+                    };
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
+                    [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
+
+                    [encoder setThreadgroupMemoryLength:8192 atIndex:0];
+                    [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                } else {
+                    int nth0 = 32;
+                    int nth1 = 1;
+                    int nrows = 1;
+                    //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
+
+                    id<MTLComputePipelineState> pipeline = nil;
 
-                            [encoder setComputePipelineState:pipeline];
-                            [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
-                            [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
-                            [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
-                            [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:3];
-                            [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:4];
-                            [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:5];
-                            [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:6];
-                            [encoder setBytes:&nb03    length:sizeof(nb03) atIndex:7];
-                            [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:8];
-                            [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:9];
-                            [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:10];
-                            [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:11];
-                            [encoder setBytes:&nb13    length:sizeof(nb13) atIndex:12];
-                            [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:13];
-                            [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:14];
-                            [encoder setBytes:&r2      length:sizeof(r2)   atIndex:15];
-                            [encoder setBytes:&r3      length:sizeof(r3)   atIndex:16];
-                            [encoder setThreadgroupMemoryLength:8192 atIndex:0];
-                            [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
-                        } else {
-                            int nth0 = 32;
-                            int nth1 = 1;
-                            int nrows = 1;
-                            //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
-
-                            id<MTLComputePipelineState> pipeline = nil;
-
-                            // use custom matrix x vector kernel
-                            switch (src0t) {
-                                case GGML_TYPE_F32:
-                                    {
-                                        GGML_ASSERT(src1t == GGML_TYPE_F32);
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
+                    // use custom matrix x vector kernel
+                    switch (src0t) {
+                        case GGML_TYPE_F32:
+                            {
+                                GGML_ASSERT(src1t == GGML_TYPE_F32);
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
+                                nrows = 4;
+                            } break;
+                        case GGML_TYPE_F16:
+                            {
+                                nth0 = 32;
+                                nth1 = 1;
+                                if (src1t == GGML_TYPE_F32) {
+                                    if (ne11 * ne12 < 4) {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline;
+                                    } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline;
+                                        nrows = ne11;
+                                    } else {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32].pipeline;
                                         nrows = 4;
-                                    } break;
-                                case GGML_TYPE_F16:
-                                    {
-                                        nth0 = 32;
-                                        nth1 = 1;
-                                        if (src1t == GGML_TYPE_F32) {
-                                            if (ne11 * ne12 < 4) {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline;
-                                            } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline;
-                                                nrows = ne11;
-                                            } else {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32].pipeline;
-                                                nrows = 4;
-                                            }
-                                        } else {
-                                            pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16].pipeline;
-                                            nrows = 4;
-                                        }
-                                    } break;
-                                case GGML_TYPE_BF16:
-                                    {
-                                        nth0 = 32;
-                                        nth1 = 1;
-                                        if (src1t == GGML_TYPE_F32) {
-                                            if (ne11 * ne12 < 4) {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW].pipeline;
-                                            } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4].pipeline;
-                                                nrows = ne11;
-                                            } else {
-                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32].pipeline;
-                                                nrows = 4;
-                                            }
-                                        } else {
-                                            pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16].pipeline;
-                                            nrows = 4;
-                                        }
-                                    } break;
-                                case GGML_TYPE_Q4_0:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q4_1:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q5_0:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q5_1:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q8_0:
-                                    {
-                                        nth0 = 8;
-                                        nth1 = 8;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q2_K:
-                                    {
-                                        nth0 = 2;
-                                        nth1 = 32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q3_K:
-                                    {
-                                        nth0 = 2;
-                                        nth1 = 32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q4_K:
-                                    {
-                                        nth0 = 4; //1;
-                                        nth1 = 8; //32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q5_K:
-                                    {
-                                        nth0 = 2;
-                                        nth1 = 32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_Q6_K:
-                                    {
-                                        nth0 = 2;
-                                        nth1 = 32;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ2_XXS:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ2_XS:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ3_XXS:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ3_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ2_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ1_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ1_M:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ4_NL:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ4_XS:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline;
-                                    } break;
-                                default:
-                                    {
-                                        GGML_LOG_ERROR("Asserting on type %d\n", (int)src0t);
-                                        GGML_ABORT("not implemented");
                                     }
-                            };
-
-                            [encoder setComputePipelineState:pipeline];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
-                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
-                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
-                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
-                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
-                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10];
-                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11];
-                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
-                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:13];
-                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:14];
-                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:15];
-                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:16];
-                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:17];
-                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:18];
-                            [encoder setBytes:&r2   length:sizeof(r2)   atIndex:19];
-                            [encoder setBytes:&r3   length:sizeof(r3)   atIndex:20];
-
-                            if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
-                                src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
-                                src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
-                                const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
-                                const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
-                                const int mem_size = 32*sizeof(float);
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_Q4_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_Q3_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_Q5_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
-                            else if (src0t == GGML_TYPE_Q6_K) {
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            } else {
-                                const int64_t ny = (ne11 + nrows - 1)/nrows;
-                                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                } else {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16].pipeline;
+                                    nrows = 4;
+                                }
+                            } break;
+                        case GGML_TYPE_BF16:
+                            {
+                                nth0 = 32;
+                                nth1 = 1;
+                                if (src1t == GGML_TYPE_F32) {
+                                    if (ne11 * ne12 < 4) {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW].pipeline;
+                                    } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4].pipeline;
+                                        nrows = ne11;
+                                    } else {
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32].pipeline;
+                                        nrows = 4;
+                                    }
+                                } else {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16].pipeline;
+                                    nrows = 4;
+                                }
+                            } break;
+                        case GGML_TYPE_Q4_0:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q4_1:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q5_0:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q5_1:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q8_0:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q2_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q3_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q4_K:
+                            {
+                                nth0 = 4; //1;
+                                nth1 = 8; //32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q5_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q6_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ2_XXS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ2_XS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ3_XXS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ3_S:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ2_S:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ1_S:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ1_M:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ4_NL:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ4_XS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline;
+                            } break;
+                        default:
+                            {
+                                GGML_LOG_ERROR("Asserting on type %d\n", (int)src0t);
+                                GGML_ABORT("not implemented");
                             }
-                        }
+                    };
+
+                    ggml_metal_kargs_mul_mv args = {
+                        /*.ne00 =*/ ne00,
+                        /*.ne01 =*/ ne01,
+                        /*.ne02 =*/ ne02,
+                        /*.nb00 =*/ nb00,
+                        /*.nb01 =*/ nb01,
+                        /*.nb02 =*/ nb02,
+                        /*.nb03 =*/ nb03,
+                        /*.ne10 =*/ ne10,
+                        /*.ne11 =*/ ne11,
+                        /*.ne12 =*/ ne12,
+                        /*.nb10 =*/ nb10,
+                        /*.nb11 =*/ nb11,
+                        /*.nb12 =*/ nb12,
+                        /*.nb13 =*/ nb13,
+                        /*.ne0  =*/ ne0,
+                        /*.ne1  =*/ ne1,
+                        /*.r2   =*/ r2,
+                        /*.r3   =*/ r3,
+                    };
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
+
+                    if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
+                        src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
+                        src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
+                        const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
+                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
+                        const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
+                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
+                        const int mem_size = 32*sizeof(float);
+                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q4_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q3_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q5_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q6_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    } else {
+                        const int64_t ny = (ne11 + nrows - 1)/nrows;
+                        [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                }
             } break;
         case GGML_OP_MUL_MAT_ID:
             {
@@ -2284,27 +2631,30 @@ static void ggml_metal_encode_node(
                         default: GGML_ABORT("MUL_MAT_ID not implemented");
                     }
 
+                    ggml_metal_kargs_mul_mm_id args = {
+                        /*.nei0 =*/ ne20,
+                        /*.nei1 =*/ ne21,
+                        /*.nbi1 =*/ nb21,
+                        /*.ne00 =*/ ne00,
+                        /*.ne02 =*/ ne02,
+                        /*.nb01 =*/ nb01,
+                        /*.nb02 =*/ nb02,
+                        /*.ne11 =*/ ne11,
+                        /*.ne12 =*/ ne12,
+                        /*.ne13 =*/ ne13,
+                        /*.nb10 =*/ nb10,
+                        /*.nb11 =*/ nb11,
+                        /*.nb12 =*/ nb12,
+                        /*.ne0  =*/ ne0,
+                        /*.ne1  =*/ ne1,
+                    };
+
                     [encoder setComputePipelineState:pipeline];
-                    [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
-                    [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
-                    [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
-                    [encoder setBuffer:id_src2 offset:offs_src2    atIndex:3];
-                    [encoder setBytes:&ne20    length:sizeof(ne20) atIndex:4];
-                    [encoder setBytes:&ne21    length:sizeof(ne21) atIndex:5];
-                    [encoder setBytes:&nb21    length:sizeof(nb21) atIndex:6];
-                    [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:7];
-                    [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:8];
-                    [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:9];
-                    [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:10];
-                    [encoder setBytes:&ne11    length:sizeof(ne11) atIndex:11];
-                    [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:12];
-                    [encoder setBytes:&ne13    length:sizeof(ne13) atIndex:13];
-                    [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:14];
-                    [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:15];
-                    [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:16];
-                    [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:17];
-                    [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:18];
-                    [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:19];
+                    [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
+                    [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
+                    [encoder setBuffer:id_src2 offset:offs_src2    atIndex:4];
 
                     [encoder setThreadgroupMemoryLength:GGML_PAD(8192 + dst_rows*4/*sizeof(ushort2)*/, 16) atIndex:0];
 
@@ -2463,30 +2813,34 @@ static void ggml_metal_encode_node(
                         GGML_ASSERT(ne00 >= nth0*nth1);
                     }
 
+                    ggml_metal_kargs_mul_mv_id args = {
+                        /*.nei0 =*/ ne20,
+                        /*.nei1 =*/ ne21,
+                        /*.nbi1 =*/ nb21,
+                        /*.ne00 =*/ ne00,
+                        /*.ne01 =*/ ne01,
+                        /*.ne02 =*/ ne02,
+                        /*.nb00 =*/ nb00,
+                        /*.nb01 =*/ nb01,
+                        /*.nb02 =*/ nb02,
+                        /*.ne10 =*/ ne10,
+                        /*.ne11 =*/ ne11,
+                        /*.ne12 =*/ ne12,
+                        /*.ne13 =*/ ne13,
+                        /*.nb10 =*/ nb10,
+                        /*.nb11 =*/ nb11,
+                        /*.nb12 =*/ nb12,
+                        /*.ne0  =*/ ne0,
+                        /*.ne1  =*/ ne1,
+                        /*.nb1  =*/ nb1,
+                    };
+
                     [encoder setComputePipelineState:pipeline];
-                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                    [encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
-                    [encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4];
-                    [encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5];
-                    [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6];
-                    [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7];
-                    [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:8];
-                    [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:9];
-                    [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:10];
-                    [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:11];
-                    [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:12];
-                    [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:13];
-                    [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:14];
-                    [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:15];
-                    [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:16];
-                    [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:17];
-                    [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:18];
-                    [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:19];
-                    [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:20];
-                    [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:21];
-                    [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:22];
+                    [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:2];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:3];
+                    [encoder setBuffer:id_src2 offset:offs_src2 atIndex:4];
 
                     const int64_t _ne1 = 1;
                     const int tgz = dst_rows;
@@ -2559,6 +2913,7 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("not implemented");
                 }
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0     offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_src1     offset:offs_src1 atIndex:1];
@@ -2582,20 +2937,28 @@ static void ggml_metal_encode_node(
                 float eps;
                 memcpy(&eps, dst->op_params, sizeof(float));
 
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM].pipeline;
+
                 int nth = 32; // SIMD width
 
-                while (nth < ne00/4 && nth < 1024) {
+                while (nth < ne00/4 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
                     nth *= 2;
                 }
 
-                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM].pipeline;
+                nth = MIN(nth, ne00/4);
+
+                ggml_metal_kargs_rms_norm args = {
+                    /*.ne00   =*/ ne00,
+                    /*.ne00_4 =*/ ne00/4,
+                    /*.nb01   =*/ nb01,
+                    /*.eps    =*/ eps,
+                };
 
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
-                [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+
                 [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
                 const int64_t nrows = ggml_nrows(src0);
@@ -2604,7 +2967,6 @@ static void ggml_metal_encode_node(
             } break;
         case GGML_OP_GROUP_NORM:
             {
-                GGML_ASSERT(ne00 % 4 == 0);
                 GGML_ASSERT(ggml_is_contiguous(src0));
 
                 float eps;
@@ -2620,6 +2982,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GROUP_NORM].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0  offset:offs_src0        atIndex:0];
                 [encoder setBuffer:id_dst   offset:offs_dst         atIndex:1];
@@ -2637,22 +3000,35 @@ static void ggml_metal_encode_node(
             } break;
         case GGML_OP_NORM:
             {
+                GGML_ASSERT(ne00 % 4 == 0);
                 GGML_ASSERT(ggml_is_contiguous_1(src0));
 
                 float eps;
                 memcpy(&eps, dst->op_params, sizeof(float));
 
-                const int nth = MIN(256, ne00);
-
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_NORM].pipeline;
 
+                int nth = 32; // SIMD width
+
+                while (nth < ne00/4 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                    nth *= 2;
+                }
+
+                nth = MIN(nth, ne00/4);
+
+                ggml_metal_kargs_norm args = {
+                    /*.ne00   =*/ ne00,
+                    /*.ne00_4 =*/ ne00/4,
+                    /*.nb01   =*/ nb01,
+                    /*.eps    =*/ eps,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
-                [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
-                [encoder setThreadgroupMemoryLength:GGML_PAD(nth*sizeof(float), 16) atIndex:0];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+
+                [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
                 const int64_t nrows = ggml_nrows(src0);
 
@@ -2660,7 +3036,9 @@ static void ggml_metal_encode_node(
             } break;
         case GGML_OP_ROPE:
             {
-                GGML_ASSERT(ne10 == ne02);
+                // make sure we have one or more position id(ne10) per token(ne02)
+                GGML_ASSERT(ne10 % ne02 == 0);
+                GGML_ASSERT(ne10 >= ne02);
 
                 const int nth = MIN(1024, ne00);
 
@@ -2702,40 +3080,44 @@ static void ggml_metal_encode_node(
                     };
                 }
 
+                ggml_metal_kargs_rope args = {
+                    /*.ne00        =*/ ne00,
+                    /*.ne01        =*/ ne01,
+                    /*.ne02        =*/ ne02,
+                    /*.ne03        =*/ ne03,
+                    /*.nb00        =*/ nb00,
+                    /*.nb01        =*/ nb01,
+                    /*.nb02        =*/ nb02,
+                    /*.nb03        =*/ nb03,
+                    /*.ne0         =*/ ne0,
+                    /*.ne1         =*/ ne1,
+                    /*.ne2         =*/ ne2,
+                    /*.ne3         =*/ ne3,
+                    /*.nb0         =*/ nb0,
+                    /*.nb1         =*/ nb1,
+                    /*.nb2         =*/ nb2,
+                    /*.nb3         =*/ nb3,
+                    /*.n_past      =*/ n_past,
+                    /*.n_dims      =*/ n_dims,
+                    /*.n_ctx_orig  =*/ n_ctx_orig,
+                    /*.freq_base   =*/ freq_base,
+                    /*.freq_scale  =*/ freq_scale,
+                    /*.ext_factor  =*/ ext_factor,
+                    /*.attn_factor =*/ attn_factor,
+                    /*.beta_fast   =*/ beta_fast,
+                    /*.beta_slow   =*/ beta_slow,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
-                [encoder setBuffer:id_src1     offset:offs_src1        atIndex:1];
+                [encoder setBytes:&args length:sizeof(args)     atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0     atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1     atIndex:2];
                 if (id_src2 != nil) {
-                    [encoder setBuffer:id_src2 offset:offs_src2        atIndex:2];
+                    [encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
                 } else {
-                    [encoder setBuffer:id_src0 offset:offs_src0        atIndex:2];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:3];
                 }
-                [encoder setBuffer:id_dst      offset:offs_dst         atIndex:3];
-                [encoder setBytes:&ne00        length:sizeof( int64_t) atIndex:4];
-                [encoder setBytes:&ne01        length:sizeof( int64_t) atIndex:5];
-                [encoder setBytes:&ne02        length:sizeof( int64_t) atIndex:6];
-                [encoder setBytes:&ne03        length:sizeof( int64_t) atIndex:7];
-                [encoder setBytes:&nb00        length:sizeof(uint64_t) atIndex:8];
-                [encoder setBytes:&nb01        length:sizeof(uint64_t) atIndex:9];
-                [encoder setBytes:&nb02        length:sizeof(uint64_t) atIndex:10];
-                [encoder setBytes:&nb03        length:sizeof(uint64_t) atIndex:11];
-                [encoder setBytes:&ne0         length:sizeof( int64_t) atIndex:12];
-                [encoder setBytes:&ne1         length:sizeof( int64_t) atIndex:13];
-                [encoder setBytes:&ne2         length:sizeof( int64_t) atIndex:14];
-                [encoder setBytes:&ne3         length:sizeof( int64_t) atIndex:15];
-                [encoder setBytes:&nb0         length:sizeof(uint64_t) atIndex:16];
-                [encoder setBytes:&nb1         length:sizeof(uint64_t) atIndex:17];
-                [encoder setBytes:&nb2         length:sizeof(uint64_t) atIndex:18];
-                [encoder setBytes:&nb3         length:sizeof(uint64_t) atIndex:19];
-                [encoder setBytes:&n_past      length:sizeof(     int) atIndex:20];
-                [encoder setBytes:&n_dims      length:sizeof(     int) atIndex:21];
-                [encoder setBytes:&n_ctx_orig  length:sizeof(     int) atIndex:22];
-                [encoder setBytes:&freq_base   length:sizeof(   float) atIndex:23];
-                [encoder setBytes:&freq_scale  length:sizeof(   float) atIndex:24];
-                [encoder setBytes:&ext_factor  length:sizeof(   float) atIndex:25];
-                [encoder setBytes:&attn_factor length:sizeof(   float) atIndex:26];
-                [encoder setBytes:&beta_fast   length:sizeof(   float) atIndex:27];
-                [encoder setBytes:&beta_slow   length:sizeof(   float) atIndex:28];
+                [encoder setBuffer:id_dst  offset:offs_dst      atIndex:4];
 
                 [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
@@ -2792,6 +3174,7 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("fatal error");
                 };
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src1 offset:offs_src1       atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst        atIndex:1];
@@ -2821,6 +3204,49 @@ static void ggml_metal_encode_node(
                     [encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)];
                 }
             } break;
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            {
+                GGML_ASSERT(ggml_is_contiguous(src0));
+                GGML_ASSERT(ggml_is_contiguous(src1));
+                GGML_ASSERT(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F32);
+                GGML_ASSERT(src1->type == GGML_TYPE_F32);
+                GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+                const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
+
+                const int32_t IC = src1->ne[1];
+                const int32_t IL = src1->ne[0];
+
+                const int32_t K  = src0->ne[0];
+
+                const int32_t OL = dst->ne[0];
+                const int32_t OC = dst->ne[1];
+
+                id<MTLComputePipelineState> pipeline;
+
+                switch (src0->type) {
+                    case GGML_TYPE_F32: {
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F32_F32].pipeline;
+                    } break;
+                    case GGML_TYPE_F16: {
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CONV_TRANSPOSE_1D_F16_F32].pipeline;
+                    } break;
+                    default: GGML_ABORT("fatal error");
+                };
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0         atIndex:0];
+                [encoder setBuffer:id_src1 offset:offs_src1         atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst          atIndex:2];
+                [encoder setBytes:&IC      length:sizeof( int32_t)  atIndex:3];
+                [encoder setBytes:&IL      length:sizeof( int32_t)  atIndex:4];
+                [encoder setBytes:&K       length:sizeof( int32_t)  atIndex:5];
+                [encoder setBytes:&s0      length:sizeof( int32_t)  atIndex:6];
+                [encoder setBytes:&nb0     length:sizeof(uint64_t)  atIndex:7];
+                [encoder setBytes:&nb1     length:sizeof(uint64_t)  atIndex:8];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(OL, OC, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+            } break;
         case GGML_OP_UPSCALE:
             {
                 GGML_ASSERT(src0->type == GGML_TYPE_F32);
@@ -2832,6 +3258,7 @@ static void ggml_metal_encode_node(
 
                 const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UPSCALE_F32].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -2866,6 +3293,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_PAD_F32].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -2888,6 +3316,38 @@ static void ggml_metal_encode_node(
 
                 const int nth = MIN(1024, ne0);
 
+                [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_PAD_REFLECT_1D:
+            {
+                GGML_ASSERT(src0->type == GGML_TYPE_F32);
+
+                const int32_t p0 = ((const int32_t *)(dst->op_params))[0];
+                const int32_t p1 = ((const int32_t *)(dst->op_params))[1];
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_PAD_REFLECT_1D_F32].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:6];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
+                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:11];
+                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:12];
+                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:13];
+                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:14];
+                [encoder setBytes:&p0   length:sizeof(p0)   atIndex:15];
+                [encoder setBytes:&p1   length:sizeof(p1)   atIndex:16];
+
+                const int nth = MIN(1024, ne0);
+
                 [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
         case GGML_OP_ARANGE:
@@ -2902,6 +3362,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARANGE_F32].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_dst  offset:offs_dst    atIndex:0];
                 [encoder setBytes:&ne0   length:sizeof(ne0)   atIndex:1];
@@ -2923,6 +3384,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
@@ -2961,6 +3423,7 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("fatal error");
                 };
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
                 [encoder setBuffer:id_dst      offset:offs_dst         atIndex:1];
@@ -2979,6 +3442,7 @@ static void ggml_metal_encode_node(
 
                 id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32].pipeline;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst    atIndex:1];
@@ -3220,37 +3684,41 @@ static void ggml_metal_encode_node(
                     }
                 }
 
+                ggml_metal_kargs_flash_attn_ext args = {
+                    /*.ne01          =*/ ne01,
+                    /*.ne02          =*/ ne02,
+                    /*.ne03          =*/ ne03,
+                    /*.nb01          =*/ nb01,
+                    /*.nb02          =*/ nb02,
+                    /*.nb03          =*/ nb03,
+                    /*.ne11          =*/ ne11,
+                    /*.ne_12_2       =*/ ne12,
+                    /*.ne_12_3       =*/ ne13,
+                    /*.nb_12_1       =*/ nb11,
+                    /*.nb_12_2       =*/ nb12,
+                    /*.nb_12_3       =*/ nb13,
+                    /*.nb31          =*/ nb31,
+                    /*.ne1           =*/ ne1,
+                    /*.ne2           =*/ ne2,
+                    /*.scale         =*/ scale,
+                    /*.max_bias      =*/ max_bias,
+                    /*.m0            =*/ m0,
+                    /*.m1            =*/ m1,
+                    /*.n_head_log2   =*/ n_head_log2,
+                    /*.logit_softcap =*/ logit_softcap,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0     offset:offs_src0           atIndex:0];
-                [encoder setBuffer:id_src1     offset:offs_src1           atIndex:1];
-                [encoder setBuffer:id_src2     offset:offs_src2           atIndex:2];
+                [encoder setBytes:&args length:sizeof(args)     atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0     atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1     atIndex:2];
+                [encoder setBuffer:id_src2 offset:offs_src2     atIndex:3];
                 if (id_src3) {
-                    [encoder setBuffer:id_src3     offset:offs_src3           atIndex:3];
+                    [encoder setBuffer:id_src3 offset:offs_src3 atIndex:4];
                 } else {
-                    [encoder setBuffer:id_src0     offset:offs_src0           atIndex:3];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:4];
                 }
-                [encoder setBuffer:id_dst        offset:offs_dst              atIndex:4];
-                [encoder setBytes:&ne01          length:sizeof( int64_t)      atIndex:5];
-                [encoder setBytes:&ne02          length:sizeof( int64_t)      atIndex:6];
-                [encoder setBytes:&ne03          length:sizeof( int64_t)      atIndex:7];
-                [encoder setBytes:&nb01          length:sizeof(uint64_t)      atIndex:8];
-                [encoder setBytes:&nb02          length:sizeof(uint64_t)      atIndex:9];
-                [encoder setBytes:&nb03          length:sizeof(uint64_t)      atIndex:10];
-                [encoder setBytes:&ne11          length:sizeof( int64_t)      atIndex:11];
-                [encoder setBytes:&ne12          length:sizeof( int64_t)      atIndex:12];
-                [encoder setBytes:&ne13          length:sizeof( int64_t)      atIndex:13];
-                [encoder setBytes:&nb11          length:sizeof(uint64_t)      atIndex:14];
-                [encoder setBytes:&nb12          length:sizeof(uint64_t)      atIndex:15];
-                [encoder setBytes:&nb13          length:sizeof(uint64_t)      atIndex:16];
-                [encoder setBytes:&nb31          length:sizeof(uint64_t)      atIndex:17];
-                [encoder setBytes:&ne1           length:sizeof( int64_t)      atIndex:18];
-                [encoder setBytes:&ne2           length:sizeof( int64_t)      atIndex:19];
-                [encoder setBytes:&scale         length:sizeof(   float)      atIndex:20];
-                [encoder setBytes:&max_bias      length:sizeof(   float)      atIndex:21];
-                [encoder setBytes:&m0            length:sizeof(m0)            atIndex:22];
-                [encoder setBytes:&m1            length:sizeof(m1)            atIndex:23];
-                [encoder setBytes:&n_head_log2   length:sizeof(n_head_log2)   atIndex:24];
-                [encoder setBytes:&logit_softcap length:sizeof(logit_softcap) atIndex:25];
+                [encoder setBuffer:id_dst offset:offs_dst       atIndex:5];
 
                 if (!use_vec_kernel) {
                     // half8x8 kernel
@@ -3385,28 +3853,94 @@ static void ggml_metal_encode_node(
                     default: GGML_ABORT("not implemented");
                 }
 
+                ggml_metal_kargs_cpy args = {
+                    /*.ne00 =*/ ne00,
+                    /*.ne01 =*/ ne01,
+                    /*.ne02 =*/ ne02,
+                    /*.ne03 =*/ ne03,
+                    /*.nb00 =*/ nb00,
+                    /*.nb01 =*/ nb01,
+                    /*.nb02 =*/ nb02,
+                    /*.nb03 =*/ nb03,
+                    /*.ne0  =*/ ne0,
+                    /*.ne1  =*/ ne1,
+                    /*.ne2  =*/ ne2,
+                    /*.ne3  =*/ ne3,
+                    /*.nb0  =*/ nb0,
+                    /*.nb1  =*/ nb1,
+                    /*.nb2  =*/ nb2,
+                    /*.nb3  =*/ nb3,
+                };
+
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
-                [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
-                [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
-                [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
-                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
-                [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
-                [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
-                [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
-                [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
-                [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
-                [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
-                [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
-                [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
-                [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
-                [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
 
                 [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
+        case GGML_OP_SET:
+            {
+                GGML_ASSERT(ggml_are_same_shape(src0, dst));
+                GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
+
+                // src0 and dst as viewed during set
+                const size_t dst_nb0 = ggml_element_size(src0);
+
+                const size_t dst_nb1 = ((int32_t *) dst->op_params)[0];
+                const size_t dst_nb2 = ((int32_t *) dst->op_params)[1];
+                const size_t dst_nb3 = ((int32_t *) dst->op_params)[2];
+                const size_t offset  = ((int32_t *) dst->op_params)[3];
+                const bool   inplace = (bool) ((int32_t *) dst->op_params)[4];
+
+                if (!inplace) {
+                    memcpy(((char *)  dst->data), ((char *) src0->data), ggml_nbytes(dst));
+                }
+
+                const int im0 = (ne10 == 0 ? 0 : ne10-1);
+                const int im1 = (ne11 == 0 ? 0 : ne11-1);
+                const int im2 = (ne12 == 0 ? 0 : ne12-1);
+                const int im3 = (ne13 == 0 ? 0 : ne13-1);
+
+                GGML_ASSERT(offset + im0*dst_nb0  + im1*dst_nb1  + im2*dst_nb2  + im3*dst_nb3  <= ggml_nbytes(dst));
+
+                id<MTLComputePipelineState> pipeline = nil;
+
+                switch (src0t) {
+                    case GGML_TYPE_F32:
+                        GGML_ASSERT(nb10 == sizeof(float));
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_F32].pipeline; break;
+                    case GGML_TYPE_I32:
+                        GGML_ASSERT(nb10 == sizeof(int32_t));
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_I32].pipeline; break;
+                    default: GGML_ABORT("fatal error");
+                }
+
+                ggml_metal_kargs_set args = {
+                    /*.ne10    =*/ ne10,
+                    /*.ne11    =*/ ne11,
+                    /*.ne12    =*/ ne12,
+                    /*.nb10    =*/ nb10,
+                    /*.nb11    =*/ nb11,
+                    /*.nb12    =*/ nb12,
+                    /*.nb13    =*/ nb13,
+                    /*.nb1     =*/ dst_nb1,
+                    /*.nb2     =*/ dst_nb2,
+                    /*.nb3     =*/ dst_nb3,
+                    /*.offs    =*/ offset,
+                    /*.inplace =*/ inplace,
+                };
+
+                const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne10);
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
+                [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne11, ne12, ne13) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
         case GGML_OP_POOL_2D:
             {
                 GGML_ASSERT(ggml_is_contiguous(src0));
@@ -3448,6 +3982,7 @@ static void ggml_metal_encode_node(
                 const int64_t n_threads = MIN((int64_t)[pipeline maxTotalThreadsPerThreadgroup], parallel_elements);
                 const int64_t n_tg = (parallel_elements + n_threads - 1) / n_threads;
 
+                // TODO: add ggml_metal_kargs struct
                 [encoder setComputePipelineState:pipeline];
                 [encoder setBuffer:id_src0 offset:offs_src0       atIndex:0];
                 [encoder setBuffer:id_dst  offset:offs_dst        atIndex:1];
@@ -3465,6 +4000,31 @@ static void ggml_metal_encode_node(
 
                 [encoder dispatchThreadgroups:MTLSizeMake(n_tg, 1, 1) threadsPerThreadgroup:MTLSizeMake(n_threads, 1, 1)];
             } break;
+            case GGML_OP_ARGMAX:
+            {
+                GGML_ASSERT(src0->type == GGML_TYPE_F32);
+                GGML_ASSERT(ggml_is_contiguous_1(src0));
+                GGML_ASSERT(nb00 == ggml_type_size(src0->type));
+
+                const int64_t nrows = ggml_nrows(src0);
+
+                int nth = 32; // SIMD width
+                while (nth < ne00 && nth*ne01*ne02*ne03 < 256) {
+                    nth *= 2;
+                }
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGMAX].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
+                [encoder setThreadgroupMemoryLength:32*sizeof(float)   atIndex:0];
+                [encoder setThreadgroupMemoryLength:32*sizeof(int32_t) atIndex:1];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
        default:
             {
                 GGML_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op));
@@ -3635,6 +4195,12 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
     return ctx->all_data;
 }
 
+static void ggml_backend_metal_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    memset((char *)tensor->data + offset, value, size);
+
+    UNUSED(buffer);
+}
+
 static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     memcpy((char *)tensor->data + offset, data, size);
 
@@ -3667,7 +4233,7 @@ static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_
     /* .free_buffer     = */ ggml_backend_metal_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_metal_buffer_get_base,
     /* .init_tensor     = */ NULL,
-    /* .memset_tensor   = */ NULL,
+    /* .memset_tensor   = */ ggml_backend_metal_buffer_memset_tensor,
     /* .set_tensor      = */ ggml_backend_metal_buffer_set_tensor,
     /* .get_tensor      = */ ggml_backend_metal_buffer_get_tensor,
     /* .cpy_tensor      = */ ggml_backend_metal_buffer_cpy_tensor,
@@ -4264,19 +4830,45 @@ static ggml_backend_dev_t ggml_backend_metal_reg_device_get(ggml_backend_reg_t r
     GGML_UNUSED(index);
 }
 
+static struct ggml_backend_feature g_ggml_backend_metal_features[] = {
+#if defined(GGML_METAL_EMBED_LIBRARY)
+    { "EMBED_LIBRARY", "1" },
+#endif
+#if defined(GGML_METAL_USE_BF16)
+    { "BF16", "1" },
+#endif
+    { nil, nil },
+};
+
+static struct ggml_backend_feature * ggml_backend_metal_get_features(ggml_backend_reg_t reg) {
+    return g_ggml_backend_metal_features;
+
+    GGML_UNUSED(reg);
+}
+
+static void * ggml_backend_metal_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (strcmp(name, "ggml_backend_get_features") == 0) {
+        return (void *)ggml_backend_metal_get_features;
+    }
+
+    return NULL;
+
+    GGML_UNUSED(reg);
+}
 static struct ggml_backend_reg_i ggml_backend_metal_reg_i = {
     /* .get_name         = */ ggml_backend_metal_reg_get_name,
     /* .device_count     = */ ggml_backend_metal_reg_device_count,
     /* .device_get       = */ ggml_backend_metal_reg_device_get,
-    /* .get_proc_address = */ NULL,
+    /* .get_proc_address = */ ggml_backend_metal_get_proc_address,
 };
 
 ggml_backend_reg_t ggml_backend_metal_reg(void) {
     // TODO: make this thread-safe somehow?
     {
         g_ggml_backend_metal_reg = (struct ggml_backend_reg) {
-            /* .iface   = */ ggml_backend_metal_reg_i,
-            /* .context = */ NULL,
+            /* .api_version = */ GGML_BACKEND_API_VERSION,
+            /* .iface       = */ ggml_backend_metal_reg_i,
+            /* .context     = */ NULL,
         };
 
         g_ggml_backend_metal_device = (struct ggml_backend_device) {
@@ -4288,3 +4880,5 @@ ggml_backend_reg_t ggml_backend_metal_reg(void) {
 
     return &g_ggml_backend_metal_reg;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_metal_reg)
diff --git a/ggml/src/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
similarity index 63%
rename from ggml/src/ggml-metal.metal
rename to ggml/src/ggml-metal/ggml-metal.metal
index e8b71a9f883..8ba43904d0c 100644
--- a/ggml/src/ggml-metal.metal
+++ b/ggml/src/ggml-metal/ggml-metal.metal
@@ -1,6 +1,12 @@
 #define GGML_COMMON_DECL_METAL
 #define GGML_COMMON_IMPL_METAL
-#include "ggml-common.h"
+#if defined(GGML_METAL_EMBED_LIBRARY)
+__embed_ggml-common.h__
+#else
+// TODO: this should not be a relative path, but can't figure out how to set Metal include paths in Package.swift
+#include "../ggml-common.h"
+#endif
+#include "ggml-metal-impl.h"
 
 #include <metal_stdlib>
 
@@ -15,8 +21,8 @@ using namespace metal;
 // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
 //
 // cmd:
-//   .../usr/bin/metal -dM -E -c                             ggml/src/ggml-metal.metal
-//   .../usr/bin/metal -dM -E -c -target air64-apple-ios14.0 ggml/src/ggml-metal.metal
+//   .../usr/bin/metal -dM -E -c                             ggml/src/ggml-metal/ggml-metal.metal
+//   .../usr/bin/metal -dM -E -c -target air64-apple-ios14.0 ggml/src/ggml-metal/ggml-metal.metal
 //
 #if __METAL_VERSION__ < 310 && defined(GGML_METAL_USE_BF16)
 #undef GGML_METAL_USE_BF16
@@ -41,6 +47,11 @@ void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg)
     reg = (type4x4)(*src);
 }
 
+template <typename type4>
+void dequantize_f16_t4(device const half4 * src, short il, thread type4 & reg) {
+    reg = (type4)(*(src + il));
+}
+
 #if defined(GGML_METAL_USE_BF16)
 template <typename type4x4>
 void dequantize_bf16(device const bfloat4x4 * src, short il, thread type4x4 & reg) {
@@ -49,7 +60,7 @@ void dequantize_bf16(device const bfloat4x4 * src, short il, thread type4x4 & re
 #endif
 
 template <typename type4x4>
-void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) {
+void dequantize_q4_0(device const block_q4_0 * xb, short il, thread type4x4 & reg) {
     device const uint16_t * qs = ((device const uint16_t *)xb + 1);
     const float d1 = il ? (xb->d / 16.h) : xb->d;
     const float d2 = d1 / 256.f;
@@ -67,8 +78,23 @@ void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg
     reg = (type4x4) reg_f;
 }
 
+template <typename type4>
+void dequantize_q4_0_t4(device const block_q4_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + md;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + md;
+    }
+}
+
 template <typename type4x4>
-void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg) {
+void dequantize_q4_1(device const block_q4_1 * xb, short il, thread type4x4 & reg) {
     device const uint16_t * qs = ((device const uint16_t *)xb + 2);
     const float d1 = il ? (xb->d / 16.h) : xb->d;
     const float d2 = d1 / 256.f;
@@ -86,8 +112,23 @@ void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg
     reg = (type4x4) reg_f;
 }
 
+template <typename type4>
+void dequantize_q4_1_t4(device const block_q4_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = (il/4) ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = (il/4) ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i = 0; i < 2; i++) {
+        reg[2*i + 0] = d1 * (qs[2*(il%4) + i] & mask0) + m;
+        reg[2*i + 1] = d2 * (qs[2*(il%4) + i] & mask1) + m;
+    }
+}
+
 template <typename type4x4>
-void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg) {
+void dequantize_q5_0(device const block_q5_0 * xb, short il, thread type4x4 & reg) {
     device const uint16_t * qs = ((device const uint16_t *)xb + 3);
     const float d = xb->d;
     const float md = -16.h * xb->d;
@@ -118,8 +159,38 @@ void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg
     reg = (type4x4) reg_f;
 }
 
+template <typename type4>
+void dequantize_q5_0_t4(device const block_q5_0 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + md;
+        reg[2*ii + 1] = d * x1 + md;
+    }
+}
+
 template <typename type4x4>
-void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg) {
+void dequantize_q5_1(device const block_q5_1 * xb, short il, thread type4x4 & reg) {
     device const uint16_t * qs = ((device const uint16_t *)xb + 4);
     const float d = xb->d;
     const float m = xb->m;
@@ -150,10 +221,40 @@ void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg
     reg = (type4x4) reg_f;
 }
 
+template <typename type4>
+void dequantize_q5_1_t4(device const block_q5_1 * xb, short il, thread type4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = (il/4) ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = (il/4) ? 4 : 0;
+
+    const int gh_mv = (il/4) ? 12 : 0;
+    const int gh_bk = (il/4) ?  0 : 4;
+
+    for (int ii = 0; ii < 2; ii++) {
+        int i = 2*(il%4) + ii;
+
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[2*ii + 0] = d * x0 + m;
+        reg[2*ii + 1] = d * x1 + m;
+    }
+}
+
 template <typename type4x4>
 void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
     device const int8_t * qs = ((device const int8_t *)xb->qs);
-    const half d = xb->d;
+    const float d = xb->d;
 
     float4x4 reg_f;
 
@@ -164,6 +265,16 @@ void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg
     reg = (type4x4) reg_f;
 }
 
+template <typename type4>
+void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const float d = xb->d;
+
+    for (int i = 0; i < 4; i++) {
+        reg[i] = (qs[4*(il%4) + i + 16*(il/4)] * d);
+    }
+}
+
 template <typename type4x4>
 void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
     const float d = xb->d;
@@ -218,7 +329,7 @@ static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q
 }
 
 template <typename type4x4>
-void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) {
+void dequantize_q4_K(device const block_q4_K * xb, short il, thread type4x4 & reg) {
     device const uchar * q = xb->qs;
 
     short is = (il/4) * 2;
@@ -230,7 +341,7 @@ void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg
     const float dl = d * sc[0];
     const float ml = min * sc[1];
 
-    const ushort mask = il<2 ? 0x0F : 0xF0;
+    const ushort mask = il < 2 ? 0x0F : 0xF0;
     for (int i = 0; i < 16; ++i) {
         reg[i/4][i%4] = dl * (q[i] & mask) - ml;
     }
@@ -463,6 +574,19 @@ void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4
     }
 }
 
+template <typename type4>
+void dequantize_iq4_nl_t4(device const block_iq4_nl * xb, short il, thread type4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    aux32 = ((q4[2*(il%4)] | (q4[2*(il%4)+1] << 16)) >> 4*(il/4)) & 0x0f0f0f0f;
+    reg[0] = d * kvalues_iq4nl_f[q8[0]];
+    reg[1] = d * kvalues_iq4nl_f[q8[1]];
+    reg[2] = d * kvalues_iq4nl_f[q8[2]];
+    reg[3] = d * kvalues_iq4nl_f[q8[3]];
+}
+
 template <typename type4x4>
 void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
     // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
@@ -492,240 +616,131 @@ enum ggml_sort_order {
 // pros: works for non-contiguous tensors, supports broadcast across all dims
 // cons: not very efficient
 kernel void kernel_add(
+        constant ggml_metal_kargs_bin & args,
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        constant  int64_t & ne00,
-        constant  int64_t & ne01,
-        constant  int64_t & ne02,
-        constant  int64_t & ne03,
-        constant uint64_t & nb00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant uint64_t & nb03,
-        constant  int64_t & ne10,
-        constant  int64_t & ne11,
-        constant  int64_t & ne12,
-        constant  int64_t & ne13,
-        constant uint64_t & nb10,
-        constant uint64_t & nb11,
-        constant uint64_t & nb12,
-        constant uint64_t & nb13,
-        constant  int64_t & ne0,
-        constant  int64_t & ne1,
-        constant  int64_t & ne2,
-        constant  int64_t & ne3,
-        constant uint64_t & nb0,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        constant uint64_t & nb3,
-        constant  int64_t & offs,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig.z;
-    const int64_t i02 = tgpig.y;
-    const int64_t i01 = tgpig.x;
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
 
-    const int64_t i13 = i03 % ne13;
-    const int64_t i12 = i02 % ne12;
-    const int64_t i11 = i01 % ne11;
+    const int i13 = i03%args.ne13;
+    const int i12 = i02%args.ne12;
+    const int i11 = i01%args.ne11;
 
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs;
-    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + offs;
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs;
+    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11;
+    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs;
 
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        const int i10 = i0 % ne10;
-        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) + *((device float *)(src1_ptr + i10*nb10));
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i10 = i0%args.ne10;
+        *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) + *((device float *)(src1_ptr + i10*args.nb10));
     }
 }
 
 kernel void kernel_sub(
+        constant ggml_metal_kargs_bin & args,
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        constant  int64_t & ne00,
-        constant  int64_t & ne01,
-        constant  int64_t & ne02,
-        constant  int64_t & ne03,
-        constant uint64_t & nb00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant uint64_t & nb03,
-        constant  int64_t & ne10,
-        constant  int64_t & ne11,
-        constant  int64_t & ne12,
-        constant  int64_t & ne13,
-        constant uint64_t & nb10,
-        constant uint64_t & nb11,
-        constant uint64_t & nb12,
-        constant uint64_t & nb13,
-        constant  int64_t & ne0,
-        constant  int64_t & ne1,
-        constant  int64_t & ne2,
-        constant  int64_t & ne3,
-        constant uint64_t & nb0,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        constant uint64_t & nb3,
-        constant  int64_t & offs,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig.z;
-    const int64_t i02 = tgpig.y;
-    const int64_t i01 = tgpig.x;
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
 
-    const int64_t i13 = i03 % ne13;
-    const int64_t i12 = i02 % ne12;
-    const int64_t i11 = i01 % ne11;
+    const int i13 = i03%args.ne13;
+    const int i12 = i02%args.ne12;
+    const int i11 = i01%args.ne11;
 
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs;
-    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + offs;
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + args.offs;
+    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11;
+    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1  + args.offs;
 
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        const int i10 = i0 % ne10;
-        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) - *((device float *)(src1_ptr + i10*nb10));
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i10 = i0%args.ne10;
+        *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) - *((device float *)(src1_ptr + i10*args.nb10));
     }
 }
 
 kernel void kernel_mul(
+        constant ggml_metal_kargs_bin & args,
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        constant  int64_t & ne00,
-        constant  int64_t & ne01,
-        constant  int64_t & ne02,
-        constant  int64_t & ne03,
-        constant uint64_t & nb00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant uint64_t & nb03,
-        constant  int64_t & ne10,
-        constant  int64_t & ne11,
-        constant  int64_t & ne12,
-        constant  int64_t & ne13,
-        constant uint64_t & nb10,
-        constant uint64_t & nb11,
-        constant uint64_t & nb12,
-        constant uint64_t & nb13,
-        constant  int64_t & ne0,
-        constant  int64_t & ne1,
-        constant  int64_t & ne2,
-        constant  int64_t & ne3,
-        constant uint64_t & nb0,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        constant uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig.z;
-    const int64_t i02 = tgpig.y;
-    const int64_t i01 = tgpig.x;
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
 
-    const int64_t i13 = i03 % ne13;
-    const int64_t i12 = i02 % ne12;
-    const int64_t i11 = i01 % ne11;
+    const int i13 = i03%args.ne13;
+    const int i12 = i02%args.ne12;
+    const int i11 = i01%args.ne11;
 
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
-    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
+    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11;
+    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1;
 
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        const int i10 = i0 % ne10;
-        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) * *((device float *)(src1_ptr + i10*nb10));
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i10 = i0%args.ne10;
+        *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) * *((device float *)(src1_ptr + i10*args.nb10));
     }
 }
 
 kernel void kernel_div(
+        constant ggml_metal_kargs_bin & args,
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        constant  int64_t & ne00,
-        constant  int64_t & ne01,
-        constant  int64_t & ne02,
-        constant  int64_t & ne03,
-        constant uint64_t & nb00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant uint64_t & nb03,
-        constant  int64_t & ne10,
-        constant  int64_t & ne11,
-        constant  int64_t & ne12,
-        constant  int64_t & ne13,
-        constant uint64_t & nb10,
-        constant uint64_t & nb11,
-        constant uint64_t & nb12,
-        constant uint64_t & nb13,
-        constant  int64_t & ne0,
-        constant  int64_t & ne1,
-        constant  int64_t & ne2,
-        constant  int64_t & ne3,
-        constant uint64_t & nb0,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        constant uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig.z;
-    const int64_t i02 = tgpig.y;
-    const int64_t i01 = tgpig.x;
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig.z;
+    const int i02 = tgpig.y;
+    const int i01 = tgpig.x;
 
-    const int64_t i13 = i03 % ne13;
-    const int64_t i12 = i02 % ne12;
-    const int64_t i11 = i01 % ne11;
+    const int i13 = i03%args.ne13;
+    const int i12 = i02%args.ne12;
+    const int i11 = i01%args.ne11;
 
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
-    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
-    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
+    device const char * src1_ptr = src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11;
+    device       char * dst_ptr  = dst  + i03*args.nb3  + i02*args.nb2  + i01*args.nb1;
 
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        const int i10 = i0 % ne10;
-        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) / *((device float *)(src1_ptr + i10*nb10));
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i10 = i0%args.ne10;
+        *((device float *)(dst_ptr + i0*args.nb0)) = *((device float *)(src0_ptr + i0*args.nb00)) / *((device float *)(src1_ptr + i10*args.nb10));
     }
 }
 
 template<typename T>
 kernel void kernel_repeat(
+        constant ggml_metal_kargs_repeat & args,
         device const char * src0,
         device       char * dst,
-        constant  int64_t & ne00,
-        constant  int64_t & ne01,
-        constant  int64_t & ne02,
-        constant  int64_t & ne03,
-        constant uint64_t & nb00,
-        constant uint64_t & nb01,
-        constant uint64_t & nb02,
-        constant uint64_t & nb03,
-        constant  int64_t & ne0,
-        constant  int64_t & ne1,
-        constant  int64_t & ne2,
-        constant  int64_t & ne3,
-        constant uint64_t & nb0,
-        constant uint64_t & nb1,
-        constant uint64_t & nb2,
-        constant uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i3 = tgpig.z;
-    const int64_t i2 = tgpig.y;
-    const int64_t i1 = tgpig.x;
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
 
-    const int64_t i03 = i3 % ne03;
-    const int64_t i02 = i2 % ne02;
-    const int64_t i01 = i1 % ne01;
+    const int i03 = i3%args.ne03;
+    const int i02 = i2%args.ne02;
+    const int i01 = i1%args.ne01;
 
-    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
-    device       char * dst_ptr  = dst  +  i3*nb3  +  i2*nb2  +  i1*nb1 ;
+    device const char * src0_ptr = src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01;
+    device       char * dst_ptr  = dst  +  i3*args.nb3  +  i2*args.nb2  +  i1*args.nb1;
 
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        const int i00 = i0 % ne00;
-        *((device T *)(dst_ptr + i0*nb0)) = *((device T *)(src0_ptr + i00*nb00));
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        const int i00 = i0%args.ne00;
+        *((device T *)(dst_ptr + i0*args.nb0)) = *((device T *)(src0_ptr + i00*args.nb00));
     }
 }
 
@@ -739,38 +754,42 @@ template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat
 // assumption: src1 is a row
 // broadcast src1 into src0
 kernel void kernel_add_row(
+        constant ggml_metal_kargs_bin & args,
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant   uint64_t & nb [[buffer(28)]],
         uint tpig[[thread_position_in_grid]]) {
+    const uint nb = args.ne00/4;
     dst[tpig] = src0[tpig] + src1[tpig % nb];
 }
 
 kernel void kernel_sub_row(
+        constant ggml_metal_kargs_bin & args,
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant   uint64_t & nb [[buffer(28)]],
         uint tpig[[thread_position_in_grid]]) {
+    const uint nb = args.ne00/4;
     dst[tpig] = src0[tpig] - src1[tpig % nb];
 }
 
 kernel void kernel_mul_row(
+        constant ggml_metal_kargs_bin & args,
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant   uint64_t & nb  [[buffer(28)]],
         uint tpig[[thread_position_in_grid]]) {
+    const uint nb = args.ne00/4;
     dst[tpig] = src0[tpig] * src1[tpig % nb];
 }
 
 kernel void kernel_div_row(
+        constant ggml_metal_kargs_bin & args,
         device const float4 * src0,
         device const float4 * src1,
         device       float4 * dst,
-        constant   uint64_t & nb  [[buffer(28)]],
         uint tpig[[thread_position_in_grid]]) {
+    const uint nb = args.ne00/4;
     dst[tpig] = src0[tpig] / src1[tpig % nb];
 }
 
@@ -881,6 +900,14 @@ kernel void kernel_silu_4(
     dst[tpig] = x / (1.0f + exp(-x));
 }
 
+kernel void kernel_elu(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+    dst[tpig] = (x > 0.0f) ? x : (exp(x) - 1.0f);
+}
+
 kernel void kernel_sqr(
         device const float * src0,
         device       float * dst,
@@ -1339,103 +1366,170 @@ kernel void kernel_ssm_scan_f32(
     }
 }
 
-kernel void kernel_norm(
-        device const  void * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant  uint64_t & nb01,
-        constant     float & eps,
-        threadgroup float  * sum [[threadgroup(0)]],
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    device const float * x = (device const float *) ((device const char *) src0 + tgpig*nb01);
-    // MEAN
-    // parallel sum
-    sum[tpitg] = 0.0f;
-    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        sum[tpitg] += x[i00];
-    }
-    // reduce
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    for (uint i = ntg/2; i > 0; i /= 2) {
-        if (tpitg < i) {
-            sum[tpitg] += sum[tpitg + i];
+kernel void kernel_argmax(
+        device   const void * x,
+        device      int32_t * dst,
+        constant    int64_t & ncols,
+        constant   uint64_t & nb01,
+        threadgroup   float * shared_maxval [[threadgroup(0)]],
+        threadgroup int32_t * shared_argmax [[threadgroup(1)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    device const float * x_row = (device const float *) ((device const char *) x + tgpig * nb01);
+
+    float   lmax = -INFINITY;
+    int32_t larg = -1;
+
+    for (int i00 = tpitg; i00 < ncols; i00 += ntg) {
+        if (x_row[i00] > lmax) {
+            lmax = x_row[i00];
+            larg = i00;
         }
-        threadgroup_barrier(mem_flags::mem_threadgroup);
     }
-    const float mean  = sum[0] / ne00;
 
-    // recenter and VARIANCE
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    device float * y = dst + tgpig*ne00;
-    sum[tpitg] = 0.0f;
-    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        y[i00] = x[i00] - mean;
-        sum[tpitg] += y[i00] * y[i00];
-    }
+    // find the argmax value in the block
+    float max_val = simd_max(lmax);
+    int32_t arg_val = simd_max(select(-1, larg, lmax == max_val));
 
-    // reduce
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-    for (uint i = ntg/2; i > 0; i /= 2) {
-        if (tpitg < i) {
-            sum[tpitg] += sum[tpitg + i];
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            shared_maxval[tiisg] = -INFINITY;
+            shared_argmax[tiisg] = -1;
         }
+
         threadgroup_barrier(mem_flags::mem_threadgroup);
-    }
-    const float variance = sum[0] / ne00;
 
-    const float scale = 1.0f/sqrt(variance + eps);
-    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
-        y[i00] = y[i00] * scale;
-    }
-}
+        if (tiisg == 0) {
+            shared_maxval[sgitg] = max_val;
+            shared_argmax[sgitg] = arg_val;
+        }
 
-kernel void kernel_rms_norm(
-        device const  void * src0,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant  uint64_t & nb01,
-        constant     float & eps,
-        threadgroup float  * buf [[threadgroup(0)]],
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    device const float4 * x = (device const float4 *) ((device const char *) src0 + tgpig*nb01);
+        threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    float4 sumf = 0;
-    float all_sum = 0;
+        max_val = shared_maxval[tiisg];
+        arg_val = shared_argmax[tiisg];
 
-    // parallel sum
-    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
-        sumf += x[i00] * x[i00];
+        float max_val_reduced   = simd_max(max_val);
+        int32_t arg_val_reduced = simd_max(select(-1, arg_val, max_val == max_val_reduced));
+
+        dst[tgpig] = arg_val_reduced;
+
+        return;
     }
-    all_sum = sumf[0] + sumf[1] + sumf[2] + sumf[3];
-    all_sum = simd_sum(all_sum);
-    if (ntg > N_SIMDWIDTH) {
-        if (sgitg == 0) {
-            buf[tiisg] = 0.0f;
-        }
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
+    dst[tgpig] = arg_val;
+}
 
-        if (tiisg == 0) {
-            buf[sgitg] = all_sum;
-        }
+kernel void kernel_norm(
+        constant ggml_metal_kargs_norm & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint   tgpig[[threadgroup_position_in_grid]],
+        ushort tpitg[[thread_position_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
+    device const float4 * x = (device const float4 *) (src0 + tgpig*args.nb01);
+
+    float4 sumf4(0.0f);
 
-        all_sum = buf[tiisg];
-        all_sum = simd_sum(all_sum);
+    float sumf = 0.0f;
+
+    for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) {
+        sumf4 += x[i00];
     }
+    sumf = sumf4[0] + sumf4[1] + sumf4[2] + sumf4[3];
+    sumf = simd_sum(sumf);
 
-    const float mean  = all_sum/ne00;
-    const float scale = 1.0f/sqrt(mean + eps);
+    threadgroup_barrier(mem_flags::mem_threadgroup);
 
-    device float4 * y = (device float4 *) (dst + tgpig*ne00);
-    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean = sumf/args.ne00;
+
+    device float4 * y = (device float4 *) dst + tgpig*args.ne00_4;
+
+    sumf = 0.0f;
+    for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) {
+        y[i00] = x[i00] - mean;
+        sumf += dot(y[i00], y[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float variance = sumf/args.ne00;
+
+    const float scale = 1.0f/sqrt(variance + args.eps);
+    for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) {
+        y[i00] = y[i00] * scale;
+    }
+}
+
+kernel void kernel_rms_norm(
+        constant ggml_metal_kargs_rms_norm & args,
+        device const char * src0,
+        device       char * dst,
+        threadgroup float * shmem_f32 [[threadgroup(0)]],
+        uint   tgpig[[threadgroup_position_in_grid]],
+        ushort tpitg[[thread_position_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort   ntg[[threads_per_threadgroup]]) {
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
+    device const float4 * x = (device const float4 *) (src0 + tgpig*args.nb01);
+
+    float sumf = 0.0f;
+
+    // parallel sum
+    for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    const float mean  = sumf/args.ne00;
+    const float scale = 1.0f/sqrt(mean + args.eps);
+
+    device float4 * y = (device float4 *) dst + tgpig*args.ne00_4;
+    for (int i00 = tpitg; i00 < args.ne00_4; i00 += ntg) {
         y[i00] = x[i00] * scale;
     }
 }
@@ -1623,31 +1717,17 @@ inline float block_q_n_dot_y(device const block_q5_1 * qb_curr, float sumy, thre
 //      quantizations where the block size is 32. It also does not
 //      guard against the number of rows not being divisible by
 //      N_DST, so this is another explicit assumption of the implementation.
-template<typename block_q_type, int nr, int nsg, int nw>
+template<typename block_q_type, int nr, int nsg, int nw, typename args_t>
 void mul_vec_q_n_f32_impl(
-        device const void  * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                     uint3   tgpig,
-                     uint    tiisg,
-                     uint    sgitg) {
-    const int nb = ne00/QK4_0;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+    const int nb = args.ne00/QK4_0;
 
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
@@ -1655,19 +1735,19 @@ void mul_vec_q_n_f32_impl(
 
     const int first_row = (r0 * nsg + sgitg) * nr;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-  //const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+  //const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-  //device const block_q_type * x = (device const block_q_type *) ((device char *) src0 + offset0);
-    device const float        * y = (device const float        *) ((device char *) src1 + offset1);
+  //device const block_q_type * x = (device const block_q_type *) (src0 + offset0);
+    device const float        * y = (device const float        *) (src1 + offset1);
 
     // pointers to src0 rows
     device const block_q_type * ax[nr];
     for (int row = 0; row < nr; ++row) {
-        const uint offset0 = (first_row + row)*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+        const uint64_t offset0 = (first_row + row)*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
 
         ax[row] = (device const block_q_type *) ((device char *) src0 + offset0);
     }
@@ -1675,10 +1755,10 @@ void mul_vec_q_n_f32_impl(
     float yl[16]; // src1 vector cache
     float sumf[nr] = {0.f};
 
-    const int ix = (tiisg/2);
-    const int il = (tiisg%2)*8;
+    const short ix = (tiisg/2);
+    const short il = (tiisg%2)*8;
 
-    device const float * yb = y + ix * QK4_0 + il;
+    device const float * yb = y + ix*QK4_0 + il;
 
     // each thread in a SIMD group deals with half a block.
     for (int ib = ix; ib < nb; ib += nw/2) {
@@ -1703,324 +1783,511 @@ void mul_vec_q_n_f32_impl(
         yb += QK4_0 * 16;
     }
 
+    device float * dst_f32 = (device float *) dst + im*args.ne0*args.ne1 + r1*args.ne0;
+
     for (int row = 0; row < nr; ++row) {
         const float tot = simd_sum(sumf[row]);
-        if (tiisg == 0 && first_row + row < ne01) {
-            dst[im*ne0*ne1 + r1*ne0 + first_row + row] = tot;
+
+        if (tiisg == 0 && first_row + row < args.ne01) {
+            dst_f32[first_row + row] = tot;
         }
     }
 }
 
 kernel void kernel_mul_mv_q4_0_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,nb01,nb02,nb03,ne10,ne12,nb11,nb12,nb13,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
 kernel void kernel_mul_mv_q4_1_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
-     mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,nb01,nb02,nb03,ne10,ne12,nb11,nb12,nb13,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+     mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
 kernel void kernel_mul_mv_q5_0_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,nb01,nb02,nb03,ne10,ne12,nb11,nb12,nb13,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
 kernel void kernel_mul_mv_q5_1_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,nb01,nb02,nb03,ne10,ne12,nb11,nb12,nb13,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
-
 #define NB_Q8_0 8
 
+template<typename args_t>
 void kernel_mul_mv_q8_0_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
     const int nr  = N_DST;
     const int nsg = N_SIMDGROUP;
     const int nw  = N_SIMDWIDTH;
 
-    const int nb = ne00/QK8_0;
+    const int nb = args.ne00/QK8_0;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
-    const int first_row = (r0 * nsg + sgitg) * nr;
+    const int first_row = (r0*nsg + sgitg)*nr;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-  //const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+  //const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-  //device const block_q8_0 * x = (device const block_q8_0 *) ((device char *) src0 + offset0);
-    device const float      * y = (device const float      *) ((device char *) src1 + offset1);
+  //device const block_q8_0 * x = (device const block_q8_0 *) (src0 + offset0);
+    device const float      * y = (device const float      *) (src1 + offset1);
 
     // pointers to src0 rows
     device const block_q8_0 * ax[nr];
     for (int row = 0; row < nr; ++row) {
-        const uint offset0 = (first_row + row)*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+        const uint64_t offset0 = (first_row + row)*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
 
         ax[row] = (device const block_q8_0 *) ((device char *) src0 + offset0);
     }
 
     float yl[NB_Q8_0];
-    float sumf[nr]={0.f};
+    float sumf[nr] = { 0.f };
 
-    const int ix = tiisg/4;
-    const int il = tiisg%4;
+    const short ix = tiisg/4;
+    const short il = tiisg%4;
 
-    device const float * yb = y + ix * QK8_0 + NB_Q8_0*il;
+    device const float * yb = y + ix*QK8_0 + il*NB_Q8_0;
 
     // each thread in a SIMD group deals with NB_Q8_0 quants at a time
     for (int ib = ix; ib < nb; ib += nw/4) {
-        for (int i = 0; i < NB_Q8_0; ++i) {
+        for (short i = 0; i < NB_Q8_0; ++i) {
             yl[i] = yb[i];
         }
 
         for (int row = 0; row < nr; row++) {
-            device const int8_t * qs = ax[row][ib].qs + NB_Q8_0*il;
+            device const int8_t * qs = ax[row][ib].qs + il*NB_Q8_0;
             float sumq = 0.f;
-            for (int iq = 0; iq < NB_Q8_0; ++iq) {
+            for (short iq = 0; iq < NB_Q8_0; ++iq) {
                 sumq += qs[iq] * yl[iq];
             }
             sumf[row] += sumq*ax[row][ib].d;
         }
 
-        yb += NB_Q8_0 * nw;
+        yb += nw*NB_Q8_0;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < nr; ++row) {
         const float tot = simd_sum(sumf[row]);
-        if (tiisg == 0 && first_row + row < ne01) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
+
+        if (tiisg == 0 && first_row + row < args.ne01) {
+            dst_f32[first_row + row] = tot;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_q8_0_f32")]]
 kernel void kernel_mul_mv_q8_0_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-    kernel_mul_mv_q8_0_f32_impl(src0,src1,dst,ne00,ne01,ne02,nb01,nb02,nb03,ne10,ne12,nb11,nb12,nb13,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    kernel_mul_mv_q8_0_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
+}
+
+// mat-vec kernel processing in chunks of float4
+// chpb - chunks per quantization block
+template<short nxpsg, short r1ptg, typename q_t, short chpb, void (*deq_t4)(device const q_t *, short, thread float4 &) >
+void kernel_mul_mv_ext_q4_f32_impl(
+        constant ggml_metal_kargs_mul_mv_ext & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]]) {
+    const short chpt = 4; // chunks per thread
+
+  //const short nxpsg = (32);
+    const short nypsg = (32/nxpsg);
+
+    const short tx = tiisg%nxpsg;
+    const short ty = tiisg/nxpsg;
+
+    const int i01 = tgpig.x*(nypsg*args.nsg) + nypsg*sgitg + ty;
+    const int i11 = tgpig.y*r1ptg;
+    const int i1m = tgpig.z;
+
+    const int i12 = i1m%args.ne12;
+    const int i13 = i1m/args.ne12;
+
+    const uint64_t offset0 = i01*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 = i11*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
+
+    device const q_t * xq = (i01 < args.ne01) ? (device const q_t *) (src0 + offset0) + tx/chpb : (device const q_t *) src0;
+
+    device const float4 * y4[r1ptg];
+
+    for (int ir1 = 0; ir1 < r1ptg; ++ir1) {
+        y4[ir1] = (i11 + ir1 < args.ne11) ? (device const float4 *) (src1 + offset1 + ir1*args.nb11) + tx : (device const float4 *) src1;
+    }
+
+    float sumf[r1ptg] = { [ 0 ... r1ptg - 1 ] = 0.0f };
+
+    short cch = tx%chpb; // current chunk index
+
+    for (int ich = tx; 4*ich < args.ne00; ich += chpt*nxpsg) {
+        float4 lx[chpt];
+
+#pragma unroll(chpt)
+        for (short ch = 0; ch < chpt; ++ch) {
+            deq_t4(xq, cch, lx[ch]);
+
+            cch += nxpsg;
+            if (cch >= chpb) {
+                xq  += cch/chpb;
+                cch %= chpb;
+            }
+        }
+
+#pragma unroll(chpt)
+        for (short ch = 0; ch < chpt; ++ch) {
+#pragma unroll(r1ptg)
+            for (short ir1 = 0; ir1 < r1ptg; ++ir1) {
+                sumf[ir1] += dot(lx[ch], y4[ir1][ch*nxpsg]);
+
+            }
+        }
+
+#pragma unroll(r1ptg)
+        for (short ir1 = 0; ir1 < r1ptg; ++ir1) {
+            y4[ir1] += chpt*nxpsg;
+        }
+    }
+
+    // reduce only the threads in each row
+    for (short ir1 = 0; ir1 < r1ptg; ++ir1) {
+        if (nxpsg >= 32) {
+            sumf[ir1] += simd_shuffle_down(sumf[ir1], 16);
+        }
+        if (nxpsg >= 16) {
+            sumf[ir1] += simd_shuffle_down(sumf[ir1],  8);
+        }
+        if (nxpsg >= 8) {
+            sumf[ir1] += simd_shuffle_down(sumf[ir1],  4);
+        }
+        if (nxpsg >= 4) {
+            sumf[ir1] += simd_shuffle_down(sumf[ir1],  2);
+        }
+        if (nxpsg >= 2) {
+            sumf[ir1] += simd_shuffle_down(sumf[ir1],  1);
+        }
+
+        //sumf[ir1] = simd_sum(sumf[ir1]);
+    }
+
+    if (tx == 0) {
+        for (short ir1 = 0; ir1 < r1ptg && i11 + ir1 < args.ne11; ++ir1) {
+            device float * dst_f32 = (device float *) dst + (uint64_t)i1m*args.ne0*args.ne1 + (uint64_t)(i11 + ir1)*args.ne0;
+
+            if (i01 < args.ne01) {
+                dst_f32[i01] = sumf[ir1];
+            }
+        }
+    }
 }
 
+// mat-vec kernel processing in chunks of float4x4
+template<short nxpsg, short r1ptg, typename q_t, short chpb, void (*deq_t4x4)(device const q_t *, short, thread float4x4 &) >
+void kernel_mul_mv_ext_q4x4_f32_impl(
+        constant ggml_metal_kargs_mul_mv_ext & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]]) {
+    const short chpt = 1;
+
+  //const short nxpsg = (32);
+    const short nypsg = (32/nxpsg);
+
+    const short tx = tiisg%nxpsg;
+    const short ty = tiisg/nxpsg;
+
+    const int i01 = tgpig.x*(nypsg*args.nsg) + nypsg*sgitg + ty;
+    const int i11 = tgpig.y*r1ptg;
+    const int i1m = tgpig.z;
+
+    const int i12 = i1m%args.ne12;
+    const int i13 = i1m/args.ne12;
+
+    const uint64_t offset0 = i01*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 = i11*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
+
+    device const q_t * xq = (i01 < args.ne01) ? (device const q_t *) (src0 + offset0) + tx/chpb : (device const q_t *) src0;
+
+    device const float4x4 * y4x4[r1ptg];
+
+    for (int ir1 = 0; ir1 < r1ptg; ++ir1) {
+        y4x4[ir1] = (i11 + ir1 < args.ne11) ? (device const float4x4 *) (src1 + offset1 + ir1*args.nb11) + tx : (device const float4x4 *) src1;
+    }
+
+    float sumf[r1ptg] = { [ 0 ... r1ptg - 1 ] = 0.0f };
+
+    short cch = tx%chpb;
+
+    for (int ich = tx; 16*ich < args.ne00; ich += chpt*nxpsg) {
+        float4x4 lx[chpt];
+
+#pragma unroll(chpt)
+        for (short ch = 0; ch < chpt; ++ch) {
+            deq_t4x4(xq, cch, lx[ch]);
+
+            cch += nxpsg;
+            if (cch >= chpb) {
+                xq  += cch/chpb;
+                cch %= chpb;
+            }
+        }
+
+#pragma unroll(chpt)
+        for (short ch = 0; ch < chpt; ++ch) {
+#pragma unroll(r1ptg)
+            for (short ir1 = 0; ir1 < r1ptg; ++ir1) {
+                sumf[ir1] +=
+                    dot(lx[ch][0], y4x4[ir1][ch*nxpsg][0]) +
+                    dot(lx[ch][1], y4x4[ir1][ch*nxpsg][1]) +
+                    dot(lx[ch][2], y4x4[ir1][ch*nxpsg][2]) +
+                    dot(lx[ch][3], y4x4[ir1][ch*nxpsg][3]);
+
+            }
+        }
+
+#pragma unroll(r1ptg)
+        for (short ir1 = 0; ir1 < r1ptg; ++ir1) {
+            y4x4[ir1] += chpt*nxpsg;
+        }
+    }
+
+    for (short ir1 = 0; ir1 < r1ptg; ++ir1) {
+        if (nxpsg >= 32) {
+            sumf[ir1] += simd_shuffle_down(sumf[ir1], 16);
+        }
+        if (nxpsg >= 16) {
+            sumf[ir1] += simd_shuffle_down(sumf[ir1],  8);
+        }
+        if (nxpsg >= 8) {
+            sumf[ir1] += simd_shuffle_down(sumf[ir1],  4);
+        }
+        if (nxpsg >= 4) {
+            sumf[ir1] += simd_shuffle_down(sumf[ir1],  2);
+        }
+        if (nxpsg >= 2) {
+            sumf[ir1] += simd_shuffle_down(sumf[ir1],  1);
+        }
+
+        //sumf[ir1] = simd_sum(sumf[ir1]);
+    }
+
+    if (tx == 0) {
+        for (short ir1 = 0; ir1 < r1ptg && i11 + ir1 < args.ne11; ++ir1) {
+            device float * dst_f32 = (device float *) dst + (uint64_t)i1m*args.ne0*args.ne1 + (uint64_t)(i11 + ir1)*args.ne0;
+
+            if (i01 < args.ne01) {
+                dst_f32[i01] = sumf[ir1];
+            }
+        }
+    }
+}
+
+// dispatchers needed for compile-time nxpsg
+// epb - elements per quantization block
+template<short r1ptg, typename q_t, short epb, void (*deq_t4)(device const q_t *, short, thread float4 &)>
+kernel void kernel_mul_mv_ext_q4_f32_disp(
+        constant ggml_metal_kargs_mul_mv_ext & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]]) {
+    switch (args.nxpsg) {
+        case 4:  kernel_mul_mv_ext_q4_f32_impl<4,  r1ptg, q_t, epb/4, deq_t4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break;
+        case 8:  kernel_mul_mv_ext_q4_f32_impl<8,  r1ptg, q_t, epb/4, deq_t4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break;
+        case 16: kernel_mul_mv_ext_q4_f32_impl<16, r1ptg, q_t, epb/4, deq_t4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break;
+        case 32: kernel_mul_mv_ext_q4_f32_impl<32, r1ptg, q_t, epb/4, deq_t4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break;
+    }
+}
+
+template<short r1ptg, typename q_t, short epb, void (*deq_t4x4)(device const q_t *, short, thread float4x4 &)>
+kernel void kernel_mul_mv_ext_q4x4_f32_disp(
+        constant ggml_metal_kargs_mul_mv_ext & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]]) {
+    switch (args.nxpsg) {
+        case 4:  kernel_mul_mv_ext_q4x4_f32_impl<4,  r1ptg, q_t, epb/16, deq_t4x4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break;
+        case 8:  kernel_mul_mv_ext_q4x4_f32_impl<8,  r1ptg, q_t, epb/16, deq_t4x4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break;
+        case 16: kernel_mul_mv_ext_q4x4_f32_impl<16, r1ptg, q_t, epb/16, deq_t4x4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break;
+        case 32: kernel_mul_mv_ext_q4x4_f32_impl<32, r1ptg, q_t, epb/16, deq_t4x4>(args, src0, src1, dst, tgpig, tiisg, sgitg); break;
+    }
+}
+
+typedef decltype(kernel_mul_mv_ext_q4_f32_disp  <2, block_q8_0, 32,  dequantize_q8_0_t4>) mul_mv_ext_q4_f32_t;
+typedef decltype(kernel_mul_mv_ext_q4x4_f32_disp<2, block_q4_K, 256, dequantize_q4_K>)    mul_mv_ext_q4x4_f32_t;
+
+template [[host_name("kernel_mul_mv_ext_f16_f32_r1_2")]]    kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, half4,        4,  dequantize_f16_t4>;
+template [[host_name("kernel_mul_mv_ext_f16_f32_r1_3")]]    kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, half4,        4,  dequantize_f16_t4>;
+template [[host_name("kernel_mul_mv_ext_f16_f32_r1_4")]]    kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, half4,        4,  dequantize_f16_t4>;
+template [[host_name("kernel_mul_mv_ext_f16_f32_r1_5")]]    kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, half4,        4,  dequantize_f16_t4>;
+
+template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_2")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_q4_0,   32, dequantize_q4_0_t4>;
+template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_3")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_q4_0,   32, dequantize_q4_0_t4>;
+template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_4")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q4_0,   32, dequantize_q4_0_t4>;
+template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_5")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_q4_0,   32, dequantize_q4_0_t4>;
+
+template [[host_name("kernel_mul_mv_ext_q4_1_f32_r1_2")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_q4_1,   32, dequantize_q4_1_t4>;
+template [[host_name("kernel_mul_mv_ext_q4_1_f32_r1_3")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_q4_1,   32, dequantize_q4_1_t4>;
+template [[host_name("kernel_mul_mv_ext_q4_1_f32_r1_4")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q4_1,   32, dequantize_q4_1_t4>;
+template [[host_name("kernel_mul_mv_ext_q4_1_f32_r1_5")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_q4_1,   32, dequantize_q4_1_t4>;
+
+template [[host_name("kernel_mul_mv_ext_q5_0_f32_r1_2")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_q5_0,   32, dequantize_q5_0_t4>;
+template [[host_name("kernel_mul_mv_ext_q5_0_f32_r1_3")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_q5_0,   32, dequantize_q5_0_t4>;
+template [[host_name("kernel_mul_mv_ext_q5_0_f32_r1_4")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q5_0,   32, dequantize_q5_0_t4>;
+template [[host_name("kernel_mul_mv_ext_q5_0_f32_r1_5")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_q5_0,   32, dequantize_q5_0_t4>;
+
+template [[host_name("kernel_mul_mv_ext_q5_1_f32_r1_2")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_q5_1,   32, dequantize_q5_1_t4>;
+template [[host_name("kernel_mul_mv_ext_q5_1_f32_r1_3")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_q5_1,   32, dequantize_q5_1_t4>;
+template [[host_name("kernel_mul_mv_ext_q5_1_f32_r1_4")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q5_1,   32, dequantize_q5_1_t4>;
+template [[host_name("kernel_mul_mv_ext_q5_1_f32_r1_5")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_q5_1,   32, dequantize_q5_1_t4>;
+
+template [[host_name("kernel_mul_mv_ext_q8_0_f32_r1_2")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_q8_0,   32, dequantize_q8_0_t4>;
+template [[host_name("kernel_mul_mv_ext_q8_0_f32_r1_3")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_q8_0,   32, dequantize_q8_0_t4>;
+template [[host_name("kernel_mul_mv_ext_q8_0_f32_r1_4")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q8_0,   32, dequantize_q8_0_t4>;
+template [[host_name("kernel_mul_mv_ext_q8_0_f32_r1_5")]]   kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_q8_0,   32, dequantize_q8_0_t4>;
+
+template [[host_name("kernel_mul_mv_ext_iq4_nl_f32_r1_2")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_iq4_nl, 32, dequantize_iq4_nl_t4>;
+template [[host_name("kernel_mul_mv_ext_iq4_nl_f32_r1_3")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_iq4_nl, 32, dequantize_iq4_nl_t4>;
+template [[host_name("kernel_mul_mv_ext_iq4_nl_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_iq4_nl, 32, dequantize_iq4_nl_t4>;
+template [[host_name("kernel_mul_mv_ext_iq4_nl_f32_r1_5")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, block_iq4_nl, 32, dequantize_iq4_nl_t4>;
+
+template [[host_name("kernel_mul_mv_ext_q4_K_f32_r1_2")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<2, block_q4_K, 256, dequantize_q4_K>;
+template [[host_name("kernel_mul_mv_ext_q4_K_f32_r1_3")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<3, block_q4_K, 256, dequantize_q4_K>;
+template [[host_name("kernel_mul_mv_ext_q4_K_f32_r1_4")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<4, block_q4_K, 256, dequantize_q4_K>;
+template [[host_name("kernel_mul_mv_ext_q4_K_f32_r1_5")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<5, block_q4_K, 256, dequantize_q4_K>;
+
+template [[host_name("kernel_mul_mv_ext_q5_K_f32_r1_2")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<2, block_q5_K, 256, dequantize_q5_K>;
+template [[host_name("kernel_mul_mv_ext_q5_K_f32_r1_3")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<3, block_q5_K, 256, dequantize_q5_K>;
+template [[host_name("kernel_mul_mv_ext_q5_K_f32_r1_4")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<4, block_q5_K, 256, dequantize_q5_K>;
+template [[host_name("kernel_mul_mv_ext_q5_K_f32_r1_5")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<5, block_q5_K, 256, dequantize_q5_K>;
+
+template [[host_name("kernel_mul_mv_ext_q6_K_f32_r1_2")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<2, block_q6_K, 256, dequantize_q6_K>;
+template [[host_name("kernel_mul_mv_ext_q6_K_f32_r1_3")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<3, block_q6_K, 256, dequantize_q6_K>;
+template [[host_name("kernel_mul_mv_ext_q6_K_f32_r1_4")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<4, block_q6_K, 256, dequantize_q6_K>;
+template [[host_name("kernel_mul_mv_ext_q6_K_f32_r1_5")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<5, block_q6_K, 256, dequantize_q6_K>;
+
 #define N_MV_T_T 4
 
-template<typename T0, typename T04, typename T1, typename T14>
+template<typename T0, typename T04, typename T1, typename T14, typename args_t>
 void kernel_mul_mv_impl(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb00,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne11,
-                   int64_t   ne12,
-                  uint64_t   nb10,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-                   uint3     tgpig,
-                   uint      tiisg) {
-    const int64_t r0 = tgpig.x;
-    const int64_t rb = tgpig.y*N_MV_T_T;
-    const int64_t im = tgpig.z;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig,
+        ushort tiisg) {
+    const int r0 = tgpig.x;
+    const int rb = tgpig.y*N_MV_T_T;
+    const int im = tgpig.z;
+
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
+
+    const uint64_t offset0 = r0*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
 
     device const T0 * x = (device const T0 *) (src0 + offset0);
 
-    if (ne00 < 128) {
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1;
+
+    if (args.ne00 < 128) {
         for (int row = 0; row < N_MV_T_T; ++row) {
             int r1 = rb + row;
-            if (r1 >= ne11) {
+            if (r1 >= args.ne11) {
                 break;
             }
 
-            const uint offset1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+            const uint64_t offset1 = r1*args.nb11 + (i12   )*args.nb12 + (i13   )*args.nb13;
 
             device const T1 * y = (device const T1 *) (src1 + offset1);
 
             float sumf = 0;
-            for (int i = tiisg; i < ne00; i += 32) {
+            for (int i = tiisg; i < args.ne00; i += 32) {
                 sumf += (T0) x[i] * (T1) y[i];
             }
 
             float all_sum = simd_sum(sumf);
             if (tiisg == 0) {
-                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+                dst_f32[(uint64_t)r1*args.ne0 + r0] = all_sum;
             }
         }
     } else {
         device const T04 * x4 = (device const T04 *) x;
         for (int row = 0; row < N_MV_T_T; ++row) {
             int r1 = rb + row;
-            if (r1 >= ne11) {
+            if (r1 >= args.ne11) {
                 break;
             }
 
-            const uint offset1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+            const uint64_t offset1 = r1*args.nb11 + (i12   )*args.nb12 + (i13   )*args.nb13;
 
             device const T1  * y  = (device const T1  *) (src1 + offset1);
             device const T14 * y4 = (device const T14 *) y;
 
             float sumf = 0;
-            for (int i = tiisg; i < ne00/4; i += 32) {
-                for (int k = 0; k < 4; ++k) sumf += (float) (x4[i][k] * y4[i][k]);
+            for (int i = tiisg; i < args.ne00/4; i += 32) {
+                sumf += dot((float4) x4[i], (float4) y4[i]);
             }
 
             float all_sum = simd_sum(sumf);
             if (tiisg == 0) {
-                for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) (x[i] * y[i]);
-                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+                for (int i = 4*(args.ne00/4); i < args.ne00; ++i) all_sum += (float) (x[i] * y[i]);
+                dst_f32[(uint64_t)r1*args.ne0 + r0] = all_sum;
             }
         }
     }
@@ -2028,51 +2295,17 @@ void kernel_mul_mv_impl(
 
 template<typename T0, typename T04, typename T1, typename T14>
 kernel void kernel_mul_mv(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]]) {
-    kernel_mul_mv_impl<T0, T04, T1, T14>(
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]]) {
+    kernel_mul_mv_impl<T0, T04, T1, T14, constant ggml_metal_kargs_mul_mv &>(
+        args,
         src0,
         src1,
         dst,
-        ne00,
-        ne01,
-        ne02,
-        nb00,
-        nb01,
-        nb02,
-        nb03,
-        ne10,
-        ne11,
-        ne12,
-        nb10,
-        nb11,
-        nb12,
-        nb13,
-        ne0,
-        ne1,
-        r2,
-        r3,
         tgpig,
         tiisg);
 }
@@ -2089,65 +2322,50 @@ template [[host_name("kernel_mul_mv_bf16_bf16")]] kernel mul_mv_t kernel_mul_mv<
 
 template<typename T, typename T4>
 kernel void kernel_mul_mv_1row(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]]) {
 
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-    const int64_t im = tgpig.z;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = r0*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 = r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
     device const T     * x = (device const T     *) (src0 + offset0);
     device const float * y = (device const float *) (src1 + offset1);
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     float sumf = 0;
-    if (ne00 < 128) {
-        for (int i = tiisg; i < ne00; i += 32) {
+    if (args.ne00 < 128) {
+        for (int i = tiisg; i < args.ne00; i += 32) {
             sumf += (float) x[i] * (float) y[i];
         }
         float all_sum = simd_sum(sumf);
         if (tiisg == 0) {
-            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            dst_f32[r0] = all_sum;
         }
     } else {
         device const T4     * x4 = (device const T4     *) x;
         device const float4 * y4 = (device const float4 *) y;
 
-        for (int i = tiisg; i < ne00/4; i += 32) {
-            for (int k = 0; k < 4; ++k) sumf += (float) (x4[i][k] * y4[i][k]);
+        for (int i = tiisg; i < args.ne00/4; i += 32) {
+            sumf += dot((float4) x4[i], y4[i]);
         }
 
         float all_sum = simd_sum(sumf);
 
         if (tiisg == 0) {
-            for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) (x[i] * y[i]);
-            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            for (int i = 4*(args.ne00/4); i < args.ne00; ++i) all_sum += (float) (x[i] * y[i]);
+            dst_f32[r0] = all_sum;
         }
     }
 }
@@ -2162,54 +2380,39 @@ template [[host_name("kernel_mul_mv_bf16_f32_1row")]] kernel mul_mv_1row_t kerne
 // Assumes row size (ne00) is a multiple of 4
 template<typename T, typename T4>
 kernel void kernel_mul_mv_l4(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]]) {
 
-    const int nrows = ne11;
-    const int64_t r0 = tgpig.x;
-    const int64_t im = tgpig.z;
+    const int nrows = args.ne11;
+    const int r0 = tgpig.x;
+    const int im = tgpig.z;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    const uint64_t offset0 = r0*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
 
     device const T4 * x4 = (device const T4 *) (src0 + offset0);
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1;
+
     for (int r1 = 0; r1 < nrows; ++r1) {
-        const uint offset1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+        const uint64_t offset1 = r1*args.nb11 + (i12   )*args.nb12 + (i13   )*args.nb13;
 
         device const float4 * y4 = (device const float4 *) (src1 + offset1);
 
         float sumf = 0;
-        for (int i = tiisg; i < ne00/4; i += 32) {
-            for (int k = 0; k < 4; ++k) sumf += (float) (x4[i][k] * y4[i][k]);
+        for (int i = tiisg; i < args.ne00/4; i += 32) {
+            sumf += dot((float4) x4[i], y4[i]);
         }
 
         float all_sum = simd_sum(sumf);
         if (tiisg == 0) {
-            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            dst_f32[(uint64_t)r1*args.ne0 + r0] = all_sum;
         }
     }
 }
@@ -2229,7 +2432,7 @@ static float rope_yarn_ramp(const float low, const float high, const int i0) {
 // YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
 // MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
 static void rope_yarn(
-    float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale,
+    float theta_extrap, float freq_scale, float corr_dims[2], int i0, float ext_factor, float mscale,
     thread float * cos_theta, thread float * sin_theta) {
     // Get n-d rotational scaling corrected for extrapolation
     float theta_interp = freq_scale * theta_extrap;
@@ -2261,65 +2464,41 @@ static void rope_yarn_corr_dims(
 
 template<typename T>
 kernel void kernel_rope_norm(
-        device const    void * src0,
-        device const int32_t * src1,
-        device const   float * src2,
-        device         float * dst,
-        constant     int64_t & ne00,
-        constant     int64_t & ne01,
-        constant     int64_t & ne02,
-        constant     int64_t & ne03,
-        constant    uint64_t & nb00,
-        constant    uint64_t & nb01,
-        constant    uint64_t & nb02,
-        constant    uint64_t & nb03,
-        constant     int64_t & ne0,
-        constant     int64_t & ne1,
-        constant     int64_t & ne2,
-        constant     int64_t & ne3,
-        constant    uint64_t & nb0,
-        constant    uint64_t & nb1,
-        constant    uint64_t & nb2,
-        constant    uint64_t & nb3,
-        constant         int & n_past,
-        constant         int & n_dims,
-        constant         int & n_ctx_orig,
-        constant       float & freq_base,
-        constant       float & freq_scale,
-        constant       float & ext_factor,
-        constant       float & attn_factor,
-        constant       float & beta_fast,
-        constant       float & beta_slow,
-        uint  tiitg[[thread_index_in_threadgroup]],
-        uint3 tptg[[threads_per_threadgroup]],
-        uint3 tgpig[[threadgroup_position_in_grid]]) {
-    const int64_t i3 = tgpig[2];
-    const int64_t i2 = tgpig[1];
-    const int64_t i1 = tgpig[0];
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
 
     float corr_dims[2];
-    rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
 
-    device const int32_t * pos = src1;
+    device const int32_t * pos = (device const int32_t *) src1;
 
     const float theta_base = (float) pos[i2];
-    const float inv_ndims = -1.f/n_dims;
+    const float inv_ndims = -1.f/args.n_dims;
 
     float cos_theta;
     float sin_theta;
 
-    for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
-        if (i0 < n_dims) {
-            const int64_t ic = i0/2;
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
 
-            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
 
-            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
 
-            rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
 
-            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
 
             const float x0 = src[0];
             const float x1 = src[1];
@@ -2327,8 +2506,8 @@ kernel void kernel_rope_norm(
             dst_data[0] = x0*cos_theta - x1*sin_theta;
             dst_data[1] = x0*sin_theta + x1*cos_theta;
         } else {
-            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
 
             dst_data[0] = src[0];
             dst_data[1] = src[1];
@@ -2338,74 +2517,50 @@ kernel void kernel_rope_norm(
 
 template<typename T>
 kernel void kernel_rope_neox(
-        device const    void * src0,
-        device const int32_t * src1,
-        device const   float * src2,
-        device         float * dst,
-        constant     int64_t & ne00,
-        constant     int64_t & ne01,
-        constant     int64_t & ne02,
-        constant     int64_t & ne03,
-        constant    uint64_t & nb00,
-        constant    uint64_t & nb01,
-        constant    uint64_t & nb02,
-        constant    uint64_t & nb03,
-        constant     int64_t & ne0,
-        constant     int64_t & ne1,
-        constant     int64_t & ne2,
-        constant     int64_t & ne3,
-        constant    uint64_t & nb0,
-        constant    uint64_t & nb1,
-        constant    uint64_t & nb2,
-        constant    uint64_t & nb3,
-        constant         int & n_past,
-        constant         int & n_dims,
-        constant         int & n_ctx_orig,
-        constant       float & freq_base,
-        constant       float & freq_scale,
-        constant       float & ext_factor,
-        constant       float & attn_factor,
-        constant       float & beta_fast,
-        constant       float & beta_slow,
-        uint  tiitg[[thread_index_in_threadgroup]],
-        uint3 tptg[[threads_per_threadgroup]],
-        uint3 tgpig[[threadgroup_position_in_grid]]) {
-    const int64_t i3 = tgpig[2];
-    const int64_t i2 = tgpig[1];
-    const int64_t i1 = tgpig[0];
+        constant ggml_metal_kargs_rope & args,
+        device const char * src0,
+        device const char * src1,
+        device const char * src2,
+        device       char * dst,
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 tptg [[threads_per_threadgroup]],
+        uint3   tgpig[[threadgroup_position_in_grid]]) {
+    const int i3 = tgpig[2];
+    const int i2 = tgpig[1];
+    const int i1 = tgpig[0];
 
     float corr_dims[2];
-    rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+    rope_yarn_corr_dims(args.n_dims, args.n_ctx_orig, args.freq_base, args.beta_fast, args.beta_slow, corr_dims);
 
-    device const int32_t * pos = src1;
+    device const int32_t * pos = (device const int32_t *) src1;
 
     const float theta_base = (float) pos[i2];
-    const float inv_ndims = -1.f/n_dims;
+    const float inv_ndims = -1.f/args.n_dims;
 
     float cos_theta;
     float sin_theta;
 
-    for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
-        if (i0 < n_dims) {
-            const int64_t ic = i0/2;
+    for (int i0 = 2*tiitg; i0 < args.ne0; i0 += 2*tptg.x) {
+        if (i0 < args.n_dims) {
+            const int ic = i0/2;
 
-            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+            const float theta = theta_base * pow(args.freq_base, inv_ndims*i0);
 
-            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+            const float freq_factor = src2 != src0 ? ((device const float *) src2)[ic] : 1.0f;
 
-            rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+            rope_yarn(theta/freq_factor, args.freq_scale, corr_dims, i0, args.ext_factor, args.attn_factor, &cos_theta, &sin_theta);
 
-            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
-            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + ic*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + ic*args.nb0);
 
             const float x0 = src[0];
-            const float x1 = src[n_dims/2];
+            const float x1 = src[args.n_dims/2];
 
-            dst_data[0]        = x0*cos_theta - x1*sin_theta;
-            dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
+            dst_data[0]             = x0*cos_theta - x1*sin_theta;
+            dst_data[args.n_dims/2] = x0*sin_theta + x1*cos_theta;
         } else {
-            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+            device const T * const src = (device T *)(src0 + i3*args.nb03 + i2*args.nb02 + i1*args.nb01 + i0*args.nb00);
+            device       T * dst_data  = (device T *)( dst + i3*args.nb3  + i2*args.nb2  + i1*args.nb1  + i0*args.nb0);
 
             dst_data[0] = src[0];
             dst_data[1] = src[1];
@@ -2460,20 +2615,34 @@ kernel void kernel_im2col(
         uint3  tgpg[[threadgroups_per_grid]],
         uint3 tpitg[[thread_position_in_threadgroup]],
         uint3   ntg[[threads_per_threadgroup]]) {
-    const int32_t iiw = tgpig[2] * s0 + tpitg[2] * d0 - p0;
-    const int32_t iih = tgpig[1] * s1 + tpitg[1] * d1 - p1;
+//    const int64_t IC = tgpg[0];
+    const int64_t OH = tgpg[1];
+    const int64_t OW = tgpg[2];
+
+//    const int64_t N  = ntg[0];
+    const int64_t KH = ntg[1];
+    const int64_t KW = ntg[2];
 
-    const int32_t offset_dst =
-        (tpitg[0] * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * CHW +
-        (tgpig[0] * (ntg[1] * ntg[2]) + tpitg[1] * ntg[2] + tpitg[2]);
+    const int64_t in  = tpitg[0];
+    const int64_t ikh = tpitg[1];
+    const int64_t ikw = tpitg[2];
+
+    const int64_t iic = tgpig[0];
+    const int64_t ioh = tgpig[1];
+    const int64_t iow = tgpig[2];
+
+    const int64_t iiw = iow*s0 + ikw*d0 - p0;
+    const int64_t iih = ioh*s1 + ikh*d1 - p1;
+
+    const int64_t offset_dst = (in*OH*OW + ioh*OW + iow)*CHW + (iic*(KH*KW) + ikh*KW + ikw);
 
     device T * pdst = (device T *) (dst);
 
     if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
         pdst[offset_dst] = 0.0f;
     } else {
-        const int32_t offset_src = tpitg[0] * ofs0 + tgpig[0] * ofs1;
-        pdst[offset_dst] = x[offset_src + iih * IW + iiw];
+        const int64_t offset_src = in*ofs0 + iic*ofs1 + iih*IW + iiw;
+        pdst[offset_dst] = x[offset_src];
     }
 }
 
@@ -2524,25 +2693,25 @@ kernel void kernel_im2col_ext(
         uint3  tgpg[[threadgroups_per_grid]],      // tgpg[0] = D x IC x KH x KW, CHW = IC x KH x KW
         uint3 tpitg[[thread_position_in_threadgroup]],
         uint3   ntg[[threads_per_threadgroup]]) {  // [M, 1, 1]
-    const int32_t KHW = KH * KW;             // KHW == ntg[1] * ntg[2], KW == ntg[2]
+    const int64_t KHW = KH * KW;             // KHW == ntg[1] * ntg[2], KW == ntg[2]
 
-    const int32_t d = tgpig[0] / CHW;
-    const int32_t chw = tgpig[0] % CHW;
-    const int32_t tgpig_0 = chw / KHW;  // 0 ~ (IC - 1)
-    const int32_t HW = tgpig[0] % KHW;
+    const int64_t d = tgpig[0] / CHW;
+    const int64_t chw = tgpig[0] % CHW;
+    const int64_t tgpig_0 = chw / KHW;  // 0 ~ (IC - 1)
+    const int64_t HW = tgpig[0] % KHW;
 
-    const int32_t tpitg_0 = (d * ntg[0]) + tpitg[0];
+    const int64_t tpitg_0 = (d * ntg[0]) + tpitg[0];
     if (tpitg_0 >= N) {
         return;
     }
 
-    const int32_t tpitg_1 = HW / KW;
-    const int32_t tpitg_2 = HW % KW;
+    const int64_t tpitg_1 = HW / KW;
+    const int64_t tpitg_2 = HW % KW;
 
-    const int32_t iiw = tgpig[2] * s0 + tpitg_2 * d0 - p0;
-    const int32_t iih = tgpig[1] * s1 + tpitg_1 * d1 - p1;
+    const int64_t iiw = tgpig[2] * s0 + tpitg_2 * d0 - p0;
+    const int64_t iih = tgpig[1] * s1 + tpitg_1 * d1 - p1;
 
-    const int32_t offset_dst =
+    const int64_t offset_dst =
         (tpitg_0 * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * CHW +
         (tgpig_0 * KHW + tpitg_1 * KW + tpitg_2);
 
@@ -2551,7 +2720,7 @@ kernel void kernel_im2col_ext(
     if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
         pdst[offset_dst] = 0.0f;
     } else {
-        const int32_t offset_src = tpitg_0 * ofs0 + tgpig_0 * ofs1;
+        const int64_t offset_src = tpitg_0 * ofs0 + tgpig_0 * ofs1;
         pdst[offset_dst] = x[offset_src + iih * IW + iiw];
     }
 }
@@ -2559,6 +2728,79 @@ kernel void kernel_im2col_ext(
 template [[host_name("kernel_im2col_ext_f32")]] kernel im2col_ext_t kernel_im2col_ext<float>;
 template [[host_name("kernel_im2col_ext_f16")]] kernel im2col_ext_t kernel_im2col_ext<half>;
 
+typedef void (conv_transpose_1d_t)(
+        device const float * src0,
+        device const float * src1,
+        device        char * dst,
+        constant   int32_t & IC,
+        constant   int32_t & IL,
+        constant   int32_t & K,
+        constant   int32_t & s0,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3    tgpg[[threadgroups_per_grid]]);
+
+template <typename T>
+kernel void kernel_conv_transpose_1d(
+        device const     T * src0,
+        device const float * src1,
+        device        char * dst,
+        constant   int32_t & IC,
+        constant   int32_t & IL,
+        constant   int32_t & K,
+        constant   int32_t & s0,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        uint3   tgpg[[threadgroups_per_grid]]) {
+
+    float v = 0.0f;
+
+    for (int64_t c = 0; c < IC; c++) {
+        const int32_t kernel_offset = c * tgpg[1] * K + K * tgpig[1];
+        const int32_t input_offset = c * IL;
+
+        for (int64_t i = 0; i < IL; i++) {
+            if (tgpig[0] >= i * s0 && tgpig[0] < i * s0 + K) {
+                v += src0[kernel_offset + tgpig[0] - i * s0] * src1[input_offset + i];
+            }
+        }
+    }
+
+    device float * dst_ptr = (device float *) (dst + tgpig[0] * nb0 + tgpig[1] * nb1);
+
+    dst_ptr[0] = v;
+}
+
+template [[host_name("kernel_conv_transpose_1d_f32_f32")]]
+kernel void kernel_conv_transpose_1d<float>(
+    device const float * src0,
+    device const float * src1,
+    device        char * dst,
+    constant   int32_t & IC,
+    constant   int32_t & IL,
+    constant   int32_t & K,
+    constant   int32_t & s0,
+    constant  uint64_t & nb0,
+    constant  uint64_t & nb1,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
+template [[host_name("kernel_conv_transpose_1d_f16_f32")]]
+kernel void kernel_conv_transpose_1d<half>(
+    device const half  * src0,
+    device const float * src1,
+    device        char * dst,
+    constant   int32_t & IC,
+    constant   int32_t & IL,
+    constant   int32_t & K,
+    constant   int32_t & s0,
+    constant  uint64_t & nb0,
+    constant  uint64_t & nb1,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    uint3    tgpg[[threadgroups_per_grid]]);
+
 kernel void kernel_upscale_f32(
     device  const char * src0,
     device        char * dst,
@@ -2655,6 +2897,53 @@ kernel void kernel_pad_f32(
     }
 }
 
+kernel void kernel_pad_reflect_1d_f32(
+    device  const char * src0,
+    device        char * dst,
+    constant   int64_t & ne00,
+    constant   int64_t & ne01,
+    constant   int64_t & ne02,
+    constant   int64_t & ne03,
+    constant   int64_t & ne0,
+    constant  uint64_t & nb00,
+    constant  uint64_t & nb01,
+    constant  uint64_t & nb02,
+    constant  uint64_t & nb03,
+    constant  uint64_t & nb0,
+    constant  uint64_t & nb1,
+    constant  uint64_t & nb2,
+    constant  uint64_t & nb3,
+    constant   int32_t & p0,
+    constant   int32_t & p1,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3  tgpg[[threadgroups_per_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3;
+    const int64_t i02 = i2;
+    const int64_t i01 = i1;
+
+    device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1);
+
+    if (i1 < ne01 && i2 < ne02 && i3 < ne03) {
+        for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+            if (i0 < p0) {
+                dst_ptr[i0] = src0_ptr[p0 - i0];
+            } else if (i0 < ne0 - p1) {
+                dst_ptr[i0] = src0_ptr[i0 - p0];
+            } else {
+                dst_ptr[i0] = src0_ptr[(ne0 - p1 - p0) - (p1 + 1 - (ne0 - i0)) - 1];
+            }
+        }
+    }
+}
+
 kernel void kernel_arange_f32(
     device        char * dst,
     constant   int64_t & ne0,
@@ -2803,37 +3092,17 @@ template<
     short KV = 8,    // key/value processed per each simdgroup
     short C  = 32>   // cache items per threadgroup
 kernel void kernel_flash_attn_ext(
-        device const  char * q,
-        device const  char * k,
-        device const  char * v,
-        device const  char * mask,
-        device       float * dst,
-        constant   int32_t & ne01,
-        constant   int32_t & ne02,
-        constant   int32_t & ne03,
-        constant  uint32_t & nb01,
-        constant  uint32_t & nb02,
-        constant  uint32_t & nb03,
-        constant   int32_t & ne11,
-        constant   int32_t & ne_12_2, // assume K and V are same shape
-        constant   int32_t & ne_12_3,
-        constant  uint32_t & nb_12_1,
-        constant  uint32_t & nb_12_2,
-        constant  uint32_t & nb_12_3,
-        constant  uint32_t & nb31,
-        constant   int32_t & ne1,
-        constant   int32_t & ne2,
-        constant     float & scale,
-        constant     float & max_bias,
-        constant     float & m0,
-        constant     float & m1,
-        constant  uint16_t & n_head_log2,
-        constant     float & logit_softcap,
-        threadgroup   half * shared [[threadgroup(0)]],
-        ushort3  tgpig[[threadgroup_position_in_grid]],
-        ushort3    ntg[[threads_per_threadgroup]],
-        ushort   tiisg[[thread_index_in_simdgroup]],
-        ushort   sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_flash_attn_ext & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * mask,
+        device       char * dst,
+        threadgroup  half * shmem_f16 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3   ntg[[threads_per_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]]) {
     const short nsg = ntg.y; // number of simdgroups
 
     const int iq3 = tgpig[2];
@@ -2849,27 +3118,27 @@ kernel void kernel_flash_attn_ext(
     const short TS = nsg*SH;   // shared memory size per query in (s_t == float)
     const short T  = D + 2*TS; // shared memory size per query in (half)
 
-    threadgroup q_t  * sq  = (threadgroup q_t  *) (shared +              0*D); // holds the query data
-    threadgroup q4_t * sq4 = (threadgroup q4_t *) (shared +              0*D); // same as above but in q4_t
-    threadgroup o_t  * so  = (threadgroup o_t  *) (shared +              0*D); // reuse query data for accumulation
-    threadgroup o4_t * so4 = (threadgroup o4_t *) (shared +              0*D); // same as above but in o4_t
-    threadgroup s_t  * ss  = (threadgroup s_t  *) (shared + 2*sgitg*SH + Q*D); // scratch buffer for attention, mask and diagonal matrix
+    threadgroup q_t  * sq  = (threadgroup q_t  *) (shmem_f16 +              0*D); // holds the query data
+    threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 +              0*D); // same as above but in q4_t
+    threadgroup o_t  * so  = (threadgroup o_t  *) (shmem_f16 +              0*D); // reuse query data for accumulation
+    threadgroup o4_t * so4 = (threadgroup o4_t *) (shmem_f16 +              0*D); // same as above but in o4_t
+    threadgroup s_t  * ss  = (threadgroup s_t  *) (shmem_f16 + 2*sgitg*SH + Q*D); // scratch buffer for attention, mask and diagonal matrix
 
-    threadgroup k_t    * sk    = (threadgroup k_t    *) (shared + sgitg*(4*16*KV) + Q*T); // scratch buffer to load K in shared memory
-    threadgroup k4x4_t * sk4x4 = (threadgroup k4x4_t *) (shared + sgitg*(4*16*KV) + Q*T); // same as above but in k4x4_t
+    threadgroup k_t    * sk    = (threadgroup k_t    *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // scratch buffer to load K in shared memory
+    threadgroup k4x4_t * sk4x4 = (threadgroup k4x4_t *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // same as above but in k4x4_t
 
-    threadgroup v_t    * sv    = (threadgroup v_t    *) (shared + sgitg*(4*16*KV) + Q*T); // scratch buffer to load V in shared memory
-    threadgroup v4x4_t * sv4x4 = (threadgroup v4x4_t *) (shared + sgitg*(4*16*KV) + Q*T); // same as above but in v4x4_t
+    threadgroup v_t    * sv    = (threadgroup v_t    *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // scratch buffer to load V in shared memory
+    threadgroup v4x4_t * sv4x4 = (threadgroup v4x4_t *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // same as above but in v4x4_t
 
     // store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper)
     o8x8_t lo[D8];
 
     // load heads from Q to shared memory
     for (short j = sgitg; j < Q; j += nsg) {
-        device const float4 * q4 = (device const float4 *) ((device const char *) q + ((iq1 + j)*nb01 + iq2*nb02 + iq3*nb03));
+        device const float4 * q4 = (device const float4 *) ((device const char *) q + ((iq1 + j)*args.nb01 + iq2*args.nb02 + iq3*args.nb03));
 
         for (short i = tiisg; i < D4; i += NW) {
-            if (iq1 + j < ne01) {
+            if (iq1 + j < args.ne01) {
                 sq4[j*D4 + i] = (q4_t) q4[i];
             } else {
                 sq4[j*D4 + i] = (q4_t) 0.0f;
@@ -2902,11 +3171,11 @@ kernel void kernel_flash_attn_ext(
         const short ty = tiisg/4;
 
         // broadcast kv
-        //const short rk2 = ne02/ne12;
-        //const short rk3 = ne03/ne13;
+        //const short rk2 = args.ne02/args.ne12;
+        //const short rk3 = args.ne03/args.ne13;
 
-        const short ikv2 = iq2/(ne02/ne_12_2);
-        const short ikv3 = iq3/(ne03/ne_12_3);
+        const short ikv2 = iq2/(args.ne02/args.ne_12_2);
+        const short ikv3 = iq3/(args.ne03/args.ne_12_3);
 
         // load the queries from shared memory into local memory
         q8x8_t mq[D8];
@@ -2920,20 +3189,20 @@ kernel void kernel_flash_attn_ext(
         half slope = 1.0f;
 
         // ALiBi
-        if (max_bias > 0.0f) {
+        if (args.max_bias > 0.0f) {
             const short h = iq2;
 
-            const half  base = h < n_head_log2 ? m0 : m1;
-            const short exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+            const half  base = h < args.n_head_log2 ? args.m0 : args.m1;
+            const short exph = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
 
             slope = pow(base, exph);
         }
 
         // loop over the KV cache
         // each simdgroup handles blocks of Q rows and C columns
-        for (int ic0 = 0; ic0 < ne11; ic0 += C*nsg) {
+        for (int ic0 = 0; ic0 < args.ne11; ic0 += C*nsg) {
             const int ic = ic0 + C*sgitg;
-            if (ic >= ne11) {
+            if (ic >= args.ne11) {
                 break;
             }
 
@@ -2944,7 +3213,7 @@ kernel void kernel_flash_attn_ext(
                 // load the mask in shared memory
                 #pragma unroll(Q)
                 for (short j = 0; j < Q; ++j) {
-                    device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*nb31);
+                    device const half * pm = (device const half *) ((device const char *) mask + (iq1 + j)*args.nb31);
 
                     const half m = pm[ic + tiisg];
 
@@ -2967,18 +3236,18 @@ kernel void kernel_flash_attn_ext(
                     // this is compile-time check, so it does not have runtime overhead
                     if (is_same<kd4x4_t, k4x4_t>::value) {
                         // we can read directly from global memory
-                        device const k_t * pk = (device const k_t *) ((device const char *) k + ((ic + 8*cc)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
+                        device const k_t * pk = (device const k_t *) ((device const char *) k + ((ic + 8*cc)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
 
                         #pragma unroll(D8)
                         for (short i = 0; i < D8; ++i) {
                             k8x8_t mk;
-                            simdgroup_load(mk, pk + i*8, nb_12_1/sizeof(k_t), 0, true); // transpose // TODO: use ne10
+                            simdgroup_load(mk, pk + i*8, args.nb_12_1/sizeof(k_t), 0, true); // transpose // TODO: use ne10
 
                             simdgroup_multiply_accumulate(mqk, mq[i], mk, mqk);
                         }
                     } else {
                         for (short ii = 0; ii < D16; ii += 4) {
-                            device const kd4x4_t * pk4x4 = (device const kd4x4_t *) ((device const char *) k + ((ic + 8*cc + ty)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
+                            device const kd4x4_t * pk4x4 = (device const kd4x4_t *) ((device const char *) k + ((ic + 8*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
 
                             if (D16%4 == 0) {
                                 // the head is evenly divisible by 4*16 = 64, so no need for bound checks
@@ -3037,10 +3306,10 @@ kernel void kernel_flash_attn_ext(
                     const half m = M[j];
 
                     // scale and apply the logitcap / mask
-                    half s = ss[j*TS + tiisg]*scale;
+                    half s = ss[j*TS + tiisg]*args.scale;
 
-                    if (logit_softcap != 0.0f) {
-                        s = logit_softcap*precise::tanh(s);
+                    if (args.logit_softcap != 0.0f) {
+                        s = args.logit_softcap*precise::tanh(s);
                     }
 
                     // mqk = mqk + mask*slope
@@ -3082,18 +3351,18 @@ kernel void kernel_flash_attn_ext(
 
                     if (is_same<vd4x4_t, v4x4_t>::value) {
                         // we can read directly from global memory
-                        device const v_t * pv = (device const v_t *) ((device const char *) v + ((ic + 8*cc)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
+                        device const v_t * pv = (device const v_t *) ((device const char *) v + ((ic + 8*cc)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
 
                         #pragma unroll(D8)
                         for (short i = 0; i < D8; ++i) {
                             v8x8_t mv;
-                            simdgroup_load(mv, pv + i*8, nb_12_1/sizeof(v_t), 0, false); // TODO: use ne20
+                            simdgroup_load(mv, pv + i*8, args.nb_12_1/sizeof(v_t), 0, false); // TODO: use ne20
 
                             simdgroup_multiply_accumulate(lo[i], ms, mv, lo[i]);
                         }
                     } else {
                         for (short ii = 0; ii < D16; ii += 4) {
-                            device const vd4x4_t * pv4x4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 8*cc + ty)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
+                            device const vd4x4_t * pv4x4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 8*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
 
                             if (D16%4 == 0) {
                                 // no need for bound checks
@@ -3222,11 +3491,11 @@ kernel void kernel_flash_attn_ext(
 
     // final rescale with 1/S and store to global memory
     if (sgitg == 0) {
-        for (short j = 0; j < Q && iq1 + j < ne01; ++j) {
+        for (short j = 0; j < Q && iq1 + j < args.ne01; ++j) {
             const float S = ss[j*TS + 0];
 
             for (short i = tiisg; i < D4; i += NW) {
-                dst4[((int64_t)iq3*ne2*ne1 + iq2 + (iq1 + j)*ne1)*D4 + i] = (float4) so4[j*D4 + i]/S;
+                dst4[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*D4 + i] = (float4) so4[j*D4 + i]/S;
             }
         }
     }
@@ -3318,38 +3587,17 @@ template<
     short Q  = 1,    // queries per threadgroup
     short C  = 32>   // cache items per threadgroup
 kernel void kernel_flash_attn_ext_vec(
-        device const  char * q,
-        device const  char * k,
-        device const  char * v,
-        device const  char * mask,
-        device       float * dst,
-        constant   int32_t & ne01,
-        constant   int32_t & ne02,
-        constant   int32_t & ne03,
-        constant  uint32_t & nb01,
-        constant  uint32_t & nb02,
-        constant  uint32_t & nb03,
-        constant   int32_t & ne11,
-        constant   int32_t & ne_12_2, // assume K and V are same shape
-        constant   int32_t & ne_12_3,
-        constant  uint32_t & nb_12_1,
-        constant  uint32_t & nb_12_2,
-        constant  uint32_t & nb_12_3,
-        constant  uint32_t & nb31,
-        constant   int32_t & ne1,
-        constant   int32_t & ne2,
-        constant     float & scale,
-        constant     float & max_bias,
-        constant     float & m0,
-        constant     float & m1,
-        constant  uint16_t & n_head_log2,
-        constant     float & logit_softcap,
-        threadgroup   half * shared [[threadgroup(0)]],
-        ushort3  tgpig[[threadgroup_position_in_grid]],
-        ushort3  tpitg[[thread_position_in_threadgroup]],
-        ushort3    ntg[[threads_per_threadgroup]],
-        ushort   tiisg[[thread_index_in_simdgroup]],
-        ushort   sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_flash_attn_ext & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * mask,
+        device       char * dst,
+        threadgroup  half * shmem_f16 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3   ntg[[threads_per_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]]) {
     const short nsg = ntg.y; // number of simdgroups
 
     const int iq3 = tgpig[2];
@@ -3364,22 +3612,22 @@ kernel void kernel_flash_attn_ext_vec(
 
     const short T = D + nsg*SH; // shared memory size per query in (half)
 
-  //threadgroup q_t    * sq    = (threadgroup q_t    *) (shared +                0*D); // holds the query data
-    threadgroup q4_t   * sq4   = (threadgroup q4_t   *) (shared +                0*D); // same as above but in q4_t
-    threadgroup q4x4_t * sq4x4 = (threadgroup q4x4_t *) (shared +                0*D); // same as above but in q4x4_t
-    threadgroup s_t    * ss    = (threadgroup s_t    *) (shared + sgitg*SH     + Q*D); // scratch buffer for attention
-    threadgroup s4_t   * ss4   = (threadgroup s4_t   *) (shared + sgitg*SH     + Q*D); // same as above but in s4_t
-    threadgroup half   * sm    = (threadgroup half   *) (shared + sgitg*SH + C + Q*D); // scratch buffer for mask
-    threadgroup o4x4_t * sr4x4 = (threadgroup o4x4_t *) (shared + sgitg*D      + Q*T); // scratch buffer for the results
+  //threadgroup q_t    * sq    = (threadgroup q_t    *) (shmem_f16 +                0*D); // holds the query data
+    threadgroup q4_t   * sq4   = (threadgroup q4_t   *) (shmem_f16 +                0*D); // same as above but in q4_t
+    threadgroup q4x4_t * sq4x4 = (threadgroup q4x4_t *) (shmem_f16 +                0*D); // same as above but in q4x4_t
+    threadgroup s_t    * ss    = (threadgroup s_t    *) (shmem_f16 + sgitg*SH     + Q*D); // scratch buffer for attention
+    threadgroup s4_t   * ss4   = (threadgroup s4_t   *) (shmem_f16 + sgitg*SH     + Q*D); // same as above but in s4_t
+    threadgroup half   * sm    = (threadgroup half   *) (shmem_f16 + sgitg*SH + C + Q*D); // scratch buffer for mask
+    threadgroup o4x4_t * sr4x4 = (threadgroup o4x4_t *) (shmem_f16 + sgitg*D      + Q*T); // scratch buffer for the results
 
     // store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper)
     o4x4_t lo[D16/NL];
 
     // load heads from Q to shared memory
-    device const float4 * q4 = (device const float4 *) ((device const char *) q + (iq1*nb01 + iq2*nb02 + iq3*nb03));
+    device const float4 * q4 = (device const float4 *) ((device const char *) q + (iq1*args.nb01 + iq2*args.nb02 + iq3*args.nb03));
 
     for (short i = tiisg; i < D4; i += NW) {
-        if (iq1 < ne01) {
+        if (iq1 < args.ne01) {
             sq4[i] = (q4_t) q4[i];
         } else {
             sq4[i] = (q4_t) 0.0f;
@@ -3407,11 +3655,11 @@ kernel void kernel_flash_attn_ext_vec(
         const short ty = tiisg/NL;
 
         // broadcast kv
-        //const short rk2 = ne02/ne12;
-        //const short rk3 = ne03/ne13;
+        //const short rk2 = args.ne02/args.ne12;
+        //const short rk3 = args.ne03/args.ne13;
 
-        const short ikv2 = iq2/(ne02/ne_12_2);
-        const short ikv3 = iq3/(ne03/ne_12_3);
+        const short ikv2 = iq2/(args.ne02/args.ne_12_2);
+        const short ikv3 = iq3/(args.ne03/args.ne_12_3);
 
         // load the queries from shared memory into local memory
         q4x4_t mq[D16/NL];
@@ -3424,25 +3672,25 @@ kernel void kernel_flash_attn_ext_vec(
         const bool has_mask = mask != q;
 
         // pointer to the mask
-        device const half * pm = (device const half *) (mask + iq1*nb31);
+        device const half * pm = (device const half *) (mask + iq1*args.nb31);
 
         half slope = 1.0f;
 
         // ALiBi
-        if (max_bias > 0.0f) {
+        if (args.max_bias > 0.0f) {
             const short h = iq2;
 
-            const half  base = h < n_head_log2 ? m0 : m1;
-            const short exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+            const half  base = h < args.n_head_log2 ? args.m0 : args.m1;
+            const short exph = h < args.n_head_log2 ? h + 1 : 2*(h - args.n_head_log2) + 1;
 
             slope = pow(base, exph);
         }
 
         // loop over the KV cache
         // each simdgroup handles blocks of Q rows and C columns
-        for (int ic0 = 0; ic0 < ne11; ic0 += C*nsg) {
+        for (int ic0 = 0; ic0 < args.ne11; ic0 += C*nsg) {
             const int ic = ic0 + C*sgitg;
-            if (ic >= ne11) {
+            if (ic >= args.ne11) {
                 break;
             }
 
@@ -3456,7 +3704,7 @@ kernel void kernel_flash_attn_ext_vec(
                 for (short cc = 0; cc < C/4; ++cc) {
                     qk_t mqka[4] = { 0.0, 0.0, 0.0, 0.0 };
 
-                    device const kd4x4_t * pk = (device const kd4x4_t *) ((device const char *) k + ((ic + 4*cc + ty)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
+                    device const kd4x4_t * pk = (device const kd4x4_t *) ((device const char *) k + ((ic + 4*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
 
                     #pragma unroll(D16/NL)
                     for (short ii = 0; ii < D16; ii += NL) {
@@ -3492,10 +3740,10 @@ kernel void kernel_flash_attn_ext_vec(
 
                     // mqk = mqk*scale + mask*slope
                     if (tx == 0) {
-                        mqk *= scale;
+                        mqk *= args.scale;
 
-                        if (logit_softcap != 0.0f) {
-                            mqk = logit_softcap*precise::tanh(mqk);
+                        if (args.logit_softcap != 0.0f) {
+                            mqk = args.logit_softcap*precise::tanh(mqk);
                         }
 
                         mqk += sm[4*cc + ty]*slope;
@@ -3534,7 +3782,7 @@ kernel void kernel_flash_attn_ext_vec(
             // O = O + (Q*K^T)*V
             {
                 for (short cc = 0; cc < C/4; ++cc) {
-                    device const vd4x4_t * pv4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 4*cc + ty)*nb_12_1 + ikv2*nb_12_2 + ikv3*nb_12_3));
+                    device const vd4x4_t * pv4 = (device const vd4x4_t *) ((device const char *) v + ((ic + 4*cc + ty)*args.nb_12_1 + ikv2*args.nb_12_2 + ikv3*args.nb_12_3));
 
                     const s4x4_t ms(ss[4*cc + ty]);
 
@@ -3639,7 +3887,7 @@ kernel void kernel_flash_attn_ext_vec(
         const float S = ss[0];
 
         for (short i = tiisg; i < D16; i += NW) {
-            dst44[((int64_t)iq3*ne2*ne1 + iq2 + (iq1)*ne1)*D16 + i] = (float4x4) sr4x4[i]/S;
+            dst44[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)iq1*args.ne1)*D16 + i] = (float4x4) sr4x4[i]/S;
         }
     }
 }
@@ -3679,44 +3927,61 @@ template [[host_name("kernel_flash_attn_ext_vec_q8_0_h256")]] kernel flash_attn_
 
 #undef FA_TYPES
 
+template<typename T>
+kernel void kernel_set(
+    constant ggml_metal_kargs_set & args,
+    device  const char * src0,
+    device  const char * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    ushort3 tpitg[[thread_position_in_threadgroup]],
+    ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i13 = tgpig[2];
+    const int i12 = tgpig[1];
+    const int i11 = tgpig[0];
+
+    const int64_t n = i13*args.ne12*args.ne11*args.ne10 + i12*args.ne11*args.ne10 + i11*args.ne10;
+
+    const int64_t i3 = n / (args.ne12*args.ne11*args.ne10);
+    const int64_t i2 = (n - i3*args.ne12*args.ne11*args.ne10) / (args.ne11*args.ne10);
+    const int64_t i1 = (n - i3*args.ne12*args.ne11*args.ne10 - i2*args.ne11*args.ne10) / args.ne10;
+
+    device T * dst_data = (device T *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + args.offs);
+
+    for (int64_t i10 = tpitg.x; i10 < args.ne10; i10 += ntg.x) {
+        device const T * src = (device T *) (src1 + i13*args.nb13 + i12*args.nb12 + i11*args.nb11 + i10*args.nb10);
+        dst_data[i10] = (T) src[0];
+    }
+}
+
+typedef decltype(kernel_set<float>) kernel_set_t;
+
+template [[host_name("kernel_set_f32")]] kernel kernel_set_t kernel_set<float>;
+template [[host_name("kernel_set_i32")]] kernel kernel_set_t kernel_set<int32_t>;
+
 template<typename T0, typename T1>
 kernel void kernel_cpy(
-        device  const void * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        constant ggml_metal_kargs_cpy & args,
+        device  const char * src0,
+        device        char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
 
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+    const int64_t i3 = n/(args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0)/(args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0)/args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0);
 
-    device T1 * dst_data = (device T1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    device T1 * dst_data = (device T1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
-        device const T0 * src = (device T0 *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+    for (int64_t i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+        device const T0 * src = (device T0 *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
         dst_data[i00] = (T1) src[0];
     }
 }
@@ -3736,42 +4001,27 @@ template [[host_name("kernel_cpy_bf16_bf16")]] kernel kernel_cpy_t kernel_cpy<bf
 #endif
 
 kernel void kernel_cpy_f32_q8_0(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
 
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK8_0;
+    const int64_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK8_0;
 
-    device block_q8_0 * dst_data = (device block_q8_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    device block_q8_0 * dst_data = (device block_q8_0 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x*QK8_0; i00 < ne00; i00 += ntg.x*QK8_0) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+    for (int64_t i00 = tpitg.x*QK8_0; i00 < args.ne00; i00 += ntg.x*QK8_0) {
+        device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
         float amax = 0.0f; // absolute max
 
@@ -3794,42 +4044,27 @@ kernel void kernel_cpy_f32_q8_0(
 }
 
 kernel void kernel_cpy_f32_q4_0(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
 
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_0;
+    const int64_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK4_0;
 
-    device block_q4_0 * dst_data = (device block_q4_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    device block_q4_0 * dst_data = (device block_q4_0 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x*QK4_0; i00 < ne00; i00 += ntg.x*QK4_0) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+    for (int64_t i00 = tpitg.x*QK4_0; i00 < args.ne00; i00 += ntg.x*QK4_0) {
+        device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
         float amax = 0.0f; // absolute max
         float max  = 0.0f;
@@ -3861,42 +4096,27 @@ kernel void kernel_cpy_f32_q4_0(
 }
 
 kernel void kernel_cpy_f32_q4_1(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
 
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_1;
+    const int64_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK4_1;
 
-    device block_q4_1 * dst_data = (device block_q4_1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    device block_q4_1 * dst_data = (device block_q4_1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x*QK4_1; i00 < ne00; i00 += ntg.x*QK4_1) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+    for (int64_t i00 = tpitg.x*QK4_1; i00 < args.ne00; i00 += ntg.x*QK4_1) {
+        device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
         float min = FLT_MAX;
         float max = -FLT_MAX;
@@ -3927,42 +4147,27 @@ kernel void kernel_cpy_f32_q4_1(
 }
 
 kernel void kernel_cpy_f32_q5_0(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
 
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK5_0;
+    const int64_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK5_0;
 
-    device block_q5_0 * dst_data = (device block_q5_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    device block_q5_0 * dst_data = (device block_q5_0 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x*QK5_0; i00 < ne00; i00 += ntg.x*QK5_0) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+    for (int64_t i00 = tpitg.x*QK5_0; i00 < args.ne00; i00 += ntg.x*QK5_0) {
+        device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
         float amax = 0.0f; // absolute max
         float max  = 0.0f;
@@ -4000,42 +4205,27 @@ kernel void kernel_cpy_f32_q5_0(
 }
 
 kernel void kernel_cpy_f32_q5_1(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
 
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK5_1;
+    const int64_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK5_1;
 
-    device block_q5_1 * dst_data = (device block_q5_1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    device block_q5_1 * dst_data = (device block_q5_1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x*QK5_1; i00 < ne00; i00 += ntg.x*QK5_1) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+    for (int64_t i00 = tpitg.x*QK5_1; i00 < args.ne00; i00 += ntg.x*QK5_1) {
+        device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
         float max = src[0];
         float min = src[0];
@@ -4083,42 +4273,27 @@ static inline int best_index_int8(int n, constant float * val, float x) {
 }
 
 kernel void kernel_cpy_f32_iq4_nl(
-        device const float * src0,
-        device        void * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne03,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   int64_t & ne2,
-        constant   int64_t & ne3,
-        constant  uint64_t & nb0,
-        constant  uint64_t & nb1,
-        constant  uint64_t & nb2,
-        constant  uint64_t & nb3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint3 tpitg[[thread_position_in_threadgroup]],
-        uint3   ntg[[threads_per_threadgroup]]) {
-    const int64_t i03 = tgpig[2];
-    const int64_t i02 = tgpig[1];
-    const int64_t i01 = tgpig[0];
+        constant ggml_metal_kargs_cpy & args,
+        device const char * src0,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i03 = tgpig[2];
+    const int i02 = tgpig[1];
+    const int i01 = tgpig[0];
 
-    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
-    const int64_t i3 = n / (ne2*ne1*ne0);
-    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
-    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
-    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_NL;
+    const int64_t i3 = n / (args.ne2*args.ne1*args.ne0);
+    const int64_t i2 = (n - i3*args.ne2*args.ne1*args.ne0) / (args.ne1*args.ne0);
+    const int64_t i1 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0) / args.ne0;
+    const int64_t i0 = (n - i3*args.ne2*args.ne1*args.ne0 - i2*args.ne1*args.ne0 - i1*args.ne0)/QK4_NL;
 
-    device block_iq4_nl * dst_data = (device block_iq4_nl *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+    device block_iq4_nl * dst_data = (device block_iq4_nl *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x*QK4_NL; i00 < ne00; i00 += ntg.x*QK4_NL) {
-        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+    for (int64_t i00 = tpitg.x*QK4_NL; i00 < args.ne00; i00 += ntg.x*QK4_NL) {
+        device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
         float amax = 0.0f; // absolute max
         float max  = 0.0f;
@@ -4153,105 +4328,67 @@ kernel void kernel_cpy_f32_iq4_nl(
 
         }
 
-        dst_data[i00/QK4_NL].d = sumq2 > 0 ? sumqx/sumq2 : d;
-
-    }
-}
-
-kernel void kernel_concat(
-    device  const char * src0,
-    device  const char * src1,
-    device        char * dst,
-    constant   int64_t & ne00,
-    constant   int64_t & ne01,
-    constant   int64_t & ne02,
-    constant   int64_t & ne03,
-    constant  uint64_t & nb00,
-    constant  uint64_t & nb01,
-    constant  uint64_t & nb02,
-    constant  uint64_t & nb03,
-    constant   int64_t & ne10,
-    constant   int64_t & ne11,
-    constant   int64_t & ne12,
-    constant   int64_t & ne13,
-    constant  uint64_t & nb10,
-    constant  uint64_t & nb11,
-    constant  uint64_t & nb12,
-    constant  uint64_t & nb13,
-    constant   int64_t & ne0,
-    constant   int64_t & ne1,
-    constant   int64_t & ne2,
-    constant   int64_t & ne3,
-    constant  uint64_t & nb0,
-    constant  uint64_t & nb1,
-    constant  uint64_t & nb2,
-    constant  uint64_t & nb3,
-    constant   int32_t & dim,
-    uint3 tgpig[[threadgroup_position_in_grid]],
-    uint3 tpitg[[thread_position_in_threadgroup]],
-    uint3   ntg[[threads_per_threadgroup]]) {
-
-    const int64_t i3 = tgpig.z;
-    const int64_t i2 = tgpig.y;
-    const int64_t i1 = tgpig.x;
+        dst_data[i00/QK4_NL].d = sumq2 > 0 ? sumqx/sumq2 : d;
+    }
+}
+
+kernel void kernel_concat(
+    constant ggml_metal_kargs_concat & args,
+    device  const char * src0,
+    device  const char * src1,
+    device        char * dst,
+    uint3   tgpig[[threadgroup_position_in_grid]],
+    ushort3 tpitg[[thread_position_in_threadgroup]],
+    ushort3   ntg[[threads_per_threadgroup]]) {
+
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
 
-    int64_t o[4] = {0, 0, 0, 0};
-    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
+    int o[4] = {0, 0, 0, 0};
+    o[args.dim] = args.dim == 0 ? args.ne00 : (args.dim == 1 ? args.ne01 : (args.dim == 2 ? args.ne02 : args.ne03));
 
     device const float * x;
 
-    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
-        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-            x = (device const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        if (i0 < args.ne00 && i1 < args.ne01 && i2 < args.ne02 && i3 < args.ne03) {
+            x = (device const float *)(src0 + (i3       )*args.nb03 + (i2       )*args.nb02 + (i1       )*args.nb01 + (i0       )*args.nb00);
         } else {
-            x = (device const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
+            x = (device const float *)(src1 + (i3 - o[3])*args.nb13 + (i2 - o[2])*args.nb12 + (i1 - o[1])*args.nb11 + (i0 - o[0])*args.nb10);
         }
 
-        device float * y = (device float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+        device float * y = (device float *)(dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
         *y = *x;
     }
 }
 
+template<typename args_t>
 void kernel_mul_mv_q2_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_q2_K * x = (device const block_q2_K *) ((device char *) src0 + offset0);
-    device const float      * y = (device const float      *) ((device char *) src1 + offset1);
+    device const block_q2_K * x = (device const block_q2_K *) (src0 + offset0);
+    device const float      * y = (device const float      *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
@@ -4300,92 +4437,64 @@ void kernel_mul_mv_q2_K_f32_impl(
                                  (acc1[3] + 1.f/256.f * acc2[3]) * (sc[6] & 0xF) * 1.f/64.f) -
                          dmin * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) + sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));
 
-            qs += nb01/2;
-            sc += nb01;
-            dh += nb01/2;
+            qs += args.nb01/2;
+            sc += args.nb01;
+            dh += args.nb01/2;
         }
 
         y4 += 4 * QK_K;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+            dst_f32[first_row + row] = all_sum;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_q2_K_f32")]]
 kernel void kernel_mul_mv_q2_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_q2_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+    kernel_mul_mv_q2_K_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
+template<typename args_t>
 void kernel_mul_mv_q3_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
-
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-    const int64_t im = tgpig.z;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
+
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_q3_K * x = (device const block_q3_K *) ((device char *) src0 + offset0);
-    device const float     * yy = (device const float      *) ((device char *) src1 + offset1);
+    device const block_q3_K * x = (device const block_q3_K *) (src0 + offset0);
+    device const float     * yy = (device const float      *) (src1 + offset1);
 
     float yl[32];
 
@@ -4415,9 +4524,10 @@ void kernel_mul_mv_q3_K_f32_impl(
 
     const ushort4 hm = mm[2*ip + il/2];
 
-    const int shift = 2*il;
-    const float    v1 = il == 0 ? 4.f : 64.f;
-    const float    v2 = 4.f * v1;
+    const short shift = 2*il;
+
+    const float v1 = il == 0 ? 4.f : 64.f;
+    const float v2 = 4.f * v1;
 
     const uint16_t s_shift1 = 4*ip;
     const uint16_t s_shift2 = s_shift1 + il;
@@ -4486,10 +4596,10 @@ void kernel_mul_mv_q3_K_f32_impl(
             sumf1[row] += d1 * (scales[1] - 32);
             sumf2[row] += d2 * (scales[3] - 32);
 
-            q  += nb01/2;
-            h  += nb01/2;
-            a  += nb01/2;
-            dh += nb01/2;
+            q  += args.nb01/2;
+            h  += args.nb01/2;
+            a  += args.nb01/2;
+            dh += args.nb01/2;
         }
 
         y1 += 4 * QK_K;
@@ -4499,66 +4609,39 @@ void kernel_mul_mv_q3_K_f32_impl(
         const float sumf = (sumf1[row] + 0.25f * sumf2[row]) / (1 << shift);
         sumf1[row] = simd_sum(sumf);
     }
+
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     if (tiisg == 0) {
         for (int row = 0; row < 2; ++row) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = sumf1[row];
+            dst_f32[first_row + row] = sumf1[row];
         }
     }
 }
 
 [[host_name("kernel_mul_mv_q3_K_f32")]]
 kernel void kernel_mul_mv_q3_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_q3_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+    kernel_mul_mv_q3_K_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
+template<typename args_t>
 void kernel_mul_mv_q4_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
 
     const uint16_t kmask1 = 0x3f3f;
     const uint16_t kmask2 = 0x0f0f;
@@ -4569,21 +4652,21 @@ void kernel_mul_mv_q4_K_f32_impl(
     const int iq = it/4;     // 0 or 1
     const int ir = it%4;     // 0...3
 
-    const int nb = ne00/QK_K;
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
     //const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
     const int first_row = r0 * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_q4_K * x = (device const block_q4_K *) ((device char *) src0 + offset0);
-    device const float      * y = (device const float      *) ((device char *) src1 + offset1);
+    device const block_q4_K * x = (device const block_q4_K *) (src0 + offset0);
+    device const float      * y = (device const float      *) (src1 + offset1);
 
     float yl[16];
     float yh[16];
@@ -4636,92 +4719,64 @@ void kernel_mul_mv_q4_K_f32_impl(
                                  (acc2[2] + 1.f/256.f * acc2[3]) * sc8[5] * 1.f/16.f) -
                          dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
 
-            q1 += nb01/2;
-            sc += nb01/2;
-            dh += nb01/2;
+            q1 += args.nb01/2;
+            sc += args.nb01/2;
+            dh += args.nb01/2;
         }
 
         y4 += 4 * QK_K;
     }
 
+    device float * dst_f32 = (device float *) dst + (int64_t)im*args.ne0*args.ne1 + (int64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+            dst_f32[first_row + row] = all_sum;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_q4_K_f32")]]
 kernel void kernel_mul_mv_q4_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_q4_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+    kernel_mul_mv_q4_K_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
+template<typename args_t>
 void kernel_mul_mv_q5_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
-
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
+
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_q5_K * x = (device const block_q5_K *) ((device char *) src0 + offset0);
-    device const float     * yy = (device const float      *) ((device char *) src1 + offset1);
+    device const block_q5_K * x = (device const block_q5_K *) (src0 + offset0);
+    device const float     * yy = (device const float      *) (src1 + offset1);
 
     float sumf[2]={0.f};
 
@@ -4795,98 +4850,70 @@ void kernel_mul_mv_q5_K_f32_impl(
                                  sc8[5] * (acc1[3]/16.f + 16.f*acc2[3])) -
                          dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
 
-            q1 += nb01;
-            qh += nb01;
-            dh += nb01/2;
-            a  += nb01/2;
+            q1 += args.nb01;
+            qh += args.nb01;
+            dh += args.nb01/2;
+            a  += args.nb01/2;
         }
 
         y1 += 4 * QK_K;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < 2; ++row) {
         const float tot = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
+            dst_f32[first_row + row] = tot;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_q5_K_f32")]]
 kernel void kernel_mul_mv_q5_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_q5_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+    kernel_mul_mv_q5_K_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
+template <typename args_t>
 void kernel_mul_mv_q6_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
 
     const uint8_t kmask1 = 0x03;
     const uint8_t kmask2 = 0x0C;
     const uint8_t kmask3 = 0x30;
     const uint8_t kmask4 = 0xC0;
 
-    const int nb = ne00/QK_K;
+    const int nb = args.ne00/QK_K;
 
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-    const int     im = tgpig.z;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
 
-    const int row = 2 * r0 + sgitg;
+    const int row = 2*r0 + sgitg;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =  r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =  r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_q6_K * x = (device const block_q6_K *) ((device char *) src0 + offset0);
-    device const float     * yy = (device const float      *) ((device char *) src1 + offset1);
+    device const block_q6_K * x = (device const block_q6_K *) (src0 + offset0);
+    device const float     * yy = (device const float      *) (src1 + offset1);
 
     float sumf = 0;
 
@@ -4903,7 +4930,6 @@ void kernel_mul_mv_q6_K_f32_impl(
     const int q_offset_h = 32*ip + l0;
 
     for (int i = ix; i < nb; i += 2) {
-
         device const uint8_t * q1 = x[i].ql + q_offset_l;
         device const uint8_t * q2 = q1 + 32;
         device const uint8_t * qh = x[i].qh + q_offset_h;
@@ -4925,98 +4951,70 @@ void kernel_mul_mv_q6_K_f32_impl(
 
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     const float tot = simd_sum(sumf);
     if (tiisg == 0) {
-        dst[r1*ne0 + im*ne0*ne1 + row] = tot;
+        dst_f32[row] = tot;
     }
 }
 
 [[host_name("kernel_mul_mv_q6_K_f32")]]
 kernel void kernel_mul_mv_q6_K_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_q6_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+    kernel_mul_mv_q6_K_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
 // ======================= "True" 2-bit
 
+template<typename args_t>
 void kernel_mul_mv_iq2_xxs_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq2_xxs * x = (device const block_iq2_xxs *) ((device char *) src0 + offset0);
-    device const float         * y = (device const float         *) ((device char *) src1 + offset1);
+    device const block_iq2_xxs * x = (device const block_iq2_xxs *) (src0 + offset0);
+    device const float         * y = (device const float         *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
 
     const int nb32 = nb * (QK_K / 32);
 
-    threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
-    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 256);
+    threadgroup uint64_t * svalues = (threadgroup uint64_t *)(shmem);
+    threadgroup uint8_t  * ssigns  = (threadgroup uint8_t  *)(svalues + 256);
     {
         int nval = 4;
         int pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) values[pos + i] = iq2xxs_grid[pos + i];
+        for (int i = 0; i < nval; ++i) svalues[pos + i] = iq2xxs_grid[pos + i];
         nval = 2;
         pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
+        for (int i = 0; i < nval; ++i) ssigns[pos+i] = ksigns_iq2xs[pos+i];
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
@@ -5046,114 +5044,85 @@ void kernel_mul_mv_iq2_xxs_f32_impl(
 
             float sum = 0;
             for (int l = 0; l < 4; ++l) {
-                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + aux8[l]);
-                const uint8_t signs = shared_signs[(aux32 >> 7*l) & 127];
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(svalues + aux8[l]);
+                const uint8_t signs = ssigns[(aux32 >> 7*l) & 127];
                 for (int j = 0; j < 8; ++j) {
                     sum += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
                 }
             }
             sumf[row] += d * sum;
 
-            dh += nb01/2;
-            q2 += nb01/2;
+            dh += args.nb01/2;
+            q2 += args.nb01/2;
         }
 
         y4 += 32 * 32;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
+            dst_f32[first_row + row] = all_sum * 0.25f;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_iq2_xxs_f32")]]
 kernel void kernel_mul_mv_iq2_xxs_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    kernel_mul_mv_iq2_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    kernel_mul_mv_iq2_xxs_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
+template<typename args_t>
 void kernel_mul_mv_iq2_xs_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq2_xs * x = (device const block_iq2_xs *) ((device char *) src0 + offset0);
-    device const float        * y = (device const float        *) ((device char *) src1 + offset1);
+    device const block_iq2_xs * x = (device const block_iq2_xs *) (src0 + offset0);
+    device const float        * y = (device const float        *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
 
     const int nb32 = nb * (QK_K / 32);
 
-    threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
-    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 512);
+    threadgroup uint64_t * svalues = (threadgroup uint64_t *)(shmem);
+    threadgroup uint8_t  * ssigns  = (threadgroup uint8_t  *)(svalues + 512);
     {
         int nval = 8;
         int pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) values[pos + i] = iq2xs_grid[pos + i];
+        for (int i = 0; i < nval; ++i) svalues[pos + i] = iq2xs_grid[pos + i];
         nval = 2;
         pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
+        for (int i = 0; i < nval; ++i) ssigns[pos+i] = ksigns_iq2xs[pos+i];
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
@@ -5185,122 +5154,94 @@ void kernel_mul_mv_iq2_xs_f32_impl(
 
             float sum1 = 0, sum2 = 0;
             for (int l = 0; l < 2; ++l) {
-                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511));
-                const uint8_t signs = shared_signs[(q2[l] >> 9)];
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(svalues + (q2[l] & 511));
+                const uint8_t signs = ssigns[(q2[l] >> 9)];
                 for (int j = 0; j < 8; ++j) {
                     sum1 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
                 }
             }
             for (int l = 2; l < 4; ++l) {
-                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511));
-                const uint8_t signs = shared_signs[(q2[l] >> 9)];
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(svalues + (q2[l] & 511));
+                const uint8_t signs = ssigns[(q2[l] >> 9)];
                 for (int j = 0; j < 8; ++j) {
                     sum2 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
                 }
             }
             sumf[row] += d1 * sum1 + d2 * sum2;
 
-            dh += nb01/2;
-            q2 += nb01/2;
-            sc += nb01;
+            dh += args.nb01/2;
+            q2 += args.nb01/2;
+            sc += args.nb01;
         }
 
         y4 += 32 * 32;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
+            dst_f32[first_row + row] = all_sum * 0.25f;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_iq2_xs_f32")]]
 kernel void kernel_mul_mv_iq2_xs_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq2_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq2_xs_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
+template <typename args_t>
 void kernel_mul_mv_iq3_xxs_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq3_xxs * x = (device const block_iq3_xxs *) ((device char *) src0 + offset0);
-    device const float         * y = (device const float         *) ((device char *) src1 + offset1);
+    device const block_iq3_xxs * x = (device const block_iq3_xxs *) (src0 + offset0);
+    device const float         * y = (device const float         *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
 
     const int nb32 = nb * (QK_K / 32);
 
-    threadgroup uint32_t * values = (threadgroup uint32_t *)shared_values;
-    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 256);
+    threadgroup uint32_t * svalues = (threadgroup uint32_t *)(shmem);
+    threadgroup uint8_t  * ssigns  = (threadgroup uint8_t  *)(svalues + 256);
     {
         int nval = 4;
         int pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) values[pos + i] = iq3xxs_grid[pos + i];
+        for (int i = 0; i < nval; ++i) svalues[pos + i] = iq3xxs_grid[pos + i];
         nval = 2;
         pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
+        for (int i = 0; i < nval; ++i) ssigns[pos+i] = ksigns_iq2xs[pos+i];
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
@@ -5309,7 +5250,6 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
     device const float * y4 = y + 32 * ix;
 
     for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
-
         for (int i = 0; i < 32; ++i) {
             yl[i] = y4[i];
         }
@@ -5323,16 +5263,15 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
         device const half * dh = &xr->d;
 
         for (int row = 0; row < N_DST; row++) {
-
             const float db = dh[0];
             const uint32_t aux32 = gas[0] | (gas[1] << 16);
             const float d = db * (0.5f + (aux32 >> 28));
 
             float2 sum = {0};
             for (int l = 0; l < 4; ++l) {
-                const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + q3[2*l+0]);
-                const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + q3[2*l+1]);
-                const uint8_t signs = shared_signs[(aux32 >> 7*l) & 127];
+                const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(svalues + q3[2*l+0]);
+                const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(svalues + q3[2*l+1]);
+                const uint8_t signs = ssigns[(aux32 >> 7*l) & 127];
                 for (int j = 0; j < 4; ++j) {
                     sum[0] += yl[8*l + j + 0] * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
                     sum[1] += yl[8*l + j + 4] * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
@@ -5340,103 +5279,75 @@ void kernel_mul_mv_iq3_xxs_f32_impl(
             }
             sumf[row] += d * (sum[0] + sum[1]);
 
-            dh  += nb01/2;
-            q3  += nb01;
-            gas += nb01/2;
+            dh  += args.nb01/2;
+            q3  += args.nb01;
+            gas += args.nb01/2;
         }
 
         y4 += 32 * 32;
     }
 
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.5f;
-        }
-    }
-}
-
-[[host_name("kernel_mul_mv_iq3_xxs_f32")]]
-kernel void kernel_mul_mv_iq3_xxs_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
 
-    kernel_mul_mv_iq3_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst_f32[first_row + row] = all_sum * 0.5f;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_iq3_xxs_f32")]]
+kernel void kernel_mul_mv_iq3_xxs_f32(
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq3_xxs_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
+template<typename args_t>
 void kernel_mul_mv_iq3_s_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq3_s * x = (device const block_iq3_s *) ((device char *) src0 + offset0);
-    device const float       * y = (device const float       *) ((device char *) src1 + offset1);
+    device const block_iq3_s * x = (device const block_iq3_s *) (src0 + offset0);
+    device const float       * y = (device const float       *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
 
     const int nb32 = nb * (QK_K / 32);
 
-    threadgroup uint32_t * values = (threadgroup uint32_t *)shared_values;
+    threadgroup uint32_t * svalues = (threadgroup uint32_t *) shmem;
     {
         int nval = 8;
         int pos  = (32*sgitg + tiisg)*nval;
-        for (int i = 0; i < nval; ++i) values[pos + i] = iq3s_grid[pos + i];
+        for (int i = 0; i < nval; ++i) svalues[pos + i] = iq3s_grid[pos + i];
         threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
@@ -5467,8 +5378,8 @@ void kernel_mul_mv_iq3_s_f32_impl(
 
             float2 sum = {0};
             for (int l = 0; l < 4; ++l) {
-                const threadgroup uint32_t * table1 = qh[0] & kmask_iq2xs[2*l+0] ? values + 256 : values;
-                const threadgroup uint32_t * table2 = qh[0] & kmask_iq2xs[2*l+1] ? values + 256 : values;
+                const threadgroup uint32_t * table1 = qh[0] & kmask_iq2xs[2*l+0] ? svalues + 256 : svalues;
+                const threadgroup uint32_t * table2 = qh[0] & kmask_iq2xs[2*l+1] ? svalues + 256 : svalues;
                 const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(table1 + qs[2*l+0]);
                 const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(table2 + qs[2*l+1]);
                 for (int j = 0; j < 4; ++j) {
@@ -5478,105 +5389,77 @@ void kernel_mul_mv_iq3_s_f32_impl(
             }
             sumf[row] += d * (sum[0] + sum[1]);
 
-            dh    += nb01/2;
-            qs    += nb01;
-            qh    += nb01;
-            sc    += nb01;
-            signs += nb01;
+            dh    += args.nb01/2;
+            qs    += args.nb01;
+            qh    += args.nb01;
+            sc    += args.nb01;
+            signs += args.nb01;
         }
 
         y4 += 32 * 32;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+            dst_f32[first_row + row] = all_sum;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_iq3_s_f32")]]
 kernel void kernel_mul_mv_iq3_s_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq3_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq3_s_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
+template <typename args_t>
 void kernel_mul_mv_iq2_s_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq2_s * x = (device const block_iq2_s *) ((device char *) src0 + offset0);
-    device const float       * y = (device const float       *) ((device char *) src1 + offset1);
+    device const block_iq2_s * x = (device const block_iq2_s *) (src0 + offset0);
+    device const float       * y = (device const float       *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
 
     const int nb32 = nb * (QK_K / 32);
 
-    //threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
+    //threadgroup uint64_t * svalues = (threadgroup uint64_t *) shmem;
     //{
     //    int nval = 32;
     //    int pos  = (32*sgitg + tiisg)*nval;
-    //    for (int i = 0; i < nval; ++i) values[pos + i] = iq2s_grid[pos + i];
+    //    for (int i = 0; i < nval; ++i) svalues[pos + i] = iq2s_grid[pos + i];
     //    threadgroup_barrier(mem_flags::mem_threadgroup);
     //}
 
@@ -5608,8 +5491,8 @@ void kernel_mul_mv_iq2_s_f32_impl(
 
             float2 sum = {0};
             for (int l = 0; l < 2; ++l) {
-                //const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
-                //const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
+                //const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(svalues + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
+                //const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(svalues + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
                 constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
                 constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
                 for (int j = 0; j < 8; ++j) {
@@ -5619,94 +5502,66 @@ void kernel_mul_mv_iq2_s_f32_impl(
             }
             sumf[row] += d1 * sum[0] + d2 * sum[1];
 
-            dh    += nb01/2;
-            qs    += nb01;
-            qh    += nb01;
-            sc    += nb01;
-            signs += nb01;
+            dh    += args.nb01/2;
+            qs    += args.nb01;
+            qh    += args.nb01;
+            sc    += args.nb01;
+            signs += args.nb01;
         }
 
         y4 += 32 * 32;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
+            dst_f32[first_row + row] = all_sum * 0.25f;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_iq2_s_f32")]]
 kernel void kernel_mul_mv_iq2_s_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq2_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq2_s_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
+template<typename args_t>
 void kernel_mul_mv_iq1_s_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_value,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq1_s * x = (device const block_iq1_s *) ((device char *) src0 + offset0);
-    device const float       * y = (device const float       *) ((device char *) src1 + offset1);
+    device const block_iq1_s * x = (device const block_iq1_s *) (src0 + offset0);
+    device const float       * y = (device const float       *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
@@ -5749,61 +5604,50 @@ void kernel_mul_mv_iq1_s_f32_impl(
             }
             sumf[row] += (float)dh[0] * (sum + sumy * (qh[0] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA)) * (2*((qh[0] >> 12) & 7) + 1);
 
-            dh += nb01/2;
-            qs += nb01;
-            qh += nb01/2;
+            dh += args.nb01/2;
+            qs += args.nb01;
+            qh += args.nb01/2;
         }
 
         y4 += 32 * 32;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+            dst_f32[first_row + row] = all_sum;
         }
     }
 }
 
+template <typename args_t>
 void kernel_mul_mv_iq1_m_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_value,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
 
     const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq1_m * x = (device const block_iq1_m *) ((device char *) src0 + offset0);
-    device const float       * y = (device const float       *) ((device char *) src1 + offset1);
+    device const block_iq1_m * x = (device const block_iq1_m *) (src0 + offset0);
+    device const float       * y = (device const float       *) (src1 + offset1);
 
     float yl[32];
     float sumf[N_DST]={0.f}, all_sum;
@@ -5855,66 +5699,55 @@ void kernel_mul_mv_iq1_m_f32_impl(
             sumf[row] += (float)scale.f16 * ((sum[0] + delta1) * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 7) + 1) +
                                              (sum[1] + delta2) * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 7) + 1));
 
-            sc += nb01/2;
-            qs += nb01;
-            qh += nb01;
+            sc += args.nb01/2;
+            qs += args.nb01;
+            qh += args.nb01;
         }
 
         y4 += 32 * 32;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+            dst_f32[first_row + row] = all_sum;
         }
     }
 }
 
+template<typename args_t>
 void kernel_mul_mv_iq4_nl_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values_i8,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    threadgroup float * shared_values = (threadgroup float *)shared_values_i8;
-    const int nb = ne00/QK4_NL;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
+    const int nb = args.ne00/QK4_NL;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
     const int first_row = (r0 * 2 + sgitg) * 2;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq4_nl * x = (device const block_iq4_nl *) ((device char *) src0 + offset0);
-    device const float        * y = (device const float        *) ((device char *) src1 + offset1);
+    device const block_iq4_nl * x = (device const block_iq4_nl *) (src0 + offset0);
+    device const float        * y = (device const float        *) (src1 + offset1);
 
     const int ix = tiisg/2;  // 0...15
     const int it = tiisg%2;  // 0 or 1
 
-    shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
+    shmem_f32[tiisg] = kvalues_iq4nl_f[tiisg%16];
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
     float4 yl[4];
@@ -5932,7 +5765,7 @@ void kernel_mul_mv_iq4_nl_f32_impl(
         device const float4 * y4 = (device const float4 *)yb;
         yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
 
-        for (int row = 0; row < 2 && first_row + row < ne01; ++row) {
+        for (int row = 0; row < 2 && first_row + row < args.ne01; ++row) {
 
             device const block_iq4_nl & xb = x[row*nb + ib];
             device const uint16_t * q4 = (device const uint16_t *)(xb.qs + 8*it);
@@ -5942,16 +5775,16 @@ void kernel_mul_mv_iq4_nl_f32_impl(
             aux32[0] = q4[0] | (q4[1] << 16);
             aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
             aux32[0] &= 0x0f0f0f0f;
-            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
-            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            qf1 = {shmem_f32[q8[0]], shmem_f32[q8[1]], shmem_f32[q8[2]], shmem_f32[q8[3]]};
+            qf2 = {shmem_f32[q8[4]], shmem_f32[q8[5]], shmem_f32[q8[6]], shmem_f32[q8[7]]};
             acc1 += yl[0] * qf1;
             acc2 += yl[1] * qf2;
 
             aux32[0] = q4[2] | (q4[3] << 16);
             aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
             aux32[0] &= 0x0f0f0f0f;
-            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
-            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            qf1 = {shmem_f32[q8[0]], shmem_f32[q8[1]], shmem_f32[q8[2]], shmem_f32[q8[3]]};
+            qf2 = {shmem_f32[q8[4]], shmem_f32[q8[5]], shmem_f32[q8[6]], shmem_f32[q8[7]]};
             acc1 += yl[2] * qf1;
             acc2 += yl[3] * qf2;
 
@@ -5964,60 +5797,49 @@ void kernel_mul_mv_iq4_nl_f32_impl(
         yb += 16 * QK4_NL;
     }
 
-    for (int row = 0; row < 2 && first_row + row < ne01; ++row) {
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
+    for (int row = 0; row < 2 && first_row + row < args.ne01; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+            dst_f32[first_row + row] = all_sum;
         }
     }
 }
 
+template<typename args_t>
 void kernel_mul_mv_iq4_xs_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values_i8,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg) {
-
-    threadgroup float * shared_values = (threadgroup float *)shared_values_i8;
-    const int nb = ne00/QK_K;
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg) {
+
+    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
+    const int nb = args.ne00/QK_K;
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
     const int first_row = (r0 * 2 + sgitg) * 2;
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
 
-    const uint offset0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
-    const uint offset1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+    const uint64_t offset0 = first_row*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const uint64_t offset1 =        r1*args.nb11 + (i12        )*args.nb12 + (i13        )*args.nb13;
 
-    device const block_iq4_xs * x = (device const block_iq4_xs *) ((device char *) src0 + offset0);
-    device const float        * y = (device const float        *) ((device char *) src1 + offset1);
+    device const block_iq4_xs * x = (device const block_iq4_xs *) (src0 + offset0);
+    device const float        * y = (device const float        *) (src1 + offset1);
 
     const int ix = tiisg/16;  // 0 or 1
     const int it = tiisg%16;  // 0...15
     const int ib = it/2;
     const int il = it%2;
 
-    shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
+    shmem_f32[tiisg] = kvalues_iq4nl_f[tiisg%16];
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
     float4 yl[4];
@@ -6031,28 +5853,26 @@ void kernel_mul_mv_iq4_xs_f32_impl(
     float4 qf1, qf2;
 
     for (int ibl = ix; ibl < nb; ibl += 2) {
-
         device const float4 * y4 = (device const float4 *)yb;
         yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
 
         for (int row = 0; row < 2; ++row) {
-
             device const block_iq4_xs & xb = x[row*nb + ibl];
             device const uint32_t * q4 = (device const uint32_t *)(xb.qs + 16*ib + 8*il);
 
             float4 acc1 = {0.f}, acc2 = {0.f};
 
-            aux32[0] = q4[0] & 0x0f0f0f0f;
+            aux32[0] = (q4[0]     ) & 0x0f0f0f0f;
             aux32[1] = (q4[0] >> 4) & 0x0f0f0f0f;
-            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
-            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            qf1 = {shmem_f32[q8[0]], shmem_f32[q8[1]], shmem_f32[q8[2]], shmem_f32[q8[3]]};
+            qf2 = {shmem_f32[q8[4]], shmem_f32[q8[5]], shmem_f32[q8[6]], shmem_f32[q8[7]]};
             acc1 += yl[0] * qf1;
             acc2 += yl[1] * qf2;
 
-            aux32[0] = q4[1] & 0x0f0f0f0f;
+            aux32[0] = (q4[1]     ) & 0x0f0f0f0f;
             aux32[1] = (q4[1] >> 4) & 0x0f0f0f0f;
-            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
-            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            qf1 = {shmem_f32[q8[0]], shmem_f32[q8[1]], shmem_f32[q8[2]], shmem_f32[q8[3]]};
+            qf2 = {shmem_f32[q8[4]], shmem_f32[q8[5]], shmem_f32[q8[6]], shmem_f32[q8[7]]};
             acc1 += yl[2] * qf1;
             acc2 += yl[3] * qf2;
 
@@ -6066,134 +5886,68 @@ void kernel_mul_mv_iq4_xs_f32_impl(
         yb += 2 * QK_K;
     }
 
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1 + (uint64_t)r1*args.ne0;
+
     for (int row = 0; row < 2; ++row) {
         all_sum = simd_sum(sumf[row]);
         if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+            dst_f32[first_row + row] = all_sum;
         }
     }
 }
 
 [[host_name("kernel_mul_mv_iq1_s_f32")]]
 kernel void kernel_mul_mv_iq1_s_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq1_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq1_s_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
 [[host_name("kernel_mul_mv_iq1_m_f32")]]
 kernel void kernel_mul_mv_iq1_m_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq1_m_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq1_m_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, nullptr, tgpig, tiisg, sgitg);
 }
 
 [[host_name("kernel_mul_mv_iq4_nl_f32")]]
 kernel void kernel_mul_mv_iq4_nl_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq4_nl_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
 [[host_name("kernel_mul_mv_iq4_xs_f32")]]
 kernel void kernel_mul_mv_iq4_xs_f32(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant  uint64_t & nb00,
-        constant  uint64_t & nb01,
-        constant  uint64_t & nb02,
-        constant  uint64_t & nb03,
-        constant   int64_t & ne10,
-        constant   int64_t & ne11,
-        constant   int64_t & ne12,
-        constant  uint64_t & nb10,
-        constant  uint64_t & nb11,
-        constant  uint64_t & nb12,
-        constant  uint64_t & nb13,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint tiisg[[thread_index_in_simdgroup]],
-        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
-    kernel_mul_mv_iq4_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, nb01, nb02, nb03, ne10, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+    kernel_mul_mv_iq4_xs_f32_impl<constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
@@ -6297,42 +6051,30 @@ kernel void kernel_get_rows_i32(
 
 // each block_q contains 16*nl weights
 template<typename T, typename T4x4, typename simdgroup_T8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
-kernel void kernel_mul_mm(device const  uchar * src0,
-                          device const  uchar * src1,
-                          device        float * dst,
-                          constant    int64_t & ne00,
-                          constant    int64_t & ne02,
-                          constant   uint64_t & nb01,
-                          constant   uint64_t & nb02,
-                          constant   uint64_t & nb03,
-                          constant    int64_t & ne12,
-                          constant   uint64_t & nb10,
-                          constant   uint64_t & nb11,
-                          constant   uint64_t & nb12,
-                          constant   uint64_t & nb13,
-                          constant    int64_t & ne0,
-                          constant    int64_t & ne1,
-                          constant       uint & r2,
-                          constant       uint & r3,
-                          threadgroup   uchar * shared_memory [[threadgroup(0)]],
-                          uint3                 tgpig[[threadgroup_position_in_grid]],
-                          uint                  tiitg[[thread_index_in_threadgroup]],
-                          uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    threadgroup T     * sa = (threadgroup T     *)(shared_memory);
-    threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
-
-    const uint r0 = tgpig.y;
-    const uint r1 = tgpig.x;
-    const uint im = tgpig.z;
+kernel void kernel_mul_mm(
+        constant ggml_metal_kargs_mul_mm & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup T     * sa = (threadgroup T     *)(shmem);
+    threadgroup float * sb = (threadgroup float *)(shmem + 4096);
+
+    const int r0 = tgpig.y;
+    const int r1 = tgpig.x;
+    const int im = tgpig.z;
 
     // if this block is of 64x32 shape or smaller
-    short n_rows = (ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
-    short n_cols = (ne1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
+    const short n_rows = (args.ne0 - r0*BLOCK_SIZE_M < BLOCK_SIZE_M) ? (args.ne0 - r0*BLOCK_SIZE_M) : BLOCK_SIZE_M;
+    const short n_cols = (args.ne1 - r1*BLOCK_SIZE_N < BLOCK_SIZE_N) ? (args.ne1 - r1*BLOCK_SIZE_N) : BLOCK_SIZE_N;
 
     // a thread shouldn't load data outside of the matrix
-    short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
-    short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+    const short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
+    const short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
 
     simdgroup_T8x8     ma[4];
     simdgroup_float8x8 mb[2];
@@ -6344,23 +6086,26 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 
     short il = (tiitg % THREAD_PER_ROW);
 
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
+    const int i12 = im%args.ne12;
+    const int i13 = im/args.ne12;
 
-    uint   offset0 = (i12/r2)*nb02 + (i13/r3)*nb03;
-    ushort offset1 = il/nl;
+    const uint64_t offset0 = (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+    const short    offset1 = il/nl;
+
+    device const block_q * x = (device const block_q *)(src0
+        + args.nb01*(r0*BLOCK_SIZE_M + thread_row) + offset0) + offset1;
 
-    device const block_q * x = (device const block_q *)(src0 + (r0*BLOCK_SIZE_M + thread_row)*nb01 + offset0) + offset1;
     device const float   * y = (device const float   *)(src1
-        + nb13 * i13
-        + nb12 * i12
-        + nb11 * (r1 * BLOCK_SIZE_N + thread_col)
-        + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
+        + args.nb13*i13
+        + args.nb12*i12
+        + args.nb11*(r1*BLOCK_SIZE_N + thread_col)
+        + args.nb10*(BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
 
-    for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
+    for (int loop_k = 0; loop_k < args.ne00; loop_k += BLOCK_SIZE_K) {
         // load data and store to threadgroup memory
         T4x4 temp_a;
         dequantize_func(x, il, temp_a);
+
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         #pragma unroll(16)
@@ -6370,51 +6115,55 @@ kernel void kernel_mul_mm(device const  uchar * src0,
             +                     (tiitg/THREAD_PER_ROW)%8  + (i&7)*8) = temp_a[i/4][i%4];
         }
 
-        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL)*8*32 + 8*(tiitg/THREAD_PER_COL)) = *((device float2x4 *) y);
+        *(threadgroup float2x4 *)(sb + 32*8*(tiitg%THREAD_PER_COL) + 8*(tiitg/THREAD_PER_COL)) = *((device float2x4 *) y);
 
         il = (il + 2 < nl) ? il + 2 : il % 2;
-        x  = (il < 2) ? x + (2+nl-1)/nl : x;
+        x  = (il < 2) ? x + (2 + nl - 1)/nl : x;
         y += BLOCK_SIZE_K;
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
         // load matrices from threadgroup memory and conduct outer products
-        threadgroup T     * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
-        threadgroup float * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
+        threadgroup const T     * lsma = (sa + THREAD_MAT_M*SG_MAT_SIZE*(sgitg%2));
+        threadgroup const float * lsmb = (sb + THREAD_MAT_N*SG_MAT_SIZE*(sgitg/2));
 
         #pragma unroll(4)
-        for (short ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
+        for (short ik = 0; ik < BLOCK_SIZE_K/8; ik++) {
             #pragma unroll(4)
             for (short i = 0; i < 4; i++) {
                 simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
             }
+
             simdgroup_barrier(mem_flags::mem_none);
+
             #pragma unroll(2)
             for (short i = 0; i < 2; i++) {
                 simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
             }
 
-            lsma += BLOCK_SIZE_M/SG_MAT_ROW * SG_MAT_SIZE;
-            lsmb += BLOCK_SIZE_N/SG_MAT_ROW * SG_MAT_SIZE;
-
             #pragma unroll(8)
             for (short i = 0; i < 8; i++){
                 simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
             }
+
+            lsma += (BLOCK_SIZE_M/SG_MAT_ROW)*SG_MAT_SIZE;
+            lsmb += (BLOCK_SIZE_N/SG_MAT_ROW)*SG_MAT_SIZE;
         }
     }
 
-    if ((r0 + 1) * BLOCK_SIZE_M <= ne0 && (r1 + 1) * BLOCK_SIZE_N <= ne1) {
-        device float * C = dst + (BLOCK_SIZE_M * r0 + 32 * (sgitg &  1)) \
-                               + (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * ne0 + im*ne1*ne0;
+    if ((r0 + 1) * BLOCK_SIZE_M <= args.ne0 && (r1 + 1) * BLOCK_SIZE_N <= args.ne1) {
+        device float * C = (device float *) dst +
+            (BLOCK_SIZE_M * r0 + 32*(sgitg &  1)) + \
+            (BLOCK_SIZE_N * r1 + 16*(sgitg >> 1)) * args.ne0 + im*args.ne1*args.ne0;
+
         for (short i = 0; i < 8; i++) {
-            simdgroup_store(mc[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
+            simdgroup_store(mc[i], C + 8 * (i%4) + 8 * args.ne0 * (i/4), args.ne0);
         }
     } else {
         // block is smaller than 64x32, we should avoid writing data outside of the matrix
         threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float * temp_str = ((threadgroup float *) shared_memory) \
-                                      + 32 * (sgitg&1) + (16 * (sgitg>>1))*BLOCK_SIZE_M;
+        threadgroup float * temp_str = ((threadgroup float *) shmem) \
+                                     + 32*(sgitg&1) + (16*(sgitg >> 1))*BLOCK_SIZE_M;
         for (short i = 0; i < 8; i++) {
             simdgroup_store(mc[i], temp_str + 8*(i%4) + 8*BLOCK_SIZE_M*(i/4), BLOCK_SIZE_M);
         }
@@ -6423,7 +6172,7 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 
         if (sgitg == 0) {
             for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
-                device float  * D  = dst + (r0*BLOCK_SIZE_M) + (r1*BLOCK_SIZE_N + j)*ne0 + im*ne1*ne0;
+                device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + (r1*BLOCK_SIZE_N + j)*args.ne0 + im*args.ne1*args.ne0;
                 device float4 * D4 = (device float4 *) D;
 
                 threadgroup float  * C  = temp_str + (j*BLOCK_SIZE_M);
@@ -6444,36 +6193,37 @@ kernel void kernel_mul_mm(device const  uchar * src0,
 }
 
 // same as kernel_mul_mm_impl, but src1 and dst are accessed via indices stored in rowids
+// TODO: this kernel needs to be reimplemented from scratch for better performance
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
 void kernel_mul_mm_id_impl(
-        device const  uchar * src0,
-        device const  uchar * src1,
+        int32_t  ne00,
+        int32_t  ne02,
+        uint64_t nb01,
+        uint64_t nb02,
+        int32_t  ne11,
+        int32_t  ne12,
+        uint64_t nb10,
+        uint64_t nb11,
+        uint64_t nb12,
+        int32_t  ne0,
+        int32_t  ne1,
+        int64_t  ne0ne1,
+        device   const char * src0,
+        device   const char * src1,
         threadgroup ushort2 * rowids,
-        device        float * dst,
-        constant    int64_t & ne00,
-        constant    int64_t & ne02,
-        constant   uint64_t & nb01,
-        constant   uint64_t & nb02,
-        constant    int64_t & ne11,
-        constant    int64_t & ne12,
-        constant   uint64_t & nb10,
-        constant   uint64_t & nb11,
-        constant   uint64_t & nb12,
-        constant    int64_t & ne0,
-                    int64_t   ne1,
-                    int64_t   ne0ne1,
-        threadgroup   uchar * shared_memory,
-        uint3                 tgpig[[threadgroup_position_in_grid]],
-        uint                  tiitg[[thread_index_in_threadgroup]],
-        uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    threadgroup half  * sa = (threadgroup half  *)(shared_memory);
-    threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
-
-    const uint r0 = tgpig.y;
-    const uint r1 = tgpig.x;
-
-    if (r1 * BLOCK_SIZE_N >= ne1) return;
+        device         char * dst,
+        threadgroup    char * shmem,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup half  * sa = (threadgroup half  *)(shmem);
+    threadgroup float * sb = (threadgroup float *)(shmem + 4096);
+
+    const int r0 = tgpig.y;
+    const int r1 = tgpig.x;
+
+    if (r1*BLOCK_SIZE_N >= ne1) return;
 
     // if this block is of 64x32 shape or smaller
     short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
@@ -6485,9 +6235,9 @@ void kernel_mul_mm_id_impl(
 
     simdgroup_half8x8  ma[4];
     simdgroup_float8x8 mb[2];
-    simdgroup_float8x8 c_res[8];
+    simdgroup_float8x8 mc[8];
     for (int i = 0; i < 8; i++){
-        c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+        mc[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
     }
     short il = (tiitg % THREAD_PER_ROW);
 
@@ -6525,11 +6275,14 @@ void kernel_mul_mm_id_impl(
         threadgroup half  * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
         threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
 
+        #pragma unroll(BLOCK_SIZE_K/8)
         for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
+            #pragma unroll(4)
             for (int i = 0; i < 4; i++) {
                 simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
             }
             simdgroup_barrier(mem_flags::mem_none);
+            #pragma unroll(2)
             for (int i = 0; i < 2; i++) {
                 simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
             }
@@ -6537,29 +6290,42 @@ void kernel_mul_mm_id_impl(
             lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
             lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
 
+            #pragma unroll(8)
             for (int i = 0; i < 8; i++){
-                simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
+                simdgroup_multiply_accumulate(mc[i], mb[i/4], ma[i%4], mc[i]);
             }
         }
     }
 
     {
         threadgroup_barrier(mem_flags::mem_threadgroup);
-        threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
+        threadgroup float * temp_str = ((threadgroup float *) shmem) \
                                       + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
         for (int i = 0; i < 8; i++) {
-            simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
+            simdgroup_store(mc[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
         }
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
-        device float * C = dst + (BLOCK_SIZE_M * r0);
         if (sgitg == 0) {
             for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
                 threadgroup const auto & jid = rowids[r1 * BLOCK_SIZE_N + j];
-                int joff =  jid[0] * ne0 + jid[1] * ne0ne1;
-                for (int i = 0; i < n_rows; i++) {
-                    *(C + i + joff) = *(temp_str + i + j * BLOCK_SIZE_M);
+                int64_t joff = jid[0]*ne0 + jid[1]*ne0ne1;
+
+                device float  * D  = (device float  *) dst + (r0*BLOCK_SIZE_M) + joff;
+                device float4 * D4 = (device float4 *) D;
+
+                threadgroup float  * C  = temp_str + (j*BLOCK_SIZE_M);
+                threadgroup float4 * C4 = (threadgroup float4 *) C;
+
+                int i = 0;
+                for (; i < n_rows/4; i++) {
+                    *(D4 + i) = *(C4 + i);
+                }
+
+                i *= 4;
+                for (; i < n_rows; i++) {
+                    *(D + i) = *(C + i);
                 }
             }
         }
@@ -6568,48 +6334,34 @@ void kernel_mul_mm_id_impl(
 
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
 kernel void kernel_mul_mm_id(
-        device const   uchar * src0s,
-        device const   uchar * src1,
-        device         float * dst,
-        device const   uchar * ids,
-        constant     int64_t & nei0,
-        constant     int64_t & nei1,
-        constant    uint64_t & nbi1,
-        constant     int64_t & ne00,
-        constant     int64_t & ne02,
-        constant    uint64_t & nb01,
-        constant    uint64_t & nb02,
-        constant     int64_t & ne11,
-        constant     int64_t & ne12,
-        constant     int64_t & ne13,
-        constant    uint64_t & nb10,
-        constant    uint64_t & nb11,
-        constant    uint64_t & nb12,
-        constant     int64_t & ne0,
-        constant     int64_t & ne1,
-        constant    uint64_t & nb1,
-        threadgroup    uchar * shared_memory [[threadgroup(0)]],
-        uint3                  tgpig[[threadgroup_position_in_grid]],
-        uint                   tiitg[[thread_index_in_threadgroup]],
-        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+        constant ggml_metal_kargs_mul_mm_id & args,
+        device const char * src0s,
+        device const char * src1,
+        device       char * dst,
+        device const char * ids,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
 
     const int32_t i02 = tgpig.z;
+
     tgpig.z = 0;
 
-    device const uchar * src0 = src0s + i02*nb02;
+    device const char * src0 = src0s + i02*args.nb02;
 
     // row indices
-    threadgroup ushort2 * rowids = (threadgroup ushort2 *)(shared_memory + 8192);
+    threadgroup ushort2 * rowids = (threadgroup ushort2 *)(shmem + 8192);
 
     // TODO: parallelize this loop
-    int64_t _ne1 = 0;
-    for (ushort ii1 = 0; ii1 < nei1; ii1++) {
-        for (ushort ii0 = 0; ii0 < nei0; ii0++) {
-            int32_t id = ((device int32_t *) (ids + ii1*nbi1))[ii0];
+    int32_t _ne1 = 0;
+    for (ushort ii1 = 0; ii1 < args.nei1; ii1++) {
+        for (ushort ii0 = 0; ii0 < args.nei0; ii0++) {
+            int32_t id = ((device int32_t *) (ids + ii1*args.nbi1))[ii0];
             if (id == i02) {
-                //if (tiitg == 0) {
+                if (tiitg == 0) {
                     rowids[_ne1] = ushort2(ii0, ii1);
-                //}
+                }
                 _ne1++;
             }
         }
@@ -6618,23 +6370,23 @@ kernel void kernel_mul_mm_id(
     threadgroup_barrier(mem_flags::mem_threadgroup);
 
     kernel_mul_mm_id_impl<block_q, nl, dequantize_func>(
+        args.ne00,
+        args.ne02,
+        args.nb01,
+        args.nb02,
+        args.ne11,
+        args.ne12,
+        args.nb10,
+        args.nb11,
+        args.nb12,
+        args.ne0,
+        _ne1,
+        (int64_t)args.ne0*args.ne1,
         src0,
         src1,
         rowids,
         dst,
-        ne00,
-        ne02,
-        nb01,
-        nb02,
-        ne11,
-        ne12,
-        nb10,
-        nb11,
-        nb12,
-        ne0,
-        _ne1,
-        ne0*ne1,
-        shared_memory,
+        shmem,
         tgpig,
         tiitg,
         sgitg);
@@ -6743,194 +6495,110 @@ template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mat_mm_id_t kernel
 //
 
 typedef void (kernel_mul_mv_impl_t)(
-        device const  char * src0,
-        device const  char * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb00,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne11,
-                   int64_t   ne12,
-                  uint64_t   nb10,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-                   uint3     tgpig,
-                   uint      tiisg);
+        ggml_metal_kargs_mul_mv args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig,
+        ushort tiisg);
 
 typedef void (kernel_mul_mv2_impl_t)(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-                   int64_t   ne00,
-                   int64_t   ne01,
-                   int64_t   ne02,
-                  uint64_t   nb01,
-                  uint64_t   nb02,
-                  uint64_t   nb03,
-                   int64_t   ne10,
-                   int64_t   ne12,
-                  uint64_t   nb11,
-                  uint64_t   nb12,
-                  uint64_t   nb13,
-                   int64_t   ne0,
-                   int64_t   ne1,
-                   uint      r2,
-                   uint      r3,
-        threadgroup int8_t * shared_values,
-                   uint3     tgpig,
-                   uint      tiisg,
-                   uint      sgitg);
+        ggml_metal_kargs_mul_mv args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiisg,
+        ushort sgitg);
 
 template<kernel_mul_mv_impl_t impl_fn>
 void mmv_fn(
-        device const    char * src0,
-        device const    char * src1,
-        device         float * dst,
-                     int64_t   ne00,
-                     int64_t   ne01,
-                     int64_t   ne02,
-                    uint64_t   nb00,
-                    uint64_t   nb01,
-                    uint64_t   nb02,
-                    uint64_t   nb03,
-                     int64_t   ne10,
-                     int64_t   ne11,
-                     int64_t   ne12,
-                     int64_t   ne13,
-                    uint64_t   nb10,
-                    uint64_t   nb11,
-                    uint64_t   nb12,
-                    uint64_t   nb13,
-                     int64_t   ne0,
-                     int64_t   ne1,
-                    uint64_t   nb1,
-                        uint   r2,
-                        uint   r3,
-        threadgroup int8_t   * shared_values,
-        uint3                  tgpig,
-        uint                   tiitg,
-        uint                   tiisg,
-        uint                   sgitg) {
-    impl_fn(src0,src1,dst,ne00,ne01,ne02,nb00,nb01,nb02,nb03,ne10,ne11,ne12,nb10,nb11,nb12,nb13,ne0,ne1,r2,r3,tgpig,tiisg);
+        ggml_metal_kargs_mul_mv args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiitg,
+        ushort tiisg,
+        ushort sgitg) {
+    impl_fn(args, src0, src1, dst, tgpig, tiisg);
 }
 
 template<kernel_mul_mv2_impl_t impl_fn>
 void mmv_fn(
-        device const    char * src0,
-        device const    char * src1,
-        device         float * dst,
-                     int64_t   ne00,
-                     int64_t   ne01,
-                     int64_t   ne02,
-                    uint64_t   nb00,
-                    uint64_t   nb01,
-                    uint64_t   nb02,
-                    uint64_t   nb03,
-                     int64_t   ne10,
-                     int64_t   ne11,
-                     int64_t   ne12,
-                     int64_t   ne13,
-                    uint64_t   nb10,
-                    uint64_t   nb11,
-                    uint64_t   nb12,
-                    uint64_t   nb13,
-                     int64_t   ne0,
-                     int64_t   ne1,
-                    uint64_t   nb1,
-                        uint   r2,
-                        uint   r3,
-        threadgroup int8_t   * shared_values,
-        uint3                  tgpig,
-        uint                   tiitg,
-        uint                   tiisg,
-        uint                   sgitg) {
-    impl_fn(src0,(const device float *)src1,dst,ne00,ne01,ne02,nb01,nb02,nb03,ne10,ne12,nb11,nb12,nb13,ne0,ne1,r2,r3,shared_values,tgpig,tiisg,sgitg);
-}
-
-typedef decltype(mmv_fn<kernel_mul_mv_impl<half, half4, half, half4>>) mul_mv_impl_fn_t;
+        ggml_metal_kargs_mul_mv args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        threadgroup  char * shmem,
+        uint3  tgpig,
+        ushort tiitg,
+        ushort tiisg,
+        ushort sgitg) {
+    impl_fn(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
+}
+
+typedef decltype(mmv_fn<kernel_mul_mv_impl<half, half4, half, half4, ggml_metal_kargs_mul_mv>>) mul_mv_impl_fn_t;
 
 template<mul_mv_impl_fn_t impl_fn>
 kernel void kernel_mul_mv_id(
-        device const    char * src0s,
-        device const    char * src1,
-        device         float * dst,
-        device const    char * ids,
-        constant     int64_t & nei0,
-        constant     int64_t & nei1,
-        constant    uint64_t & nbi1,
-        constant     int64_t & ne00,
-        constant     int64_t & ne01,
-        constant     int64_t & ne02,
-        constant    uint64_t & nb00,
-        constant    uint64_t & nb01,
-        constant    uint64_t & nb02,
-        constant     int64_t & ne10,
-        constant     int64_t & ne11,
-        constant     int64_t & ne12,
-        constant     int64_t & ne13,
-        constant    uint64_t & nb10,
-        constant    uint64_t & nb11,
-        constant    uint64_t & nb12,
-        constant     int64_t & ne0,
-        constant     int64_t & ne1,
-        constant    uint64_t & nb1,
-        threadgroup int8_t   * shared_values [[threadgroup(0)]],
-        uint3                  tgpig[[threadgroup_position_in_grid]],
-        uint                   tiitg[[thread_index_in_threadgroup]],
-        uint                   tiisg[[thread_index_in_simdgroup]],
-        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
-    const int iid1 = tgpig.z/nei0;
-    const int idx = tgpig.z%nei0;
+        constant ggml_metal_kargs_mul_mv_id & args,
+        device const char * src0s,
+        device const char * src1,
+        device       char * dst,
+        device const char * ids,
+        threadgroup  char * shmem [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    const int iid1 = tgpig.z/args.nei0;
+    const int idx  = tgpig.z%args.nei0;
 
     tgpig.z = 0;
 
-    const int32_t i02 = ((device const int32_t *) (ids + iid1*nbi1))[idx];
+    const int32_t i02 = ((device const int32_t *) (ids + iid1*args.nbi1))[idx];
 
-    const int64_t i11 = idx % ne11;
+    const int64_t i11 = idx % args.ne11;
     const int64_t i12 = iid1;
 
     const int64_t i1 = idx;
     const int64_t i2 = i12;
 
-    device const char * src0_cur = src0s + i02*nb02;
-    device const char * src1_cur = src1  + i11*nb11 + i12*nb12;
-    device      float *  dst_cur = dst   + i1*ne0   + i2*ne1*ne0;
+    device const char * src0_cur = src0s + i02*args.nb02;
+    device const char * src1_cur = src1  + i11*args.nb11 + i12*args.nb12;
+
+    device char * dst_cur = dst + (i1*args.ne0 + i2*args.ne1*args.ne0)*sizeof(float);
+
+    ggml_metal_kargs_mul_mv args0 = {
+        /*.ne00 =*/ args.ne00,
+        /*.ne01 =*/ args.ne01,
+        /*.ne02 =*/ 1, // args.ne02,
+        /*.nb00 =*/ args.nb00,
+        /*.nb01 =*/ args.nb01,
+        /*.nb02 =*/ args.nb02,
+        /*.nb03 =*/ args.nb02, // args.ne02 == 1
+        /*.ne10 =*/ args.ne10,
+        /*.ne11 =*/ 1, // args.ne11,
+        /*.ne12 =*/ 1, // args.ne12,
+        /*.nb10 =*/ args.nb10,
+        /*.nb11 =*/ args.nb11,
+        /*.nb12 =*/ args.nb12,
+        /*.nb13 =*/ args.nb12, // ne12 == 1
+        /*.ne0  =*/ args.ne0,
+        /*.ne1  =*/ 1, // args.ne1,
+        /*.r2   =*/ 1,
+        /*.r3   =*/ 1,
+    };
 
     impl_fn(
+        args0,
         /* src0 */ src0_cur,
         /* src1 */ src1_cur,
         /* dst  */ dst_cur,
-        /* ne00 */ ne00,
-        /* ne01 */ ne01,
-        /* ne02 */ 1, // ne02,
-        /* nb00 */ nb00,
-        /* nb01 */ nb01,
-        /* nb02 */ nb02,
-        /* nb03 */ nb02, // ne02 == 1
-        /* ne10 */ ne10,
-        /* ne11 */ 1, // ne11,
-        /* ne12 */ 1, // ne12,
-        /* ne13 */ 1, // ne13,
-        /* nb10 */ nb10,
-        /* nb11 */ nb11,
-        /* nb12 */ nb12,
-        /* ne13 */ nb12, // ne12 == 1
-        /* ne0  */ ne0,
-        /* ne1  */ 1, // ne1,
-        /* nb1  */ nb1,
-        /* r2   */ 1,
-        /* r3   */ 1,
-        shared_values,
+        shmem,
         tgpig,
         tiitg,
         tiisg,
diff --git a/ggml/src/ggml-musa/CMakeLists.txt b/ggml/src/ggml-musa/CMakeLists.txt
new file mode 100644
index 00000000000..415b2b2e09c
--- /dev/null
+++ b/ggml/src/ggml-musa/CMakeLists.txt
@@ -0,0 +1,107 @@
+if (NOT EXISTS $ENV{MUSA_PATH})
+    if (NOT EXISTS /opt/musa)
+        set(MUSA_PATH /usr/local/musa)
+    else()
+        set(MUSA_PATH /opt/musa)
+    endif()
+else()
+    set(MUSA_PATH $ENV{MUSA_PATH})
+endif()
+
+set(CMAKE_C_COMPILER "${MUSA_PATH}/bin/clang")
+set(CMAKE_C_EXTENSIONS OFF)
+set(CMAKE_CXX_COMPILER "${MUSA_PATH}/bin/clang++")
+set(CMAKE_CXX_EXTENSIONS OFF)
+
+list(APPEND CMAKE_MODULE_PATH "${MUSA_PATH}/cmake")
+
+find_package(MUSAToolkit)
+
+if (MUSAToolkit_FOUND)
+    message(STATUS "MUSA Toolkit found")
+
+    if (NOT DEFINED MUSA_ARCHITECTURES)
+        set(MUSA_ARCHITECTURES "21;22")
+    endif()
+    message(STATUS "Using MUSA architectures: ${MUSA_ARCHITECTURES}")
+
+    file(GLOB   GGML_HEADERS_MUSA "../ggml-cuda/*.cuh")
+    list(APPEND GGML_HEADERS_MUSA "../../include/ggml-cuda.h")
+
+    file(GLOB   GGML_SOURCES_MUSA "../ggml-cuda/*.cu")
+    file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-wmma*.cu")
+    list(APPEND GGML_SOURCES_MUSA ${SRCS})
+    file(GLOB   SRCS "../ggml-cuda/template-instances/mmq*.cu")
+    list(APPEND GGML_SOURCES_MUSA ${SRCS})
+
+    if (GGML_CUDA_FA_ALL_QUANTS)
+        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*.cu")
+        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+    else()
+        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
+        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
+        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+        file(GLOB   SRCS "../ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
+        list(APPEND GGML_SOURCES_MUSA ${SRCS})
+    endif()
+
+    set_source_files_properties(${GGML_SOURCES_MUSA} PROPERTIES LANGUAGE CXX)
+    foreach(SOURCE ${GGML_SOURCES_MUSA})
+        set(COMPILE_FLAGS "-x musa -mtgpu")
+        foreach(ARCH ${MUSA_ARCHITECTURES})
+            set(COMPILE_FLAGS "${COMPILE_FLAGS} --cuda-gpu-arch=mp_${ARCH}")
+        endforeach()
+        set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS ${COMPILE_FLAGS})
+    endforeach()
+
+    ggml_add_backend_library(ggml-musa
+                             ${GGML_HEADERS_MUSA}
+                             ${GGML_SOURCES_MUSA}
+                            )
+
+    # TODO: do not use CUDA definitions for MUSA
+    target_compile_definitions(ggml PUBLIC GGML_USE_CUDA)
+
+    add_compile_definitions(GGML_USE_MUSA)
+    add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
+
+    if (GGML_CUDA_GRAPHS)
+        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
+    endif()
+
+    if (GGML_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    if (GGML_CUDA_FORCE_CUBLAS)
+        add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        add_compile_definitions(GGML_CUDA_NO_VMM)
+    endif()
+
+    if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+        add_compile_definitions(GGML_CUDA_F16)
+    endif()
+
+    if (GGML_CUDA_NO_PEER_COPY)
+        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+    endif()
+
+    if (GGML_STATIC)
+        target_link_libraries(ggml-musa PRIVATE MUSA::musart_static MUSA::mublas_static)
+    else()
+        target_link_libraries(ggml-musa PRIVATE MUSA::musart MUSA::mublas)
+    endif()
+
+    if (GGML_CUDA_NO_VMM)
+        # No VMM requested, no need to link directly with the musa driver lib (libmusa.so)
+    else()
+        target_link_libraries(ggml-musa PRIVATE MUSA::musa_driver)
+    endif()
+else()
+    message(FATAL_ERROR "MUSA Toolkit not found")
+endif()
diff --git a/ggml/src/ggml-opt.cpp b/ggml/src/ggml-opt.cpp
new file mode 100644
index 00000000000..7c3e24103a2
--- /dev/null
+++ b/ggml/src/ggml-opt.cpp
@@ -0,0 +1,854 @@
+#include "ggml-opt.h"
+
+#include "ggml.h"
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <cinttypes>
+#include <map>
+#include <random>
+#include <vector>
+
+struct ggml_opt_dataset {
+    struct ggml_context   * ctx    = nullptr;
+    ggml_backend_buffer_t   buf    = nullptr;
+    struct ggml_tensor    * data   = nullptr;
+    struct ggml_tensor    * labels = nullptr;
+
+    int64_t ndata       = -1;
+    int64_t ndata_shard = -1;
+    size_t  nbs_data    = -1;
+    size_t  nbs_labels  = -1;
+
+    std::vector<int64_t> permutation;
+};
+
+struct ggml_opt_context {
+    ggml_backend_sched_t    backend_sched        = nullptr;
+    ggml_cgraph           * allocated_graph      = nullptr;
+    ggml_cgraph           * allocated_graph_copy = nullptr;
+    struct ggml_context   * ctx_static           = nullptr;
+    struct ggml_context   * ctx_static_cpu       = nullptr;
+    struct ggml_context   * ctx_compute          = nullptr;
+    struct ggml_context   * ctx_copy             = nullptr;
+    ggml_backend_buffer_t   buf_static           = nullptr;
+    ggml_backend_buffer_t   buf_static_cpu       = nullptr;
+    std::mt19937            rng;
+
+    struct ggml_tensor * inputs  = nullptr;
+    struct ggml_tensor * outputs = nullptr;
+    struct ggml_tensor * labels  = nullptr;
+
+    struct ggml_tensor * loss     = nullptr;
+    struct ggml_tensor * pred     = nullptr;
+    struct ggml_tensor * ncorrect = nullptr;
+
+    struct ggml_cgraph * gf      = nullptr;
+    struct ggml_cgraph * gb_grad = nullptr;
+    struct ggml_cgraph * gb_opt  = nullptr;
+
+    int64_t iter               = 1;
+    int32_t opt_period         = 1;
+    int32_t opt_i              = 0;
+    bool    loss_per_datapoint = false;
+
+    ggml_opt_get_optimizer_params get_opt_pars = nullptr;
+    void * get_opt_pars_ud                     = nullptr;
+    struct ggml_tensor * adamw_params          = nullptr;
+};
+
+struct ggml_opt_result {
+    int64_t              ndata    = 0;
+    std::vector<float>   loss;
+    std::vector<int32_t> pred;
+    int64_t              ncorrect = 0;
+
+    int64_t opt_period         = -1;
+    bool    loss_per_datapoint = false;
+};
+
+// ====== Dataset ======
+
+ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label, int64_t ndata, int64_t ndata_shard) {
+    GGML_ASSERT(ne_datapoint >  0);
+    GGML_ASSERT(ne_label     >= 0);
+    GGML_ASSERT(ndata        >  0);
+    GGML_ASSERT(ndata_shard  >  0);
+
+    ggml_opt_dataset_t result = new ggml_opt_dataset;
+    result->ndata       = ndata;
+    result->ndata_shard = ndata_shard;
+
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ 2*ggml_tensor_overhead(),
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        result->ctx = ggml_init(params);
+    }
+
+    result->data = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_datapoint, ndata);
+    result->nbs_data = ggml_nbytes(result->data) * ndata_shard/ndata;
+
+    if (ne_label > 0) {
+        result->labels = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_label, ndata);
+        result->nbs_labels = ggml_nbytes(result->labels) * ndata_shard/ndata;
+    } else {
+        result->labels = nullptr;
+        result->nbs_labels = 0;
+    }
+
+    result->buf = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx, ggml_backend_cpu_buffer_type());
+
+    const int64_t nshards = ndata/ndata_shard;
+    result->permutation.resize(nshards);
+    for (int64_t i = 0; i < nshards; ++i) {
+        result->permutation[i] = i;
+    }
+    return result;
+}
+
+void ggml_opt_dataset_free(ggml_opt_dataset_t dataset) {
+    ggml_backend_buffer_free(dataset->buf);
+    ggml_free(dataset->ctx);
+    delete dataset;
+}
+
+struct ggml_tensor * ggml_opt_dataset_data(ggml_opt_dataset_t dataset) {
+    return dataset->data;
+}
+
+struct ggml_tensor * ggml_opt_dataset_labels(ggml_opt_dataset_t dataset) {
+    return dataset->labels;
+}
+
+void ggml_opt_dataset_shuffle(ggml_opt_context_t opt_ctx, ggml_opt_dataset_t dataset, int64_t idata) {
+    GGML_ASSERT(idata <= dataset->ndata);
+
+    if (idata < 0) {
+        std::shuffle(dataset->permutation.begin(), dataset->permutation.end(), opt_ctx->rng);
+        return;
+    }
+
+    GGML_ASSERT(idata % dataset->ndata_shard == 0);
+    const int64_t ishard_max = idata / dataset->ndata_shard;
+    std::shuffle(dataset->permutation.begin(), dataset->permutation.begin() + ishard_max, opt_ctx->rng);
+}
+
+void ggml_opt_dataset_get_batch(ggml_opt_dataset_t dataset, struct ggml_tensor * data_batch, struct ggml_tensor * labels_batch, int64_t ibatch) {
+    GGML_ASSERT(   data_batch && ggml_is_contiguous(data_batch));
+    GGML_ASSERT(!labels_batch || ggml_is_contiguous(labels_batch));
+    GGML_ASSERT((labels_batch == nullptr) == (dataset->labels == nullptr));
+
+    const size_t nb_data_batch = ggml_nbytes(data_batch);
+    GGML_ASSERT(nb_data_batch % dataset->nbs_data == 0);
+    const int64_t shards_per_batch = nb_data_batch / dataset->nbs_data;
+
+    if (labels_batch) {
+        const size_t nb_labels_batch = ggml_nbytes(labels_batch);
+        GGML_ASSERT(nb_labels_batch == shards_per_batch*dataset->nbs_labels);
+    }
+
+    GGML_ASSERT((ibatch + 1)*shards_per_batch <= int64_t(dataset->permutation.size()));
+
+    for (int64_t ishard_batch = 0; ishard_batch < shards_per_batch; ++ishard_batch) {
+        const int64_t ishard = dataset->permutation[ibatch*shards_per_batch + ishard_batch];
+
+        const char * ptr_data = (const char *) dataset->data->data + ishard*dataset->nbs_data;
+        ggml_backend_tensor_set(data_batch, ptr_data, ishard_batch*dataset->nbs_data, dataset->nbs_data);
+
+        if (!labels_batch) {
+            continue;
+        }
+
+        const char * ptr_labels = (const char *) dataset->labels->data + ishard*dataset->nbs_labels;
+        ggml_backend_tensor_set(labels_batch, ptr_labels, ishard_batch*dataset->nbs_labels, dataset->nbs_labels);
+    }
+}
+
+// ====== Model / Context ======
+
+struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * userdata) {
+    GGML_UNUSED(userdata);
+
+    ggml_opt_optimizer_params result;
+
+    result.adamw.alpha = 0.001f;
+    result.adamw.beta1 = 0.9f;
+    result.adamw.beta2 = 0.999f;
+    result.adamw.eps   = 1e-8f;
+    result.adamw.wd    = 0.0f;
+
+    return result;
+}
+
+struct ggml_opt_params ggml_opt_default_params(
+        ggml_backend_sched_t      backend_sched,
+        struct ggml_context     * ctx_compute,
+        struct ggml_tensor      * inputs,
+        struct ggml_tensor      * outputs,
+        enum ggml_opt_loss_type   loss_type) {
+    return {
+        /*backend_sched   =*/ backend_sched,
+        /*ctx_compute     =*/ ctx_compute,
+        /*inputs          =*/ inputs,
+        /*logits          =*/ outputs,
+        /*loss_type       =*/ loss_type,
+        /*build_type      =*/ GGML_OPT_BUILD_TYPE_OPT,
+        /*opt_period      =*/ 1,
+        /*get_opt_pars    =*/ ggml_opt_get_default_optimizer_params,
+        /*get_opt_pars_ud =*/ nullptr,
+    };
+}
+
+static ggml_tensor * map_tensor(std::map<ggml_tensor *, ggml_tensor *> & tensor_map, ggml_context * ctx, ggml_tensor * tensor) {
+    if (!tensor) {
+        return nullptr;
+    }
+
+    if (tensor_map.find(tensor) != tensor_map.end()) {
+        return tensor_map[tensor];
+    }
+
+    ggml_tensor * new_tensor = ggml_dup_tensor(ctx, tensor);
+    tensor_map[tensor] = new_tensor;
+
+    new_tensor->op = tensor->op;
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        new_tensor->nb[i] = tensor->nb[i];
+    }
+    new_tensor->flags = tensor->flags;
+    memcpy(new_tensor->op_params, tensor->op_params, sizeof(tensor->op_params));
+    strcpy(new_tensor->name, tensor->name);
+    new_tensor->data = tensor->data;
+    new_tensor->buffer = tensor->buffer;
+    new_tensor->extra = tensor->extra;
+    new_tensor->view_offs = tensor->view_offs;
+    new_tensor->view_src = map_tensor(tensor_map, ctx, tensor->view_src);
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        new_tensor->src[i] = map_tensor(tensor_map, ctx, tensor->src[i]);
+    }
+
+    return new_tensor;
+}
+
+static ggml_cgraph * dup_graph(ggml_context * ctx, ggml_cgraph * src) {
+    std::map<ggml_tensor *, ggml_tensor *> tensor_map;
+
+    ggml_cgraph * dst = ggml_new_graph_custom(ctx, src->size, /*grads =*/ true);
+
+    for (int i = 0; i < src->n_leafs; i++) {
+        ggml_build_forward_expand(dst, map_tensor(tensor_map, ctx, src->leafs[i]));
+    }
+    GGML_ASSERT(dst->n_leafs == src->n_leafs);
+    for (int i = 0; i < src->n_nodes; i++) {
+        ggml_build_forward_expand(dst, map_tensor(tensor_map, ctx, src->nodes[i]));
+    }
+    GGML_ASSERT(dst->n_nodes == src->n_nodes);
+    for (int i = 0; i < src->n_nodes; ++i) {
+        const size_t igrad_src = ggml_hash_find(&src->visited_hash_set, src->nodes[i]);
+        const size_t igrad_dst = ggml_hash_find(&dst->visited_hash_set, dst->nodes[i]);
+
+        GGML_ASSERT(igrad_src != GGML_HASHSET_FULL);
+        GGML_ASSERT(ggml_bitset_get(src->visited_hash_set.used, igrad_src));
+        GGML_ASSERT(igrad_dst != GGML_HASHSET_FULL);
+        GGML_ASSERT(ggml_bitset_get(dst->visited_hash_set.used, igrad_dst));
+
+        dst->grads[igrad_dst]     = src->grads[igrad_src];
+        dst->grad_accs[igrad_dst] = src->grad_accs[igrad_src];
+    }
+
+    return dst;
+}
+
+static void ggml_opt_alloc_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph) {
+    GGML_ASSERT(graph);
+    if (opt_ctx->allocated_graph == graph) {
+        return;
+    }
+
+    ggml_backend_sched_reset(opt_ctx->backend_sched); // clear allocation of previous graph
+
+    {
+        ggml_init_params params = {
+            /*.mem_size   =*/ ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE,
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        ggml_free(opt_ctx->ctx_copy);
+        opt_ctx->ctx_copy = ggml_init(params);
+    }
+
+    opt_ctx->allocated_graph_copy = dup_graph(opt_ctx->ctx_copy, graph);
+
+    ggml_backend_sched_alloc_graph(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
+    opt_ctx->allocated_graph = graph;
+}
+
+ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
+    ggml_opt_context_t result = new struct ggml_opt_context;
+    result->backend_sched   = params.backend_sched;
+    result->ctx_compute     = params.ctx_compute;
+    result->inputs          = params.inputs;
+    result->outputs         = params.outputs;
+    result->opt_period      = params.opt_period;
+    result->get_opt_pars    = params.get_opt_pars;
+    result->get_opt_pars_ud = params.get_opt_pars_ud;
+
+    GGML_ASSERT(result->inputs->data && "the inputs must be allocated statically");
+    GGML_ASSERT(result->opt_period >= 1);
+
+    const bool accumulate = params.build_type == GGML_OPT_BUILD_TYPE_GRAD ||
+        (params.build_type == GGML_OPT_BUILD_TYPE_OPT && result->opt_period > 1);
+
+    ggml_set_input(result->inputs);
+    ggml_set_output(result->outputs);
+
+    result->gf = ggml_new_graph_custom(result->ctx_compute, GGML_DEFAULT_GRAPH_SIZE, /*grads =*/ true); // Forward pass.
+    ggml_build_forward_expand(result->gf, result->outputs);
+
+    int n_param = 0;
+    for (int i = 0; i < result->gf->n_nodes; ++i) {
+        if (result->gf->nodes[i]->flags & GGML_TENSOR_FLAG_PARAM) {
+            n_param++;
+        }
+    }
+
+    {
+        // The static context is used for:
+        //   - gradients (1 tensor per param if using gradient accumulation)
+        //   - optimizer momenta (2 tensors per param)
+        //   - labels
+        //   - loss + its gradient (up to 5 tensors)
+        //   - pred
+        //   - ncorrect (2 tensors).
+        const size_t tensors_per_param = (accumulate ? 1 : 0) + (params.build_type == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0);
+        const size_t size_meta = (tensors_per_param*n_param + 9) * ggml_tensor_overhead();
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ size_meta,
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        result->ctx_static = ggml_init(params);
+    }
+    {
+        // The static cpu context is used for:
+        //   - optimizer parameters (1 for the entire context)
+        const size_t size_meta = 1 * ggml_tensor_overhead();
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ size_meta,
+            /*.mem_buffer =*/ nullptr,
+            /*.no_alloc   =*/ true,
+        };
+        result->ctx_static_cpu = ggml_init(params);
+    }
+
+
+    switch (params.loss_type) {
+        case GGML_OPT_LOSS_TYPE_MEAN: {
+            result->loss = ggml_sum(result->ctx_static, result->outputs);
+            ggml_set_name(result->loss, "loss_sum");
+            const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
+            result->loss = ggml_scale(result->ctx_static, result->loss, scale);
+            ggml_set_name(result->loss, "loss_mean");
+            result->loss_per_datapoint = true;
+            break;
+        }
+        case GGML_OPT_LOSS_TYPE_SUM: {
+            result->loss = ggml_sum(result->ctx_static, result->outputs);
+            ggml_set_name(result->loss, "loss_sum");
+            result->loss_per_datapoint = false;
+            break;
+        }
+        case GGML_OPT_LOSS_TYPE_CROSS_ENTROPY: {
+            result->labels = ggml_dup_tensor(result->ctx_static, result->outputs);
+            ggml_set_input(result->labels);
+            ggml_set_name(result->labels, "labels");
+            result->loss = ggml_cross_entropy_loss(result->ctx_static, result->outputs, result->labels);
+            ggml_set_name(result->loss, "loss_cross_entropy");
+            if (result->opt_period > 1) {
+                result->loss = ggml_scale(result->ctx_static, result->loss, 1.0f / result->opt_period);
+                ggml_set_name(result->loss, "loss_cross_entropy_scaled");
+            }
+            result->loss_per_datapoint = true;
+            break;
+        }
+        case GGML_OPT_LOSS_TYPE_MEAN_SQUARED_ERROR: {
+            result->labels = ggml_dup_tensor(result->ctx_static, result->outputs);
+            ggml_set_input(result->labels);
+            ggml_set_name(result->labels, "labels");
+            result->loss = ggml_sub(result->ctx_static, result->outputs, result->labels);
+            ggml_set_name(result->loss, "loss_error");
+            result->loss = ggml_sqr(result->ctx_static, result->loss);
+            ggml_set_name(result->loss, "loss_squared_error");
+            result->loss = ggml_sum(result->ctx_static, result->loss);
+            ggml_set_name(result->loss, "loss_sum_squared_error");
+            const float scale = 1.0f / (result->opt_period * ggml_nelements(result->outputs));
+            result->loss = ggml_scale(result->ctx_static, result->loss, scale);
+            ggml_set_name(result->loss, "loss_mean_squared_error");
+            result->loss_per_datapoint = true;
+            break;
+        }
+    }
+    ggml_set_output(result->loss);
+    ggml_set_loss(result->loss);
+    ggml_build_forward_expand(result->gf, result->loss);
+
+    result->pred = ggml_argmax(result->ctx_static, result->outputs);
+    ggml_set_name(result->pred, "pred");
+    ggml_set_output(result->pred);
+    ggml_build_forward_expand(result->gf, result->pred);
+
+    if (result->labels) {
+        result->ncorrect = ggml_count_equal(result->ctx_static, result->pred, ggml_argmax(result->ctx_static, result->labels));
+        ggml_set_name(result->ncorrect, "ncorrect");
+        ggml_set_output(result->ncorrect);
+        ggml_build_forward_expand(result->gf, result->ncorrect);
+    } else {
+        result->ncorrect = nullptr;
+    }
+
+    if (params.build_type == GGML_OPT_BUILD_TYPE_FORWARD) {
+        result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
+        return result;
+    }
+
+    // gb_grad == graph backward gradients, forward pass, then backward pass to calculate gradients.
+    result->gb_grad = ggml_graph_dup(result->ctx_compute, result->gf);
+    ggml_build_backward_expand(result->ctx_static, result->ctx_compute, result->gb_grad, accumulate);
+
+    if (params.build_type == GGML_OPT_BUILD_TYPE_GRAD) {
+        result->buf_static = ggml_backend_alloc_ctx_tensors(result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
+        ggml_graph_reset(result->gb_grad);
+        return result;
+    }
+
+    GGML_ASSERT(params.build_type == GGML_OPT_BUILD_TYPE_OPT);
+
+    // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step.
+    result->gb_opt = ggml_graph_dup(result->ctx_compute, result->gb_grad);
+
+    result->adamw_params = ggml_new_tensor_1d(result->ctx_static_cpu, GGML_TYPE_F32, 7);
+    ggml_set_input(result->adamw_params);
+    ggml_set_name(result->adamw_params, "adamw_params");
+
+    for (int i = result->gf->n_nodes-1; i >= 0; --i) {
+        struct ggml_tensor * node = result->gb_opt->nodes[i];
+        struct ggml_tensor * grad = ggml_graph_get_grad(result->gb_opt, node);
+
+        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
+            struct ggml_tensor * m        = ggml_dup_tensor(result->ctx_static, node);
+            struct ggml_tensor * v        = ggml_dup_tensor(result->ctx_static, node);
+            struct ggml_tensor * opt_step = ggml_opt_step_adamw(result->ctx_compute, node, grad, m, v, result->adamw_params);
+            ggml_build_forward_expand(result->gb_opt, opt_step);
+        }
+    }
+
+    result->buf_static = ggml_backend_alloc_ctx_tensors(
+        result->ctx_static, ggml_backend_sched_get_backend(result->backend_sched, 0));
+
+    result->buf_static_cpu = ggml_backend_alloc_ctx_tensors_from_buft(result->ctx_static_cpu, ggml_backend_cpu_buffer_type());
+
+    ggml_graph_reset(result->gb_opt);
+
+    return result;
+}
+
+void ggml_opt_free(ggml_opt_context_t opt_ctx) {
+    if (opt_ctx == nullptr) {
+        return;
+    }
+    ggml_backend_buffer_free(opt_ctx->buf_static);
+    ggml_backend_buffer_free(opt_ctx->buf_static_cpu);
+    ggml_free(opt_ctx->ctx_static);
+    ggml_free(opt_ctx->ctx_static_cpu);
+    delete opt_ctx;
+}
+
+void ggml_opt_reset(ggml_opt_context_t opt_ctx, bool optimizer) {
+    if (optimizer) {
+        ggml_graph_reset(opt_ctx->gb_opt);
+        opt_ctx->iter = 1;
+    } else {
+        ggml_graph_reset(opt_ctx->gb_grad);
+    }
+}
+
+struct ggml_tensor * ggml_opt_inputs(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->inputs;
+}
+
+struct ggml_tensor * ggml_opt_outputs(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->outputs;
+}
+
+struct ggml_tensor * ggml_opt_labels(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->labels;
+}
+
+struct ggml_tensor * ggml_opt_loss(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->loss;
+}
+
+struct ggml_tensor * ggml_opt_pred(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->pred;
+}
+
+struct ggml_tensor * ggml_opt_ncorrect(ggml_opt_context_t opt_ctx) {
+    return opt_ctx->ncorrect;
+}
+
+struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node) {
+    return ggml_graph_get_grad_acc(opt_ctx->gb_opt, node);
+}
+
+// ====== Optimization Result ======
+
+ggml_opt_result_t ggml_opt_result_init() {
+    return new ggml_opt_result;
+}
+
+void ggml_opt_result_free(ggml_opt_result_t result) {
+    delete result;
+}
+
+void ggml_opt_result_reset(ggml_opt_result_t result) {
+    result->ndata = 0;
+    result->loss.clear();
+    result->pred.clear();
+    result->ncorrect = 0;
+}
+
+void ggml_opt_result_ndata(ggml_opt_result_t result, int64_t * ndata) {
+    *ndata = result->ndata;
+}
+
+void ggml_opt_result_loss(ggml_opt_result_t result, double * loss, double * unc) {
+    const int64_t nbatches = result->loss.size(); // Number of physical batches.
+
+    if (nbatches == 0) {
+        *loss = 0.0;
+        *unc  = NAN;
+        return;
+    }
+
+    double sum         = 0.0;
+    double sum_squared = 0.0;
+
+    for (const float & loss : result->loss) {
+        // If the loss is per datapoint it was scaled by 1.0f/opt_period for each physical batch.
+        const float loss_scaled = result->loss_per_datapoint ? loss*result->opt_period : loss;
+        sum         += loss_scaled;
+        sum_squared += loss_scaled*loss_scaled;
+    }
+
+    const double mean = sum/nbatches;
+    *loss = result->loss_per_datapoint ? mean : sum;
+
+    if (!unc) {
+        return;
+    }
+
+    if (nbatches < 2) {
+        *unc = NAN;
+        return;
+    }
+
+    const double var_sum = sum_squared/nbatches - mean*mean; // variance without Bessel's correction, i.e. nbatches/(nbatches-1)
+    *unc = result->loss_per_datapoint ? sqrt(var_sum / (nbatches - 1)) : sqrt(var_sum * nbatches/(nbatches - 1));
+}
+
+void ggml_opt_result_pred(ggml_opt_result_t result, int32_t * pred) {
+    for (size_t i = 0; i < result->pred.size(); ++i) {
+        pred[i] = result->pred[i];
+    }
+}
+
+void ggml_opt_result_accuracy(ggml_opt_result_t result, double * accuracy, double * unc) {
+    *accuracy = result->ncorrect >= 0 ? double(result->ncorrect) / double(result->ndata) : NAN;
+
+    if (!unc) {
+        return;
+    }
+
+    *unc = result->ncorrect >= 0 && result->ndata >= 2 ?
+        sqrt((*accuracy) * (1.0 - (*accuracy)) / double(result->ndata - 1)) : NAN;
+}
+
+// ====== Computation ======
+
+static void ggml_opt_eval_graph(ggml_opt_context_t opt_ctx, ggml_cgraph * graph, ggml_opt_result * result) {
+    if (graph != opt_ctx->gf) {
+        struct ggml_opt_optimizer_params opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud);
+
+        GGML_ASSERT(opt_pars.adamw.alpha >  0.0f);
+        GGML_ASSERT(opt_pars.adamw.beta1 >= 0.0f);
+        GGML_ASSERT(opt_pars.adamw.beta1 <= 1.0f);
+        GGML_ASSERT(opt_pars.adamw.beta2 >= 0.0f);
+        GGML_ASSERT(opt_pars.adamw.beta2 <= 1.0f);
+        GGML_ASSERT(opt_pars.adamw.eps   >= 0.0f);
+        GGML_ASSERT(opt_pars.adamw.wd    >= 0.0f);
+        GGML_ASSERT(opt_pars.adamw.wd    <= 1.0f);
+
+        // beta1, beta2 after applying warmup
+        const float beta1h = 1.0f/(1.0f - powf(opt_pars.adamw.beta1, opt_ctx->iter));
+        const float beta2h = 1.0f/(1.0f - powf(opt_pars.adamw.beta2, opt_ctx->iter));
+
+        float * adamw_par_data = ggml_get_data_f32(opt_ctx->adamw_params);
+        adamw_par_data[0] = opt_pars.adamw.alpha;
+        adamw_par_data[1] = opt_pars.adamw.beta1;
+        adamw_par_data[2] = opt_pars.adamw.beta2;
+        adamw_par_data[3] = opt_pars.adamw.eps;
+        adamw_par_data[4] = opt_pars.adamw.wd;
+        adamw_par_data[5] = beta1h;
+        adamw_par_data[6] = beta2h;
+    }
+
+    ggml_opt_alloc_graph(opt_ctx, graph);
+    ggml_backend_sched_graph_compute(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
+    opt_ctx->iter += opt_ctx->allocated_graph == opt_ctx->gb_opt;
+
+    if (!result) {
+        return;
+    }
+
+    if (result->ndata == 0) {
+        result->loss_per_datapoint = opt_ctx->loss_per_datapoint;
+        result->opt_period         = opt_ctx->opt_period;
+    } else {
+        GGML_ASSERT(result->loss_per_datapoint == opt_ctx->loss_per_datapoint);
+        GGML_ASSERT(result->opt_period         == opt_ctx->opt_period);
+    }
+
+    const int64_t ndata = opt_ctx->outputs->ne[1];
+    GGML_ASSERT(result->ndata == ndata*int64_t(result->loss.size()) && "varying batch size not supported");
+    result->ndata += ndata;
+
+    GGML_ASSERT(ggml_is_scalar(opt_ctx->loss));
+    GGML_ASSERT(opt_ctx->loss->type == GGML_TYPE_F32);
+    float loss;
+    ggml_backend_tensor_get(opt_ctx->loss, &loss, 0, ggml_nbytes(opt_ctx->loss));
+    result->loss.push_back(loss);
+
+    GGML_ASSERT(opt_ctx->pred->type == GGML_TYPE_I32);
+    std::vector<int32_t> pred(ndata);
+    ggml_backend_tensor_get(opt_ctx->pred, pred.data(), 0, ggml_nbytes(opt_ctx->pred));
+    result->pred.insert(result->pred.end(), pred.begin(), pred.end());
+
+    if (!opt_ctx->labels || result->ncorrect < 0) {
+        result->ncorrect = -1;
+        return;
+    }
+
+    GGML_ASSERT(ggml_is_scalar(opt_ctx->ncorrect));
+    GGML_ASSERT(opt_ctx->ncorrect->type == GGML_TYPE_I64);
+    int64_t ncorrect;
+    ggml_backend_tensor_get(opt_ctx->ncorrect, &ncorrect, 0, ggml_nbytes(opt_ctx->ncorrect));
+    result->ncorrect += ncorrect;
+}
+
+void ggml_opt_forward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) {
+    ggml_opt_eval_graph(opt_ctx, opt_ctx->gf, result);
+}
+
+void ggml_opt_forward_backward(ggml_opt_context_t opt_ctx, ggml_opt_result * result) {
+    if (opt_ctx->opt_period == 1) {
+        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result);
+        return;
+    }
+
+    const int32_t opt_i_next = (opt_ctx->opt_i + 1) % opt_ctx->opt_period;
+    if (opt_i_next == 0) {
+        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_opt, result);
+        ggml_opt_reset(opt_ctx, /*optimizer =*/ false);
+    } else {
+        ggml_opt_eval_graph(opt_ctx, opt_ctx->gb_grad, result);
+    }
+    opt_ctx->opt_i = opt_i_next;
+}
+
+// ====== High-Level Functions ======
+
+void ggml_opt_epoch(
+        ggml_opt_context_t      opt_ctx,
+        ggml_opt_dataset_t      dataset,
+        ggml_opt_result_t       result_train,
+        ggml_opt_result_t       result_eval,
+        int64_t                 idata_split,
+        ggml_opt_epoch_callback callback_train,
+        ggml_opt_epoch_callback callback_eval) {
+    struct ggml_tensor * inputs = ggml_opt_inputs(opt_ctx);
+    struct ggml_tensor * labels = ggml_opt_labels(opt_ctx);
+    struct ggml_tensor * data   = ggml_opt_dataset_data(dataset);
+    GGML_ASSERT(data->ne[0] == inputs->ne[0]);
+
+    const int64_t ndata       =   data->ne[1];
+    const int64_t ndata_batch = inputs->ne[1];
+
+    GGML_ASSERT(data->ne[1] % inputs->ne[1] == 0);
+    const int64_t nbatches = ndata/ndata_batch;
+
+    idata_split = idata_split < 0 ? ndata : idata_split;
+    GGML_ASSERT(idata_split % ndata_batch == 0);
+    const int64_t ibatch_split = idata_split / ndata_batch;
+
+    int64_t ibatch = 0;
+    int64_t t_loop_start = ggml_time_us();
+    for (; ibatch < ibatch_split; ++ibatch) {
+        ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
+        ggml_opt_forward_backward(opt_ctx, result_train);
+        if (callback_train) {
+            callback_train(true, opt_ctx, dataset, result_train, ibatch+1, ibatch_split, t_loop_start);
+        }
+    }
+    t_loop_start = ggml_time_us();
+    for (; ibatch < nbatches; ++ibatch) {
+        ggml_opt_dataset_get_batch(dataset, inputs, labels, ibatch);
+        ggml_opt_forward(opt_ctx, result_eval);
+        if (callback_eval) {
+            callback_eval(false, opt_ctx, dataset, result_eval, ibatch+1-ibatch_split, nbatches-ibatch_split, t_loop_start);
+        }
+    }
+}
+
+void ggml_opt_epoch_callback_progress_bar(
+        bool               train,
+        ggml_opt_context_t opt_ctx,
+        ggml_opt_dataset_t dataset,
+        ggml_opt_result_t  result,
+        int64_t            ibatch,
+        int64_t            ibatch_max,
+        int64_t            t_start_us) {
+    fprintf(stderr, "%s[", train ? "train: " : "val:   ");
+
+    constexpr int64_t bar_length = 25;
+    for (int64_t j = 0; j < bar_length; ++j) {
+        const int64_t ibatch_j = ibatch_max * j/bar_length;
+        if (ibatch_j < ibatch) {
+            fprintf(stderr, "=");
+        } else if (ibatch_max * (j - 1)/bar_length < ibatch) {
+            fprintf(stderr, ">");
+        } else {
+            fprintf(stderr, " ");
+        }
+    }
+
+    const int64_t batch_size = ggml_opt_inputs(opt_ctx)->ne[1];
+    const int64_t idata      = ibatch*batch_size;
+    const int64_t idata_max  = ibatch_max*batch_size;
+
+    double loss;
+    double loss_unc;
+    ggml_opt_result_loss(result, &loss, &loss_unc);
+
+    double accuracy;
+    double accuracy_unc;
+    ggml_opt_result_accuracy(result, &accuracy, &accuracy_unc);
+
+    const int64_t t_ibatch_us = ggml_time_us() - t_start_us;
+    int64_t t_ibatch_s = t_ibatch_us / 1000000;
+    const int64_t t_ibatch_h = t_ibatch_s / 3600;
+    t_ibatch_s -= t_ibatch_h * 3600;
+    const int64_t t_ibatch_m = t_ibatch_s / 60;
+    t_ibatch_s -= t_ibatch_m * 60;
+
+    const int64_t t_eta_us = t_ibatch_us * (ibatch_max - ibatch)/ibatch;
+    int64_t t_eta_s = t_eta_us / 1000000;
+    const int64_t t_eta_h = t_eta_s / 3600;
+    t_eta_s -= t_eta_h * 3600;
+    const int64_t t_eta_m = t_eta_s / 60;
+    t_eta_s -= t_eta_m * 60;
+
+    fprintf(stderr, "| data=%06" PRId64 "/%06" PRId64 ", loss=%.6lf+-%.6lf, accuracy=%.2lf+-%.2lf%%, "
+            "t=%02" PRId64 ":%02" PRId64 ":%02" PRId64 ", ETA=%02" PRId64 ":%02" PRId64 ":%02" PRId64 "]\r",
+            idata, idata_max, loss, loss_unc, 100.0*accuracy, 100.0*accuracy_unc,
+            t_ibatch_h, t_ibatch_m, t_ibatch_s, t_eta_h, t_eta_m, t_eta_s);
+    if (ibatch == ibatch_max) {
+        fprintf(stderr, "\n");
+    }
+    fflush(stderr);
+
+    GGML_UNUSED(dataset);
+}
+
+void ggml_opt_fit(
+        ggml_backend_sched_t            backend_sched,
+        ggml_context                  * ctx_compute,
+        ggml_tensor                   * inputs,
+        ggml_tensor                   * outputs,
+        ggml_opt_dataset_t              dataset,
+        enum ggml_opt_loss_type         loss_type,
+        ggml_opt_get_optimizer_params   get_opt_pars,
+        int64_t                         nepoch,
+        int64_t                         nbatch_logical,
+        float                           val_split,
+        bool                            silent) {
+    ggml_time_init();
+    const int64_t t_start_us = ggml_time_us();
+
+    const int64_t ndata           = ggml_opt_dataset_data(dataset)->ne[1];
+    const int64_t nbatch_physical = inputs->ne[1];
+    GGML_ASSERT(ndata          % nbatch_logical  == 0);
+    GGML_ASSERT(nbatch_logical % nbatch_physical == 0);
+
+    const int64_t opt_period       = nbatch_logical / nbatch_physical;
+    const int64_t nbatches_logical = ndata / nbatch_logical;
+
+    GGML_ASSERT(val_split >= 0.0f);
+    GGML_ASSERT(val_split <  1.0f);
+    const int64_t ibatch_split = int64_t(((1.0f - val_split) * nbatches_logical)) * opt_period; // train <-> val split index (physical)
+    const int64_t idata_split  = ibatch_split * nbatch_physical;
+
+    int64_t epoch = 1;
+
+    ggml_opt_params params = ggml_opt_default_params(backend_sched, ctx_compute, inputs, outputs, loss_type);
+    params.opt_period      = opt_period;
+    params.get_opt_pars    = get_opt_pars;
+    params.get_opt_pars_ud = &epoch;
+    ggml_opt_context_t opt_ctx = ggml_opt_init(params);
+
+    // Shuffling the data is generally useful but there is only a point if not all data is used in a single batch.
+    if (nbatch_logical < ndata) {
+        ggml_opt_dataset_shuffle(opt_ctx, dataset, -1); // Shuffle all data (train + validation).
+    }
+
+    ggml_opt_result_t result_train = ggml_opt_result_init();
+    ggml_opt_result_t result_val   = ggml_opt_result_init();
+
+    ggml_opt_epoch_callback epoch_callback = silent ? nullptr : ggml_opt_epoch_callback_progress_bar;
+
+    for (; epoch <= nepoch; ++epoch) {
+        if (nbatch_logical < idata_split) {
+            ggml_opt_dataset_shuffle(opt_ctx, dataset, idata_split);
+        }
+
+        ggml_opt_result_reset(result_train);
+        ggml_opt_result_reset(result_val);
+
+        if (!silent) {
+            fprintf(stderr, "%s: epoch %04" PRId64 "/%04" PRId64 ":\n", __func__, epoch, nepoch);
+        }
+        ggml_opt_epoch(opt_ctx, dataset, result_train, result_val, idata_split, epoch_callback, epoch_callback);
+        if (!silent) {
+            fprintf(stderr, "\n");
+        }
+    }
+
+    if (!silent) {
+        int64_t t_total_s = (ggml_time_us() - t_start_us) / 1000000;
+        const int64_t t_total_h = t_total_s / 3600;
+        t_total_s -= t_total_h * 3600;
+        const int64_t t_total_m = t_total_s / 60;
+        t_total_s -= t_total_m * 60;
+        fprintf(stderr, "%s: training took %02" PRId64 ":%02" PRId64 ":%02" PRId64 "\n", __func__, t_total_h, t_total_m, t_total_s);
+    }
+
+    ggml_opt_free(opt_ctx);
+    ggml_opt_result_free(result_train);
+    ggml_opt_result_free(result_val);
+}
diff --git a/ggml/src/ggml-quants.c b/ggml/src/ggml-quants.c
index 82a463f27cc..7918388ae9f 100644
--- a/ggml/src/ggml-quants.c
+++ b/ggml/src/ggml-quants.c
@@ -3,7 +3,7 @@
 
 #include "ggml-quants.h"
 #include "ggml-impl.h"
-#include "ggml-cpu-impl.h"
+#include "ggml-cpu/ggml-cpu-impl.h"
 #include "ggml-cpu.h"
 
 #include <math.h>
@@ -27,643 +27,6 @@
 
 #define UNUSED GGML_UNUSED
 
-// some compilers don't provide _mm256_set_m128i, e.g. gcc 7
-#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
-
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
-// multiply int8_t, add results pairwise twice
-static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
-    // Get absolute values of x vectors
-    const __m128i ax = _mm_sign_epi8(x, x);
-    // Sign the values of the y vectors
-    const __m128i sy = _mm_sign_epi8(y, x);
-    // Perform multiplication and create 16-bit values
-    const __m128i dot = _mm_maddubs_epi16(ax, sy);
-    const __m128i ones = _mm_set1_epi16(1);
-    return _mm_madd_epi16(ones, dot);
-}
-
-#if __AVX__ || __AVX2__ || __AVX512F__
-// horizontally add 8 floats
-static inline float hsum_float_8(const __m256 x) {
-    __m128 res = _mm256_extractf128_ps(x, 1);
-    res = _mm_add_ps(res, _mm256_castps256_ps128(x));
-    res = _mm_add_ps(res, _mm_movehl_ps(res, res));
-    res = _mm_add_ss(res, _mm_movehdup_ps(res));
-    return _mm_cvtss_f32(res);
-}
-
-// horizontally add 8 int32_t
-static inline int hsum_i32_8(const __m256i a) {
-    const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1));
-    const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128);
-    const __m128i sum64 = _mm_add_epi32(hi64, sum128);
-    const __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
-    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
-}
-
-// horizontally add 4 int32_t
-static inline int hsum_i32_4(const __m128i a) {
-    const __m128i hi64 = _mm_unpackhi_epi64(a, a);
-    const __m128i sum64 = _mm_add_epi32(hi64, a);
-    const __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
-    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
-}
-
-#if defined(__AVX2__) || defined(__AVX512F__)
-// spread 32 bits to 32 bytes { 0x00, 0xFF }
-static inline __m256i bytes_from_bits_32(const uint8_t * x) {
-    uint32_t x32;
-    memcpy(&x32, x, sizeof(uint32_t));
-    const __m256i shuf_mask = _mm256_set_epi64x(
-            0x0303030303030303, 0x0202020202020202,
-            0x0101010101010101, 0x0000000000000000);
-    __m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask);
-    const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe);
-    bytes = _mm256_or_si256(bytes, bit_mask);
-    return _mm256_cmpeq_epi8(bytes, _mm256_set1_epi64x(-1));
-}
-
-// Unpack 32 4-bit fields into 32 bytes
-// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
-static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
-{
-    const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi);
-    const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp);
-    const __m256i lowMask = _mm256_set1_epi8( 0xF );
-    return _mm256_and_si256(lowMask, bytes);
-}
-
-// add int16_t pairwise and return as float vector
-static inline __m256 sum_i16_pairs_float(const __m256i x) {
-    const __m256i ones = _mm256_set1_epi16(1);
-    const __m256i summed_pairs = _mm256_madd_epi16(ones, x);
-    return _mm256_cvtepi32_ps(summed_pairs);
-}
-
-static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
-#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
-    const __m256i zero = _mm256_setzero_si256();
-    const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy);
-    return _mm256_cvtepi32_ps(summed_pairs);
-#else
-    // Perform multiplication and create 16-bit values
-    const __m256i dot = _mm256_maddubs_epi16(ax, sy);
-    return sum_i16_pairs_float(dot);
-#endif
-}
-
-// multiply int8_t, add results pairwise twice and return as float vector
-static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
-#if __AVXVNNIINT8__
-    const __m256i zero = _mm256_setzero_si256();
-    const __m256i summed_pairs = _mm256_dpbssd_epi32(zero, x, y);
-    return _mm256_cvtepi32_ps(summed_pairs);
-#else
-    // Get absolute values of x vectors
-    const __m256i ax = _mm256_sign_epi8(x, x);
-    // Sign the values of the y vectors
-    const __m256i sy = _mm256_sign_epi8(y, x);
-    return mul_sum_us8_pairs_float(ax, sy);
-#endif
-}
-
-static inline __m128i packNibbles( __m256i bytes )
-{
-    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
-#if __AVX512F__
-    const __m256i bytes_srli_4 = _mm256_srli_epi16(bytes, 4);   // 0000_0000_abcd_0000
-    bytes = _mm256_or_si256(bytes, bytes_srli_4);               // 0000_abcd_abcd_efgh
-    return _mm256_cvtepi16_epi8(bytes);                         // abcd_efgh
-#else
-    const __m256i lowByte = _mm256_set1_epi16( 0xFF );
-    __m256i high = _mm256_andnot_si256( lowByte, bytes );
-    __m256i low = _mm256_and_si256( lowByte, bytes );
-    high = _mm256_srli_epi16( high, 4 );
-    bytes = _mm256_or_si256( low, high );
-
-    // Compress uint16_t lanes into bytes
-    __m128i r0 = _mm256_castsi256_si128( bytes );
-    __m128i r1 = _mm256_extracti128_si256( bytes, 1 );
-    return _mm_packus_epi16( r0, r1 );
-#endif
-}
-#elif defined(__AVX__)
-// spread 32 bits to 32 bytes { 0x00, 0xFF }
-static inline __m256i bytes_from_bits_32(const uint8_t * x) {
-    uint32_t x32;
-    memcpy(&x32, x, sizeof(uint32_t));
-    const __m128i shuf_maskl = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
-    const __m128i shuf_maskh = _mm_set_epi64x(0x0303030303030303, 0x0202020202020202);
-    __m128i bytesl = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskl);
-    __m128i bytesh = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskh);
-    const __m128i bit_mask = _mm_set1_epi64x(0x7fbfdfeff7fbfdfe);
-    bytesl = _mm_or_si128(bytesl, bit_mask);
-    bytesh = _mm_or_si128(bytesh, bit_mask);
-    bytesl = _mm_cmpeq_epi8(bytesl, _mm_set1_epi64x(-1));
-    bytesh = _mm_cmpeq_epi8(bytesh, _mm_set1_epi64x(-1));
-    return MM256_SET_M128I(bytesh, bytesl);
-}
-
-// Unpack 32 4-bit fields into 32 bytes
-// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
-static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
-{
-    // Load 16 bytes from memory
-    __m128i tmpl = _mm_loadu_si128((const __m128i *)rsi);
-    __m128i tmph = _mm_srli_epi16(tmpl, 4);
-    const __m128i lowMask = _mm_set1_epi8(0xF);
-    tmpl = _mm_and_si128(lowMask, tmpl);
-    tmph = _mm_and_si128(lowMask, tmph);
-    return MM256_SET_M128I(tmph, tmpl);
-}
-
-// add int16_t pairwise and return as float vector
-static inline __m256 sum_i16_pairs_float(const __m128i xh, const __m128i xl) {
-    const __m128i ones = _mm_set1_epi16(1);
-    const __m128i summed_pairsl = _mm_madd_epi16(ones, xl);
-    const __m128i summed_pairsh = _mm_madd_epi16(ones, xh);
-    const __m256i summed_pairs = MM256_SET_M128I(summed_pairsh, summed_pairsl);
-    return _mm256_cvtepi32_ps(summed_pairs);
-}
-
-static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
-    const __m128i axl = _mm256_castsi256_si128(ax);
-    const __m128i axh = _mm256_extractf128_si256(ax, 1);
-    const __m128i syl = _mm256_castsi256_si128(sy);
-    const __m128i syh = _mm256_extractf128_si256(sy, 1);
-    // Perform multiplication and create 16-bit values
-    const __m128i dotl = _mm_maddubs_epi16(axl, syl);
-    const __m128i doth = _mm_maddubs_epi16(axh, syh);
-    return sum_i16_pairs_float(doth, dotl);
-}
-
-// multiply int8_t, add results pairwise twice and return as float vector
-static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
-    const __m128i xl = _mm256_castsi256_si128(x);
-    const __m128i xh = _mm256_extractf128_si256(x, 1);
-    const __m128i yl = _mm256_castsi256_si128(y);
-    const __m128i yh = _mm256_extractf128_si256(y, 1);
-    // Get absolute values of x vectors
-    const __m128i axl = _mm_sign_epi8(xl, xl);
-    const __m128i axh = _mm_sign_epi8(xh, xh);
-    // Sign the values of the y vectors
-    const __m128i syl = _mm_sign_epi8(yl, xl);
-    const __m128i syh = _mm_sign_epi8(yh, xh);
-    // Perform multiplication and create 16-bit values
-    const __m128i dotl = _mm_maddubs_epi16(axl, syl);
-    const __m128i doth = _mm_maddubs_epi16(axh, syh);
-    return sum_i16_pairs_float(doth, dotl);
-}
-
-static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 )
-{
-    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
-    const __m128i lowByte = _mm_set1_epi16( 0xFF );
-    __m128i high = _mm_andnot_si128( lowByte, bytes1 );
-    __m128i low = _mm_and_si128( lowByte, bytes1 );
-    high = _mm_srli_epi16( high, 4 );
-    bytes1 = _mm_or_si128( low, high );
-    high = _mm_andnot_si128( lowByte, bytes2 );
-    low = _mm_and_si128( lowByte, bytes2 );
-    high = _mm_srli_epi16( high, 4 );
-    bytes2 = _mm_or_si128( low, high );
-
-    return _mm_packus_epi16( bytes1, bytes2);
-}
-
-static inline __m128i mul_add_epi8_sse(const __m128i x, const __m128i y) {
-    const __m128i ax = _mm_sign_epi8(x, x);
-    const __m128i sy = _mm_sign_epi8(y, x);
-    return _mm_maddubs_epi16(ax, sy);
-}
-#endif
-#elif defined(__SSSE3__)
-// horizontally add 4x4 floats
-static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
-    __m128 res_0 =_mm_hadd_ps(a, b);
-    __m128 res_1 =_mm_hadd_ps(c, d);
-    __m128 res =_mm_hadd_ps(res_0, res_1);
-    res =_mm_hadd_ps(res, res);
-    res =_mm_hadd_ps(res, res);
-
-    return _mm_cvtss_f32(res);
-}
-#endif // __AVX__ || __AVX2__ || __AVX512F__
-#endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
-
-#if defined(__ARM_NEON) || defined(__wasm_simd128__) || defined(__POWER9_VECTOR__)
-#define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
-#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
-#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
-#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
-#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
-#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
-#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
-#define B8(c,s  ) B7(c,s,     c), B7(c,s,     s)
-
-// precomputed tables for expanding 8bits to 8 bytes:
-static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4
-static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
-#endif
-
-#if defined(__loongarch_asx)
-
-#ifdef __clang__
-#define VREGS_PREFIX "$vr"
-#define XREGS_PREFIX "$xr"
-#else // GCC
-#define VREGS_PREFIX "$f"
-#define XREGS_PREFIX "$f"
-#endif
-#define __ALL_REGS "0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31"
-// Convert __m128i to __m256i
-static inline __m256i ____m256i(__m128i in) {
-    __m256i out = __lasx_xvldi(0);
-    __asm__ volatile (
-        ".irp i," __ALL_REGS                "\n\t"
-        " .ifc %[out], " XREGS_PREFIX"\\i    \n\t"
-        "  .irp j," __ALL_REGS              "\n\t"
-        "   .ifc %[in], " VREGS_PREFIX "\\j  \n\t"
-        "    xvpermi.q $xr\\i, $xr\\j, 0x20  \n\t"
-        "   .endif                           \n\t"
-        "  .endr                             \n\t"
-        " .endif                             \n\t"
-        ".endr                               \n\t"
-        : [out] "+f" (out) : [in] "f" (in)
-    );
-    return out;
-}
-// Convert two __m128i to __m256i
-static inline __m256i lasx_set_q(__m128i inhi, __m128i inlo) {
-    __m256i out;
-    __asm__ volatile (
-        ".irp i," __ALL_REGS                "\n\t"
-        " .ifc %[hi], " VREGS_PREFIX "\\i    \n\t"
-        "  .irp j," __ALL_REGS              "\n\t"
-        "   .ifc %[lo], " VREGS_PREFIX "\\j  \n\t"
-        "    xvpermi.q $xr\\i, $xr\\j, 0x20  \n\t"
-        "   .endif                           \n\t"
-        "  .endr                             \n\t"
-        " .endif                             \n\t"
-        ".endr                               \n\t"
-        ".ifnc %[out], %[hi]                 \n\t"
-        ".irp i," __ALL_REGS                "\n\t"
-        " .ifc %[out], " XREGS_PREFIX "\\i   \n\t"
-        "  .irp j," __ALL_REGS              "\n\t"
-        "   .ifc %[hi], " VREGS_PREFIX "\\j  \n\t"
-        "    xvori.b $xr\\i, $xr\\j, 0       \n\t"
-        "   .endif                           \n\t"
-        "  .endr                             \n\t"
-        " .endif                             \n\t"
-        ".endr                               \n\t"
-        ".endif                              \n\t"
-        : [out] "=f" (out), [hi] "+f" (inhi)
-        : [lo] "f" (inlo)
-    );
-    return out;
-}
-// Convert __m256i low part to __m128i
-static inline __m128i lasx_extracti128_lo(__m256i in) {
-    __m128i out;
-    __asm__ volatile (
-        ".ifnc %[out], %[in]                 \n\t"
-        ".irp i," __ALL_REGS                "\n\t"
-        " .ifc %[out], " VREGS_PREFIX "\\i   \n\t"
-        "  .irp j," __ALL_REGS              "\n\t"
-        "   .ifc %[in], " XREGS_PREFIX "\\j  \n\t"
-        "    vori.b $vr\\i, $vr\\j, 0        \n\t"
-        "   .endif                           \n\t"
-        "  .endr                             \n\t"
-        " .endif                             \n\t"
-        ".endr                               \n\t"
-        ".endif                              \n\t"
-        : [out] "=f" (out) : [in] "f" (in)
-    );
-    return out;
-}
-// Convert __m256i high part to __m128i
-static inline __m128i lasx_extracti128_hi(__m256i in) {
-    __m128i out;
-    __asm__ volatile (
-        ".irp i," __ALL_REGS                "\n\t"
-        " .ifc %[out], " VREGS_PREFIX "\\i   \n\t"
-        "  .irp j," __ALL_REGS              "\n\t"
-        "   .ifc %[in], " XREGS_PREFIX "\\j  \n\t"
-        "    xvpermi.q $xr\\i, $xr\\j, 0x11  \n\t"
-        "   .endif                           \n\t"
-        "  .endr                             \n\t"
-        " .endif                             \n\t"
-        ".endr                               \n\t"
-        : [out] "=f" (out) : [in] "f" (in)
-    );
-    return out;
-}
-
-static __m256i lasx_set_w(int e7, int e6, int e5, int e4, int e3, int e2, int e1, int e0) {
-    v8i32 __ret = {e0, e1, e2, e3, e4, e5, e6, e7};
-    return (__m256i)__ret;
-}
-
-static __m128i lsx_set_w(int32_t a, int32_t b, int32_t c, int32_t d) {
-    v4i32 __ret = {d, c, b, a};
-    return (__m128i)__ret;
-}
-
-static __m256i lasx_set_d(int64_t a, int64_t b, int64_t c, int64_t d) {
-    v4i64 __ret = {d, c, b, a};
-    return (__m256i)__ret;
-}
-
-static __m256i lasx_insertf128( __m128i x, __m128i y) {
-    return lasx_set_q(x, y);
-}
-
-static __m128i lsx_shuffle_b(__m128i a, __m128i b) {
-    __m128i mask_f, zero, tmp0, tmp2, mask;
-    int f = 0x8f;
-    mask_f = __lsx_vreplgr2vr_b(f);
-    zero = __lsx_vldi(0);
-    tmp0 = __lsx_vand_v(b, mask_f); // get mask with low 4 bit and sign bits
-    tmp0 = __lsx_vori_b(tmp0, 0x10); // make each mask or  with 0x10 prepare for positive
-    mask = __lsx_vsle_b(zero, tmp0); // if mask >= 0, set mask
-    tmp2 = __lsx_vand_v(tmp0, mask); // maskout the in2 < ones
-    return __lsx_vshuf_b(a, zero, tmp2);
-}
-
-static __m256i lasx_shuffle_b(__m256i a, __m256i b) {
-    __m256i mask_f, zero, tmp0, tmp2, mask;
-    int f = 0x8f;
-    mask_f = __lasx_xvreplgr2vr_b(f);
-    zero = __lasx_xvldi(0);
-    tmp0 = __lasx_xvand_v(b, mask_f); // get mask with low 4 bit and sign bits
-    tmp0 = __lasx_xvori_b(tmp0, 0x10); // make each mask or  with 0x10 prepare for positive
-    mask = __lasx_xvsle_b(zero, tmp0); // if mask >= 0, set mask
-    tmp2 = __lasx_xvand_v(tmp0, mask); // maskout the in2 < ones
-    return __lasx_xvshuf_b(a, zero, tmp2);
-}
-
-static __m256i lasx_extu8_16(__m128i a) {
-    __m128i zero = __lsx_vldi(0);
-    __m128i vlo = __lsx_vilvl_b(zero, a);
-    __m128i vhi = __lsx_vilvh_b(zero, a);
-    return lasx_set_q(vhi, vlo);
-}
-
-static __m256i lasx_ext8_16(__m128i a) {
-     __m128i sign = __lsx_vslti_b(a, 0);
-     __m128i vlo = __lsx_vilvl_b(sign, a);
-     __m128i vhi = __lsx_vilvh_b(sign, a);
-     return lasx_set_q(vhi, vlo);
-}
-
-static __m256i lasx_ext16_32(__m128i a) {
-    __m256i tmp1;
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 0), 0);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 1), 1);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 2), 2);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 3), 3);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 4), 4);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 5), 5);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 6), 6);
-    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 7), 7);
-    return tmp1;
-}
-
-static __m128i lasx_extracti128( __m256i a, int pos) {
-    __m128i ret;
-    if( pos == 0)
-    {
-       ret = lasx_extracti128_lo(a);
-    } else {
-       ret = lasx_extracti128_hi(a);
-    }
-    return ret;
-}
-
-static __m128 lasx_extractf128( __m256 a, int pos) {
-    __m128 ret;
-    if( pos == 0)
-    {
-       ret = (__m128)lasx_extracti128_lo((__m256i)a);
-    } else {
-       ret = (__m128)lasx_extracti128_hi((__m256i)a);
-    }
-    return ret;
-}
-
-static __m128i lsx_hadd_h(__m128i a, __m128i b) {
-    __m128i tmp1 = __lsx_vpickev_h(b, a);
-    __m128i tmp2 = __lsx_vpickod_h(b, a);
-    return __lsx_vadd_h(tmp1, tmp2);
-}
-
-static __m128i lsx_hadd_w(__m128i a, __m128i b) {
-    __m128i tmp1 = __lsx_vpickev_w(b, a);
-    __m128i tmp2 = __lsx_vpickod_w(b, a);
-    return __lsx_vadd_w(tmp1, tmp2);
-}
-
-static __m128 lsx_hadd_s(__m128 a, __m128 b) {
-    __m128 tmp1 = (__m128)__lsx_vpickev_w((__m128i)b, (__m128i)a);
-    __m128 tmp2 = (__m128)__lsx_vpickod_w((__m128i)b, (__m128i)a);
-
-    return __lsx_vfadd_s(tmp1, tmp2);
-}
-
-static __m256i lasx_maddubs_h(__m256i a, __m256i b) {
-    __m256i tmp1, tmp2;
-    tmp1 = __lasx_xvmulwev_h_b(a, b);
-    tmp2 = __lasx_xvmulwod_h_b(a, b);
-    return __lasx_xvsadd_h(tmp1, tmp2);
-}
-
-static __m256i lasx_madd_h(__m256i a, __m256i b) {
-    __m256i tmp1, tmp2;
-    tmp1 = __lasx_xvmulwev_w_h(a, b);
-    tmp2 = __lasx_xvmulwod_w_h(a, b);
-    return __lasx_xvadd_w(tmp1, tmp2);
-}
-
-static __m256i lasx_packs_w(__m256i a, __m256i b) {
-    __m256i tmp, tmp1;
-    tmp = __lasx_xvsat_w(a, 15);
-    tmp1 = __lasx_xvsat_w(b, 15);
-    return __lasx_xvpickev_h(tmp1, tmp);
-}
-
-static __m256i lasx_packs_h(__m256i a, __m256i b) {
-    __m256i tmp, tmp1;
-    tmp = __lasx_xvsat_h(a, 7);
-    tmp1 = __lasx_xvsat_h(b, 7);
-    return __lasx_xvpickev_b(tmp1, tmp);
-}
-
-static __m128i lsx_packs_w(__m128i a, __m128i b) {
-    __m128i tmp, tmp1;
-    tmp = __lsx_vsat_w(a, 15);
-    tmp1 = __lsx_vsat_w(b, 15);
-    return __lsx_vpickev_h(tmp1, tmp);
-}
-
-static __m128i lsx_packs_h(__m128i a, __m128i b) {
-    __m128i tmp, tmp1;
-    tmp = __lsx_vsat_h(a, 7);
-    tmp1 = __lsx_vsat_h(b, 7);
-    return __lsx_vpickev_b(tmp1, tmp);
-}
-
-static __m128i lsx_packus_h(__m128i a, __m128i b) {
-    __m128i tmp, tmp1;
-    tmp = __lsx_vsat_hu(a, 7);
-    tmp1 = __lsx_vsat_hu(b, 7);
-    return __lsx_vpickev_b(tmp1, tmp);
-}
-
-
-static __m128i lsx_maddubs_h(__m128i a, __m128i b) {
-    __m128i tmp1, tmp2;
-    tmp1 = __lsx_vmulwev_h_b(a, b);
-    tmp2 = __lsx_vmulwod_h_b(a, b);
-    return __lsx_vsadd_h(tmp1, tmp2);
-}
-
-static __m128i lsx_madd_h(__m128i a, __m128i b) {
-    __m128i tmp1, tmp2;
-    tmp1 = __lsx_vmulwev_w_h(a, b);
-    tmp2 = __lsx_vmulwod_w_h(a, b);
-    return __lsx_vadd_w(tmp1, tmp2);
-}
-
-// multiply int8_t, add results pairwise twice
-static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
-    // Get absolute values of x vectors
-    const __m128i ax = __lsx_vsigncov_b(x, x);
-    // Sign the values of the y vectors
-    const __m128i sy = __lsx_vsigncov_b(x, y);
-    // Perform multiplication and create 16-bit values
-    const __m128i dot = lsx_maddubs_h(ax, sy);
-    const __m128i ones = __lsx_vreplgr2vr_h(1);
-    return lsx_madd_h(ones, dot);
-}
-
-// horizontally add 8 floats
-static inline float hsum_float_8(const __m256 x) {
-    __m128 res = lasx_extractf128(x, 1);
-    ft_union tmp;
-    res = __lsx_vfadd_s(res, lasx_extractf128(x, 0));
-    res = __lsx_vfadd_s(res, (__m128)__lsx_vpickod_d((__m128i)res, (__m128i)res));
-    res = __lsx_vfadd_s(res, (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w(res, 1), 0));
-    tmp.i = __lsx_vpickve2gr_w(res, 0);
-    return tmp.f;
-}
-
-// horizontally add 8 int32_t
-static inline int hsum_i32_8(const __m256i a) {
-
-    __m256i tmp1 = __lasx_xvpermi_q(a, a, 0x11);
-    __m256i tmp2 = __lasx_xvpermi_q(a, a, 0x00);
-
-    __m128i  tmp1_128 = lasx_extracti128_lo(tmp1);
-    __m128i  tmp2_128 = lasx_extracti128_lo(tmp2);
-
-    __m128i sum128 = __lsx_vadd_w(tmp1_128, tmp2_128);
-
-    __m128i ev = __lsx_vpickev_w(sum128, sum128);
-    __m128i od = __lsx_vpickod_w(sum128, sum128);
-    __m128i sum64 = __lsx_vadd_w(ev, od);
-
-    int sum64_1, sum64_2;
-    sum64_1 = __lsx_vpickve2gr_w(sum64, 0);
-    sum64_2 = __lsx_vpickve2gr_w(sum64, 1);
-
-    return  sum64_1 + sum64_2;
-}
-
-// horizontally add 4 int32_t
-static inline int hsum_i32_4(const __m128i a) {
-    __m128i ev = __lsx_vpickev_w(a, a);
-    __m128i od = __lsx_vpickod_w(a, a);
-    __m128i sum64 = __lsx_vadd_w(ev, od);
-
-    int sum64_1, sum64_2;
-    sum64_1 = __lsx_vpickve2gr_w(sum64, 0);
-    sum64_2 = __lsx_vpickve2gr_w(sum64, 1);
-
-    return  sum64_1 + sum64_2;
-}
-
-// spread 32 bits to 32 bytes { 0x00, 0xFF }
-static inline __m256i bytes_from_bits_32(const uint8_t * x) {
-
-    uint32_t x32;
-    memcpy(&x32, x, sizeof(uint32_t));
-    const __m256i shuf_mask = lasx_set_d(
-            0x0303030303030303, 0x0202020202020202,
-            0x0101010101010101, 0x0000000000000000);
-
-    __m256i bytes = lasx_shuffle_b(__lasx_xvreplgr2vr_w(x32), shuf_mask);
-    const __m256i bit_mask = __lasx_xvreplgr2vr_d(0x7fbfdfeff7fbfdfe);
-    bytes = __lasx_xvor_v(bytes, bit_mask);
-    return __lasx_xvseq_b(bytes, __lasx_xvreplgr2vr_d(-1));
-}
-
-// Unpack 32 4-bit fields into 32 bytes
-// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
-static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) {
-    const __m128i lo = __lsx_vld((const __m128i *)rsi, 0);
-    __m128i hi = __lsx_vsrli_h(lo, 4);
-    return __lasx_xvandi_b(lasx_insertf128(hi, lo), 0xf);
-}
-
-// add int16_t pairwise and return as float vector
-static inline __m256 sum_i16_pairs_float(const __m256i x) {
-    __m256i v = __lasx_xvpackod_h(x, x);
-    __m256i summed_pairs = __lasx_xvaddwev_w_h(x, v);
-    return __lasx_xvffint_s_w(summed_pairs);
-}
-
-static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
-    // Perform multiplication and create 16-bit values
-    const __m256i dot = lasx_maddubs_h(ax, sy);
-    return sum_i16_pairs_float(dot);
-}
-
-// multiply int8_t, add results pairwise twice and return as float vector
-static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
-
-    // Get absolute values of x vectors
-    const __m256i ax = __lasx_xvsigncov_b(x, x);
-    // Sign the values of the y vectors
-    const __m256i sy = __lasx_xvsigncov_b(x, y);
-
-    return mul_sum_us8_pairs_float(ax, sy);
-}
-
-static inline __m128i packNibbles( __m256i bytes ) {
-    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
-    const __m256i lowByte = __lasx_xvreplgr2vr_h(0xFF);
-     __m256i high = __lasx_xvandn_v(lowByte, bytes);
-    __m256i low = __lasx_xvand_v(lowByte, bytes);
-    high = __lasx_xvsrli_h(high, 4);
-    bytes = __lasx_xvor_v(low, high);
-    // Compress uint16_t lanes into bytes
-    __m128i *r0 = (__m128i *)&bytes;
-    __m256i tmp_h128 = __lasx_xvpermi_q(bytes, bytes, 0x11);
-    __m128i *r1 = (__m128i *)&tmp_h128;
-
-    __m128i zero = __lsx_vldi(0);
-    __m128i tmp, tmp2, tmp3;
-
-    tmp = __lsx_vmax_h(zero, *r0);
-    tmp2 = __lsx_vsat_hu(tmp, 7);
-
-    tmp = __lsx_vmax_h(zero, *r1);
-    tmp3 = __lsx_vsat_hu(tmp, 7);
-    return  __lsx_vpickev_b(tmp3, tmp2);
-}
-#endif  //__loongarch_asx
-
 // reference implementation for deterministic creation of model files
 void quantize_row_q4_0_ref(const float * restrict x, block_q4_0 * restrict y, int64_t k) {
     static const int qk = QK4_0;
@@ -702,11 +65,6 @@ void quantize_row_q4_0_ref(const float * restrict x, block_q4_0 * restrict y, in
     }
 }
 
-void quantize_row_q4_0(const float * restrict x, void * restrict y, int64_t k) {
-    quantize_row_q4_0_ref(x, y, k);
-}
-
-
 void quantize_row_q4_1_ref(const float * restrict x, block_q4_1 * restrict y, int64_t k) {
     const int qk = QK4_1;
 
@@ -744,10 +102,6 @@ void quantize_row_q4_1_ref(const float * restrict x, block_q4_1 * restrict y, in
     }
 }
 
-void quantize_row_q4_1(const float * restrict x, void * restrict y, int64_t k) {
-    quantize_row_q4_1_ref(x, y, k);
-}
-
 void quantize_row_q5_0_ref(const float * restrict x, block_q5_0 * restrict y, int64_t k) {
     static const int qk = QK5_0;
 
@@ -792,10 +146,6 @@ void quantize_row_q5_0_ref(const float * restrict x, block_q5_0 * restrict y, in
     }
 }
 
-void quantize_row_q5_0(const float * restrict x, void * restrict y, int64_t k) {
-    quantize_row_q5_0_ref(x, y, k);
-}
-
 void quantize_row_q5_1_ref(const float * restrict x, block_q5_1 * restrict y, int64_t k) {
     const int qk = QK5_1;
 
@@ -840,10 +190,6 @@ void quantize_row_q5_1_ref(const float * restrict x, block_q5_1 * restrict y, in
     }
 }
 
-void quantize_row_q5_1(const float * restrict x, void * restrict y, int64_t k) {
-    quantize_row_q5_1_ref(x, y, k);
-}
-
 // reference implementation for deterministic creation of model files
 void quantize_row_q8_0_ref(const float * restrict x, block_q8_0 * restrict y, int64_t k) {
     assert(k % QK8_0 == 0);
@@ -870,291 +216,6 @@ void quantize_row_q8_0_ref(const float * restrict x, block_q8_0 * restrict y, in
     }
 }
 
-void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k) {
-    assert(QK8_0 == 32);
-    assert(k % QK8_0 == 0);
-    const int nb = k / QK8_0;
-
-    block_q8_0 * restrict y = vy;
-
-#if defined(__ARM_NEON)
-    for (int i = 0; i < nb; i++) {
-        float32x4_t srcv [8];
-        float32x4_t asrcv[8];
-        float32x4_t amaxv[8];
-
-        for (int j = 0; j < 8; j++) srcv[j]  = vld1q_f32(x + i*32 + 4*j);
-        for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]);
-
-        for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]);
-        for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]);
-        for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]);
-
-        const float amax = vmaxvq_f32(amaxv[0]);
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        for (int j = 0; j < 8; j++) {
-            const float32x4_t v  = vmulq_n_f32(srcv[j], id);
-            const int32x4_t   vi = vcvtnq_s32_f32(v);
-
-            y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0);
-            y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1);
-            y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2);
-            y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
-        }
-    }
-#elif defined(__wasm_simd128__)
-    for (int i = 0; i < nb; i++) {
-        v128_t srcv [8];
-        v128_t asrcv[8];
-        v128_t amaxv[8];
-
-        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
-        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
-
-        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
-        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
-        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
-
-        const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
-                                   wasm_f32x4_extract_lane(amaxv[0], 1)),
-                               MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
-                                   wasm_f32x4_extract_lane(amaxv[0], 3)));
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        for (int j = 0; j < 8; j++) {
-            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
-            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
-
-            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
-            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
-            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
-            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
-        }
-    }
-#elif defined(__AVX2__) || defined(__AVX__)
-    for (int i = 0; i < nb; i++) {
-        // Load elements into 4 AVX vectors
-        __m256 v0 = _mm256_loadu_ps( x );
-        __m256 v1 = _mm256_loadu_ps( x + 8 );
-        __m256 v2 = _mm256_loadu_ps( x + 16 );
-        __m256 v3 = _mm256_loadu_ps( x + 24 );
-        x += 32;
-
-        // Compute max(abs(e)) for the block
-        const __m256 signBit = _mm256_set1_ps( -0.0f );
-        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
-
-        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
-        max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
-        max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
-        const float maxScalar = _mm_cvtss_f32( max4 );
-
-        // Quantize these floats
-        const float d = maxScalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
-        const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
-        const __m256 mul = _mm256_set1_ps( id );
-
-        // Apply the multiplier
-        v0 = _mm256_mul_ps( v0, mul );
-        v1 = _mm256_mul_ps( v1, mul );
-        v2 = _mm256_mul_ps( v2, mul );
-        v3 = _mm256_mul_ps( v3, mul );
-
-        // Round to nearest integer
-        v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
-        v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
-        v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
-        v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
-
-        // Convert floats to integers
-        __m256i i0 = _mm256_cvtps_epi32( v0 );
-        __m256i i1 = _mm256_cvtps_epi32( v1 );
-        __m256i i2 = _mm256_cvtps_epi32( v2 );
-        __m256i i3 = _mm256_cvtps_epi32( v3 );
-
-#if defined(__AVX2__)
-        // Convert int32 to int16
-        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
-        i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
-                                            // Convert int16 to int8
-        i0 = _mm256_packs_epi16( i0, i2 );	// 0, 1, 2, 3,  8, 9, 10, 11,  16, 17, 18, 19,  24, 25, 26, 27,  4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
-
-        // We got our precious signed bytes, but the order is now wrong
-        // These AVX2 pack instructions process 16-byte pieces independently
-        // The following instruction is fixing the order
-        const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
-        i0 = _mm256_permutevar8x32_epi32( i0, perm );
-
-        _mm256_storeu_si256((__m256i *)y[i].qs, i0);
-#else
-        // Since we don't have in AVX some necessary functions,
-        // we split the registers in half and call AVX2 analogs from SSE
-        __m128i ni0 = _mm256_castsi256_si128( i0 );
-        __m128i ni1 = _mm256_extractf128_si256( i0, 1);
-        __m128i ni2 = _mm256_castsi256_si128( i1 );
-        __m128i ni3 = _mm256_extractf128_si256( i1, 1);
-        __m128i ni4 = _mm256_castsi256_si128( i2 );
-        __m128i ni5 = _mm256_extractf128_si256( i2, 1);
-        __m128i ni6 = _mm256_castsi256_si128( i3 );
-        __m128i ni7 = _mm256_extractf128_si256( i3, 1);
-
-        // Convert int32 to int16
-        ni0 = _mm_packs_epi32( ni0, ni1 );
-        ni2 = _mm_packs_epi32( ni2, ni3 );
-        ni4 = _mm_packs_epi32( ni4, ni5 );
-        ni6 = _mm_packs_epi32( ni6, ni7 );
-        // Convert int16 to int8
-        ni0 = _mm_packs_epi16( ni0, ni2 );
-        ni4 = _mm_packs_epi16( ni4, ni6 );
-
-        _mm_storeu_si128((__m128i *)(y[i].qs +  0), ni0);
-        _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
-#endif
-    }
-#elif defined(__riscv_v_intrinsic)
-
-    size_t vl = __riscv_vsetvl_e32m4(QK8_0);
-
-    for (int i = 0; i < nb; i++) {
-        // load elements
-        vfloat32m4_t v_x   = __riscv_vle32_v_f32m4(x+i*QK8_0, vl);
-
-        vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl);
-        vfloat32m1_t tmp   = __riscv_vfmv_v_f_f32m1(0.0f, vl);
-        vfloat32m1_t vmax  = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl);
-        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
-
-        // convert to integer
-        vint16m2_t   vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl);
-        vint8m1_t    vs = __riscv_vncvt_x_x_w_i8m1(vi, vl);
-
-        // store result
-        __riscv_vse8_v_i8m1(y[i].qs , vs, vl);
-    }
-
-#elif defined(__POWER9_VECTOR__)
-    for (int i = 0; i < nb; i++) {
-        vector float srcv [8];
-        vector float asrcv[8];
-        vector float amaxv[8];
-        vector signed int vi[8];
-
-        for (int j = 0; j < 8; j++) srcv[j]  = vec_xl(0, x + i*32 + 4*j);
-        for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]);
-
-        for (int j = 0; j < 4; j++) amaxv[2*j] = vec_max(asrcv[2*j], asrcv[2*j+1]);
-        for (int j = 0; j < 2; j++) amaxv[4*j] = vec_max(amaxv[4*j], amaxv[4*j+2]);
-        for (int j = 0; j < 1; j++) amaxv[8*j] = vec_max(amaxv[8*j], amaxv[8*j+4]);
-
-        const float amax = MAX(MAX(vec_extract(amaxv[0], 0),
-                                   vec_extract(amaxv[0], 1)),
-                               MAX(vec_extract(amaxv[0], 2),
-                                   vec_extract(amaxv[0], 3)));
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-        const vector float vid = vec_splats(id);
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        for (int j = 0; j < 8; j++) {
-            const vector float v  = vec_round(vec_mul(srcv[j], vid));
-            vi[j] = vec_cts(v, 0);
-        }
-        vec_xst(vec_pack(vec_pack(vi[0], vi[1]), vec_pack(vi[2], vi[3])),  0, &y[i].qs[0]);
-        vec_xst(vec_pack(vec_pack(vi[4], vi[5]), vec_pack(vi[6], vi[7])), 16, &y[i].qs[0]);
-    }
-
-#elif defined(__loongarch_asx)
-    for (int i = 0; i < nb; i++) {
-        ft_union fi;
-        __m256 v0 = (__m256)__lasx_xvld( x , 0);
-        __m256 v1 = (__m256)__lasx_xvld( x , 32);
-        __m256 v2 = (__m256)__lasx_xvld( x , 64);
-        __m256 v3 = (__m256)__lasx_xvld( x , 96);
-        x += 32;
-
-        // Compute max(abs(e)) for the block
-        const __m256 sign_bit = __lasx_xvreplfr2vr_s( -0.0f );
-        __m256 max_abs = (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v0 );
-        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v1 ) );
-        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v2 ) );
-        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v3 ) );
-
-        __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs , 0) );
-        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
-        __m128 tmp = max4;
-        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vinsgr2vr_w(tmp, __lsx_vpickve2gr_w( max4, 1 ), 0 ));
-        fi.i = __lsx_vpickve2gr_w( (__m128i)max4, 0 );
-        const float max_scalar = fi.f;
-
-        // Quantize these floats
-        const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
-        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
-        const __m256 mul = (__m256)__lasx_xvreplfr2vr_s( id );
-
-        // Apply the multiplier
-        v0 = __lasx_xvfmul_s( v0, mul );
-        v1 = __lasx_xvfmul_s( v1, mul );
-        v2 = __lasx_xvfmul_s( v2, mul );
-        v3 = __lasx_xvfmul_s( v3, mul );
-
-        // Round to nearest integer
-        __m256i i0 = __lasx_xvftintrne_w_s( v0 );
-        __m256i i1 = __lasx_xvftintrne_w_s( v1 );
-        __m256i i2 = __lasx_xvftintrne_w_s( v2 );
-        __m256i i3 = __lasx_xvftintrne_w_s( v3 );
-
-        __m128i ni0 = lasx_extracti128( i0, 0 );
-        __m128i ni1 = lasx_extracti128( i0, 1);
-        __m128i ni2 = lasx_extracti128( i1, 0);
-        __m128i ni3 = lasx_extracti128( i1, 1);
-        __m128i ni4 = lasx_extracti128( i2, 0);
-        __m128i ni5 = lasx_extracti128( i2, 1);
-        __m128i ni6 = lasx_extracti128( i3, 0);
-        __m128i ni7 = lasx_extracti128( i3, 1);
-
-        // Convert int32 to int16
-        ni0 = lsx_packs_w( ni0, ni1 );
-        ni2 = lsx_packs_w( ni2, ni3 );
-        ni4 = lsx_packs_w( ni4, ni5 );
-        ni6 = lsx_packs_w( ni6, ni7 );
-        // Convert int16 to int8
-        ni0 = lsx_packs_h( ni0, ni2 );
-        ni4 = lsx_packs_h( ni4, ni6 );
-
-        __lsx_vst(ni0, (__m128i *)(y[i].qs +  0), 0);
-        __lsx_vst(ni4, (__m128i *)(y[i].qs + 16), 0);
-
-    }
-#else
-    GGML_UNUSED(nb);
-    // scalar
-    quantize_row_q8_0_ref(x, y, k);
-#endif
-}
-
 // reference implementation for deterministic creation of model files
 void quantize_row_q8_1_ref(const float * restrict x, block_q8_1 * restrict y, int64_t k) {
     assert(QK8_1 == 32);
@@ -1191,381 +252,53 @@ void quantize_row_q8_1_ref(const float * restrict x, block_q8_1 * restrict y, in
     }
 }
 
-void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k) {
-    assert(k % QK8_1 == 0);
-    const int nb = k / QK8_1;
+void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int64_t k) {
+    static const int qk = QK4_0;
 
-    block_q8_1 * restrict y = vy;
+    assert(k % qk == 0);
 
-#if defined(__ARM_NEON)
-    for (int i = 0; i < nb; i++) {
-        float32x4_t srcv [8];
-        float32x4_t asrcv[8];
-        float32x4_t amaxv[8];
+    const int nb = k / qk;
 
-        for (int j = 0; j < 8; j++) srcv[j]  = vld1q_f32(x + i*32 + 4*j);
-        for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]);
+    for (int i = 0; i < nb; i++) {
+        const float d = GGML_FP16_TO_FP32(x[i].d);
 
-        for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]);
-        for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]);
-        for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]);
+        for (int j = 0; j < qk/2; ++j) {
+            const int x0 = (x[i].qs[j] & 0x0F) - 8;
+            const int x1 = (x[i].qs[j] >>   4) - 8;
 
-        const float amax = vmaxvq_f32(amaxv[0]);
+            y[i*qk + j + 0   ] = x0*d;
+            y[i*qk + j + qk/2] = x1*d;
+        }
+    }
+}
 
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
+void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int64_t k) {
+    static const int qk = QK4_1;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+    assert(k % qk == 0);
 
-        int32x4_t accv = vdupq_n_s32(0);
+    const int nb = k / qk;
 
-        for (int j = 0; j < 8; j++) {
-            const float32x4_t v  = vmulq_n_f32(srcv[j], id);
-            const int32x4_t   vi = vcvtnq_s32_f32(v);
+    for (int i = 0; i < nb; i++) {
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        const float m = GGML_FP16_TO_FP32(x[i].m);
 
-            y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0);
-            y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1);
-            y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2);
-            y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
+        for (int j = 0; j < qk/2; ++j) {
+            const int x0 = (x[i].qs[j] & 0x0F);
+            const int x1 = (x[i].qs[j] >>   4);
 
-            accv = vaddq_s32(accv, vi);
+            y[i*qk + j + 0   ] = x0*d + m;
+            y[i*qk + j + qk/2] = x1*d + m;
         }
-
-        y[i].s = GGML_FP32_TO_FP16(d * vaddvq_s32(accv));
     }
-#elif defined(__wasm_simd128__)
-    for (int i = 0; i < nb; i++) {
-        v128_t srcv [8];
-        v128_t asrcv[8];
-        v128_t amaxv[8];
-
-        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
-        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
+}
 
-        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
-        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
-        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
+void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int64_t k) {
+    static const int qk = QK5_0;
 
-        const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
-                                   wasm_f32x4_extract_lane(amaxv[0], 1)),
-                               MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
-                                   wasm_f32x4_extract_lane(amaxv[0], 3)));
+    assert(k % qk == 0);
 
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        v128_t accv = wasm_i32x4_splat(0);
-
-        for (int j = 0; j < 8; j++) {
-            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
-            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
-
-            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
-            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
-            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
-            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
-
-            accv = wasm_i32x4_add(accv, vi);
-        }
-
-        y[i].s = GGML_FP32_TO_FP16(
-                d * (wasm_i32x4_extract_lane(accv, 0) +
-                     wasm_i32x4_extract_lane(accv, 1) +
-                     wasm_i32x4_extract_lane(accv, 2) +
-                     wasm_i32x4_extract_lane(accv, 3)));
-    }
-#elif defined(__AVX2__) || defined(__AVX__)
-    for (int i = 0; i < nb; i++) {
-        // Load elements into 4 AVX vectors
-        __m256 v0 = _mm256_loadu_ps( x );
-        __m256 v1 = _mm256_loadu_ps( x + 8 );
-        __m256 v2 = _mm256_loadu_ps( x + 16 );
-        __m256 v3 = _mm256_loadu_ps( x + 24 );
-        x += 32;
-
-        // Compute max(abs(e)) for the block
-        const __m256 signBit = _mm256_set1_ps( -0.0f );
-        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
-
-        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
-        max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
-        max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
-        const float max_scalar = _mm_cvtss_f32( max4 );
-
-        // Quantize these floats
-        const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
-        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
-        const __m256 mul = _mm256_set1_ps( id );
-
-        // Apply the multiplier
-        v0 = _mm256_mul_ps( v0, mul );
-        v1 = _mm256_mul_ps( v1, mul );
-        v2 = _mm256_mul_ps( v2, mul );
-        v3 = _mm256_mul_ps( v3, mul );
-
-        // Round to nearest integer
-        v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
-        v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
-        v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
-        v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
-
-        // Convert floats to integers
-        __m256i i0 = _mm256_cvtps_epi32( v0 );
-        __m256i i1 = _mm256_cvtps_epi32( v1 );
-        __m256i i2 = _mm256_cvtps_epi32( v2 );
-        __m256i i3 = _mm256_cvtps_epi32( v3 );
-
-#if defined(__AVX2__)
-        // Compute the sum of the quants and set y[i].s
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));
-
-        // Convert int32 to int16
-        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
-        i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
-                                            // Convert int16 to int8
-        i0 = _mm256_packs_epi16( i0, i2 );	// 0, 1, 2, 3,  8, 9, 10, 11,  16, 17, 18, 19,  24, 25, 26, 27,  4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
-
-        // We got our precious signed bytes, but the order is now wrong
-        // These AVX2 pack instructions process 16-byte pieces independently
-        // The following instruction is fixing the order
-        const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
-        i0 = _mm256_permutevar8x32_epi32( i0, perm );
-
-        _mm256_storeu_si256((__m256i *)y[i].qs, i0);
-#else
-        // Since we don't have in AVX some necessary functions,
-        // we split the registers in half and call AVX2 analogs from SSE
-        __m128i ni0 = _mm256_castsi256_si128( i0 );
-        __m128i ni1 = _mm256_extractf128_si256( i0, 1);
-        __m128i ni2 = _mm256_castsi256_si128( i1 );
-        __m128i ni3 = _mm256_extractf128_si256( i1, 1);
-        __m128i ni4 = _mm256_castsi256_si128( i2 );
-        __m128i ni5 = _mm256_extractf128_si256( i2, 1);
-        __m128i ni6 = _mm256_castsi256_si128( i3 );
-        __m128i ni7 = _mm256_extractf128_si256( i3, 1);
-
-        // Compute the sum of the quants and set y[i].s
-        const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
-        const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
-
-        // Convert int32 to int16
-        ni0 = _mm_packs_epi32( ni0, ni1 );
-        ni2 = _mm_packs_epi32( ni2, ni3 );
-        ni4 = _mm_packs_epi32( ni4, ni5 );
-        ni6 = _mm_packs_epi32( ni6, ni7 );
-        // Convert int16 to int8
-        ni0 = _mm_packs_epi16( ni0, ni2 );
-        ni4 = _mm_packs_epi16( ni4, ni6 );
-
-        _mm_storeu_si128((__m128i *)(y[i].qs +  0), ni0);
-        _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
-#endif
-    }
-#elif defined(__riscv_v_intrinsic)
-
-    size_t vl = __riscv_vsetvl_e32m4(QK8_1);
-
-    for (int i = 0; i < nb; i++) {
-        // load elements
-        vfloat32m4_t v_x   = __riscv_vle32_v_f32m4(x+i*QK8_1, vl);
-
-        vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl);
-        vfloat32m1_t tmp   = __riscv_vfmv_v_f_f32m1(0.0, vl);
-        vfloat32m1_t vmax  = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl);
-        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
-
-        const float d  = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
-
-        // convert to integer
-        vint16m2_t   vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl);
-        vint8m1_t    vs = __riscv_vncvt_x_x_w_i8m1(vi, vl);
-
-        // store result
-        __riscv_vse8_v_i8m1(y[i].qs , vs, vl);
-
-        // compute sum for y[i].s
-        vint16m1_t tmp2 = __riscv_vmv_v_x_i16m1(0, vl);
-        vint16m1_t vwrs = __riscv_vwredsum_vs_i8m1_i16m1(vs, tmp2, vl);
-
-        // set y[i].s
-        int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
-        y[i].s = GGML_FP32_TO_FP16(sum*d);
-    }
-
-#elif defined(__POWER9_VECTOR__)
-    for (int i = 0; i < nb; i++) {
-        vector float srcv [8];
-        vector float asrcv[8];
-        vector float amaxv[8];
-        vector signed int vi[8];
-
-        for (int j = 0; j < 8; j++) srcv[j]  = vec_xl(0, x + i*32 + 4*j);
-        for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]);
-
-        for (int j = 0; j < 4; j++) amaxv[2*j] = vec_max(asrcv[2*j], asrcv[2*j+1]);
-        for (int j = 0; j < 2; j++) amaxv[4*j] = vec_max(amaxv[4*j], amaxv[4*j+2]);
-        for (int j = 0; j < 1; j++) amaxv[8*j] = vec_max(amaxv[8*j], amaxv[8*j+4]);
-
-        const float amax = MAX(MAX(vec_extract(amaxv[0], 0),
-                                   vec_extract(amaxv[0], 1)),
-                               MAX(vec_extract(amaxv[0], 2),
-                                   vec_extract(amaxv[0], 3)));
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-        const vector float vid = vec_splats(id);
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        vector int accv = vec_splats(0);
-
-        for (int j = 0; j < 8; j++) {
-            const vector float v  = vec_round(vec_mul(srcv[j], vid));
-            vi[j] = vec_cts(v, 0);
-
-            accv = vec_add(accv, vi[j]);
-        }
-        vec_xst(vec_pack(vec_pack(vi[0], vi[1]), vec_pack(vi[2], vi[3])),  0, &y[i].qs[0]);
-        vec_xst(vec_pack(vec_pack(vi[4], vi[5]), vec_pack(vi[6], vi[7])), 16, &y[i].qs[0]);
-
-        accv = vec_add(accv, vec_sld(accv, accv, 4));
-        accv = vec_add(accv, vec_sld(accv, accv, 8));
-        y[i].s = GGML_FP32_TO_FP16(d * vec_extract(accv, 0));
-    }
-
-#elif defined(__loongarch_asx)
-    for (int i = 0; i < nb; i++) {
-        ft_union ft;
-        __m256 v0 = (__m256)__lasx_xvld( x , 0 );
-        __m256 v1 = (__m256)__lasx_xvld( x , 32 );
-        __m256 v2 = (__m256)__lasx_xvld( x , 64 );
-        __m256 v3 = (__m256)__lasx_xvld( x , 96 );
-        x += 32;
-
-        // Compute max(abs(e)) for the block
-        const __m256 sign_bit = __lasx_xvreplfr2vr_s( -0.0f );
-        __m256 max_abs = (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v0 );
-        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v1 ) );
-        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v2 ) );
-        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v3 ) );
-
-        __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs, 0) );
-        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
-        __m128 tmp = max4;
-        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x10 ));
-        ft.i = __lsx_vpickve2gr_w( (__m128i)max4, 0 );
-        const float max_scalar = ft.f;
-
-        // Quantize these floats
-        const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
-        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
-        const __m256 mul = __lasx_xvreplfr2vr_s( id );
-
-        // Apply the multiplier
-        v0 = __lasx_xvfmul_s( v0, mul );
-        v1 = __lasx_xvfmul_s( v1, mul );
-        v2 = __lasx_xvfmul_s( v2, mul );
-        v3 = __lasx_xvfmul_s( v3, mul );
-
-        // Round to nearest integer
-        __m256i i0 = __lasx_xvftintrne_w_s( v0 );
-        __m256i i1 = __lasx_xvftintrne_w_s( v1 );
-        __m256i i2 = __lasx_xvftintrne_w_s( v2 );
-        __m256i i3 = __lasx_xvftintrne_w_s( v3 );
-
-        __m128i ni0 = lasx_extracti128(i0, 0);
-        __m128i ni1 = lasx_extracti128( i0, 1);
-        __m128i ni2 = lasx_extracti128( i1, 0);
-        __m128i ni3 = lasx_extracti128( i1, 1);
-        __m128i ni4 = lasx_extracti128( i2, 0 );
-        __m128i ni5 = lasx_extracti128( i2, 1);
-        __m128i ni6 = lasx_extracti128( i3, 0);
-        __m128i ni7 = lasx_extracti128( i3, 1);
-
-        // Compute the sum of the quants and set y[i].s
-        const __m128i s0 = __lsx_vadd_w(__lsx_vadd_w(ni0, ni1), __lsx_vadd_w(ni2, ni3));
-        const __m128i s1 = __lsx_vadd_w(__lsx_vadd_w(ni4, ni5), __lsx_vadd_w(ni6, ni7));
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1)));
-
-        // Convert int32 to int16
-        ni0 = lsx_packs_w( ni0, ni1 );
-        ni2 = lsx_packs_w( ni2, ni3 );
-        ni4 = lsx_packs_w( ni4, ni5 );
-        ni6 = lsx_packs_w( ni6, ni7 );
-        // Convert int16 to int8
-        ni0 = lsx_packs_h( ni0, ni2 );
-        ni4 = lsx_packs_h( ni4, ni6 );
-
-        __lsx_vst(ni0, (__m128i *)(y[i].qs +  0), 0);
-        __lsx_vst(ni4, (__m128i *)(y[i].qs + 16), 0);
-    }
-#else
-    GGML_UNUSED(nb);
-    // scalar
-    quantize_row_q8_1_ref(x, y, k);
-#endif
-}
-
-void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int64_t k) {
-    static const int qk = QK4_0;
-
-    assert(k % qk == 0);
-
-    const int nb = k / qk;
-
-    for (int i = 0; i < nb; i++) {
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int x0 = (x[i].qs[j] & 0x0F) - 8;
-            const int x1 = (x[i].qs[j] >>   4) - 8;
-
-            y[i*qk + j + 0   ] = x0*d;
-            y[i*qk + j + qk/2] = x1*d;
-        }
-    }
-}
-
-void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int64_t k) {
-    static const int qk = QK4_1;
-
-    assert(k % qk == 0);
-
-    const int nb = k / qk;
-
-    for (int i = 0; i < nb; i++) {
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-        const float m = GGML_FP16_TO_FP32(x[i].m);
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int x0 = (x[i].qs[j] & 0x0F);
-            const int x1 = (x[i].qs[j] >>   4);
-
-            y[i*qk + j + 0   ] = x0*d + m;
-            y[i*qk + j + qk/2] = x1*d + m;
-        }
-    }
-}
-
-void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int64_t k) {
-    static const int qk = QK5_0;
-
-    assert(k % qk == 0);
-
-    const int nb = k / qk;
+    const int nb = k / qk;
 
     for (int i = 0; i < nb; i++) {
         const float d = GGML_FP16_TO_FP32(x[i].d);
@@ -2008,10 +741,6 @@ void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int6
     }
 }
 
-void quantize_row_q2_K(const float * restrict x, void * restrict vy, int64_t k) {
-    quantize_row_q2_K_ref(x, vy, k);
-}
-
 static float make_qkx3_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
         uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux,
         float rmin, float rdelta, int nstep, bool use_mad) {
@@ -2374,10 +1103,6 @@ void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int6
     }
 }
 
-void quantize_row_q3_K(const float * restrict x, void * restrict vy, int64_t k) {
-    quantize_row_q3_K_ref(x, vy, k);
-}
-
 static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restrict y, int64_t n_per_row, const float * restrict quant_weights) {
     assert(n_per_row % QK_K == 0);
     const int nb = n_per_row / QK_K;
@@ -2576,12 +1301,6 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int6
     }
 }
 
-void quantize_row_q4_K(const float * restrict x, void * restrict vy, int64_t k) {
-    assert(k % QK_K == 0);
-    block_q4_K * restrict y = vy;
-    quantize_row_q4_K_ref(x, y, k);
-}
-
 static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restrict y, int64_t n_per_row, const float * quant_weights) {
     assert(n_per_row % QK_K == 0);
     const int64_t nb = n_per_row / QK_K;
@@ -2787,12 +1506,6 @@ void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int6
     }
 }
 
-void quantize_row_q5_K(const float * restrict x, void * restrict vy, int64_t k) {
-    assert(k % QK_K == 0);
-    block_q5_K * restrict y = vy;
-    quantize_row_q5_K_ref(x, y, k);
-}
-
 static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restrict y, int64_t n_per_row, const float * quant_weights) {
     assert(n_per_row % QK_K == 0);
     const int64_t nb = n_per_row / QK_K;
@@ -3005,12 +1718,6 @@ void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int6
     }
 }
 
-void quantize_row_q6_K(const float * restrict x, void * restrict vy, int64_t k) {
-    assert(k % QK_K == 0);
-    block_q6_K * restrict y = vy;
-    quantize_row_q6_K_ref(x, y, k);
-}
-
 static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restrict y, int64_t n_per_row, const float * quant_weights) {
     assert(n_per_row % QK_K == 0);
     const int64_t nb = n_per_row / QK_K;
@@ -3413,33 +2120,20 @@ void quantize_row_tq2_0_ref(const float * restrict x, block_tq2_0 * restrict y,
     }
 }
 
-void quantize_row_tq1_0(const float * restrict x, void * restrict vy, int64_t k) {
-    assert(k % QK_K == 0);
-    block_tq1_0 * restrict y = vy;
-    quantize_row_tq1_0_ref(x, y, k);
-}
-
-void quantize_row_tq2_0(const float * restrict x, void * restrict vy, int64_t k) {
-    assert(k % QK_K == 0);
-    block_tq2_0 * restrict y = vy;
-    quantize_row_tq2_0_ref(x, y, k);
-}
-
 size_t quantize_tq1_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     (void)quant_weights; // not used
     const size_t row_size = ggml_row_size(GGML_TYPE_TQ1_0, n_per_row);
-    quantize_row_tq1_0(src, dst, (int64_t)nrow*n_per_row);
+    quantize_row_tq1_0_ref(src, dst, (int64_t)nrow*n_per_row);
     return nrow * row_size;
 }
 
 size_t quantize_tq2_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     (void)quant_weights; // not used
     const size_t row_size = ggml_row_size(GGML_TYPE_TQ2_0, n_per_row);
-    quantize_row_tq2_0(src, dst, (int64_t)nrow*n_per_row);
+    quantize_row_tq2_0_ref(src, dst, (int64_t)nrow*n_per_row);
     return nrow * row_size;
 }
 
-
 void dequantize_row_tq1_0(const block_tq1_0 * restrict x, float * restrict y, int64_t k) {
     assert(k % QK_K == 0);
     const int64_t nb = k / QK_K;
@@ -3612,9384 +2306,224 @@ void dequantize_row_iq3_xxs(const block_iq3_xxs * restrict x, float * restrict y
                 }
                 y += 8;
             }
-            qs += 8;
-        }
-    }
-}
-
-// ====================== 3.3125 bpw (de)-quantization
-
-void dequantize_row_iq3_s(const block_iq3_s * restrict x, float * restrict y, int64_t k) {
-    assert(k % QK_K == 0);
-    const int64_t nb = k / QK_K;
-
-    for (int i = 0; i < nb; i++) {
-
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-        const uint8_t * qs = x[i].qs;
-        const uint8_t * qh = x[i].qh;
-        const uint8_t * signs = x[i].signs;
-
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const float db1 = d * (1 + 2*(x[i].scales[ib32/2] & 0xf));
-            const float db2 = d * (1 + 2*(x[i].scales[ib32/2] >>  4));
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[0] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[0] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    y[j+0] = db1 * grid1[j] * (signs[l] & kmask_iq2xs[j+0] ? -1.f : 1.f);
-                    y[j+4] = db1 * grid2[j] * (signs[l] & kmask_iq2xs[j+4] ? -1.f : 1.f);
-                }
-                y += 8;
-            }
-            qs += 8;
-            signs += 4;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[1] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[1] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    y[j+0] = db2 * grid1[j] * (signs[l] & kmask_iq2xs[j+0] ? -1.f : 1.f);
-                    y[j+4] = db2 * grid2[j] * (signs[l] & kmask_iq2xs[j+4] ? -1.f : 1.f);
-                }
-                y += 8;
-            }
-            qh += 2;
-            qs += 8;
-            signs += 4;
-        }
-    }
-}
-
-// ====================== 1.5625 bpw (de)-quantization
-
-void dequantize_row_iq1_s(const block_iq1_s * restrict x, float * restrict y, int64_t k) {
-    assert(k % QK_K == 0);
-    const int64_t nb = k / QK_K;
-
-    for (int i = 0; i < nb; i++) {
-
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            const float dl = d * (2*((qh[ib] >> 12) & 7) + 1);
-            const float delta = qh[ib] & 0x8000 ? -IQ1S_DELTA : IQ1S_DELTA;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
-                for (int j = 0; j < 8; ++j) {
-                    y[j] = dl * (grid[j] + delta);
-                }
-                y += 8;
-            }
-            qs += 4;
-        }
-    }
-}
-
-void dequantize_row_iq1_m(const block_iq1_m * restrict x, float * restrict y, int64_t k) {
-    assert(k % QK_K == 0);
-    const int64_t nb = k / QK_K;
-
-    float delta[4];
-    uint16_t idx[4];
-
-    iq1m_scale_t scale;
-
-    for (int i = 0; i < nb; i++) {
-
-        const uint16_t * sc = (const uint16_t *)x[i].scales;
-        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
-        const float d = GGML_FP16_TO_FP32(scale.f16);
-
-        const uint8_t * qs = x[i].qs;
-        const uint8_t * qh = x[i].qh;
-
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            const float dl1 = d * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1);
-            const float dl2 = d * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1);
-
-            idx[0] = qs[0] | ((qh[0] << 8) & 0x700);
-            idx[1] = qs[1] | ((qh[0] << 4) & 0x700);
-            idx[2] = qs[2] | ((qh[1] << 8) & 0x700);
-            idx[3] = qs[3] | ((qh[1] << 4) & 0x700);
-            delta[0] = qh[0] & 0x08 ? -IQ1S_DELTA : IQ1S_DELTA;
-            delta[1] = qh[0] & 0x80 ? -IQ1S_DELTA : IQ1S_DELTA;
-            delta[2] = qh[1] & 0x08 ? -IQ1S_DELTA : IQ1S_DELTA;
-            delta[3] = qh[1] & 0x80 ? -IQ1S_DELTA : IQ1S_DELTA;
-            for (int l = 0; l < 2; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + idx[l]);
-                for (int j = 0; j < 8; ++j) {
-                    y[j] = dl1 * (grid[j] + delta[l]);
-                }
-                y += 8;
-            }
-            for (int l = 2; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + idx[l]);
-                for (int j = 0; j < 8; ++j) {
-                    y[j] = dl2 * (grid[j] + delta[l]);
-                }
-                y += 8;
-            }
-            qs += 4;
-            qh += 2;
-        }
-    }
-}
-
-static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
-void dequantize_row_iq4_nl(const block_iq4_nl * restrict x, float * restrict y, int64_t k) {
-    assert(k % QK4_NL == 0);
-    const int64_t nb = k / QK4_NL;
-
-    for (int i = 0; i < nb; i++) {
-
-        const uint8_t * qs = x[i].qs;
-
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-        for (int j = 0; j < QK4_NL/2; ++j) {
-            y[j+       0] = d * kvalues_iq4nl[qs[j] & 0xf];
-            y[j+QK4_NL/2] = d * kvalues_iq4nl[qs[j] >>  4];
-        }
-        y  += QK4_NL;
-        qs += QK4_NL/2;
-    }
-}
-
-void dequantize_row_iq4_xs(const block_iq4_xs * restrict x, float * restrict y, int64_t k) {
-    assert(k % QK_K == 0);
-    const int64_t nb = k / QK_K;
-
-    for (int i = 0; i < nb; i++) {
-
-        const uint8_t * qs = x[i].qs;
-
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-
-        for (int ib = 0; ib < QK_K/32; ++ib) {
-            const int ls = ((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4);
-            const float dl = d * (ls - 32);
-            for (int j = 0; j < 16; ++j) {
-                y[j+ 0] = dl * kvalues_iq4nl[qs[j] & 0xf];
-                y[j+16] = dl * kvalues_iq4nl[qs[j] >>  4];
-            }
-            y  += 32;
-            qs += 16;
-        }
-    }
-}
-
-//===================================== Q8_K ==============================================
-
-void quantize_row_q8_K_ref(const float * restrict x, block_q8_K * restrict y, int64_t k) {
-    assert(k % QK_K == 0);
-    const int64_t nb = k / QK_K;
-
-    for (int i = 0; i < nb; i++) {
-
-        float max = 0;
-        float amax = 0;
-        for (int j = 0; j < QK_K; ++j) {
-            float ax = fabsf(x[j]);
-            if (ax > amax) {
-                amax = ax; max = x[j];
-            }
-        }
-        if (!amax) {
-            y[i].d = 0;
-            memset(y[i].qs, 0, QK_K);
-            x += QK_K;
-            continue;
-        }
-        //const float iscale = -128.f/max;
-        // We need this change for IQ2_XXS, else the AVX implementation becomes very awkward
-        const float iscale = -127.f/max;
-        for (int j = 0; j < QK_K; ++j) {
-            int v = nearest_int(iscale*x[j]);
-            y[i].qs[j] = MIN(127, v);
-        }
-        for (int j = 0; j < QK_K/16; ++j) {
-            int sum = 0;
-            for (int ii = 0; ii < 16; ++ii) {
-                sum += y[i].qs[j*16 + ii];
-            }
-            y[i].bsums[j] = sum;
-        }
-        y[i].d = 1/iscale;
-        x += QK_K;
-    }
-}
-
-void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int64_t k) {
-    assert(k % QK_K == 0);
-    const int64_t nb = k / QK_K;
-
-    for (int i = 0; i < nb; i++) {
-        for (int j = 0; j < QK_K; ++j) {
-            *y++ = x[i].d * x[i].qs[j];
-        }
-    }
-}
-
-void quantize_row_q8_K(const float * restrict x, void * restrict y, int64_t k) {
-    quantize_row_q8_K_ref(x, y, k);
-}
-
-//===================================== Dot products =================================
-
-//
-// Helper functions
-//
-#if __AVX__ || __AVX2__ || __AVX512F__
-
-// shuffles to pick the required scales in dot products
-static inline __m256i get_scale_shuffle_q3k(int i) {
-    static const uint8_t k_shuffle[128] = {
-         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,     2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
-         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,     6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
-         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,    10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
-        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,    14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,
-    };
-    return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
-}
-static inline __m256i get_scale_shuffle_k4(int i) {
-    static const uint8_t k_shuffle[256] = {
-         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
-         2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
-         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
-         6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
-         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,
-        10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
-        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,
-        14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15
-    };
-    return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
-}
-static inline __m128i get_scale_shuffle(int i) {
-    static const uint8_t k_shuffle[128] = {
-         0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
-         2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
-         4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5,
-         6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7,
-         8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
-        10,10,10,10,10,10,10,10, 11,11,11,11,11,11,11,11,
-        12,12,12,12,12,12,12,12, 13,13,13,13,13,13,13,13,
-        14,14,14,14,14,14,14,14, 15,15,15,15,15,15,15,15
-    };
-    return _mm_loadu_si128((const __m128i*)k_shuffle + i);
-}
-#elif defined(__loongarch_asx)
-// shuffles to pick the required scales in dot products
-static inline __m256i get_scale_shuffle_q3k(int i) {
-    static const uint8_t k_shuffle[128] = {
-         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,     2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
-         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,     6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
-         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,    10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
-        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,    14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,
-    };
-    return __lasx_xvld((const __m256i*)k_shuffle + i, 0);
-}
-static inline __m256i get_scale_shuffle_k4(int i) {
-    static const uint8_t k_shuffle[256] = {
-         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
-         2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
-         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
-         6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
-         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,
-        10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
-        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,
-        14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15
-    };
-    return __lasx_xvld((const __m256i*)k_shuffle + i, 0);
-}
-static inline __m128i get_scale_shuffle(int i) {
-    static const uint8_t k_shuffle[128] = {
-         0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
-         2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
-         4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5,
-         6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7,
-         8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
-        10,10,10,10,10,10,10,10, 11,11,11,11,11,11,11,11,
-        12,12,12,12,12,12,12,12, 13,13,13,13,13,13,13,13,
-        14,14,14,14,14,14,14,14, 15,15,15,15,15,15,15,15
-    };
-    return __lsx_vld((const __m128i*)k_shuffle + i, 0);
-}
-#endif
-
-void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-
-    assert(n % qk == 0);
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    assert((nrc == 2) || (nrc == 1));
-#else
-    assert(nrc == 1);
-#endif
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_q4_0 * restrict x = vx;
-    const block_q8_0 * restrict y = vy;
-
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    if (nrc == 2) {
-        const block_q4_0 * restrict vx0 = vx;
-        const block_q4_0 * restrict vx1 = (const block_q4_0 *) ((const uint8_t*)vx + bx);
-        const block_q8_0 * restrict vy0 = vy;
-        const block_q8_0 * restrict vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by);
-
-        float32x4_t sumv0 = vdupq_n_f32(0.0f);
-
-        for (int i = 0; i < nb; i++) {
-            const block_q4_0 * restrict b_x0 = &vx0[i];
-            const block_q4_0 * restrict b_x1 = &vx1[i];
-            const block_q8_0 * restrict b_y0 = &vy0[i];
-            const block_q8_0 * restrict b_y1 = &vy1[i];
-
-            const uint8x16_t m4b = vdupq_n_u8(0x0F);
-            const int8x16_t  s8b = vdupq_n_s8(0x8);
-
-            const uint8x16_t v0_0 = vld1q_u8(b_x0->qs);
-            const uint8x16_t v0_1 = vld1q_u8(b_x1->qs);
-
-            // 4-bit -> 8-bit
-            const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-            const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-            const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-            const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-            // sub 8
-            const int8x16_t x0_l = vsubq_s8(v0_0l, s8b);
-            const int8x16_t x0_h = vsubq_s8(v0_0h, s8b);
-            const int8x16_t x1_l = vsubq_s8(v0_1l, s8b);
-            const int8x16_t x1_h = vsubq_s8(v0_1h, s8b);
-
-            // load y
-            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
-            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
-            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
-            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
-
-            float32_t _scale[4] = { GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                                    GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                                    GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                                    GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
-
-            float32x4_t scale = vld1q_f32(_scale);
-
-            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
-            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
-
-            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
-            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
-
-            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
-            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
-
-            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
-            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
-
-            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
-                                                                                l1, r1)), l2, r2)), l3, r3))), scale);
-        }
-        float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
-        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
-
-        vst1_f32(s,      vget_low_f32(sumv2));
-        vst1_f32(s + bs, vget_high_f32(sumv2));
-        return;
-    }
-#endif
-
-    int ib = 0;
-    float sumf = 0;
-
-#if defined(__ARM_FEATURE_SVE)
-    svfloat32_t sumv0 = svdup_n_f32(0.0f);
-    svfloat32_t sumv1 = svdup_n_f32(0.0f);
-
-    const int vector_length = ggml_cpu_get_sve_cnt()*8;
-
-    // VLA Implementation using switch case
-    switch (vector_length) {
-        case 128:
-            {
-                // predicate for activating higher lanes for 4 float32 elements
-                const svbool_t ph4 = svptrue_pat_b32(SV_VL4);
-
-                for (; ib + 1 < nb; ib += 2) {
-                    const block_q4_0 * restrict x0 = &x[ib + 0];
-                    const block_q4_0 * restrict x1 = &x[ib + 1];
-                    const block_q8_0 * restrict y0 = &y[ib + 0];
-                    const block_q8_0 * restrict y1 = &y[ib + 1];
-
-                    // load x
-                    const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
-                    const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
-
-                    // 4-bit -> 8-bit
-                    const svint8_t qx0l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx0r, 0x0F));
-                    const svint8_t qx0h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx0r, 0x04));
-                    const svint8_t qx1l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx1r, 0x0F));
-                    const svint8_t qx1h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx1r, 0x04));
-
-                    // sub 8
-                    const svint8_t qx0ls = svsub_n_s8_x(svptrue_b8(), qx0h, 8);
-                    const svint8_t qx0hs = svsub_n_s8_x(svptrue_b8(), qx0l, 8);
-                    const svint8_t qx1ls = svsub_n_s8_x(svptrue_b8(), qx1h, 8);
-                    const svint8_t qx1hs = svsub_n_s8_x(svptrue_b8(), qx1l, 8);
-
-                    // load y
-                    const svint8_t qy0h = svld1_s8(svptrue_b8(), y0->qs);
-                    const svint8_t qy0l = svld1_s8(svptrue_b8(), y0->qs + 16);
-                    const svint8_t qy1h = svld1_s8(svptrue_b8(), y1->qs);
-                    const svint8_t qy1l = svld1_s8(svptrue_b8(), y1->qs + 16);
-
-                    // dot product
-                    sumv0 = svmla_n_f32_x(ph4, sumv0, svcvt_f32_s32_x(ph4, svadd_x(ph4,
-                                    svdot_s32(svdup_n_s32(0), qx0ls, qy0l),
-                                    svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-                    sumv1 = svmla_n_f32_x(ph4, sumv1, svcvt_f32_s32_x(ph4, svadd_x(ph4,
-                                    svdot_s32(svdup_n_s32(0), qx1ls, qy1l),
-                                    svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-                }
-
-                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
-            } break;
-        case 256:
-            {
-                // predicate for activating higher lanes for 16 int8 elements
-                const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
-                // predicate for activating lower lanes for  16 int8 elements
-                const svbool_t pl16 = svnot_b_z(svptrue_b8(), ph16);
-
-                for (; ib + 1 < nb; ib += 2) {
-                    const block_q4_0 * restrict x0 = &x[ib + 0];
-                    const block_q4_0 * restrict x1 = &x[ib + 1];
-                    const block_q8_0 * restrict y0 = &y[ib + 0];
-                    const block_q8_0 * restrict y1 = &y[ib + 1];
-
-                    // load x
-                    const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
-                    const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
-
-                    // 4-bit -> 8-bit
-                    const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04));
-                    const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04));
-
-                    // sub 8
-                    const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8);
-                    const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8);
-
-                    // load y
-                    const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
-                    const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
-
-                    // dot product
-                    sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-                    sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-                }
-
-                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
-            } break;
-        case 512:
-            {
-                // predicate for activating higher lanes for 32 int8 elements
-                const svbool_t ph32 = svptrue_pat_b8(SV_VL32);
-
-                // predicate for activating higher lanes for 16 int8 elements
-                const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
-                // predicate for activating lower lanes for 16 int8 elements from first 32 int8 activated lanes
-                const svbool_t pl16 = svnot_b_z(ph32, ph16);
-
-                for (; ib + 1 < nb; ib += 2) {
-                    const block_q4_0 * restrict x0 = &x[ib + 0];
-                    const block_q4_0 * restrict x1 = &x[ib + 1];
-                    const block_q8_0 * restrict y0 = &y[ib + 0];
-                    const block_q8_0 * restrict y1 = &y[ib + 1];
-
-                    // load x
-                    const svuint8_t qx0r = svld1rq_u8(ph32, x0->qs);
-                    const svuint8_t qx1r = svld1rq_u8(ph32, x1->qs);
-
-                    // 4-bit -> 8-bit
-                    const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04));
-                    const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04));
-
-                    // sub 8
-                    const svint8_t qx0s = svsub_n_s8_x(ph32, qx0, 8);
-                    const svint8_t qx1s = svsub_n_s8_x(ph32, qx1, 8);
-
-                    // load y
-                    const svint8_t qy0 = svld1_s8(ph32, y0->qs);
-                    const svint8_t qy1 = svld1_s8(ph32, y1->qs);
-
-                    // dot product
-                    sumv0 = svmla_n_f32_x(ph32, sumv0, svcvt_f32_s32_x(ph32,
-                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-                    sumv1 = svmla_n_f32_x(ph32, sumv1, svcvt_f32_s32_x(ph32,
-                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-                }
-
-                sumf = svaddv_f32(ph32, svadd_f32_x(ph32, sumv0, sumv1));
-            } break;
-        default:
-            assert(false && "Unsupported vector length");
-            break;
-    }
-
-#elif defined(__ARM_NEON)
-    float32x4_t sumv0 = vdupq_n_f32(0.0f);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    for (; ib + 1 < nb; ib += 2) {
-        const block_q4_0 * restrict x0 = &x[ib + 0];
-        const block_q4_0 * restrict x1 = &x[ib + 1];
-        const block_q8_0 * restrict y0 = &y[ib + 0];
-        const block_q8_0 * restrict y1 = &y[ib + 1];
-
-        const uint8x16_t m4b = vdupq_n_u8(0x0F);
-        const int8x16_t  s8b = vdupq_n_s8(0x8);
-
-        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
-        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
-
-        // 4-bit -> 8-bit
-        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-        // sub 8
-        const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b);
-        const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b);
-        const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b);
-        const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b);
-
-        // load y
-        const int8x16_t v1_0l = vld1q_s8(y0->qs);
-        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
-        const int8x16_t v1_1l = vld1q_s8(y1->qs);
-        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
-
-        // dot product into int32x4_t
-        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
-        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-    }
-
-    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
-#elif defined(__AVX2__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
-
-        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-
-        // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
-        const __m256i off = _mm256_set1_epi8( 8 );
-        qx = _mm256_sub_epi8( qx, off );
-
-        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
-
-        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
-
-        /* Multiply q with scale and accumulate */
-        acc = _mm256_fmadd_ps( d, q, acc );
-    }
-
-    sumf = hsum_float_8(acc);
-#elif defined(__AVX__)
-    const __m128i mone = _mm_set1_epi16(1);
-
-    __m256 accum1 = _mm256_setzero_ps();
-    __m256 accum2 = _mm256_setzero_ps();
-    for (; ib + 1 < nb; ib += 2) {
-        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
-        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
-        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
-        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
-        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
-        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
-
-        const __m128i q4b_1_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_1), _mm_set1_epi8(8));
-        const __m128i q4b_1_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_1, 4)), _mm_set1_epi8(8));
-        const __m128i q4b_2_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_2), _mm_set1_epi8(8));
-        const __m128i q4b_2_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_2, 4)), _mm_set1_epi8(8));
-        const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
-        const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
-        const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
-        const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
-        const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, mone);
-        const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, mone);
-        const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, mone);
-        const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, mone);
-        accum1 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
-                _mm256_cvtepi32_ps(MM256_SET_M128I(p_1_1, p_1_0))), accum1);
-        accum2 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
-                _mm256_cvtepi32_ps(MM256_SET_M128I(p_2_1, p_2_0))), accum2);
-    }
-
-    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
-#elif defined(__SSSE3__)
-    // set constants
-    const __m128i lowMask = _mm_set1_epi8(0xF);
-    const __m128i off = _mm_set1_epi8(8);
-
-    // Initialize accumulator with zeros
-    __m128 acc_0 = _mm_setzero_ps();
-    __m128 acc_1 = _mm_setzero_ps();
-    __m128 acc_2 = _mm_setzero_ps();
-    __m128 acc_3 = _mm_setzero_ps();
-
-    for (; ib + 1 < nb; ib += 2) {
-        _mm_prefetch(&x[ib] + sizeof(block_q4_0), _MM_HINT_T0);
-        _mm_prefetch(&y[ib] + sizeof(block_q8_0), _MM_HINT_T0);
-
-        // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
-
-        const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[ib].qs);
-
-        __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1);
-        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
-        bx_0 = _mm_sub_epi8(bx_0, off);
-        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
-
-        __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4));
-        __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[ib].qs + 16));
-        bx_1 = _mm_sub_epi8(bx_1, off);
-        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
-
-        _mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
-        _mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
-
-        // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
-
-        const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
-
-        __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3);
-        __m128i by_2 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
-        bx_2 = _mm_sub_epi8(bx_2, off);
-        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
-
-        __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4));
-        __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[ib + 1].qs + 16));
-        bx_3 = _mm_sub_epi8(bx_3, off);
-        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
-
-        // Convert int32_t to float
-        __m128 p0 = _mm_cvtepi32_ps(i32_0);
-        __m128 p1 = _mm_cvtepi32_ps(i32_1);
-        __m128 p2 = _mm_cvtepi32_ps(i32_2);
-        __m128 p3 = _mm_cvtepi32_ps(i32_3);
-
-        // Apply the scale
-        __m128 p0_d = _mm_mul_ps( d_0_1, p0 );
-        __m128 p1_d = _mm_mul_ps( d_0_1, p1 );
-        __m128 p2_d = _mm_mul_ps( d_2_3, p2 );
-        __m128 p3_d = _mm_mul_ps( d_2_3, p3 );
-
-        // Acummulate
-        acc_0 = _mm_add_ps(p0_d, acc_0);
-        acc_1 = _mm_add_ps(p1_d, acc_1);
-        acc_2 = _mm_add_ps(p2_d, acc_2);
-        acc_3 = _mm_add_ps(p3_d, acc_3);
-    }
-
-    sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
-#elif defined(__riscv_v_intrinsic)
-    size_t vl = __riscv_vsetvl_e8m1(qk/2);
-
-    for (; ib < nb; ++ib) {
-        // load elements
-        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
-
-        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
-        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
-
-        // mask and store lower part of x, and then upper part
-        vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
-        vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
-
-        vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
-        vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
-
-        // subtract offset
-        vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 8, vl);
-        vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 8, vl);
-
-        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
-        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
-
-        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
-
-        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
-        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
-
-        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
-
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
-    }
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector signed int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-    const vector signed char v8 = vec_splats((signed char)0x8);
-
-    vector float vsumf0 = vec_splats(0.0f);
-
-#pragma GCC unroll 8
-    for (; ib < nb; ++ib) {
-        __builtin_prefetch(x[ib].qs, 0, 1);
-        __builtin_prefetch(y[ib].qs, 0, 1);
-
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
-        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
-        vector signed char q8y1 = vec_xl(16, y[ib].qs);
-
-        vector signed char q4x0 = vec_and(qxs, lowMask);
-        vector signed char q4x1 = vec_sr(qxs, v4);
-
-        q4x0 = vec_sub(q4x0, v8);
-        q4x1 = vec_sub(q4x1, v8);
-
-        vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
-        vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
-
-        vector signed int vsumi0 = v0;
-
-        vsumi0 = vec_sum4s(qv0, vsumi0);
-        vsumi0 = vec_sum4s(qv1, vsumi0);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-    }
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    sumf = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-    // Initialize accumulator with zeros
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        /* Compute combined scale for the block */
-        const __m256 d = __lasx_xvreplfr2vr_s( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
-
-        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-
-        // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
-        const __m256i off = __lasx_xvreplgr2vr_b( 8 );
-        qx = __lasx_xvsub_b( qx, off );
-
-        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
-
-        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
-
-        /* Multiply q with scale and accumulate */
-        acc = __lasx_xvfmadd_s( d, q, acc );
-    }
-
-    sumf = hsum_float_8(acc);
-#elif defined(__loongarch_sx)
-    // set constants
-    const __m128i low_mask = __lsx_vreplgr2vr_b(0xF);
-    const __m128i off = __lsx_vreplgr2vr_b(8);
-
-    // Initialize accumulator with zeros
-    __m128 acc_0 = __lsx_vldi(0);
-    __m128 acc_1 = __lsx_vldi(0);
-    __m128 acc_2 = __lsx_vldi(0);
-    __m128 acc_3 = __lsx_vldi(0);
-
-    for (; ib + 1 < nb; ib += 2) {
-
-        // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = __lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
-
-        const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);
-
-        __m128i bx_0 = __lsx_vand_v(low_mask, tmp_0_1);
-        __m128i by_0 = __lsx_vld((const __m128i *)y[ib].qs, 0);
-        bx_0 = __lsx_vsub_b(bx_0, off);
-        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
-
-        __m128i bx_1 = __lsx_vand_v(low_mask, __lsx_vsrli_d(tmp_0_1, 4));
-        __m128i by_1 = __lsx_vld((const __m128i *)(y[ib].qs + 16), 0);
-        bx_1 = __lsx_vsub_b(bx_1, off);
-        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
-
-        //_mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
-        //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
-
-        // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = __lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
-
-        const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);
-
-        __m128i bx_2 = __lsx_vand_v(low_mask, tmp_2_3);
-        __m128i by_2 = __lsx_vld((const __m128i *)y[ib + 1].qs, 0);
-        bx_2 = __lsx_vsub_b(bx_2, off);
-        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
-
-        __m128i bx_3 = __lsx_vand_v(low_mask, __lsx_vsrli_d(tmp_2_3, 4));
-        __m128i by_3 = __lsx_vld((const __m128i *)(y[ib + 1].qs + 16), 0);
-        bx_3 = __lsx_vsub_b(bx_3, off);
-        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
-
-        // Convert int32_t to float
-        __m128 p0 = __lsx_vffint_s_w(i32_0);
-        __m128 p1 = __lsx_vffint_s_w(i32_1);
-        __m128 p2 = __lsx_vffint_s_w(i32_2);
-        __m128 p3 = __lsx_vffint_s_w(i32_3);
-
-        // Apply the scale
-        __m128 p0_d = __lsx_vfmul_s( d_0_1, p0 );
-        __m128 p1_d = __lsx_vfmul_s( d_0_1, p1 );
-        __m128 p2_d = __lsx_vfmul_s( d_2_3, p2 );
-        __m128 p3_d = __lsx_vfmul_s( d_2_3, p3 );
-
-        // Acummulate
-        acc_0 = __lsx_vfadd_s(p0_d, acc_0);
-        acc_1 = __lsx_vfadd_s(p1_d, acc_1);
-        acc_2 = __lsx_vfadd_s(p2_d, acc_2);
-        acc_3 = __lsx_vfadd_s(p3_d, acc_3);
-    }
-
-    sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
-            const int v1 = (x[ib].qs[j] >>   4) - 8;
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
-    }
-
-    *s = sumf;
-}
-
-void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    const int qk = QK8_1;
-    const int nb = n / qk;
-
-    assert(n % qk == 0);
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    assert((nrc == 2) || (nrc == 1));
-#else
-    assert(nrc == 1);
-#endif
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_q4_1 * restrict x = vx;
-    const block_q8_1 * restrict y = vy;
-
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    if (nrc == 2) {
-        const block_q4_1 * restrict vx0 = vx;
-        const block_q4_1 * restrict vx1 = (const block_q4_1 *) ((const uint8_t*)vx + bx);
-        const block_q8_1 * restrict vy0 = vy;
-        const block_q8_1 * restrict vy1 = (const block_q8_1 *) ((const uint8_t*)vy + by);
-
-        float32x4_t sumv0 = vdupq_n_f32(0.0f);
-        float32x4_t summs0 = vdupq_n_f32(0.0f);
-
-        for (int i = 0; i < nb; i++) {
-            const block_q4_1 * restrict b_x0 = &vx0[i];
-            const block_q4_1 * restrict b_x1 = &vx1[i];
-            const block_q8_1 * restrict b_y0 = &vy0[i];
-            const block_q8_1 * restrict b_y1 = &vy1[i];
-
-            float32_t summs_t[4] = {GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
-                                    GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
-                                    GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
-                                    GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)};
-            summs0 = vaddq_f32(summs0, vld1q_f32(summs_t));
-
-            const uint8x16_t m4b = vdupq_n_u8(0x0F);
-
-            const uint8x16_t v0_0 = vld1q_u8(b_x0->qs);
-            const uint8x16_t v0_1 = vld1q_u8(b_x1->qs);
-
-            // 4-bit -> 8-bit
-            const int8x16_t x0_l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-            const int8x16_t x0_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-            const int8x16_t x1_l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-            const int8x16_t x1_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-            // load y
-            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
-            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
-            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
-            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
-
-            // mmla into int32x4_t
-            float32_t _scale[4] = {GGML_FP16_TO_FP32(b_x0->d)*b_y0->d,
-                                   GGML_FP16_TO_FP32(b_x0->d)*b_y1->d,
-                                   GGML_FP16_TO_FP32(b_x1->d)*b_y0->d,
-                                   GGML_FP16_TO_FP32(b_x1->d)*b_y1->d};
-            float32x4_t scale = vld1q_f32(_scale);
-
-            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
-            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
-
-            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
-            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
-
-            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
-            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
-
-            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
-            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
-            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
-                                                                                l1, r1)), l2, r2)), l3, r3))), scale);
-        }
-
-        float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
-        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
-        sumv2 = vaddq_f32(sumv2, summs0);
-
-        vst1_f32(s,      vget_low_f32 (sumv2));
-        vst1_f32(s + bs, vget_high_f32(sumv2));
-        return;
-    }
-#endif
-
-    int ib = 0;
-    float sumf = 0;
-
-    // TODO: add WASM SIMD
-#if defined(__ARM_NEON)
-    float32x4_t sumv0 = vdupq_n_f32(0.0f);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    float summs = 0;
-
-    for (; ib + 1 < nb; ib += 2) {
-        const block_q4_1 * restrict x0 = &x[ib + 0];
-        const block_q4_1 * restrict x1 = &x[ib + 1];
-        const block_q8_1 * restrict y0 = &y[ib + 0];
-        const block_q8_1 * restrict y1 = &y[ib + 1];
-
-        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
-
-        const uint8x16_t m4b = vdupq_n_u8(0x0F);
-
-        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
-        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
-
-        // 4-bit -> 8-bit
-        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-        // load y
-        const int8x16_t v1_0l = vld1q_s8(y0->qs);
-        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
-        const int8x16_t v1_1l = vld1q_s8(y1->qs);
-        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
-
-        // dot product into int32x4_t
-        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
-        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-    }
-
-    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
-#elif defined(__AVX2__) || defined(__AVX__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-
-    float summs = 0;
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
-        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
-
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
-
-        const __m256 d0v = _mm256_set1_ps( d0 );
-        const __m256 d1v = _mm256_set1_ps( d1 );
-
-        // Compute combined scales
-        const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
-
-        // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
-        const __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-        const __m256i qy = _mm256_loadu_si256( (const __m256i *)y[ib].qs );
-
-        const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
-
-        // Accumulate d0*d1*x*y
-#if defined(__AVX2__)
-        acc = _mm256_fmadd_ps( d0d1, xy, acc );
-#else
-        acc = _mm256_add_ps( _mm256_mul_ps( d0d1, xy ), acc );
-#endif
-    }
-
-    sumf = hsum_float_8(acc) + summs;
-#elif defined(__riscv_v_intrinsic)
-    size_t vl = __riscv_vsetvl_e8m1(qk/2);
-
-    for (; ib < nb; ++ib) {
-        // load elements
-        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
-
-        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
-        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
-
-        // mask and store lower part of x, and then upper part
-        vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
-        vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
-
-        vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
-        vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
-
-        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
-        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
-
-        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
-
-        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
-        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
-
-        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
-
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
-    }
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector signed int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-
-#pragma GCC unroll 4
-    for (; ib < nb; ++ib) {
-        __builtin_prefetch(x[ib].qs, 0, 1);
-        __builtin_prefetch(y[ib].qs, 0, 1);
-
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
-        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f};
-        vsumf0 = vec_madd(vxmin, vys, vsumf0);
-
-        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
-        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
-        vector signed char q8y1 = vec_xl(16, y[ib].qs);
-
-        vector unsigned char q4x0 = (vector unsigned char)vec_and(qxs, lowMask);
-        vector unsigned char q4x1 = (vector unsigned char)vec_sr(qxs, v4);
-
-        vector signed int vsumi0 = v0;
-
-        vsumi0 = vec_msum(q8y0, q4x0, vsumi0);
-        vsumi0 = vec_msum(q8y1, q4x1, vsumi0);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-    }
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    sumf = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-    // Initialize accumulator with zeros
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    float summs = 0;
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
-        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
-
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
-
-        const __m256 d0v = __lasx_xvreplfr2vr_s( d0 );
-        const __m256 d1v = __lasx_xvreplfr2vr_s( d1 );
-
-        // Compute combined scales
-        const __m256 d0d1 = __lasx_xvfmul_s( d0v, d1v );
-
-        // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
-        const __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-        const __m256i qy = __lasx_xvld( (const __m256i *)y[ib].qs, 0);
-
-        const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
-
-        // Accumulate d0*d1*x*y
-        acc = __lasx_xvfmadd_s( d0d1, xy, acc );
-    }
-
-    sumf = hsum_float_8(acc) + summs;
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[ib].qs[j] & 0x0F);
-            const int v1 = (x[ib].qs[j] >>   4);
-
-            sumi0 += (v0 * y[ib].qs[j]);
-            sumi1 += (v1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
-    }
-
-    *s = sumf;
-}
-
-void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-
-    int ib = 0;
-    float sumf = 0;
-
-    assert(n % qk == 0);
-    assert(qk == QK5_0);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_q5_0 * restrict x = vx;
-    const block_q8_0 * restrict y = vy;
-
-#if defined(__ARM_NEON)
-    float32x4_t sumv0 = vdupq_n_f32(0.0f);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    uint32_t qh0;
-    uint32_t qh1;
-
-    uint64_t tmp0[4];
-    uint64_t tmp1[4];
-
-    for (; ib + 1 < nb; ib += 2) {
-        const block_q5_0 * restrict x0 = &x[ib];
-        const block_q5_0 * restrict x1 = &x[ib + 1];
-        const block_q8_0 * restrict y0 = &y[ib];
-        const block_q8_0 * restrict y1 = &y[ib + 1];
-
-        const uint8x16_t m4b = vdupq_n_u8(0x0F);
-
-        // extract the 5th bit via lookup table ((!b) << 4)
-        memcpy(&qh0, x0->qh, sizeof(qh0));
-        memcpy(&qh1, x1->qh, sizeof(qh1));
-
-        tmp0[0] = table_b2b_1[(qh0 >>  0) & 0xFF];
-        tmp0[1] = table_b2b_1[(qh0 >>  8) & 0xFF];
-        tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF];
-        tmp0[3] = table_b2b_1[(qh0 >> 24)       ];
-
-        tmp1[0] = table_b2b_1[(qh1 >>  0) & 0xFF];
-        tmp1[1] = table_b2b_1[(qh1 >>  8) & 0xFF];
-        tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF];
-        tmp1[3] = table_b2b_1[(qh1 >> 24)       ];
-
-        const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
-        const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
-        const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
-        const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));
-
-        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
-        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
-
-        // 4-bit -> 8-bit
-        int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-        int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-        int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-        int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
-        const int8x16_t v0_0lf = vsubq_s8(v0_0l, qhl0);
-        const int8x16_t v0_0hf = vsubq_s8(v0_0h, qhh0);
-        const int8x16_t v0_1lf = vsubq_s8(v0_1l, qhl1);
-        const int8x16_t v0_1hf = vsubq_s8(v0_1h, qhh1);
-
-        // load y
-        const int8x16_t v1_0l = vld1q_s8(y0->qs);
-        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
-        const int8x16_t v1_1l = vld1q_s8(y1->qs);
-        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-    }
-
-    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
-#elif defined(__wasm_simd128__)
-    v128_t sumv = wasm_f32x4_splat(0.0f);
-
-    uint32_t qh;
-    uint64_t tmp[4];
-
-    // TODO: check if unrolling this is better
-    for (; ib < nb; ++ib) {
-        const block_q5_0 * restrict x0 = &x[ib];
-        const block_q8_0 * restrict y0 = &y[ib];
-
-        const v128_t m4b  = wasm_i8x16_splat(0x0F);
-
-        // extract the 5th bit
-        memcpy(&qh, x0->qh, sizeof(qh));
-
-        tmp[0] = table_b2b_1[(qh >>  0) & 0xFF];
-        tmp[1] = table_b2b_1[(qh >>  8) & 0xFF];
-        tmp[2] = table_b2b_1[(qh >> 16) & 0xFF];
-        tmp[3] = table_b2b_1[(qh >> 24)       ];
-
-        const v128_t qhl = wasm_v128_load(tmp + 0);
-        const v128_t qhh = wasm_v128_load(tmp + 2);
-
-        const v128_t v0 = wasm_v128_load(x0->qs);
-
-        // 4-bit -> 8-bit
-        const v128_t v0l = wasm_v128_and (v0, m4b);
-        const v128_t v0h = wasm_u8x16_shr(v0, 4);
-
-        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
-        const v128_t v0lf = wasm_i8x16_sub(v0l, qhl);
-        const v128_t v0hf = wasm_i8x16_sub(v0h, qhh);
-
-        // load y
-        const v128_t v1l = wasm_v128_load(y0->qs);
-        const v128_t v1h = wasm_v128_load(y0->qs + 16);
-
-        // int8x16 -> int16x8
-        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
-        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
-        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
-        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
-
-        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
-        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
-        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
-        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
-
-        // dot product
-        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
-                        wasm_i32x4_add(
-                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
-                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
-                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
-                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
-                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
-    }
-
-    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
-           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
-#elif defined(__AVX2__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
-
-        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
-        bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0));
-        qx = _mm256_or_si256(qx, bxhi);
-
-        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
-
-        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
-
-        /* Multiply q with scale and accumulate */
-        acc = _mm256_fmadd_ps(d, q, acc);
-    }
-
-    sumf = hsum_float_8(acc);
-#elif defined(__AVX__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-    __m128i mask = _mm_set1_epi8((char)0xF0);
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
-
-        __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
-        const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
-        __m128i bxhil = _mm256_castsi256_si128(bxhi);
-        __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
-        bxhil = _mm_andnot_si128(bxhil, mask);
-        bxhih = _mm_andnot_si128(bxhih, mask);
-        __m128i bxl = _mm256_castsi256_si128(bx_0);
-        __m128i bxh = _mm256_extractf128_si256(bx_0, 1);
-        bxl = _mm_or_si128(bxl, bxhil);
-        bxh = _mm_or_si128(bxh, bxhih);
-        bx_0 = MM256_SET_M128I(bxh, bxl);
-
-        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
-
-        const __m256 q = mul_sum_i8_pairs_float(bx_0, by_0);
-
-        /* Multiply q with scale and accumulate */
-        acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc);
-    }
-
-    sumf = hsum_float_8(acc);
-#elif defined(__riscv_v_intrinsic)
-    uint32_t qh;
-
-    size_t vl = __riscv_vsetvl_e8m1(qk/2);
-
-    // These temporary registers are for masking and shift operations
-    vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
-    vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl);
-
-    vuint32m2_t vt_3 = __riscv_vsll_vx_u32m2(vt_2, 16, vl);
-    vuint32m2_t vt_4 = __riscv_vadd_vx_u32m2(vt_1, 12, vl);
-
-    for (; ib < nb; ++ib) {
-        memcpy(&qh, x[ib].qh, sizeof(uint32_t));
-
-        // ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-        vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(vt_2, qh, vl);
-        vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(xha_0, vt_1, vl);
-        vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl);
-
-        // ((qh & (1u << (j + 16))) >> (j + 12));
-        vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(vt_3, qh, vl);
-        vuint32m2_t xhl_1 = __riscv_vsrl_vv_u32m2(xha_1, vt_4, vl);
-
-        // narrowing
-        vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xhl_0, vl);
-        vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl);
-
-        vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xhl_1, vl);
-        vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl);
-
-        // load
-        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
-
-        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
-        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
-
-        vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
-        vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
-
-        vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl);
-        vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl);
-
-        vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
-        vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
-
-        vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 16, vl);
-        vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 16, vl);
-
-        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
-        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
-
-        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
-
-        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
-        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
-
-        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
-
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
-    }
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector unsigned char v4 = vec_splats((unsigned char)4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-
-#pragma GCC unroll 4
-    for (; ib < nb; ++ib) {
-        __builtin_prefetch(x[ib].qs, 0, 1);
-        __builtin_prefetch(y[ib].qs, 0, 1);
-
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed long long aux64x2_0 = {(uint64_t)(table_b2b_1[x[ib].qh[0]]), (uint64_t)(table_b2b_1[x[ib].qh[1]])};
-        vector signed long long aux64x2_1 = {(uint64_t)(table_b2b_1[x[ib].qh[2]]), (uint64_t)(table_b2b_1[x[ib].qh[3]])};
-
-        vector signed char qh0 = (vector signed char)aux64x2_0;
-        vector signed char qh1 = (vector signed char)aux64x2_1;
-
-        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
-
-        vector signed char q5x0 = vec_sub(vec_and (qxs, lowMask), qh0);
-        vector signed char q5x1 = vec_sub(vec_sr(qxs, v4), qh1);
-
-        vector signed char q8y0 = vec_xl(  0, y[ib].qs);
-        vector signed char q8y1 = vec_xl( 16, y[ib].qs);
-
-        vector signed short qv0 = vec_add(vec_mule(q5x0, q8y0), vec_mulo(q5x0, q8y0));
-        vector signed short qv1 = vec_add(vec_mule(q5x1, q8y1), vec_mulo(q5x1, q8y1));
-
-        qv0 = vec_add(qv0, qv1);
-
-        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0));
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-    }
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    sumf = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-    // Initialize accumulator with zeros
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        /* Compute combined scale for the block */
-        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); //FIXME
-
-        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
-        bxhi = __lasx_xvandn_v(bxhi, __lasx_xvreplgr2vr_b((char)0xF0));
-        qx = __lasx_xvor_v(qx, bxhi);
-
-        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
-
-        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
-
-        /* Multiply q with scale and accumulate */
-        acc = __lasx_xvfmadd_s(d, q, acc);
-    }
-
-    sumf = hsum_float_8(acc);
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
-
-            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
-            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
-    }
-
-    *s = sumf;
-}
-
-void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    const int qk = QK8_1;
-    const int nb = n / qk;
-
-    int ib = 0;
-    float sumf = 0;
-
-    assert(n % qk == 0);
-    assert(qk == QK5_1);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_q5_1 * restrict x = vx;
-    const block_q8_1 * restrict y = vy;
-
-#if defined(__ARM_NEON)
-    float32x4_t sumv0 = vdupq_n_f32(0.0f);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    float summs0 = 0.0f;
-    float summs1 = 0.0f;
-
-    uint32_t qh0;
-    uint32_t qh1;
-
-    uint64_t tmp0[4];
-    uint64_t tmp1[4];
-
-    for (; ib + 1 < nb; ib += 2) {
-        const block_q5_1 * restrict x0 = &x[ib];
-        const block_q5_1 * restrict x1 = &x[ib + 1];
-        const block_q8_1 * restrict y0 = &y[ib];
-        const block_q8_1 * restrict y1 = &y[ib + 1];
-
-        const uint8x16_t m4b = vdupq_n_u8(0x0F);
-
-        summs0 += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
-        summs1 += GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
-
-        // extract the 5th bit via lookup table ((b) << 4)
-        memcpy(&qh0, x0->qh, sizeof(qh0));
-        memcpy(&qh1, x1->qh, sizeof(qh1));
-
-        tmp0[0] = table_b2b_0[(qh0 >>  0) & 0xFF];
-        tmp0[1] = table_b2b_0[(qh0 >>  8) & 0xFF];
-        tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF];
-        tmp0[3] = table_b2b_0[(qh0 >> 24)       ];
-
-        tmp1[0] = table_b2b_0[(qh1 >>  0) & 0xFF];
-        tmp1[1] = table_b2b_0[(qh1 >>  8) & 0xFF];
-        tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF];
-        tmp1[3] = table_b2b_0[(qh1 >> 24)       ];
-
-        const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
-        const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
-        const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
-        const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));
-
-        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
-        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
-
-        // 4-bit -> 8-bit
-        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-        // add high bit
-        const int8x16_t v0_0lf = vorrq_s8(v0_0l, qhl0);
-        const int8x16_t v0_0hf = vorrq_s8(v0_0h, qhh0);
-        const int8x16_t v0_1lf = vorrq_s8(v0_1l, qhl1);
-        const int8x16_t v0_1hf = vorrq_s8(v0_1h, qhh1);
-
-        // load y
-        const int8x16_t v1_0l = vld1q_s8(y0->qs);
-        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
-        const int8x16_t v1_1l = vld1q_s8(y1->qs);
-        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-    }
-
-    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
-#elif defined(__wasm_simd128__)
-    v128_t sumv = wasm_f32x4_splat(0.0f);
-
-    float summs = 0.0f;
-
-    uint32_t qh;
-    uint64_t tmp[4];
-
-    // TODO: check if unrolling this is better
-    for (; ib < nb; ++ib) {
-        const block_q5_1 * restrict x0 = &x[ib];
-        const block_q8_1 * restrict y0 = &y[ib];
-
-        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
-
-        const v128_t m4b = wasm_i8x16_splat(0x0F);
-
-        // extract the 5th bit
-        memcpy(&qh, x0->qh, sizeof(qh));
-
-        tmp[0] = table_b2b_0[(qh >>  0) & 0xFF];
-        tmp[1] = table_b2b_0[(qh >>  8) & 0xFF];
-        tmp[2] = table_b2b_0[(qh >> 16) & 0xFF];
-        tmp[3] = table_b2b_0[(qh >> 24)       ];
-
-        const v128_t qhl = wasm_v128_load(tmp + 0);
-        const v128_t qhh = wasm_v128_load(tmp + 2);
-
-        const v128_t v0 = wasm_v128_load(x0->qs);
-
-        // 4-bit -> 8-bit
-        const v128_t v0l = wasm_v128_and (v0, m4b);
-        const v128_t v0h = wasm_u8x16_shr(v0, 4);
-
-        // add high bit
-        const v128_t v0lf = wasm_v128_or(v0l, qhl);
-        const v128_t v0hf = wasm_v128_or(v0h, qhh);
-
-        // load y
-        const v128_t v1l = wasm_v128_load(y0->qs);
-        const v128_t v1h = wasm_v128_load(y0->qs + 16);
-
-        // int8x16 -> int16x8
-        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
-        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
-        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
-        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
-
-        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
-        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
-        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
-        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
-
-        // dot product
-        sumv = wasm_f32x4_add(sumv,
-                wasm_f32x4_mul(wasm_f32x4_convert_i32x4(wasm_i32x4_add(
-                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
-                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
-                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
-                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
-                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
-    }
-
-    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
-           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
-#elif defined(__AVX2__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-
-    float summs = 0.0f;
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
-
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
-
-        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
-        bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
-        qx = _mm256_or_si256(qx, bxhi);
-
-        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
-        const __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
-
-        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
-
-        acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
-    }
-
-    sumf = hsum_float_8(acc) + summs;
-#elif defined(__AVX__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-    __m128i mask = _mm_set1_epi8(0x10);
-
-    float summs = 0.0f;
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
-
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
-
-        __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
-        const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
-        __m128i bxhil = _mm256_castsi256_si128(bxhi);
-        __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
-        bxhil = _mm_and_si128(bxhil, mask);
-        bxhih = _mm_and_si128(bxhih, mask);
-        __m128i bxl = _mm256_castsi256_si128(bx_0);
-        __m128i bxh = _mm256_extractf128_si256(bx_0, 1);
-        bxl = _mm_or_si128(bxl, bxhil);
-        bxh = _mm_or_si128(bxh, bxhih);
-        bx_0 = MM256_SET_M128I(bxh, bxl);
-
-        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
-        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
-
-        const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
-
-        acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc);
-    }
-
-    sumf = hsum_float_8(acc) + summs;
-#elif defined(__riscv_v_intrinsic)
-    uint32_t qh;
-
-    size_t vl = __riscv_vsetvl_e8m1(qk/2);
-
-    // temporary registers for shift operations
-    vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
-    vuint32m2_t vt_2 = __riscv_vadd_vx_u32m2(vt_1, 12, vl);
-
-    for (; ib < nb; ++ib) {
-        memcpy(&qh, x[ib].qh, sizeof(uint32_t));
-
-        // load qh
-        vuint32m2_t vqh = __riscv_vmv_v_x_u32m2(qh, vl);
-
-        // ((qh >> (j +  0)) << 4) & 0x10;
-        vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(vqh, vt_1, vl);
-        vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl);
-        vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(xhl_0, 0x10, vl);
-
-        // ((qh >> (j + 12))     ) & 0x10;
-        vuint32m2_t xhr_1 = __riscv_vsrl_vv_u32m2(vqh, vt_2, vl);
-        vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(xhr_1, 0x10, vl);
-
-        // narrowing
-        vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xha_0, vl);
-        vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl);
-
-        vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xha_1, vl);
-        vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl);
-
-        // load
-        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
-
-        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
-        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
-
-        vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
-        vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
-
-        vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl);
-        vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl);
-
-        vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
-        vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
-
-        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
-        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
-
-        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
-
-        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
-        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
-
-        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
-
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
-    }
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector signed int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-
-#pragma GCC unroll 4
-    for (; ib < nb; ++ib) {
-        __builtin_prefetch(x[ib].qs, 0, 1);
-        __builtin_prefetch(y[ib].qs, 0, 1);
-
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
-        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f};
-        vsumf0 = vec_madd(vxmin, vys, vsumf0);
-
-        vector unsigned long long aux64x2_0 = {(uint64_t)(table_b2b_0[x[ib].qh[0]]), (uint64_t)(table_b2b_0[x[ib].qh[1]])};
-        vector unsigned long long aux64x2_1 = {(uint64_t)(table_b2b_0[x[ib].qh[2]]), (uint64_t)(table_b2b_0[x[ib].qh[3]])};
-
-        vector signed char qh0 = (vector signed char)aux64x2_0;
-        vector signed char qh1 = (vector signed char)aux64x2_1;
-
-        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
-
-        vector unsigned char q5x0 = (vector unsigned char)vec_or(vec_and(qxs, lowMask), qh0);
-        vector unsigned char q5x1 = (vector unsigned char)vec_or(vec_sr(qxs, v4), qh1);
-
-        vector signed char q8y0 = vec_xl(  0, y[ib].qs);
-        vector signed char q8y1 = vec_xl( 16, y[ib].qs);
-
-        vector signed int vsumi0 = v0;
-
-        vsumi0 = vec_msum(q8y0, q5x0, vsumi0);
-        vsumi0 = vec_msum(q8y1, q5x1, vsumi0);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-    }
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    sumf = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-    // Initialize accumulator with zeros
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    float summs = 0.0f;
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        const __m256 dx = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d));
-
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
-
-        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
-        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
-        bxhi = __lasx_xvand_v(bxhi, __lasx_xvreplgr2vr_b(0x10));
-        qx = __lasx_xvor_v(qx, bxhi);
-
-        const __m256 dy = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib].d));
-        const __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
-
-        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
-
-        acc = __lasx_xvfmadd_s(q, __lasx_xvfmul_s(dx, dy), acc);
-    }
-
-    sumf = hsum_float_8(acc) + summs;
-#endif
-    for (; ib < nb; ++ib) {
-        uint32_t qh;
-        memcpy(&qh, x[ib].qh, sizeof(qh));
-
-        int sumi0 = 0;
-        int sumi1 = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
-            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
-
-            sumi0 += (x0 * y[ib].qs[j]);
-            sumi1 += (x1 * y[ib].qs[j + qk/2]);
-        }
-
-        int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
-    }
-
-    *s = sumf;
-}
-
-void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-
-    assert(n % qk == 0);
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    assert((nrc == 2) || (nrc == 1));
-#else
-    assert(nrc == 1);
-#endif
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_q8_0 * restrict x = vx;
-    const block_q8_0 * restrict y = vy;
-
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    if (nrc == 2) {
-        const block_q8_0 * restrict vx0 = vx;
-        const block_q8_0 * restrict vx1 = (const block_q8_0 *) ((const uint8_t*)vx + bx);
-        const block_q8_0 * restrict vy0 = vy;
-        const block_q8_0 * restrict vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by);
-
-        float32x4_t sumv0 = vdupq_n_f32(0.0f);
-
-        for (int i = 0; i < nb; i++) {
-            const block_q8_0 * restrict b_x0 = &vx0[i];
-            const block_q8_0 * restrict b_y0 = &vy0[i];
-
-            const block_q8_0 * restrict b_x1 = &vx1[i];
-            const block_q8_0 * restrict b_y1 = &vy1[i];
-
-            const int8x16_t x0_l = vld1q_s8(b_x0->qs);
-            const int8x16_t x0_h = vld1q_s8(b_x0->qs + 16);
-            const int8x16_t x1_l = vld1q_s8(b_x1->qs);
-            const int8x16_t x1_h = vld1q_s8(b_x1->qs + 16);
-
-            // load y
-            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
-            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
-            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
-            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
-
-            float32_t _scale[4] = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                                   GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                                   GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                                   GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
-            float32x4_t scale = vld1q_f32(_scale);
-
-            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
-            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
-
-            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
-            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
-
-            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
-            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
-
-            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
-            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
-
-            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
-                                                                                       l1, r1)), l2, r2)), l3, r3))), scale);
-        }
-        float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
-        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
-
-        vst1_f32(s, vget_low_f32(sumv2));
-        vst1_f32(s + bs, vget_high_f32(sumv2));
-        return;
-    }
-#endif
-
-    int ib = 0;
-    float sumf = 0;
-
-#if defined(__ARM_FEATURE_SVE)
-    svfloat32_t sumv0 = svdup_n_f32(0.0f);
-    svfloat32_t sumv1 = svdup_n_f32(0.0f);
-
-    const int vector_length = ggml_cpu_get_sve_cnt()*8;
-
-    //VLA Implemenation for SVE
-    switch (vector_length) {
-        case 128:
-            {
-                // predicate for activating lanes for 16 Int8 elements
-                const svbool_t ph16 = svptrue_pat_b8 (SV_VL16);
-                const svbool_t pl16 = svptrue_pat_b32(SV_VL4);
-
-                for (; ib + 1 < nb; ib += 2) {
-                    const block_q8_0 * restrict x0 = &x[ib + 0];
-                    const block_q8_0 * restrict x1 = &x[ib + 1];
-                    const block_q8_0 * restrict y0 = &y[ib + 0];
-                    const block_q8_0 * restrict y1 = &y[ib + 1];
-
-                    // load x
-                    const svint8_t qx0_0 = svld1_s8(ph16, x0->qs);
-                    const svint8_t qx0_1 = svld1_s8(ph16, x0->qs+16);
-                    const svint8_t qx1_0 = svld1_s8(ph16, x1->qs);
-                    const svint8_t qx1_1 = svld1_s8(ph16, x1->qs+16);
-
-                    // load y
-                    const svint8_t qy0_0 = svld1_s8(ph16, y0->qs);
-                    const svint8_t qy0_1 = svld1_s8(ph16, y0->qs+16);
-                    const svint8_t qy1_0 = svld1_s8(ph16, y1->qs);
-                    const svint8_t qy1_1 = svld1_s8(ph16, y1->qs+16);
-
-                    sumv0 = svmla_n_f32_x(pl16, sumv0, svcvt_f32_s32_x(pl16, svadd_x(pl16,
-                                    svdot_s32(svdup_n_s32(0), qx0_0, qy0_0),
-                                    svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-                    sumv1 = svmla_n_f32_x(pl16, sumv1, svcvt_f32_s32_x(pl16, svadd_x(pl16,
-                                    svdot_s32(svdup_n_s32(0), qx1_0, qy1_0),
-                                    svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-                }
-
-                sumf = svaddv_f32(pl16, svadd_f32_x(pl16, sumv0, sumv1));
-            } break;
-        case 256:
-            {
-                //printf("sve256");
-                for (; ib + 1 < nb; ib += 2) {
-                    const block_q8_0 * restrict x0 = &x[ib + 0];
-                    const block_q8_0 * restrict x1 = &x[ib + 1];
-                    const block_q8_0 * restrict y0 = &y[ib + 0];
-                    const block_q8_0 * restrict y1 = &y[ib + 1];
-
-                    // load x
-                    const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs);
-                    const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs);
-
-                    // load y
-                    const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
-                    const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
-
-                    sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-                    sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-                }
-
-                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
-            } break;
-        case 512:
-            {
-                // predicate for activating high 256 bit
-                const svbool_t ph32 = svptrue_pat_b8(SV_VL32);
-                // predicate for activating low 256 bit
-                const svbool_t pl32 = svnot_b_z(svptrue_b8(), ph32);
-
-                // predicate for activating high lanes for 8 float32 elements
-                const svbool_t ph8 = svptrue_pat_b32(SV_VL8);
-                // predicate for activating low lanes for 8 float32 elements
-                const svbool_t pl8 = svnot_b_z(svptrue_b32(), ph8);
-
-                svfloat32_t sumv00 = svdup_n_f32(0.0f);
-
-                for (; ib + 1 < nb; ib += 2) {
-                    const block_q8_0 * restrict x0 = &x[ib + 0];
-                    const block_q8_0 * restrict x1 = &x[ib + 1];
-                    const block_q8_0 * restrict y0 = &y[ib + 0];
-                    const block_q8_0 * restrict y1 = &y[ib + 1];
-
-                    //load 32 int8_t in first half of vector and put another 32 int8_t in second vector lower bits
-                    // and add them to make one 64 element vector
-                    // load x
-                    const svint8_t qx_32 = svld1_s8(ph32, x0->qs);
-                          svint8_t qx_64 = svld1_s8(pl32, x0->qs + 2);
-
-                    qx_64 = svadd_s8_x(svptrue_b8(), qx_32, qx_64);
-
-                    // load y
-                    const svint8_t qy_32 = svld1_s8(ph32, y0->qs);
-                          svint8_t qy_64 = svld1_s8(pl32, y0->qs + 2);
-
-                    qy_64 = svadd_s8_x(svptrue_b8(), qy_32, qy_64);
-
-                    // scale creation
-                    const float32_t deq1 = GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d);
-                    const float32_t deq2 = GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d);
-
-                    // duplicate deq1 in first half of vector and deq2 in second half of vector
-                    const svfloat32_t temp = svdup_f32_m(svdup_f32_z(ph8, deq1), pl8, deq2);
-
-                    const svfloat32_t sumvt = svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx_64, qy_64));
-
-                    sumv00 = svmla_f32_m(svptrue_b32(), sumv00, sumvt, temp);
-                }
-
-                sumf = svaddv_f32(svptrue_b32(), sumv00);
-                break;
-            }
-        default:
-            assert(false && "Unsupported vector length");
-            break;
-    }
-#elif defined(__ARM_NEON)
-    float32x4_t sumv0 = vdupq_n_f32(0.0f);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    for (; ib + 1 < nb; ib += 2) {
-        const block_q8_0 * restrict x0 = &x[ib + 0];
-        const block_q8_0 * restrict x1 = &x[ib + 1];
-        const block_q8_0 * restrict y0 = &y[ib + 0];
-        const block_q8_0 * restrict y1 = &y[ib + 1];
-
-        const int8x16_t x0_0 = vld1q_s8(x0->qs);
-        const int8x16_t x0_1 = vld1q_s8(x0->qs + 16);
-        const int8x16_t x1_0 = vld1q_s8(x1->qs);
-        const int8x16_t x1_1 = vld1q_s8(x1->qs + 16);
-
-        // load y
-        const int8x16_t y0_0 = vld1q_s8(y0->qs);
-        const int8x16_t y0_1 = vld1q_s8(y0->qs + 16);
-        const int8x16_t y1_0 = vld1q_s8(y1->qs);
-        const int8x16_t y1_1 = vld1q_s8(y1->qs + 16);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
-                        ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
-                        ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-    }
-
-    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
-#elif defined(__AVX2__) || defined(__AVX__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        // Compute combined scale for the block
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
-        __m256i qx = _mm256_loadu_si256((const __m256i *)x[ib].qs);
-        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
-
-        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
-
-        // Multiply q with scale and accumulate
-#if defined(__AVX2__)
-        acc = _mm256_fmadd_ps( d, q, acc );
-#else
-        acc = _mm256_add_ps( _mm256_mul_ps( d, q ), acc );
-#endif
-    }
-
-    sumf = hsum_float_8(acc);
-#elif defined(__riscv_v_intrinsic)
-    size_t vl = __riscv_vsetvl_e8m1(qk);
-
-    for (; ib < nb; ++ib) {
-        // load elements
-        vint8m1_t bx_0 = __riscv_vle8_v_i8m1(x[ib].qs, vl);
-        vint8m1_t by_0 = __riscv_vle8_v_i8m1(y[ib].qs, vl);
-
-        vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx_0, by_0, vl);
-
-        vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, vl);
-        vint32m1_t v_sum = __riscv_vwredsum_vs_i16m2_i32m1(vw_mul, v_zero, vl);
-
-        int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum);
-
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
-    }
-#elif defined(__POWER9_VECTOR__)
-    const vector signed int v0 = vec_splats((int32_t)0);
-    vector float vsumf0 = vec_splats(0.0f);
-
-#pragma GCC unroll 8
-    for (; ib < nb; ++ib) {
-        __builtin_prefetch(x[ib].qs, 0, 1);
-        __builtin_prefetch(y[ib].qs, 0, 1);
-
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed char q8x0 = vec_xl( 0, x[ib].qs);
-        vector signed char q8x1 = vec_xl(16, x[ib].qs);
-        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
-        vector signed char q8y1 = vec_xl(16, y[ib].qs);
-
-        vector signed short qv0 = vec_mule(q8x0, q8y0);
-        vector signed short qv1 = vec_mulo(q8x0, q8y0);
-        vector signed short qv2 = vec_mule(q8x1, q8y1);
-        vector signed short qv3 = vec_mulo(q8x1, q8y1);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-
-        vsumi0 = vec_sum4s(qv0, vsumi0);
-        vsumi1 = vec_sum4s(qv1, vsumi1);
-        vsumi0 = vec_sum4s(qv2, vsumi0);
-        vsumi1 = vec_sum4s(qv3, vsumi1);
-
-        vsumi0 = vec_add(vsumi0, vsumi1);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-    }
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    sumf = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-    // Initialize accumulator with zeros
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    // Main loop
-    for (; ib < nb; ++ib) {
-        // Compute combined scale for the block
-        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
-        __m256i qx = __lasx_xvld((const __m256i *)x[ib].qs, 0);
-        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
-
-        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
-
-        // Multiply q with scale and accumulate
-        acc = __lasx_xvfmadd_s( d, q, acc );
-    }
-
-    sumf = hsum_float_8(acc);
-#endif
-    for (; ib < nb; ++ib) {
-        int sumi = 0;
-
-        for (int j = 0; j < qk; j++) {
-            sumi += x[ib].qs[j]*y[ib].qs[j];
-        }
-
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
-    }
-
-    *s = sumf;
-}
-
-void ggml_vec_dot_tq1_0_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_tq1_0 * restrict x = vx;
-    const block_q8_K  * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_NEON)
-    float sumf = 0.0f;
-
-    uint8_t k_shift[16] = {1, 1, 1, 1, 3, 3, 3, 3, 9, 9, 9, 9, 27, 27, 27, 27};
-
-    const uint8x16_t shift = vld1q_u8(k_shift);
-
-    for (int i = 0; i < nb; ++i) {
-#if defined(__ARM_FEATURE_DOTPROD)
-        int32x4_t sumi0 = vdupq_n_s32(0);
-        int32x4_t sumi1 = vdupq_n_s32(0);
-#else
-        int16x8_t sumi0 = vdupq_n_s16(0);
-        int16x8_t sumi1 = vdupq_n_s16(0);
-#endif
-
-        // first 32 bytes of 5 elements
-        {
-            uint8x16_t qx0 = vld1q_u8(x[i].qs + 0);
-            uint8x16_t qx1 = vld1q_u8(x[i].qs + 16);
-            uint8x16_t qx2 = vmulq_u8(qx0, vdupq_n_u8(3));
-            uint8x16_t qx3 = vmulq_u8(qx1, vdupq_n_u8(3));
-            uint8x16_t qx4 = vmulq_u8(qx0, vdupq_n_u8(9));
-            uint8x16_t qx5 = vmulq_u8(qx1, vdupq_n_u8(9));
-            uint8x16_t qx6 = vmulq_u8(qx0, vdupq_n_u8(27));
-            uint8x16_t qx7 = vmulq_u8(qx1, vdupq_n_u8(27));
-            uint8x16_t qx8 = vmulq_u8(qx0, vdupq_n_u8(81));
-            uint8x16_t qx9 = vmulq_u8(qx1, vdupq_n_u8(81));
-
-            // multiply by 3 and keep the 2 bits above 8 bits
-            int8x16_t sqx0 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx0, vshrq_n_u8(qx0, 1)), 6));
-            int8x16_t sqx1 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx1, vshrq_n_u8(qx1, 1)), 6));
-            int8x16_t sqx2 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx2, vshrq_n_u8(qx2, 1)), 6));
-            int8x16_t sqx3 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx3, vshrq_n_u8(qx3, 1)), 6));
-            int8x16_t sqx4 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx4, vshrq_n_u8(qx4, 1)), 6));
-            int8x16_t sqx5 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx5, vshrq_n_u8(qx5, 1)), 6));
-            int8x16_t sqx6 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx6, vshrq_n_u8(qx6, 1)), 6));
-            int8x16_t sqx7 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx7, vshrq_n_u8(qx7, 1)), 6));
-            int8x16_t sqx8 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx8, vshrq_n_u8(qx8, 1)), 6));
-            int8x16_t sqx9 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx9, vshrq_n_u8(qx9, 1)), 6));
-
-            const int8x16_t qy0 = vld1q_s8(y[i].qs +   0);
-            const int8x16_t qy1 = vld1q_s8(y[i].qs +  16);
-            const int8x16_t qy2 = vld1q_s8(y[i].qs +  32);
-            const int8x16_t qy3 = vld1q_s8(y[i].qs +  48);
-            const int8x16_t qy4 = vld1q_s8(y[i].qs +  64);
-            const int8x16_t qy5 = vld1q_s8(y[i].qs +  80);
-            const int8x16_t qy6 = vld1q_s8(y[i].qs +  96);
-            const int8x16_t qy7 = vld1q_s8(y[i].qs + 112);
-            const int8x16_t qy8 = vld1q_s8(y[i].qs + 128);
-            const int8x16_t qy9 = vld1q_s8(y[i].qs + 144);
-
-#if defined(__ARM_FEATURE_DOTPROD)
-            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
-            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
-            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
-            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
-            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
-            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
-            sumi0 = vdotq_s32(sumi0, sqx6, qy6);
-            sumi1 = vdotq_s32(sumi1, sqx7, qy7);
-            sumi0 = vdotq_s32(sumi0, sqx8, qy8);
-            sumi1 = vdotq_s32(sumi1, sqx9, qy9);
-#else
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx6), vget_low_s8(qy6));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx6), vget_high_s8(qy6));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx7), vget_low_s8(qy7));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx7), vget_high_s8(qy7));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx8), vget_low_s8(qy8));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx8), vget_high_s8(qy8));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx9), vget_low_s8(qy9));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx9), vget_high_s8(qy9));
-#endif
-        }
-
-        // last 16 bytes of 5-element, along with the 4 bytes of 4 elements
-        {
-            uint8x16_t qx0 = vld1q_u8(x[i].qs + 32);
-            uint8x16_t qx1 = vmulq_u8(qx0, vdupq_n_u8(3));
-            uint8x16_t qx2 = vmulq_u8(qx0, vdupq_n_u8(9));
-            uint8x16_t qx3 = vmulq_u8(qx0, vdupq_n_u8(27));
-            uint8x16_t qx4 = vmulq_u8(qx0, vdupq_n_u8(81));
-            uint32_t qh;
-            memcpy(&qh, x[i].qh, sizeof(qh)); // potentially unaligned
-            uint8x16_t qx5 = vreinterpretq_u8_u32(vdupq_n_u32(qh));
-            qx5 = vmulq_u8(qx5, shift);
-
-            // multiply by 3 and keep the 2 bits above 8 bits
-            int8x16_t sqx0 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx0, vshrq_n_u8(qx0, 1)), 6));
-            int8x16_t sqx1 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx1, vshrq_n_u8(qx1, 1)), 6));
-            int8x16_t sqx2 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx2, vshrq_n_u8(qx2, 1)), 6));
-            int8x16_t sqx3 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx3, vshrq_n_u8(qx3, 1)), 6));
-            int8x16_t sqx4 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx4, vshrq_n_u8(qx4, 1)), 6));
-            int8x16_t sqx5 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx5, vshrq_n_u8(qx5, 1)), 6));
-
-            const int8x16_t qy0 = vld1q_s8(y[i].qs + 160);
-            const int8x16_t qy1 = vld1q_s8(y[i].qs + 176);
-            const int8x16_t qy2 = vld1q_s8(y[i].qs + 192);
-            const int8x16_t qy3 = vld1q_s8(y[i].qs + 208);
-            const int8x16_t qy4 = vld1q_s8(y[i].qs + 224);
-            const int8x16_t qy5 = vld1q_s8(y[i].qs + 240);
-
-#if defined(__ARM_FEATURE_DOTPROD)
-            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
-            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
-            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
-            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
-            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
-            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
-#else
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
-#endif
-        }
-
-        const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
-        const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
-
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-
-#if defined(__ARM_FEATURE_DOTPROD)
-        sumi0 = vaddq_s32(sumi0, sumi1);
-        sumi0 = vsubq_s32(sumi0, vpaddlq_s16(vaddq_s16(ysum0, ysum1)));
-
-        sumf += d * (float) vaddvq_s32(sumi0);
-#else
-        sumi0 = vaddq_s16(sumi0, sumi1);
-        sumi0 = vsubq_s16(sumi0, vaddq_s16(ysum0, ysum1));
-
-        sumf += d * (float) vaddlvq_s16(sumi0);
-#endif
-    }
-
-    *s = sumf;
-
-#elif defined(__AVX2__)
-    __m256 sumf = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-        // 16-bit sums
-        __m256i sumi0 = _mm256_setzero_si256();
-        __m256i sumi1 = _mm256_setzero_si256();
-        __m256i sumi2 = _mm256_setzero_si256();
-
-        // first 32 bytes of 5 elements
-        {
-            __m256i qx0 = _mm256_loadu_si256((const __m256i *) (x[i].qs));
-            // 8-bit multiplies with shifts, masks and adds
-            __m256i qx1 = _mm256_add_epi8(qx0, _mm256_add_epi8(qx0, qx0)); // 1 * 3
-            __m256i qx2 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx0, 3), _mm256_set1_epi8(-8)), qx0); // 1 * 9
-            __m256i qx3 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx1, 3), _mm256_set1_epi8(-8)), qx1); // 3 * 9
-            __m256i qx4 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx2, 3), _mm256_set1_epi8(-8)), qx2); // 9 * 9
-
-            // TODO: can _mm256_mulhi_epu16 be faster even if 16-bits?
-
-            // Cancel the +1 from avg so that it behaves like a halving add
-            qx0 = _mm256_subs_epu8(qx0, _mm256_set1_epi8(1));
-            qx1 = _mm256_subs_epu8(qx1, _mm256_set1_epi8(1));
-            qx2 = _mm256_subs_epu8(qx2, _mm256_set1_epi8(1));
-            qx3 = _mm256_subs_epu8(qx3, _mm256_set1_epi8(1));
-            qx4 = _mm256_subs_epu8(qx4, _mm256_set1_epi8(1));
-            // Multiply by 3 and get the top 2 bits
-            qx0 = _mm256_avg_epu8(qx0, _mm256_avg_epu8(qx0, _mm256_setzero_si256()));
-            qx1 = _mm256_avg_epu8(qx1, _mm256_avg_epu8(qx1, _mm256_setzero_si256()));
-            qx2 = _mm256_avg_epu8(qx2, _mm256_avg_epu8(qx2, _mm256_setzero_si256()));
-            qx3 = _mm256_avg_epu8(qx3, _mm256_avg_epu8(qx3, _mm256_setzero_si256()));
-            qx4 = _mm256_avg_epu8(qx4, _mm256_avg_epu8(qx4, _mm256_setzero_si256()));
-            qx0 = _mm256_and_si256(_mm256_srli_epi16(qx0, 6), _mm256_set1_epi8(3));
-            qx1 = _mm256_and_si256(_mm256_srli_epi16(qx1, 6), _mm256_set1_epi8(3));
-            qx2 = _mm256_and_si256(_mm256_srli_epi16(qx2, 6), _mm256_set1_epi8(3));
-            qx3 = _mm256_and_si256(_mm256_srli_epi16(qx3, 6), _mm256_set1_epi8(3));
-            qx4 = _mm256_and_si256(_mm256_srli_epi16(qx4, 6), _mm256_set1_epi8(3));
-
-            const __m256i qy0 = _mm256_loadu_si256((const __m256i *) (y[i].qs +   0));
-            const __m256i qy1 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  32));
-            const __m256i qy2 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  64));
-            const __m256i qy3 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  96));
-            const __m256i qy4 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 128));
-
-            qx0 = _mm256_maddubs_epi16(qx0, qy0);
-            qx1 = _mm256_maddubs_epi16(qx1, qy1);
-            qx2 = _mm256_maddubs_epi16(qx2, qy2);
-            qx3 = _mm256_maddubs_epi16(qx3, qy3);
-            qx4 = _mm256_maddubs_epi16(qx4, qy4);
-
-            sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(qx0, qx1));
-            sumi1 = _mm256_add_epi16(sumi1, _mm256_add_epi16(qx2, qx3));
-            sumi2 = _mm256_add_epi16(sumi2, qx4);
-        }
-
-        // last 16 bytes of 5-element, along with the 4 bytes of 4 elements
-        {
-            __m128i qx0 = _mm_loadu_si128((const __m128i *) (x[i].qs + 32));
-            uint32_t qh;
-            memcpy(&qh, x[i].qh, sizeof(qh)); // potentially unaligned
-            __m256i qx5_l = _mm256_cvtepu8_epi16(_mm_set1_epi32(qh));
-            __m128i qx1 = _mm_add_epi8(qx0, _mm_add_epi8(qx0, qx0)); // 1 * 3
-            __m128i qx2 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx0, 3), _mm_set1_epi8(-8)), qx0); // 1 * 9
-            __m128i qx3 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx1, 3), _mm_set1_epi8(-8)), qx1); // 3 * 9
-            __m128i qx4 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx2, 3), _mm_set1_epi8(-8)), qx2); // 9 * 9
-            __m256i qx01 = MM256_SET_M128I(qx1, qx0);
-            __m256i qx23 = MM256_SET_M128I(qx3, qx2);
-
-            // avx2 does not have 8-bit multiplies, so 16-bit it is.
-            qx5_l = _mm256_mullo_epi16(qx5_l, _mm256_set_epi16(27, 27, 27, 27, 9, 9, 9, 9, 3, 3, 3, 3, 1, 1, 1, 1));
-            qx5_l = _mm256_and_si256(qx5_l, _mm256_set1_epi16(0xFF));
-            __m128i qx5 = _mm_packus_epi16(_mm256_castsi256_si128(qx5_l), _mm256_extracti128_si256(qx5_l, 1));
-
-            __m256i qx45 = MM256_SET_M128I(qx5, qx4);
-
-            // Cancel the +1 from avg so that it behaves like a halving add
-            qx01 = _mm256_subs_epu8(qx01, _mm256_set1_epi8(1));
-            qx23 = _mm256_subs_epu8(qx23, _mm256_set1_epi8(1));
-            qx45 = _mm256_subs_epu8(qx45, _mm256_set1_epi8(1));
-            // Multiply by 3 and get the top 2 bits
-            qx01 = _mm256_avg_epu8(qx01, _mm256_avg_epu8(qx01, _mm256_setzero_si256()));
-            qx23 = _mm256_avg_epu8(qx23, _mm256_avg_epu8(qx23, _mm256_setzero_si256()));
-            qx45 = _mm256_avg_epu8(qx45, _mm256_avg_epu8(qx45, _mm256_setzero_si256()));
-            qx01 = _mm256_and_si256(_mm256_srli_epi16(qx01, 6), _mm256_set1_epi8(3));
-            qx23 = _mm256_and_si256(_mm256_srli_epi16(qx23, 6), _mm256_set1_epi8(3));
-            qx45 = _mm256_and_si256(_mm256_srli_epi16(qx45, 6), _mm256_set1_epi8(3));
-
-            const __m256i qy01 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 160));
-            const __m256i qy23 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 192));
-            const __m256i qy45 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 224));
-
-            qx01 = _mm256_maddubs_epi16(qx01, qy01);
-            qx23 = _mm256_maddubs_epi16(qx23, qy23);
-            qx45 = _mm256_maddubs_epi16(qx45, qy45);
-
-            sumi0 = _mm256_add_epi16(sumi0, qx01);
-            sumi1 = _mm256_add_epi16(sumi1, qx23);
-            sumi2 = _mm256_add_epi16(sumi2, qx45);
-        }
-
-        const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
-
-        sumi0 = _mm256_sub_epi16(sumi0, ysum);
-        sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(sumi1, sumi2));
-        sumi0 = _mm256_madd_epi16(sumi0, _mm256_set1_epi16(1));
-
-        sumf = _mm256_add_ps(_mm256_mul_ps(_mm256_cvtepi32_ps(sumi0), d), sumf);
-    }
-
-    *s = hsum_float_8(sumf);
-
-#else
-    const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
-
-    float sumf = 0.0f;
-
-    for (int i = 0; i < nb; ++i) {
-        int sum = 0;
-
-        for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
-            for (size_t l = 0; l < 5; ++l) {
-                for (size_t m = 0; m < 32; ++m) {
-                    uint8_t q = x[i].qs[j + m] * pow3[l];
-                    uint16_t xi = ((uint16_t) q * 3) >> 8;
-                    sum += (xi - 1) * y[i].qs[j*5 + l*32 + m];
-                }
-            }
-        }
-        for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
-            for (size_t l = 0; l < 5; ++l) {
-                for (size_t m = 0; m < 16; ++m) {
-                    uint8_t q = x[i].qs[j + m] * pow3[l];
-                    uint16_t xi = ((uint16_t) q * 3) >> 8;
-                    sum += (xi - 1) * y[i].qs[j*5 + l*16 + m];
-                }
-            }
-        }
-
-        for (size_t l = 0; l < 4; ++l) {
-            for (size_t j = 0; j < sizeof(x->qh); ++j) {
-                uint8_t q = x[i].qh[j] * pow3[l];
-                uint16_t xi = ((uint16_t) q * 3) >> 8;
-                sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j];
-            }
-        }
-
-        sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d);
-    }
-
-    *s = sumf;
-#endif
-}
-
-void ggml_vec_dot_tq2_0_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_tq2_0 * restrict x = vx;
-    const block_q8_K  * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_NEON)
-    float sumf = 0.0f;
-
-    const uint8x16_t m3 = vdupq_n_u8(3);
-
-    for (int i = 0; i < nb; ++i) {
-#if defined(__ARM_FEATURE_DOTPROD)
-        int32x4_t sumi0 = vdupq_n_s32(0);
-        int32x4_t sumi1 = vdupq_n_s32(0);
-#else
-        int16x8_t sumi0 = vdupq_n_s16(0);
-        int16x8_t sumi1 = vdupq_n_s16(0);
-#endif
-
-        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
-            uint8x16_t qx0 = vld1q_u8(x[i].qs + j);
-            uint8x16_t qx1 = vld1q_u8(x[i].qs + j + 16);
-            uint8x16_t qx2 = vshrq_n_u8(qx0, 2);
-            uint8x16_t qx3 = vshrq_n_u8(qx1, 2);
-            uint8x16_t qx4 = vshrq_n_u8(qx0, 4);
-            uint8x16_t qx5 = vshrq_n_u8(qx1, 4);
-            uint8x16_t qx6 = vshrq_n_u8(qx0, 6);
-            uint8x16_t qx7 = vshrq_n_u8(qx1, 6);
-
-            int8x16_t sqx0 = vreinterpretq_s8_u8(vandq_u8(qx0, m3));
-            int8x16_t sqx1 = vreinterpretq_s8_u8(vandq_u8(qx1, m3));
-            int8x16_t sqx2 = vreinterpretq_s8_u8(vandq_u8(qx2, m3));
-            int8x16_t sqx3 = vreinterpretq_s8_u8(vandq_u8(qx3, m3));
-            int8x16_t sqx4 = vreinterpretq_s8_u8(vandq_u8(qx4, m3));
-            int8x16_t sqx5 = vreinterpretq_s8_u8(vandq_u8(qx5, m3));
-            int8x16_t sqx6 = vreinterpretq_s8_u8(vandq_u8(qx6, m3));
-            int8x16_t sqx7 = vreinterpretq_s8_u8(vandq_u8(qx7, m3));
-
-            const int8x16_t qy0 = vld1q_s8(y[i].qs + j*4 +   0);
-            const int8x16_t qy1 = vld1q_s8(y[i].qs + j*4 +  16);
-            const int8x16_t qy2 = vld1q_s8(y[i].qs + j*4 +  32);
-            const int8x16_t qy3 = vld1q_s8(y[i].qs + j*4 +  48);
-            const int8x16_t qy4 = vld1q_s8(y[i].qs + j*4 +  64);
-            const int8x16_t qy5 = vld1q_s8(y[i].qs + j*4 +  80);
-            const int8x16_t qy6 = vld1q_s8(y[i].qs + j*4 +  96);
-            const int8x16_t qy7 = vld1q_s8(y[i].qs + j*4 + 112);
-
-#if defined(__ARM_FEATURE_DOTPROD)
-            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
-            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
-            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
-            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
-            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
-            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
-            sumi0 = vdotq_s32(sumi0, sqx6, qy6);
-            sumi1 = vdotq_s32(sumi1, sqx7, qy7);
-#else
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx6), vget_low_s8(qy6));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx6), vget_high_s8(qy6));
-            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx7), vget_low_s8(qy7));
-            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx7), vget_high_s8(qy7));
-#endif
-        }
-
-        const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
-        const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
-
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-
-#if defined(__ARM_FEATURE_DOTPROD)
-        sumi0 = vaddq_s32(sumi0, sumi1);
-        sumi0 = vsubq_s32(sumi0, vpaddlq_s16(vaddq_s16(ysum0, ysum1)));
-
-        sumf += d * (float) vaddvq_s32(sumi0);
-#else
-        sumi0 = vaddq_s16(sumi0, sumi1);
-        sumi0 = vsubq_s16(sumi0, vaddq_s16(ysum0, ysum1));
-
-        sumf += d * (float) vaddlvq_s16(sumi0);
-#endif
-    }
-
-    *s = sumf;
-
-#elif defined(__AVX2__)
-    __m256 sumf = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-        // 16-bit sums, because 256*127 still fits
-        __m256i sumi0 = _mm256_setzero_si256();
-        __m256i sumi1 = _mm256_setzero_si256();
-
-        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
-            __m256i qx0 = _mm256_loadu_si256((const __m256i *) (x[i].qs + j));
-            __m256i qx1 = _mm256_srli_epi16(qx0, 2);
-            __m256i qx2 = _mm256_srli_epi16(qx0, 4);
-            __m256i qx3 = _mm256_srli_epi16(qx0, 6);
-
-            // 0, 1, 2 (should not be 3)
-            qx0 = _mm256_and_si256(qx0, _mm256_set1_epi8(3));
-            qx1 = _mm256_and_si256(qx1, _mm256_set1_epi8(3));
-            qx2 = _mm256_and_si256(qx2, _mm256_set1_epi8(3));
-            qx3 = _mm256_and_si256(qx3, _mm256_set1_epi8(3));
-
-            const __m256i qy0 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 +  0));
-            const __m256i qy1 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 32));
-            const __m256i qy2 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 64));
-            const __m256i qy3 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 96));
-
-            qx0 = _mm256_maddubs_epi16(qx0, qy0);
-            qx1 = _mm256_maddubs_epi16(qx1, qy1);
-            qx2 = _mm256_maddubs_epi16(qx2, qy2);
-            qx3 = _mm256_maddubs_epi16(qx3, qy3);
-
-            sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(qx0, qx1));
-            sumi1 = _mm256_add_epi16(sumi1, _mm256_add_epi16(qx2, qx3));
-        }
-
-        const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
-
-        sumi0 = _mm256_add_epi16(sumi0, sumi1);
-        sumi0 = _mm256_sub_epi16(sumi0, ysum);
-        sumi0 = _mm256_madd_epi16(sumi0, _mm256_set1_epi16(1));
-
-        sumf = _mm256_add_ps(_mm256_mul_ps(_mm256_cvtepi32_ps(sumi0), d), sumf);
-    }
-
-    *s = hsum_float_8(sumf);
-
-#else
-    float sumf = 0.0f;
-
-    for (int i = 0; i < nb; ++i) {
-        int32_t sumi = 0;
-
-        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
-            for (size_t l = 0; l < 4; ++l) {
-                for (size_t k = 0; k < 32; ++k) {
-                    sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1);
-                }
-            }
-        }
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        sumf += (float) sumi * d;
-    }
-
-    *s = sumf;
-#endif
-}
-
-void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_q2_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#ifdef __ARM_NEON
-    const uint8x16_t m3 = vdupq_n_u8(0x3);
-    const uint8x16_t m4 = vdupq_n_u8(0xF);
-
-    const int32x4_t vzero = vdupq_n_s32(0);
-
-    ggml_int8x16x2_t q2bytes;
-    uint8_t aux[16];
-
-    float sum = 0;
-
-    for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-        const uint8_t * restrict sc = x[i].scales;
-
-        const uint8x16_t mins_and_scales = vld1q_u8(sc);
-        const uint8x16_t scales = vandq_u8(mins_and_scales, m4);
-        vst1q_u8(aux, scales);
-
-        const uint8x16_t mins = vshrq_n_u8(mins_and_scales, 4);
-        const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
-        const ggml_int16x8x2_t mins16 = {{vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}};
-        const int32x4_t s0 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[0]), vget_low_s16 (q8sums.val[0])),
-                                       vmull_s16(vget_high_s16(mins16.val[0]), vget_high_s16(q8sums.val[0])));
-        const int32x4_t s1 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[1]), vget_low_s16 (q8sums.val[1])),
-                                       vmull_s16(vget_high_s16(mins16.val[1]), vget_high_s16(q8sums.val[1])));
-        sum += dmin * vaddvq_s32(vaddq_s32(s0, s1));
-
-        int isum = 0;
-        int is = 0;
-
-// We use this macro instead of a function call because for some reason
-// the code runs 2-3% slower, even if the function is declared inline
-#define MULTIPLY_ACCUM_WITH_SCALE(index)\
-        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * aux[is+(index)];\
-        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * aux[is+1+(index)];
-
-#define SHIFT_MULTIPLY_ACCUM_WITH_SCALE(shift, index)\
-        q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;\
-        q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[0], (shift)), m3));\
-        q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[1], (shift)), m3));\
-        MULTIPLY_ACCUM_WITH_SCALE((index));
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            const ggml_uint8x16x2_t q2bits = ggml_vld1q_u8_x2(q2); q2 += 32;
-
-            ggml_int8x16x2_t q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
-            q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[0], m3));
-            q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[1], m3));
-
-            MULTIPLY_ACCUM_WITH_SCALE(0);
-
-            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(2, 2);
-            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(4, 4);
-            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(6, 6);
-
-            is += 8;
-        }
-
-        sum += d * isum;
-    }
-
-    *s = sum;
-
-#elif defined __AVX2__
-
-    const __m256i m3 = _mm256_set1_epi8(3);
-    const __m128i m4 = _mm_set1_epi8(0xF);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
-        const __m128i scales8 = _mm_and_si128(mins_and_scales, m4);
-        const __m128i mins8 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
-        const __m256i mins = _mm256_cvtepi8_epi16(mins8);
-        const __m256i prod = _mm256_madd_epi16(mins, _mm256_loadu_si256((const __m256i*)y[i].bsums));
-
-        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(prod), acc);
-
-        const __m256i all_scales = _mm256_cvtepi8_epi16(scales8);
-        const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
-        const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
-        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
-
-        __m256i sumi = _mm256_setzero_si256();
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const __m256i q2bits = _mm256_loadu_si256((const __m256i*)q2); q2 += 32;
-
-            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            const __m256i q2_0 = _mm256_and_si256(q2bits, m3);
-            const __m256i q2_1 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 2), m3);
-            const __m256i q2_2 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 4), m3);
-            const __m256i q2_3 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 6), m3);
-
-            __m256i p0 = _mm256_maddubs_epi16(q2_0, q8_0);
-            __m256i p1 = _mm256_maddubs_epi16(q2_1, q8_1);
-            __m256i p2 = _mm256_maddubs_epi16(q2_2, q8_2);
-            __m256i p3 = _mm256_maddubs_epi16(q2_3, q8_3);
-
-            p0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(0)), p0);
-            p1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(1)), p1);
-            p2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(2)), p2);
-            p3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(3)), p3);
-
-            p0 = _mm256_add_epi32(p0, p1);
-            p2 = _mm256_add_epi32(p2, p3);
-
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p0, p2));
-        }
-
-        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
-
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __AVX__
-
-    const __m128i m3 = _mm_set1_epi8(0x3);
-    const __m128i m4 = _mm_set1_epi8(0xF);
-    const __m128i m2 = _mm_set1_epi8(0x2);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        // load mins and scales from block_q2_K.scales[QK_K/16]
-        const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
-        const __m128i scales16 = _mm_and_si128(mins_and_scales, m4);
-        const __m128i mins16 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
-        const __m128i mins_0 = _mm_cvtepi8_epi16(mins16);
-        const __m128i mins_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(mins16, mins16));
-
-        // summs = y[i].bsums * (x[i].scales >> 4) in 16bits*8*2 to 32bits*4*2
-        const __m128i summs_0 = _mm_madd_epi16(mins_0, _mm_loadu_si128((const __m128i*)&y[i].bsums[0]));
-        const __m128i summs_1 = _mm_madd_epi16(mins_1, _mm_loadu_si128((const __m128i*)&y[i].bsums[8]));
-
-        // sumf += -dmin * summs in 32bits*8
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(MM256_SET_M128I(summs_1, summs_0))), acc);
-
-        const __m128i scales_0 = _mm_cvtepi8_epi16(scales16);
-        const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales16, scales16));
-        const __m128i scales[2] = { scales_0, scales_1 };
-
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            // load Q8 quants int8*16*8 from block_q8_K.qs[QK_K]
-            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-
-            // load 2bits*16*8 from block_q2_K.qs[QK_K/4]
-            __m128i q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
-            const __m128i q2_0 = _mm_and_si128(q2bits, m3);
-            const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
-            const __m128i q2_4 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
-            const __m128i q2_6 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
-            q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
-            const __m128i q2_1 = _mm_and_si128(q2bits, m3);
-            const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
-            const __m128i q2_5 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
-            const __m128i q2_7 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
-
-            // isuml = q8[l] * ((q2[l] >> shift) & 3) in 8bits*16*8 to 16bits*8*8
-            __m128i p0 = _mm_maddubs_epi16(q2_0, q8_0);
-            __m128i p1 = _mm_maddubs_epi16(q2_1, q8_1);
-            __m128i p2 = _mm_maddubs_epi16(q2_2, q8_2);
-            __m128i p3 = _mm_maddubs_epi16(q2_3, q8_3);
-            __m128i p4 = _mm_maddubs_epi16(q2_4, q8_4);
-            __m128i p5 = _mm_maddubs_epi16(q2_5, q8_5);
-            __m128i p6 = _mm_maddubs_epi16(q2_6, q8_6);
-            __m128i p7 = _mm_maddubs_epi16(q2_7, q8_7);
-
-            // isum += (x[i].scales[is++] & 0xF) * isuml in 16bits*8*8 to 32bits*4*8
-            __m128i shuffle = _mm_set1_epi16(0x0100);
-            p0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p0);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p1);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p2);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p3);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p4);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p5);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p6);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p7);
-
-            p0 = _mm_add_epi32(p0, p1);
-            p2 = _mm_add_epi32(p2, p3);
-            p4 = _mm_add_epi32(p4, p5);
-            p6 = _mm_add_epi32(p6, p7);
-
-            // isum in 32bits*4*2
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p0, p2));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p4, p6));
-        }
-
-        // sumf += dall * isum - dmin * summs in 32bits
-        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dall), _mm256_cvtepi32_ps(sumi)), acc);
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __riscv_v_intrinsic
-
-    float sumf = 0;
-    uint8_t temp_01[32] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-                            1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1};
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        size_t vl = 16;
-
-        vuint8m1_t scales = __riscv_vle8_v_u8m1(sc, vl);
-        vuint8m1_t aux = __riscv_vand_vx_u8m1(scales, 0x0F, vl);
-
-        vint16m1_t q8sums = __riscv_vle16_v_i16m1(y[i].bsums, vl);
-
-        vuint8mf2_t scales_2 = __riscv_vle8_v_u8mf2(sc, vl);
-        vuint8mf2_t mins8 = __riscv_vsrl_vx_u8mf2(scales_2, 0x4, vl);
-        vint16m1_t mins = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(mins8, vl));
-        vint32m2_t prod = __riscv_vwmul_vv_i32m2(q8sums, mins, vl);
-        vint32m1_t vsums = __riscv_vredsum_vs_i32m2_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
-
-        sumf  += dmin * __riscv_vmv_x_s_i32m1_i32(vsums);
-
-        vl = 32;
-
-        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
-        vuint8m1_t v_b = __riscv_vle8_v_u8m1(temp_01, vl);
-
-        uint8_t is=0;
-        int isum=0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            // load Q2
-            vuint8m1_t q2_x = __riscv_vle8_v_u8m1(q2, vl);
-
-            vuint8m1_t q2_0 = __riscv_vand_vx_u8m1(q2_x, 0x03, vl);
-            vuint8m1_t q2_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x2, vl), 0x03 , vl);
-            vuint8m1_t q2_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x4, vl), 0x03 , vl);
-            vuint8m1_t q2_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x6, vl), 0x03 , vl);
-
-            // duplicate scale elements for product
-            vuint8m1_t sc0 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 0+is, vl), vl);
-            vuint8m1_t sc1 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 2+is, vl), vl);
-            vuint8m1_t sc2 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 4+is, vl), vl);
-            vuint8m1_t sc3 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 6+is, vl), vl);
-
-            vint16m2_t p0 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_0, sc0, vl));
-            vint16m2_t p1 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_1, sc1, vl));
-            vint16m2_t p2 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_2, sc2, vl));
-            vint16m2_t p3 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_3, sc3, vl));
-
-            // load Q8
-            vint8m1_t q8_0 = __riscv_vle8_v_i8m1(q8, vl);
-            vint8m1_t q8_1 = __riscv_vle8_v_i8m1(q8+32, vl);
-            vint8m1_t q8_2 = __riscv_vle8_v_i8m1(q8+64, vl);
-            vint8m1_t q8_3 = __riscv_vle8_v_i8m1(q8+96, vl);
-
-            vint32m4_t s0 = __riscv_vwmul_vv_i32m4(p0, __riscv_vwcvt_x_x_v_i16m2(q8_0, vl), vl);
-            vint32m4_t s1 = __riscv_vwmul_vv_i32m4(p1, __riscv_vwcvt_x_x_v_i16m2(q8_1, vl), vl);
-            vint32m4_t s2 = __riscv_vwmul_vv_i32m4(p2, __riscv_vwcvt_x_x_v_i16m2(q8_2, vl), vl);
-            vint32m4_t s3 = __riscv_vwmul_vv_i32m4(p3, __riscv_vwcvt_x_x_v_i16m2(q8_3, vl), vl);
-
-            vint32m1_t isum0 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s0, s1, vl), vzero, vl);
-            vint32m1_t isum1 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s2, s3, vl), isum0, vl);
-
-            isum += __riscv_vmv_x_s_i32m1_i32(isum1);
-
-            q2+=32;  q8+=128;  is=8;
-
-        }
-
-        sumf += dall * isum;
-
-    }
-
-    *s = sumf;
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0x3);
-    const vector signed char lowScaleMask = vec_splats((signed char)0xF);
-    const vector int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
-    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
-        vector float vdmin = vec_mul(vxmin, vyd);
-
-        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
-        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
-
-        vector signed char q2xmins = (vector signed char)vec_xl( 0, x[i].scales);
-        vector signed char vscales = vec_and(q2xmins, lowScaleMask);
-
-        q2xmins = vec_sr(q2xmins, v4);
-        vector signed short q2xmins0 = vec_unpackh(q2xmins);
-        vector signed short q2xmins1 = vec_unpackl(q2xmins);
-
-        vector signed int prod0 = vec_mule(q2xmins0, q8ysums0);
-        vector signed int prod1 = vec_mulo(q2xmins0, q8ysums0);
-        vector signed int prod2 = vec_mule(q2xmins1, q8ysums1);
-        vector signed int prod3 = vec_mulo(q2xmins1, q8ysums1);
-
-        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
-        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
-        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
-        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-        vector signed int vsumi4 = v0;
-        vector signed int vsumi5 = v0;
-        vector signed int vsumi6 = v0;
-        vector signed int vsumi7 = v0;
-
-        const uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            __builtin_prefetch(q2, 0, 1);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed char qxs0 = (vector signed char)vec_xl( 0, q2);
-            vector signed char qxs1 = (vector signed char)vec_xl(16, q2);
-            q2 += 32;
-
-            vector unsigned char q2x00 = (vector unsigned char)vec_and(qxs0, lowMask);
-            vector unsigned char q2x01 = (vector unsigned char)vec_and(vec_sr(qxs0, v2), lowMask);
-            vector unsigned char q2x02 = (vector unsigned char)vec_and(vec_sr(qxs0, v4), lowMask);
-            vector unsigned char q2x03 = (vector unsigned char)vec_and(vec_sr(qxs0, v6), lowMask);
-            vector unsigned char q2x10 = (vector unsigned char)vec_and(qxs1, lowMask);
-            vector unsigned char q2x11 = (vector unsigned char)vec_and(vec_sr(qxs1, v2), lowMask);
-            vector unsigned char q2x12 = (vector unsigned char)vec_and(vec_sr(qxs1, v4), lowMask);
-            vector unsigned char q2x13 = (vector unsigned char)vec_and(vec_sr(qxs1, v6), lowMask);
-
-            vector signed char q8y00 = vec_xl(  0, q8);
-            vector signed char q8y10 = vec_xl( 16, q8);
-            vector signed char q8y01 = vec_xl( 32, q8);
-            vector signed char q8y11 = vec_xl( 48, q8);
-            vector signed char q8y02 = vec_xl( 64, q8);
-            vector signed char q8y12 = vec_xl( 80, q8);
-            vector signed char q8y03 = vec_xl( 96, q8);
-            vector signed char q8y13 = vec_xl(112, q8);
-            q8 += 128;
-
-            vector signed int qv0 = vec_msum(q8y00, q2x00, v0);
-            vector signed int qv1 = vec_msum(q8y01, q2x01, v0);
-            vector signed int qv2 = vec_msum(q8y02, q2x02, v0);
-            vector signed int qv3 = vec_msum(q8y03, q2x03, v0);
-            vector signed int qv4 = vec_msum(q8y10, q2x10, v0);
-            vector signed int qv5 = vec_msum(q8y11, q2x11, v0);
-            vector signed int qv6 = vec_msum(q8y12, q2x12, v0);
-            vector signed int qv7 = vec_msum(q8y13, q2x13, v0);
-
-            vector signed short vscales_07 = vec_unpackh(vscales);
-            vector signed int vscales_03 = vec_unpackh(vscales_07);
-            vector signed int vscales_47 = vec_unpackl(vscales_07);
-            vector signed int vs0 = vec_splat(vscales_03, 0);
-            vector signed int vs1 = vec_splat(vscales_03, 1);
-            vector signed int vs2 = vec_splat(vscales_03, 2);
-            vector signed int vs3 = vec_splat(vscales_03, 3);
-            vector signed int vs4 = vec_splat(vscales_47, 0);
-            vector signed int vs5 = vec_splat(vscales_47, 1);
-            vector signed int vs6 = vec_splat(vscales_47, 2);
-            vector signed int vs7 = vec_splat(vscales_47, 3);
-            vscales = vec_sld(vscales, vscales, 8);
-
-            vsumi0 = vec_add(vec_mul(qv0, vs0), vsumi0);
-            vsumi1 = vec_add(vec_mul(qv1, vs2), vsumi1);
-            vsumi2 = vec_add(vec_mul(qv2, vs4), vsumi2);
-            vsumi3 = vec_add(vec_mul(qv3, vs6), vsumi3);
-            vsumi4 = vec_add(vec_mul(qv4, vs1), vsumi4);
-            vsumi5 = vec_add(vec_mul(qv5, vs3), vsumi5);
-            vsumi6 = vec_add(vec_mul(qv6, vs5), vsumi6);
-            vsumi7 = vec_add(vec_mul(qv7, vs7), vsumi7);
-        }
-
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined __loongarch_asx
-
-    const __m256i m3 = __lasx_xvreplgr2vr_b(3);
-    const __m128i m4 = __lsx_vreplgr2vr_b(0xF);
-
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const __m128i mins_and_scales = __lsx_vld((const __m128i*)x[i].scales, 0);
-        const __m128i scales8 = __lsx_vand_v(mins_and_scales, m4);
-        const __m128i mins8 = __lsx_vand_v(__lsx_vsrli_h(mins_and_scales, 4), m4);
-        const __m256i mins = lasx_ext8_16(mins8);
-        const __m256i prod = lasx_madd_h(mins, __lasx_xvld((const __m256i*)y[i].bsums, 0));
-
-        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(dmin), __lasx_xvffint_s_w(prod), acc);
-
-        const __m256i all_scales = lasx_ext8_16(scales8);
-        const __m128i l_scales = lasx_extracti128(all_scales, 0);
-        const __m128i h_scales = lasx_extracti128(all_scales, 1);
-        const __m256i scales[2] = {lasx_insertf128(l_scales, l_scales), lasx_insertf128(h_scales, h_scales)};
-
-        __m256i sumi = __lasx_xvldi(0);
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const __m256i q2bits = __lasx_xvld((const __m256i*)q2, 0); q2 += 32;
-
-            const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-
-            const __m256i q2_0 = __lasx_xvand_v(q2bits, m3);
-            const __m256i q2_1 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 2), m3);
-            const __m256i q2_2 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 4), m3);
-            const __m256i q2_3 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 6), m3);
-
-            __m256i p0 = lasx_maddubs_h(q2_0, q8_0);
-            __m256i p1 = lasx_maddubs_h(q2_1, q8_1);
-            __m256i p2 = lasx_maddubs_h(q2_2, q8_2);
-            __m256i p3 = lasx_maddubs_h(q2_3, q8_3);
-
-            p0 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(0)), p0);
-            p1 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(1)), p1);
-            p2 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(2)), p2);
-            p3 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(3)), p3);
-
-            p0 = __lasx_xvadd_w(p0, p1);
-            p2 = __lasx_xvadd_w(p2, p3);
-
-            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p0, p2));
-        }
-
-        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);
-
-    }
-
-    *s = hsum_float_8(acc);
-
-#else
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
-#endif
-}
-
-void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(n % QK_K == 0);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const uint32_t kmask1 = 0x03030303;
-    const uint32_t kmask2 = 0x0f0f0f0f;
-
-    const block_q3_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#ifdef __ARM_NEON
-
-    uint32_t aux[3];
-    uint32_t utmp[4];
-
-    const uint8x16_t m3b = vdupq_n_u8(0x3);
-    const int32x4_t  vzero = vdupq_n_s32(0);
-
-    const uint8x16_t m0 = vdupq_n_u8(1);
-    const uint8x16_t m1 = vshlq_n_u8(m0, 1);
-    const uint8x16_t m2 = vshlq_n_u8(m0, 2);
-    const uint8x16_t m3 = vshlq_n_u8(m0, 3);
-    const int8_t m32 = 32;
-
-    ggml_int8x16x4_t q3bytes;
-
-    float sum = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict qh = x[i].hmask;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh);
-
-        ggml_uint8x16x4_t q3h;
-
-        int32_t isum = 0;
-
-        // Set up scales
-        memcpy(aux, x[i].scales, 12);
-        utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
-        utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
-        utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
-        utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
-
-        int8_t * scale = (int8_t *)utmp;
-        for (int j = 0; j < 16; ++j) scale[j] -= m32;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const ggml_uint8x16x2_t q3bits = ggml_vld1q_u8_x2(q3); q3 += 32;
-            const ggml_int8x16x4_t q8bytes_1 = ggml_vld1q_s8_x4(q8); q8 += 64;
-            const ggml_int8x16x4_t q8bytes_2 = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            q3h.val[0] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[0]), 2);
-            q3h.val[1] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[1]), 2);
-            q3h.val[2] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[0]), 1);
-            q3h.val[3] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[1]), 1);
-
-            q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[0], m3b)), vreinterpretq_s8_u8(q3h.val[0]));
-            q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[1], m3b)), vreinterpretq_s8_u8(q3h.val[1]));
-            q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 2), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
-            q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 2), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
-
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes_1.val[0])) * scale[0];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes_1.val[1])) * scale[1];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes_1.val[2])) * scale[2];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes_1.val[3])) * scale[3];
-
-            scale += 4;
-
-            q3h.val[0] = vbicq_u8(m2, qhbits.val[0]);
-            q3h.val[1] = vbicq_u8(m2, qhbits.val[1]);
-            q3h.val[2] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[0]), 1);
-            q3h.val[3] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[1]), 1);
-
-            q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 4), m3b)), vreinterpretq_s8_u8(q3h.val[0]));
-            q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 4), m3b)), vreinterpretq_s8_u8(q3h.val[1]));
-            q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 6), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
-            q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 6), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
-
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes_2.val[0])) * scale[0];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes_2.val[1])) * scale[1];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes_2.val[2])) * scale[2];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes_2.val[3])) * scale[3];
-
-            scale += 4;
-
-            if (j == 0) {
-                qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 4);
-                qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 4);
-            }
-
-        }
-        sum += d * isum;
-
-    }
-
-    *s = sum;
-
-#elif defined __AVX2__
-
-    const __m256i m3 = _mm256_set1_epi8(3);
-    const __m256i mone = _mm256_set1_epi8(1);
-    const __m128i m32 = _mm_set1_epi8(32);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    uint32_t aux[3];
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q3 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        // Set up scales
-        memcpy(aux, x[i].scales, 12);
-        __m128i scales128 = _mm_set_epi32(
-                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
-                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
-                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
-                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
-        scales128 = _mm_sub_epi8(scales128, m32);
-        const __m256i all_scales = _mm256_cvtepi8_epi16(scales128);
-        const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
-        const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
-        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
-
-        // high bit
-        const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].hmask);
-
-        // integer accumulator
-        __m256i sumi = _mm256_setzero_si256();
-
-        int bit = 0;
-        int is  = 0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            // load low 2 bits
-            const __m256i q3bits = _mm256_loadu_si256((const __m256i*)q3); q3 += 32;
-
-            // prepare low and high bits
-            const __m256i q3l_0 = _mm256_and_si256(q3bits, m3);
-            const __m256i q3h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
-            ++bit;
-
-            const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 2), m3);
-            const __m256i q3h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
-            ++bit;
-
-            const __m256i q3l_2 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 4), m3);
-            const __m256i q3h_2 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
-            ++bit;
-
-            const __m256i q3l_3 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 6), m3);
-            const __m256i q3h_3 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
-            ++bit;
-
-            // load Q8 quants
-            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
-            // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
-            // and 2 if the high bit was set)
-            __m256i q8s_0 = _mm256_maddubs_epi16(q3h_0, q8_0);
-            __m256i q8s_1 = _mm256_maddubs_epi16(q3h_1, q8_1);
-            __m256i q8s_2 = _mm256_maddubs_epi16(q3h_2, q8_2);
-            __m256i q8s_3 = _mm256_maddubs_epi16(q3h_3, q8_3);
-
-            __m256i p16_0 = _mm256_maddubs_epi16(q3l_0, q8_0);
-            __m256i p16_1 = _mm256_maddubs_epi16(q3l_1, q8_1);
-            __m256i p16_2 = _mm256_maddubs_epi16(q3l_2, q8_2);
-            __m256i p16_3 = _mm256_maddubs_epi16(q3l_3, q8_3);
-
-            p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
-            p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
-            p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
-            p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
-
-            // multiply with scales
-            p16_0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 0)), p16_0);
-            p16_1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 1)), p16_1);
-            p16_2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 2)), p16_2);
-            p16_3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 3)), p16_3);
-
-            // accumulate
-            p16_0 = _mm256_add_epi32(p16_0, p16_1);
-            p16_2 = _mm256_add_epi32(p16_2, p16_3);
-            sumi  = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_2));
-
-        }
-
-        // multiply with block scale and accumulate
-        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
-
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __AVX__
-
-    const __m128i m3 = _mm_set1_epi8(3);
-    const __m128i mone = _mm_set1_epi8(1);
-    const __m128i m32 = _mm_set1_epi8(32);
-    const __m128i m2 = _mm_set1_epi8(2);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    const uint32_t *aux;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q3 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        // Set up scales
-        aux = (const uint32_t *)x[i].scales;
-        __m128i scales128 = _mm_set_epi32(
-                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
-                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
-                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
-                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
-        scales128 = _mm_sub_epi8(scales128, m32);
-        const __m128i scales_0 = _mm_cvtepi8_epi16(scales128);
-        const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales128, scales128));
-        const __m128i scales[2] = { scales_0, scales_1 };
-
-        // high bit *128*2 from block_q3_K.hmask[QK_K/8]
-        const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].hmask[0]);
-        const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].hmask[16]);
-
-        // integer accumulator
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            // load low 2 bits *64*2 from block_q3_K.qs[QK_K/4]
-            const __m128i q3bits_0 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
-            const __m128i q3bits_1 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
-
-            // prepare low and high bits
-            const int bit = j << 2;
-
-            const __m128i q3l_0 = _mm_and_si128(q3bits_0, m3);
-            const __m128i q3l_1 = _mm_and_si128(q3bits_1, m3);
-            const __m128i q3h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit)), bit), 2);
-            const __m128i q3h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit)), bit), 2);
-
-            const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 2), m3);
-            const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 2), m3);
-            const __m128i q3h_2 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
-            const __m128i q3h_3 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
-
-            const __m128i q3l_4 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 4), m3);
-            const __m128i q3l_5 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 4), m3);
-            const __m128i q3h_4 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
-            const __m128i q3h_5 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
-
-            const __m128i q3l_6 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 6), m3);
-            const __m128i q3l_7 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 6), m3);
-            const __m128i q3h_6 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
-            const __m128i q3h_7 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
-
-            // load Q8 quants from block_q8_K.qs[QK_K]
-            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-
-            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
-            // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
-            // and 2 if the high bit was set)
-            __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, q8_0);
-            __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, q8_1);
-            __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, q8_2);
-            __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, q8_3);
-            __m128i q8s_4 = _mm_maddubs_epi16(q3h_4, q8_4);
-            __m128i q8s_5 = _mm_maddubs_epi16(q3h_5, q8_5);
-            __m128i q8s_6 = _mm_maddubs_epi16(q3h_6, q8_6);
-            __m128i q8s_7 = _mm_maddubs_epi16(q3h_7, q8_7);
-
-            __m128i p16_0 = _mm_maddubs_epi16(q3l_0, q8_0);
-            __m128i p16_1 = _mm_maddubs_epi16(q3l_1, q8_1);
-            __m128i p16_2 = _mm_maddubs_epi16(q3l_2, q8_2);
-            __m128i p16_3 = _mm_maddubs_epi16(q3l_3, q8_3);
-            __m128i p16_4 = _mm_maddubs_epi16(q3l_4, q8_4);
-            __m128i p16_5 = _mm_maddubs_epi16(q3l_5, q8_5);
-            __m128i p16_6 = _mm_maddubs_epi16(q3l_6, q8_6);
-            __m128i p16_7 = _mm_maddubs_epi16(q3l_7, q8_7);
-
-            p16_0 = _mm_sub_epi16(p16_0, q8s_0);
-            p16_1 = _mm_sub_epi16(p16_1, q8s_1);
-            p16_2 = _mm_sub_epi16(p16_2, q8s_2);
-            p16_3 = _mm_sub_epi16(p16_3, q8s_3);
-            p16_4 = _mm_sub_epi16(p16_4, q8s_4);
-            p16_5 = _mm_sub_epi16(p16_5, q8s_5);
-            p16_6 = _mm_sub_epi16(p16_6, q8s_6);
-            p16_7 = _mm_sub_epi16(p16_7, q8s_7);
-
-            // multiply with scales
-            __m128i shuffle = _mm_set1_epi16(0x0100);
-            p16_0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_0);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_1);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_2);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_3);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_4);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_5);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_6);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_7);
-
-            // accumulate
-            p16_0 = _mm_add_epi32(p16_0, p16_1);
-            p16_2 = _mm_add_epi32(p16_2, p16_3);
-            p16_4 = _mm_add_epi32(p16_4, p16_5);
-            p16_6 = _mm_add_epi32(p16_6, p16_7);
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_4, p16_6));
-
-        }
-
-        // multiply with block scale and accumulate
-        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
-
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __riscv_v_intrinsic
-
-    uint32_t aux[3];
-    uint32_t utmp[4];
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict qh = x[i].hmask;
-        const  int8_t * restrict q8 = y[i].qs;
-
-        memcpy(aux, x[i].scales, 12);
-        utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
-        utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
-        utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
-        utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
-
-        int8_t * scale = (int8_t *)utmp;
-        for (int j = 0; j < 16; ++j) scale[j] -= 32;
-
-
-        size_t vl = 32;
-        uint8_t m =  1;
-
-        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
-        vuint8m1_t vqh = __riscv_vle8_v_u8m1(qh, vl);
-
-        int sum_t = 0;
-
-        for (int j = 0; j < QK_K; j += 128) {
-
-            vl = 32;
-
-            // load Q3
-            vuint8m1_t q3_x = __riscv_vle8_v_u8m1(q3, vl);
-
-            vint8m1_t q3_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q3_x, 0x03, vl));
-            vint8m1_t q3_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x2, vl), 0x03 , vl));
-            vint8m1_t q3_2 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x4, vl), 0x03 , vl));
-            vint8m1_t q3_3 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x6, vl), 0x03 , vl));
-
-            // compute mask for subtraction
-            vuint8m1_t qh_m0 = __riscv_vand_vx_u8m1(vqh, m, vl);
-            vbool8_t vmask_0 = __riscv_vmseq_vx_u8m1_b8(qh_m0, 0, vl);
-            vint8m1_t q3_m0 = __riscv_vsub_vx_i8m1_mu(vmask_0, q3_0, q3_0, 0x4, vl);
-            m <<= 1;
-
-            vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
-            vbool8_t vmask_1 = __riscv_vmseq_vx_u8m1_b8(qh_m1, 0, vl);
-            vint8m1_t q3_m1 = __riscv_vsub_vx_i8m1_mu(vmask_1, q3_1, q3_1, 0x4, vl);
-            m <<= 1;
-
-            vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
-            vbool8_t vmask_2 = __riscv_vmseq_vx_u8m1_b8(qh_m2, 0, vl);
-            vint8m1_t q3_m2 = __riscv_vsub_vx_i8m1_mu(vmask_2, q3_2, q3_2, 0x4, vl);
-            m <<= 1;
-
-            vuint8m1_t qh_m3 = __riscv_vand_vx_u8m1(vqh, m, vl);
-            vbool8_t vmask_3 = __riscv_vmseq_vx_u8m1_b8(qh_m3, 0, vl);
-            vint8m1_t q3_m3 = __riscv_vsub_vx_i8m1_mu(vmask_3, q3_3, q3_3, 0x4, vl);
-            m <<= 1;
-
-            // load Q8 and take product with Q3
-            vint16m2_t a0 = __riscv_vwmul_vv_i16m2(q3_m0, __riscv_vle8_v_i8m1(q8, vl), vl);
-            vint16m2_t a1 = __riscv_vwmul_vv_i16m2(q3_m1, __riscv_vle8_v_i8m1(q8+32, vl), vl);
-            vint16m2_t a2 = __riscv_vwmul_vv_i16m2(q3_m2, __riscv_vle8_v_i8m1(q8+64, vl), vl);
-            vint16m2_t a3 = __riscv_vwmul_vv_i16m2(q3_m3, __riscv_vle8_v_i8m1(q8+96, vl), vl);
-
-            vl = 16;
-
-            // retrieve lane to multiply with scale
-            vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl);
-            vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl);
-            vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl);
-            vint32m2_t aux1_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 1), (scale[3]), vl);
-            vint32m2_t aux2_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 0), (scale[4]), vl);
-            vint32m2_t aux2_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 1), (scale[5]), vl);
-            vint32m2_t aux3_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 0), (scale[6]), vl);
-            vint32m2_t aux3_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 1), (scale[7]), vl);
-
-            vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux0_0, aux0_1, vl), vzero, vl);
-            vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux1_0, aux1_1, vl), isum0, vl);
-            vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux2_0, aux2_1, vl), isum1, vl);
-            vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux3_0, aux3_1, vl), isum2, vl);
-
-            sum_t +=  __riscv_vmv_x_s_i32m1_i32(isum3);
-
-            q3 += 32;    q8 += 128;   scale += 8;
-
-        }
-
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-
-        sumf += d*sum_t;
-
-    }
-
-    *s = sumf;
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0x3);
-    const vector signed char lowMask1 = vec_splats((int8_t)0xf);
-    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
-    const vector int v0 = vec_splats((int32_t)0);
-    const vector signed char v1 = vec_splats((signed char)0x1);
-    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
-    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
-    const vector signed char off = vec_splats((signed char)0x20);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        UNUSED(kmask1);
-        UNUSED(kmask2);
-
-        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
-        vector signed char u1 = vec_and(u0, lowMask1);
-        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
-        vector signed char u3 = (vector signed char)vec_mergeh((vector signed int)u2, (vector signed int)vec_sr(u2, v2));
-        vector signed char u30 = vec_sl(vec_and(u3, lowMask), v4);
-        vector signed char u31 = vec_and(u3, lowMask2);
-
-        u1 = vec_or(u1, u30);
-        u2 = vec_or(vec_sr(u0, v4), u31);
-
-        vector signed char vscales = (vector signed char)vec_mergeh((vector signed long long)u1, (vector signed long long)u2);
-        vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].hmask);
-        vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].hmask);
-
-        vscales = vec_sub(vscales, off);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-        vector signed int vsumi4 = v0;
-        vector signed int vsumi5 = v0;
-        vector signed int vsumi6 = v0;
-        vector signed int vsumi7 = v0;
-
-        const uint8_t * restrict q3 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            __builtin_prefetch(q3, 0, 1);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed char qxs0 = (vector signed char)vec_xl( 0, q3);
-            vector signed char qxs1 = (vector signed char)vec_xl(16, q3);
-            q3 += 32;
-
-            //the low 2 bits
-            vector signed char qxs00 = vec_and(qxs0, lowMask);
-            vector signed char qxs01 = vec_and(vec_sr(qxs0, v2), lowMask);
-            vector signed char qxs02 = vec_and(vec_sr(qxs0, v4), lowMask);
-            vector signed char qxs03 = vec_and(vec_sr(qxs0, v6), lowMask);
-            vector signed char qxs10 = vec_and(qxs1, lowMask);
-            vector signed char qxs11 = vec_and(vec_sr(qxs1, v2), lowMask);
-            vector signed char qxs12 = vec_and(vec_sr(qxs1, v4), lowMask);
-            vector signed char qxs13 = vec_and(vec_sr(qxs1, v6), lowMask);
-
-            //the 3rd bit
-            vector signed char qxh00 = vec_sl(vec_andc(v1, qxhs0), v2);
-            vector signed char qxh01 = vec_sl(vec_andc(v1, vec_sr(qxhs0, (vector unsigned char)v1)), v2);
-            vector signed char qxh02 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v2)), v2);
-            vector signed char qxh03 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v3)), v2);
-            vector signed char qxh10 = vec_sl(vec_andc(v1, qxhs1), v2);
-            vector signed char qxh11 = vec_sl(vec_andc(v1, vec_sr(qxhs1, (vector unsigned char)v1)), v2);
-            vector signed char qxh12 = vec_sl(vec_andc(v1, vec_sr(qxhs1, v2)), v2);
-            vector signed char qxh13 = vec_sl(vec_andc(v1, vec_sr(qxhs1, v3)), v2);
-            qxhs0 = vec_sr(qxhs0, v4);
-            qxhs1 = vec_sr(qxhs1, v4);
-
-            vector signed char q3x00 = vec_sub(qxs00, qxh00);
-            vector signed char q3x01 = vec_sub(qxs01, qxh01);
-            vector signed char q3x02 = vec_sub(qxs02, qxh02);
-            vector signed char q3x03 = vec_sub(qxs03, qxh03);
-            vector signed char q3x10 = vec_sub(qxs10, qxh10);
-            vector signed char q3x11 = vec_sub(qxs11, qxh11);
-            vector signed char q3x12 = vec_sub(qxs12, qxh12);
-            vector signed char q3x13 = vec_sub(qxs13, qxh13);
-
-            vector signed char q8y00 = vec_xl(  0, q8);
-            vector signed char q8y10 = vec_xl( 16, q8);
-            vector signed char q8y01 = vec_xl( 32, q8);
-            vector signed char q8y11 = vec_xl( 48, q8);
-            vector signed char q8y02 = vec_xl( 64, q8);
-            vector signed char q8y12 = vec_xl( 80, q8);
-            vector signed char q8y03 = vec_xl( 96, q8);
-            vector signed char q8y13 = vec_xl(112, q8);
-            q8 += 128;
-
-            vector signed short vscales_h = vec_unpackh(vscales);
-            vector signed short vs0 = vec_splat(vscales_h, 0);
-            vector signed short vs1 = vec_splat(vscales_h, 1);
-            vector signed short vs2 = vec_splat(vscales_h, 2);
-            vector signed short vs3 = vec_splat(vscales_h, 3);
-            vector signed short vs4 = vec_splat(vscales_h, 4);
-            vector signed short vs5 = vec_splat(vscales_h, 5);
-            vector signed short vs6 = vec_splat(vscales_h, 6);
-            vector signed short vs7 = vec_splat(vscales_h, 7);
-            vscales = vec_sld(vscales, vscales, 8);
-
-            vector signed short qv00 = vec_add(vec_mule(q3x00, q8y00), vec_mulo(q3x00, q8y00));
-            vector signed short qv01 = vec_add(vec_mule(q3x01, q8y01), vec_mulo(q3x01, q8y01));
-            vector signed short qv02 = vec_add(vec_mule(q3x02, q8y02), vec_mulo(q3x02, q8y02));
-            vector signed short qv03 = vec_add(vec_mule(q3x03, q8y03), vec_mulo(q3x03, q8y03));
-            vector signed short qv10 = vec_add(vec_mule(q3x10, q8y10), vec_mulo(q3x10, q8y10));
-            vector signed short qv11 = vec_add(vec_mule(q3x11, q8y11), vec_mulo(q3x11, q8y11));
-            vector signed short qv12 = vec_add(vec_mule(q3x12, q8y12), vec_mulo(q3x12, q8y12));
-            vector signed short qv13 = vec_add(vec_mule(q3x13, q8y13), vec_mulo(q3x13, q8y13));
-
-            vsumi0 = vec_msum(qv00, vs0, vsumi0);
-            vsumi1 = vec_msum(qv01, vs2, vsumi1);
-            vsumi2 = vec_msum(qv02, vs4, vsumi2);
-            vsumi3 = vec_msum(qv03, vs6, vsumi3);
-            vsumi4 = vec_msum(qv10, vs1, vsumi4);
-            vsumi5 = vec_msum(qv11, vs3, vsumi5);
-            vsumi6 = vec_msum(qv12, vs5, vsumi6);
-            vsumi7 = vec_msum(qv13, vs7, vsumi7);
-        }
-
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined __loongarch_asx
-
-    const __m256i m3 = __lasx_xvreplgr2vr_b(3);
-    const __m256i mone = __lasx_xvreplgr2vr_b(1);
-    const __m128i m32 = __lsx_vreplgr2vr_b(32);
-
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    uint32_t aux[3];
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const uint8_t * restrict q3 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-        // Set up scales
-        memcpy(aux, x[i].scales, 12);
-        __m128i scales128 = lsx_set_w(
-                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
-                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
-                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
-                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
-        scales128 = __lsx_vsub_b(scales128, m32);
-        const __m256i all_scales = lasx_ext8_16(scales128);
-        const __m128i l_scales = lasx_extracti128(all_scales, 0);
-        const __m128i h_scales = lasx_extracti128(all_scales, 1);
-        const __m256i scales[2] = {lasx_insertf128(l_scales, l_scales), lasx_insertf128(h_scales, h_scales)};
-
-        // high bit
-        const __m256i hbits = __lasx_xvld((const __m256i*)x[i].hmask, 0);
-
-        // integer accumulator
-        __m256i sumi = __lasx_xvldi(0);
-
-        int bit = 0;
-        int is  = 0;
-        __m256i xvbit;
-
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            // load low 2 bits
-            const __m256i q3bits = __lasx_xvld((const __m256i*)q3, 0); q3 += 32;
-
-            xvbit = __lasx_xvreplgr2vr_h(bit);
-            // prepare low and high bits
-            const __m256i q3l_0 = __lasx_xvand_v(q3bits, m3);
-            const __m256i q3h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
-            ++bit;
-
-            xvbit = __lasx_xvreplgr2vr_h(bit);
-            const __m256i q3l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 2), m3);
-            const __m256i q3h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
-            ++bit;
-
-            xvbit = __lasx_xvreplgr2vr_h(bit);
-            const __m256i q3l_2 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 4), m3);
-            const __m256i q3h_2 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
-            ++bit;
-
-            xvbit = __lasx_xvreplgr2vr_h(bit);
-            const __m256i q3l_3 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 6), m3);
-            const __m256i q3h_3 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
-            ++bit;
-
-            // load Q8 quants
-            const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-
-            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use lasx_maddubs_h,
-            // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
-            // and 2 if the high bit was set)
-            __m256i q8s_0 = lasx_maddubs_h(q3h_0, q8_0);
-            __m256i q8s_1 = lasx_maddubs_h(q3h_1, q8_1);
-            __m256i q8s_2 = lasx_maddubs_h(q3h_2, q8_2);
-            __m256i q8s_3 = lasx_maddubs_h(q3h_3, q8_3);
-
-            __m256i p16_0 = lasx_maddubs_h(q3l_0, q8_0);
-            __m256i p16_1 = lasx_maddubs_h(q3l_1, q8_1);
-            __m256i p16_2 = lasx_maddubs_h(q3l_2, q8_2);
-            __m256i p16_3 = lasx_maddubs_h(q3l_3, q8_3);
-
-            p16_0 = __lasx_xvsub_h(p16_0, q8s_0);
-            p16_1 = __lasx_xvsub_h(p16_1, q8s_1);
-            p16_2 = __lasx_xvsub_h(p16_2, q8s_2);
-            p16_3 = __lasx_xvsub_h(p16_3, q8s_3);
-
-            // multiply with scales
-            p16_0 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 0)), p16_0);
-            p16_1 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 1)), p16_1);
-            p16_2 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 2)), p16_2);
-            p16_3 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 3)), p16_3);
-
-            // accumulate
-            p16_0 = __lasx_xvadd_w(p16_0, p16_1);
-            p16_2 = __lasx_xvadd_w(p16_2, p16_3);
-            sumi  = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_2));
-        }
-        // multiply with block scale and accumulate
-        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);//FIXME
-    }
-
-    *s = hsum_float_8(acc);
-
-#else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict hm = x[i].hmask;
-        const  int8_t * restrict q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * restrict a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
-#endif
-
-}
-
-void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(n % QK_K == 0);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_q4_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-    static const uint32_t kmask1 = 0x3f3f3f3f;
-    static const uint32_t kmask2 = 0x0f0f0f0f;
-    static const uint32_t kmask3 = 0x03030303;
-
-    uint32_t utmp[4];
-
-#ifdef __ARM_NEON
-    const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const int32x4_t mzero = vdupq_n_s32(0);
-
-    ggml_int8x16x2_t q4bytes;
-    ggml_int8x16x2_t q8bytes;
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
-
-        memcpy(utmp, x[i].scales, 12);
-
-        uint32x2_t mins8 = { 0 };
-        mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
-        mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);
-
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[0] &= kmask1;
-
-        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins8)));
-        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
-                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
-        sumf -= dmin * vaddvq_s32(prod);
-
-        const uint8_t * scales = (const uint8_t *)utmp;
-
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        int32_t sumi1 = 0;
-        int32_t sumi2 = 0;
-
-        for (int j = 0; j < QK_K/64; ++j) {
-            const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); q4 += 32;
-
-            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
-            q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[0], m4b));
-            q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[1], m4b));
-
-            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
-            sumi1 += vaddvq_s32(p1) * scales[2*j+0];
-
-            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
-            q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
-            q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
-
-            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
-
-            sumi2 += vaddvq_s32(p2) * scales[2*j+1];
-        }
-
-        sumf += d * (sumi1 + sumi2);
-
-    }
-
-    *s = sumf;
-
-#elif defined __AVX2__
-
-    const __m256i m4 = _mm256_set1_epi8(0xF);
-
-    __m256 acc = _mm256_setzero_ps();
-    __m128 acc_m = _mm_setzero_ps();
-
-   for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
-
-        const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
-        const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
-        const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
-        acc_m = _mm_fmadd_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod), acc_m);
-
-        const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
-        const __m256i scales = MM256_SET_M128I(sc128, sc128);
-
-        __m256i sumi = _mm256_setzero_si256();
-
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const __m256i scale_l = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
-            const __m256i scale_h = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
-
-            const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
-            const __m256i q4l = _mm256_and_si256(q4bits, m4);
-            const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(q4bits, 4), m4);
-
-            const __m256i q8l = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            __m256i p16l = _mm256_maddubs_epi16(q4l, q8l);
-            p16l = _mm256_madd_epi16(scale_l, p16l);
-
-            const __m256i q8h = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            __m256i p16h = _mm256_maddubs_epi16(q4h, q8h);
-            p16h = _mm256_madd_epi16(scale_h, p16h);
-            const __m256i sumj = _mm256_add_epi32(p16l, p16h);
-
-            sumi = _mm256_add_epi32(sumi, sumj);
-        }
-
-        __m256 vd = _mm256_set1_ps(d);
-        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
-
-    }
-
-    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
-    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
-
-    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
-
-#elif defined __AVX__
-
-    const __m128i m4 = _mm_set1_epi8(0xF);
-    const __m128i m2 = _mm_set1_epi8(0x2);
-
-    __m256 acc = _mm256_setzero_ps();
-    __m128 acc_m = _mm_setzero_ps();
-
-   for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
-        const __m128i scales = _mm_cvtepu8_epi16(utmps);
-        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
-
-        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
-        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
-        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
-        const __m128i prod = _mm_madd_epi16(mins, q8s);
-        acc_m = _mm_add_ps(_mm_mul_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod)), acc_m);
-
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        __m128i shuffle = _mm_set1_epi16(0x0100);
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const __m128i scale_l = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            const __m128i scale_h = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi16(shuffle, m2);
-
-            __m128i q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4l_0 = _mm_and_si128(q4bits, m4);
-            const __m128i q4h_0 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
-            q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4l_1 = _mm_and_si128(q4bits, m4);
-            const __m128i q4h_1 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
-
-            const __m128i q8l_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i p16l = _mm_maddubs_epi16(q4l_0, q8l_0);
-            p16l = _mm_madd_epi16(scale_l, p16l);
-            sumi_0 = _mm_add_epi32(sumi_0, p16l);
-            const __m128i q8l_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            p16l = _mm_maddubs_epi16(q4l_1, q8l_1);
-            p16l = _mm_madd_epi16(scale_l, p16l);
-            sumi_1 = _mm_add_epi32(sumi_1, p16l);
-
-            const __m128i q8h_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i p16h = _mm_maddubs_epi16(q4h_0, q8h_0);
-            p16h = _mm_madd_epi16(scale_h, p16h);
-            sumi_0 = _mm_add_epi32(sumi_0, p16h);
-            const __m128i q8h_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            p16h = _mm_maddubs_epi16(q4h_1, q8h_1);
-            p16h = _mm_madd_epi16(scale_h, p16h);
-            sumi_1 = _mm_add_epi32(sumi_1, p16h);
-
-        }
-
-        __m256 vd = _mm256_set1_ps(d);
-        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
-
-    }
-
-    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
-    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
-
-    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
-
-#elif defined __riscv_v_intrinsic
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        size_t vl = 8;
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
-        vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
-        vint16mf2_t q8sums   = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl);
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        vuint8mf4_t mins8  = __riscv_vle8_v_u8mf4(mins, vl);
-        vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl));
-        vint32m1_t  prod   = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl);
-
-        vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
-        sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
-
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        vl = 32;
-
-        int32_t sum_1 = 0;
-        int32_t sum_2 = 0;
-
-        vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1);
-
-        for (int j = 0; j < QK_K/64; ++j) {
-            // load Q4
-            vuint8m1_t q4_x = __riscv_vle8_v_u8m1(q4, vl);
-
-            // load Q8 and multiply it with lower Q4 nibble
-            vint8m1_t  q8_0 = __riscv_vle8_v_i8m1(q8, vl);
-            vint8m1_t  q4_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q4_x, 0x0F, vl));
-            vint16m2_t qv_0 = __riscv_vwmul_vv_i16m2(q4_0, q8_0, vl);
-            vint16m1_t vs_0 = __riscv_vredsum_vs_i16m2_i16m1(qv_0, vzero, vl);
-
-            sum_1 += __riscv_vmv_x_s_i16m1_i16(vs_0) * scales[2*j+0];
-
-            // load Q8 and multiply it with upper Q4 nibble
-            vint8m1_t  q8_1 = __riscv_vle8_v_i8m1(q8+32, vl);
-            vint8m1_t  q4_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q4_x, 0x04, vl));
-            vint16m2_t qv_1 = __riscv_vwmul_vv_i16m2(q4_1, q8_1, vl);
-            vint16m1_t vs_1 = __riscv_vredsum_vs_i16m2_i16m1(qv_1, vzero, vl);
-
-            sum_2 += __riscv_vmv_x_s_i16m1_i16(vs_1) * scales[2*j+1];
-
-            q4 += 32;    q8 += 64;
-
-        }
-
-        sumf += d*(sum_1 + sum_2);
-
-    }
-
-    *s = sumf;
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector signed char lowMask1 = vec_splats((int8_t)0x3f);
-    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
-    const vector int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v2 = vec_splats((uint8_t)2);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
-        vector float vdmin = vec_mul(vxmin, vyd);
-
-        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
-        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
-
-        UNUSED(kmask1);
-        UNUSED(kmask2);
-        UNUSED(kmask3);
-        UNUSED(utmp);
-
-        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
-        vector signed char u1 = vec_and(vec_sr(u0, v2), lowMask2);
-        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
-        vector signed char u3 = vec_sr(u2, v4);
-
-        vector signed char u30 = u1;
-        vector signed char u31 = (vector signed char)vec_mergeh((vector signed int)vec_and(u2, lowMask), (vector signed int)u3);
-
-        u1 = vec_and(u0, lowMask1);
-        u2 = vec_or(u30, u31);
-
-        vector signed char utmps = (vector signed char)vec_mergeh((vector signed int)u1, (vector signed int)u2);
-
-        vector signed short vscales = vec_unpackh(utmps);
-        vector signed short q4xmins = vec_unpackl(utmps);
-        vector signed short q4xmins0 = vec_mergeh(q4xmins, q4xmins);
-        vector signed short q4xmins1 = vec_mergel(q4xmins, q4xmins);
-
-        vector signed int prod0 = vec_mule(q4xmins0, q8ysums0);
-        vector signed int prod1 = vec_mule(q4xmins1, q8ysums1);
-        vector signed int prod2 = vec_mulo(q4xmins0, q8ysums0);
-        vector signed int prod3 = vec_mulo(q4xmins1, q8ysums1);
-
-        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
-        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
-        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
-        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        for (int j = 0; j < QK_K/64; j+=2) {
-            __builtin_prefetch(q4, 0, 1);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed char qxs0 = (vector signed char)vec_xl( 0, q4);
-            vector signed char qxs1 = (vector signed char)vec_xl(16, q4);
-            vector signed char qxs2 = (vector signed char)vec_xl(32, q4);
-            vector signed char qxs3 = (vector signed char)vec_xl(48, q4);
-            q4 += 64;
-
-            vector unsigned char q4x00 = (vector unsigned char)vec_and(qxs0, lowMask);
-            vector unsigned char q4x01 = (vector unsigned char)vec_sr(qxs0, v4);
-            vector unsigned char q4x10 = (vector unsigned char)vec_and(qxs1, lowMask);
-            vector unsigned char q4x11 = (vector unsigned char)vec_sr(qxs1, v4);
-            vector unsigned char q4x20 = (vector unsigned char)vec_and(qxs2, lowMask);
-            vector unsigned char q4x21 = (vector unsigned char)vec_sr(qxs2, v4);
-            vector unsigned char q4x30 = (vector unsigned char)vec_and(qxs3, lowMask);
-            vector unsigned char q4x31 = (vector unsigned char)vec_sr(qxs3, v4);
-
-            vector signed char q8y00 = vec_xl(  0, q8);
-            vector signed char q8y10 = vec_xl( 16, q8);
-            vector signed char q8y01 = vec_xl( 32, q8);
-            vector signed char q8y11 = vec_xl( 48, q8);
-            vector signed char q8y20 = vec_xl( 64, q8);
-            vector signed char q8y30 = vec_xl( 80, q8);
-            vector signed char q8y21 = vec_xl( 96, q8);
-            vector signed char q8y31 = vec_xl(112, q8);
-            q8 += 128;
-
-            vector signed int qv00 = vec_msum(q8y00, q4x00, v0);
-            vector signed int qv01 = vec_msum(q8y01, q4x01, v0);
-            vector signed int qv10 = vec_msum(q8y10, q4x10, v0);
-            vector signed int qv11 = vec_msum(q8y11, q4x11, v0);
-            vector signed int qv20 = vec_msum(q8y20, q4x20, v0);
-            vector signed int qv21 = vec_msum(q8y21, q4x21, v0);
-            vector signed int qv30 = vec_msum(q8y30, q4x30, v0);
-            vector signed int qv31 = vec_msum(q8y31, q4x31, v0);
-
-            vector signed int vscales_h = vec_unpackh(vscales);
-            vector signed int vs0 = vec_splat(vscales_h, 0);
-            vector signed int vs1 = vec_splat(vscales_h, 1);
-            vector signed int vs2 = vec_splat(vscales_h, 2);
-            vector signed int vs3 = vec_splat(vscales_h, 3);
-            vscales = vec_sld(vscales, vscales, 8);
-
-            vsumi0 = vec_add(vec_mul(qv00, vs0), vsumi0);
-            vsumi1 = vec_add(vec_mul(qv01, vs1), vsumi1);
-            vsumi2 = vec_add(vec_mul(qv20, vs2), vsumi2);
-            vsumi3 = vec_add(vec_mul(qv21, vs3), vsumi3);
-
-            vsumi0 = vec_add(vec_mul(qv10, vs0), vsumi0);
-            vsumi1 = vec_add(vec_mul(qv11, vs1), vsumi1);
-            vsumi2 = vec_add(vec_mul(qv30, vs2), vsumi2);
-            vsumi3 = vec_add(vec_mul(qv31, vs3), vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined __loongarch_asx
-    GGML_UNUSED(kmask1);
-    GGML_UNUSED(kmask2);
-    GGML_UNUSED(kmask3);
-
-    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
-
-    __m256 acc = (__m256)__lasx_xvldi(0);
-    __m128 acc_m = (__m128)__lsx_vldi(0);
-
-   for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]));
-
-        const __m256i q8sums = __lasx_xvld((const __m256i*)y[i].bsums, 0);
-        const __m128i q8s = lsx_hadd_h(lasx_extracti128(q8sums, 0), lasx_extracti128(q8sums, 1));
-        const __m128i prod = lsx_madd_h(lasx_extracti128(mins_and_scales, 1), q8s);
-        acc_m = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(dmin), __lsx_vffint_s_w(prod), acc_m);
-
-        const __m128i sc128  = lasx_extracti128(mins_and_scales, 0);
-        const __m256i scales = lasx_insertf128(sc128, sc128);
-
-        __m256i sumi = __lasx_xvldi(0);
-
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const __m256i scale_l = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+0));
-            const __m256i scale_h = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+1));
-
-            const __m256i q4bits = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
-            const __m256i q4l = __lasx_xvand_v(q4bits, m4);
-            const __m256i q4h = __lasx_xvand_v(__lasx_xvsrli_h(q4bits, 4), m4);
-
-            const __m256i q8l = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            __m256i p16l = lasx_maddubs_h(q4l, q8l);
-            p16l = lasx_madd_h(scale_l, p16l);
-
-            const __m256i q8h = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            __m256i p16h = lasx_maddubs_h(q4h, q8h);
-            p16h = lasx_madd_h(scale_h, p16h);
-            const __m256i sumj = __lasx_xvadd_w(p16l, p16h);
-
-            sumi = __lasx_xvadd_w(sumi, sumj);
-        }
-
-        __m256 vd = __lasx_xvreplfr2vr_s(d);
-        acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc);
-
-    }
-
-    acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vpermi_w((__m128i)acc_m, (__m128i)acc_m, 0xee));
-    __m128i tmp1 = __lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w((__m128i)acc_m, 1), 0);
-    acc_m = __lsx_vfadd_s(acc_m, (__m128)tmp1);
-
-
-    ft_union fi;
-    fi.i = __lsx_vpickve2gr_w(acc_m, 0);
-    *s = hsum_float_8(acc) + fi.f ;
-#else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q4 = x[i].qs;
-        const  int8_t * restrict q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * restrict a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-#endif
-}
-
-void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy,  size_t by, int nrc) {
-    assert(n % QK_K == 0);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_q5_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-    static const uint32_t kmask1 = 0x3f3f3f3f;
-    static const uint32_t kmask2 = 0x0f0f0f0f;
-    static const uint32_t kmask3 = 0x03030303;
-
-    uint32_t utmp[4];
-
-#ifdef __ARM_NEON
-    const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const uint8x16_t mone = vdupq_n_u8(1);
-    const uint8x16_t mtwo = vdupq_n_u8(2);
-    const int32x4_t mzero = vdupq_n_s32(0);
-
-    ggml_int8x16x4_t q5bytes;
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        const uint8x8_t mins8 = vld1_u8((const uint8_t*)utmp + 8);
-        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(mins8));
-        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
-                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
-        int32_t sumi_mins = vaddvq_s32(prod);
-
-        const uint8_t * scales = (const uint8_t *)utmp;
-
-        const uint8_t * restrict q5 = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh);
-
-        ggml_uint8x16x4_t q5h;
-
-        int32_t sumi = 0;
-
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const ggml_uint8x16x2_t q5bits = ggml_vld1q_u8_x2(q5); q5 += 32;
-            const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            q5h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
-            q5h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
-            q5h.val[2] = vshlq_n_u8(vandq_u8(mtwo, qhbits.val[0]), 3);
-            q5h.val[3] = vshlq_n_u8(vandq_u8(mtwo, qhbits.val[1]), 3);
-            qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 2);
-            qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 2);
-
-            q5bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.val[0], m4b), q5h.val[0]));
-            q5bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.val[1], m4b), q5h.val[1]));
-            q5bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[0], 4), q5h.val[2]));
-            q5bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[1], 4), q5h.val[3]));
-
-            sumi += vaddvq_s32(ggml_vdotq_s32(ggml_vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0]), q5bytes.val[1], q8bytes.val[1])) * *scales++;
-            sumi += vaddvq_s32(ggml_vdotq_s32(ggml_vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2]), q5bytes.val[3], q8bytes.val[3])) * *scales++;
-        }
-
-        sumf += d * sumi - dmin * sumi_mins;
-    }
-
-    *s = sumf;
-
-#elif defined __AVX2__
-
-    const __m256i m4 = _mm256_set1_epi8(0xF);
-    const __m128i mzero = _mm_setzero_si128();
-    const __m256i mone  = _mm256_set1_epi8(1);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    float summs = 0.f;
-
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q5 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
-
-        const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
-        const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
-        const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
-        const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
-        summs += dmin * _mm_extract_epi32(hsum, 0);
-
-        const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
-        const __m256i scales = MM256_SET_M128I(sc128, sc128);
-
-        const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].qh);
-        __m256i hmask = mone;
-
-        __m256i sumi = _mm256_setzero_si256();
-
-        int bit = 0;
-
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const __m256i scale_0 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
-            const __m256i scale_1 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
-
-            const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5); q5 += 32;
-
-            const __m256i q5l_0 = _mm256_and_si256(q5bits, m4);
-            const __m256i q5h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
-            const __m256i q5_0  = _mm256_add_epi8(q5l_0, q5h_0);
-            hmask = _mm256_slli_epi16(hmask, 1);
-
-            const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4);
-            const __m256i q5h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
-            const __m256i q5_1  = _mm256_add_epi8(q5l_1, q5h_1);
-            hmask = _mm256_slli_epi16(hmask, 1);
-
-            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            __m256i p16_0 = _mm256_maddubs_epi16(q5_0, q8_0);
-            __m256i p16_1 = _mm256_maddubs_epi16(q5_1, q8_1);
-
-            p16_0 = _mm256_madd_epi16(scale_0, p16_0);
-            p16_1 = _mm256_madd_epi16(scale_1, p16_1);
-
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
-
-        }
-
-        __m256 vd = _mm256_set1_ps(d);
-        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
-
-    }
-
-    *s = hsum_float_8(acc) + summs;
-
-#elif defined __AVX__
-
-    const __m128i m4 = _mm_set1_epi8(0xF);
-    const __m128i mzero = _mm_setzero_si128();
-    const __m128i mone  = _mm_set1_epi8(1);
-    const __m128i m2 = _mm_set1_epi8(2);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    float summs = 0.f;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const uint8_t * restrict q5 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
-        const __m128i scales = _mm_cvtepu8_epi16(utmps);
-        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
-
-        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
-        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
-        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
-        const __m128i prod = _mm_madd_epi16(mins, q8s);
-        const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
-        summs += dmin * _mm_extract_epi32(hsum, 0);
-
-        const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].qh[0]);
-        const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].qh[16]);
-        __m128i hmask = mone;
-
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        int bit = 0;
-
-        __m128i shuffle = _mm_set1_epi16(0x0100);
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const __m128i scale_0 = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            const __m128i scale_1 = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi16(shuffle, m2);
-
-            const __m128i q5bits_0 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
-            const __m128i q5bits_1 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
-
-            __m128i q5l_0 = _mm_and_si128(q5bits_0, m4);
-            __m128i q5l_1 = _mm_and_si128(q5bits_1, m4);
-            __m128i q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
-            __m128i q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
-            __m128i q5_0  = _mm_add_epi8(q5l_0, q5h_0);
-            __m128i q5_1  = _mm_add_epi8(q5l_1, q5h_1);
-            hmask = _mm_slli_epi16(hmask, 1);
-
-            __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i p16_0 = _mm_maddubs_epi16(q5_0, q8_0);
-            __m128i p16_1 = _mm_maddubs_epi16(q5_1, q8_1);
-            p16_0 = _mm_madd_epi16(scale_0, p16_0);
-            p16_1 = _mm_madd_epi16(scale_0, p16_1);
-
-            q5l_0 = _mm_and_si128(_mm_srli_epi16(q5bits_0, 4), m4);
-            q5l_1 = _mm_and_si128(_mm_srli_epi16(q5bits_1, 4), m4);
-            q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
-            q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
-            q5_0  = _mm_add_epi8(q5l_0, q5h_0);
-            q5_1  = _mm_add_epi8(q5l_1, q5h_1);
-            hmask = _mm_slli_epi16(hmask, 1);
-
-            q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i p16_2 = _mm_maddubs_epi16(q5_0, q8_0);
-            __m128i p16_3 = _mm_maddubs_epi16(q5_1, q8_1);
-            p16_2 = _mm_madd_epi16(scale_1, p16_2);
-            p16_3 = _mm_madd_epi16(scale_1, p16_3);
-
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
-
-        }
-
-        __m256 vd = _mm256_set1_ps(d);
-        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
-
-    }
-
-    *s = hsum_float_8(acc) + summs;
-
-#elif defined __riscv_v_intrinsic
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    float sumf = 0;
-    float sums = 0.0;
-
-    size_t vl;
-
-    for (int i = 0; i < nb; ++i) {
-
-        vl = 8;
-
-        const uint8_t * restrict q5 = x[i].qs;
-        const uint8_t * restrict hm = x[i].qh;
-        const  int8_t * restrict q8 = y[i].qs;
-
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
-
-        vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
-        vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
-        vint16mf2_t q8sums = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl);
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        vuint8mf4_t mins8 = __riscv_vle8_v_u8mf4(mins, vl);
-        vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl));
-        vint32m1_t prod = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl);
-
-        vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
-        sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
-
-        vl = 32;
-        int32_t aux32 = 0;
-        int is = 0;
-
-        uint8_t m = 1;
-        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
-        vuint8m1_t vqh = __riscv_vle8_v_u8m1(hm, vl);
-
-        for (int j = 0; j < QK_K/64; ++j) {
-            // load Q5 and Q8
-            vuint8m1_t q5_x = __riscv_vle8_v_u8m1(q5, vl);
-            vint8m1_t  q8_y1 = __riscv_vle8_v_i8m1(q8, vl);
-            vint8m1_t  q8_y2 = __riscv_vle8_v_i8m1(q8+32, vl);
-
-            // compute mask for addition
-            vint8m1_t q5_a = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q5_x, 0x0F, vl));
-            vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
-            vbool8_t vmask_1 = __riscv_vmsne_vx_u8m1_b8(qh_m1, 0, vl);
-            vint8m1_t q5_m1 = __riscv_vadd_vx_i8m1_mu(vmask_1, q5_a, q5_a, 16, vl);
-            m <<= 1;
-
-            vint8m1_t q5_l = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q5_x, 0x04, vl));
-            vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
-            vbool8_t vmask_2 = __riscv_vmsne_vx_u8m1_b8(qh_m2, 0, vl);
-            vint8m1_t q5_m2 = __riscv_vadd_vx_i8m1_mu(vmask_2, q5_l, q5_l, 16, vl);
-            m <<= 1;
-
-            vint16m2_t v0 = __riscv_vwmul_vv_i16m2(q5_m1, q8_y1, vl);
-            vint16m2_t v1 = __riscv_vwmul_vv_i16m2(q5_m2, q8_y2, vl);
-
-            vint32m4_t vs1 = __riscv_vwmul_vx_i32m4(v0, scales[is++], vl);
-            vint32m4_t vs2 = __riscv_vwmul_vx_i32m4(v1, scales[is++], vl);
-
-            vint32m1_t vacc1 = __riscv_vredsum_vs_i32m4_i32m1(vs1, vzero, vl);
-            vint32m1_t vacc2 = __riscv_vredsum_vs_i32m4_i32m1(vs2, vzero, vl);
-
-            aux32 += __riscv_vmv_x_s_i32m1_i32(vacc1) + __riscv_vmv_x_s_i32m1_i32(vacc2);
-            q5 += 32;    q8 += 64;
-
-        }
-
-        vfloat32m1_t vaux = __riscv_vfmul_vf_f32m1(__riscv_vfmv_v_f_f32m1(aux32, 1), d, 1);
-        sums += __riscv_vfmv_f_s_f32m1_f32(vaux);
-
-    }
-
-    *s = sumf+sums;
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector signed char lowMask1 = vec_splats((int8_t)0x3f);
-    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
-    const vector int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v1 = vec_splats((unsigned char)0x1);
-    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
-    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
-        vector float vdmin = vec_mul(vxmin, vyd);
-
-        UNUSED(kmask1);
-        UNUSED(kmask2);
-        UNUSED(kmask3);
-        UNUSED(utmp);
-
-        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
-        vector signed char u1 = vec_and(vec_sr(u0, v2), lowMask2);
-        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
-        vector signed char u3 = vec_sr(u2, v4);
-
-        vector signed char u30 = u1;
-        vector signed char u31 = (vector signed char)vec_mergeh((vector signed int)vec_and(u2, lowMask), (vector signed int)u3);
-
-        u1 = vec_and(u0, lowMask1);
-        u2 = vec_or(u30, u31);
-
-        vector signed char utmps = (vector signed char)vec_mergeh((vector signed int)u1, (vector signed int)u2);
-
-        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
-        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
-
-        vector signed short vscales = vec_unpackh(utmps);
-
-        vector signed short q5xmins = vec_unpackl(utmps);
-        vector signed short q5xmins0 = vec_mergeh(q5xmins, q5xmins);
-        vector signed short q5xmins1 = vec_mergel(q5xmins, q5xmins);
-
-        vector signed int prod0 = vec_mule(q5xmins0, q8ysums0);
-        vector signed int prod1 = vec_mule(q5xmins1, q8ysums1);
-        vector signed int prod2 = vec_mulo(q5xmins0, q8ysums0);
-        vector signed int prod3 = vec_mulo(q5xmins1, q8ysums1);
-
-        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
-        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
-        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
-        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
-
-        vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].qh);
-        vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].qh);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        const uint8_t * restrict q5 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        for (int j = 0; j < QK_K/64; ++j) {
-            __builtin_prefetch(q5, 0, 1);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed char qxs0 = (vector signed char)vec_xl( 0, q5);
-            vector signed char qxs1 = (vector signed char)vec_xl(16, q5);
-            q5 += 32;
-
-            vector signed char qxs00 = vec_and(qxs0, lowMask);
-            vector signed char qxs01 = vec_sr(qxs0, v4);
-            vector signed char qxs10 = vec_and(qxs1, lowMask);
-            vector signed char qxs11 = vec_sr(qxs1, v4);
-
-            vector signed char q5h00 = vec_sl(vec_and((vector signed char)v1, qxhs0), v4);
-            vector signed char q5h01 = vec_sl(vec_and((vector signed char)v2, qxhs0), v3);
-            vector signed char q5h10 = vec_sl(vec_and((vector signed char)v1, qxhs1), v4);
-            vector signed char q5h11 = vec_sl(vec_and((vector signed char)v2, qxhs1), v3);
-            qxhs0 = vec_sr(qxhs0, v2);
-            qxhs1 = vec_sr(qxhs1, v2);
-
-            vector unsigned char q5x00 = (vector unsigned char)vec_or(q5h00, qxs00);
-            vector unsigned char q5x01 = (vector unsigned char)vec_or(q5h01, qxs01);
-            vector unsigned char q5x10 = (vector unsigned char)vec_or(q5h10, qxs10);
-            vector unsigned char q5x11 = (vector unsigned char)vec_or(q5h11, qxs11);
-
-            vector signed char q8y00 = vec_xl( 0, q8);
-            vector signed char q8y10 = vec_xl(16, q8);
-            vector signed char q8y01 = vec_xl(32, q8);
-            vector signed char q8y11 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed int qv00 = vec_msum(q8y00, q5x00, v0);
-            vector signed int qv01 = vec_msum(q8y01, q5x01, v0);
-            vector signed int qv10 = vec_msum(q8y10, q5x10, v0);
-            vector signed int qv11 = vec_msum(q8y11, q5x11, v0);
-
-            vector signed int vscales_h = vec_unpackh(vscales);
-            vector signed int vs0 = vec_splat(vscales_h, 0);
-            vector signed int vs1 = vec_splat(vscales_h, 1);
-            vscales = vec_sld(vscales, vscales, 12);
-
-            vsumi0 = vec_add(vec_mul(qv00, vs0), vsumi0);
-            vsumi1 = vec_add(vec_mul(qv10, vs0), vsumi1);
-            vsumi2 = vec_add(vec_mul(qv01, vs1), vsumi2);
-            vsumi3 = vec_add(vec_mul(qv11, vs1), vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined __loongarch_asx
-    GGML_UNUSED(kmask1);
-    GGML_UNUSED(kmask2);
-    GGML_UNUSED(kmask3);
-
-    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
-    const __m128i mzero = __lsx_vldi(0);
-    const __m256i mone  = __lasx_xvreplgr2vr_b(1);
-
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    float summs = 0.f;
-
-   for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * restrict q5 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]));
-
-        const __m256i q8sums = __lasx_xvld((const __m256i*)y[i].bsums, 0);
-        const __m128i q8s = lsx_hadd_h(lasx_extracti128(q8sums, 0), lasx_extracti128(q8sums, 1));
-        const __m128i prod = lsx_madd_h(lasx_extracti128(mins_and_scales, 1), q8s);
-        const __m128i hsum = lsx_hadd_w(lsx_hadd_w(prod, mzero), mzero);
-        summs += dmin * __lsx_vpickve2gr_w(hsum, 0);    //TODO check
-
-        const __m128i sc128  = lasx_extracti128(mins_and_scales, 0);
-        const __m256i scales = lasx_insertf128(sc128, sc128);
-
-        const __m256i hbits = __lasx_xvld((const __m256i*)x[i].qh, 0);
-        __m256i hmask = mone;
-
-        __m256i sumi = __lasx_xvldi(0);
-
-        int bit = 0;
-        __m256i xvbit;
-
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const __m256i scale_0 = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+0));
-            const __m256i scale_1 = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+1));
-
-            const __m256i q5bits = __lasx_xvld((const __m256i*)q5, 0); q5 += 32;
-
-            xvbit = __lasx_xvreplgr2vr_h(bit++);
-            const __m256i q5l_0 = __lasx_xvand_v(q5bits, m4);
-            const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4);
-            const __m256i q5_0  = __lasx_xvadd_b(q5l_0, q5h_0);
-            hmask = __lasx_xvslli_h(hmask, 1);
-
-            xvbit = __lasx_xvreplgr2vr_h(bit++);
-            const __m256i q5l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q5bits, 4), m4);
-            const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4);
-            const __m256i q5_1  = __lasx_xvadd_b(q5l_1, q5h_1);
-            hmask = __lasx_xvslli_h(hmask, 1);
-
-            const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-
-            __m256i p16_0 = lasx_maddubs_h(q5_0, q8_0);
-            __m256i p16_1 = lasx_maddubs_h(q5_1, q8_1);
-
-            p16_0 = lasx_madd_h(scale_0, p16_0);
-            p16_1 = lasx_madd_h(scale_1, p16_1);
-
-            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1));
-
-        }
-
-        __m256 vd = __lasx_xvreplfr2vr_s(d);
-        acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc);
-
-    }
-
-    *s = hsum_float_8(acc) + summs;
-
-#else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q4 = x[i].qs;
-        const uint8_t * restrict hm = x[i].qh;
-        const  int8_t * restrict q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * restrict a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-#endif
-}
-
-void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(n % QK_K == 0);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_q6_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#ifdef __ARM_NEON
-    float sum = 0;
-
-    const uint8x16_t m4b = vdupq_n_u8(0xF);
-    const int32x4_t  vzero = vdupq_n_s32(0);
-    //const int8x16_t  m32s = vdupq_n_s8(32);
-
-    const uint8x16_t mone = vdupq_n_u8(3);
-
-    ggml_int8x16x4_t q6bytes;
-    ggml_uint8x16x4_t q6h;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d_all = GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q6 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const int8_t * restrict scale = x[i].scales;
-
-        const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
-        const int8x16_t scales = vld1q_s8(scale);
-        const ggml_int16x8x2_t q6scales = {{vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}};
-
-        const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])),
-                                                   vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))),
-                                         vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[1]), vget_low_s16 (q6scales.val[1])),
-                                                   vmull_s16(vget_high_s16(q8sums.val[1]), vget_high_s16(q6scales.val[1]))));
-        int32_t isum_mins = vaddvq_s32(prod);
-
-        int32_t isum = 0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); qh += 32;
-            ggml_uint8x16x4_t q6bits = ggml_vld1q_u8_x4(q6); q6 += 64;
-            ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
-            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
-            uint8x16_t shifted = vshrq_n_u8(qhbits.val[0], 2);
-            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-            shifted = vshrq_n_u8(qhbits.val[1], 2);
-            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-
-            //q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0])), m32s);
-            //q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1])), m32s);
-            //q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2])), m32s);
-            //q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3])), m32s);
-            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0]));
-            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1]));
-            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2]));
-            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3]));
-
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
-
-            scale += 4;
-
-            q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            shifted = vshrq_n_u8(qhbits.val[0], 4);
-            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-            shifted = vshrq_n_u8(qhbits.val[1], 4);
-            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-            shifted = vshrq_n_u8(qhbits.val[0], 6);
-            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-            shifted = vshrq_n_u8(qhbits.val[1], 6);
-            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-
-            //q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0])), m32s);
-            //q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1])), m32s);
-            //q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2])), m32s);
-            //q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3])), m32s);
-            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0]));
-            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1]));
-            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2]));
-            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3]));
-
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
-            scale += 4;
-        }
-        //sum += isum * d_all * y[i].d;
-        sum += d_all * y[i].d * (isum - 32 * isum_mins);
-
-    }
-    *s = sum;
-
-#elif defined __AVX2__
-
-    const __m256i m4 = _mm256_set1_epi8(0xF);
-    const __m256i m2 = _mm256_set1_epi8(3);
-    const __m256i m32s = _mm256_set1_epi8(32);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
-
-        __m256i sumi = _mm256_setzero_si256();
-
-        int is = 0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
-            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
-            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
-            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
-            is += 4;
-
-            const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
-            const __m256i q4bits2 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
-            const __m256i q4bitsH = _mm256_loadu_si256((const __m256i*)qh); qh += 32;
-
-            const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, m2), 4);
-            const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 2), m2), 4);
-            const __m256i q4h_2 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 4), m2), 4);
-            const __m256i q4h_3 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 6), m2), 4);
-
-            const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
-            const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(q4bits2, m4), q4h_1);
-            const __m256i q4_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_2);
-            const __m256i q4_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits2, 4), m4), q4h_3);
-
-            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            __m256i q8s_0 = _mm256_maddubs_epi16(m32s, q8_0);
-            __m256i q8s_1 = _mm256_maddubs_epi16(m32s, q8_1);
-            __m256i q8s_2 = _mm256_maddubs_epi16(m32s, q8_2);
-            __m256i q8s_3 = _mm256_maddubs_epi16(m32s, q8_3);
-
-            __m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
-            __m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
-            __m256i p16_2 = _mm256_maddubs_epi16(q4_2, q8_2);
-            __m256i p16_3 = _mm256_maddubs_epi16(q4_3, q8_3);
-
-            p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
-            p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
-            p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
-            p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
-
-            p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
-            p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
-            p16_2 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_2), p16_2);
-            p16_3 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_3), p16_3);
-
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_2, p16_3));
-
-        }
-
-        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __AVX__
-
-    const __m128i m3 = _mm_set1_epi8(3);
-    const __m128i m15 = _mm_set1_epi8(15);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        // handle the q6_k -32 offset separately using bsums
-        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)y[i].bsums);
-        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)y[i].bsums + 1);
-        const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
-        const __m128i scales_16_0 = _mm_cvtepi8_epi16(scales);
-        const __m128i scales_16_1 = _mm_cvtepi8_epi16(_mm_bsrli_si128(scales, 8));
-        const __m128i q8sclsub_0 = _mm_slli_epi32(_mm_madd_epi16(q8sums_0, scales_16_0), 5);
-        const __m128i q8sclsub_1 = _mm_slli_epi32(_mm_madd_epi16(q8sums_1, scales_16_1), 5);
-
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        int is = 0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const __m128i q4bitsH_0 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
-            const __m128i q4bitsH_1 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
-
-            const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, m3), 4);
-            const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, m3), 4);
-            const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(12)), 2);
-            const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(12)), 2);
-            const __m128i q4h_4 = _mm_and_si128(q4bitsH_0, _mm_set1_epi8(48));
-            const __m128i q4h_5 = _mm_and_si128(q4bitsH_1, _mm_set1_epi8(48));
-            const __m128i q4h_6 = _mm_srli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(-64)), 2);
-            const __m128i q4h_7 = _mm_srli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(-64)), 2);
-
-            const __m128i q4bits1_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4bits1_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4bits2_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4bits2_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-
-            const __m128i q4_0 = _mm_or_si128(_mm_and_si128(q4bits1_0, m15), q4h_0);
-            const __m128i q4_1 = _mm_or_si128(_mm_and_si128(q4bits1_1, m15), q4h_1);
-            const __m128i q4_2 = _mm_or_si128(_mm_and_si128(q4bits2_0, m15), q4h_2);
-            const __m128i q4_3 = _mm_or_si128(_mm_and_si128(q4bits2_1, m15), q4h_3);
-            const __m128i q4_4 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_0, 4), m15), q4h_4);
-            const __m128i q4_5 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_1, 4), m15), q4h_5);
-            const __m128i q4_6 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_0, 4), m15), q4h_6);
-            const __m128i q4_7 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_1, 4), m15), q4h_7);
-
-            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-
-            __m128i p16_0 = _mm_maddubs_epi16(q4_0, q8_0);
-            __m128i p16_1 = _mm_maddubs_epi16(q4_1, q8_1);
-            __m128i p16_2 = _mm_maddubs_epi16(q4_2, q8_2);
-            __m128i p16_3 = _mm_maddubs_epi16(q4_3, q8_3);
-            __m128i p16_4 = _mm_maddubs_epi16(q4_4, q8_4);
-            __m128i p16_5 = _mm_maddubs_epi16(q4_5, q8_5);
-            __m128i p16_6 = _mm_maddubs_epi16(q4_6, q8_6);
-            __m128i p16_7 = _mm_maddubs_epi16(q4_7, q8_7);
-
-            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
-            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
-            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
-            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
-            is += 4;
-
-            p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
-            p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_0, 8)), p16_1);
-            p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
-            p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_1, 8)), p16_3);
-            p16_4 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_2), p16_4);
-            p16_5 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_2, 8)), p16_5);
-            p16_6 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_3), p16_6);
-            p16_7 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_3, 8)), p16_7);
-
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_4, p16_6));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_5, p16_7));
-
-        }
-
-        sumi_0 = _mm_sub_epi32(sumi_0, q8sclsub_0);
-        sumi_1 = _mm_sub_epi32(sumi_1, q8sclsub_1);
-        const __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi)), acc);
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __riscv_v_intrinsic
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-
-        const uint8_t * restrict q6 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const  int8_t * restrict q8 = y[i].qs;
-
-        const int8_t * restrict scale = x[i].scales;
-
-        size_t vl;
-
-        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
-
-        int sum_t = 0;
-        int is = 0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            vl = 32;
-
-            // load qh
-            vuint8m1_t qh_x = __riscv_vle8_v_u8m1(qh, vl);
-
-            // load Q6
-            vuint8m1_t q6_0 = __riscv_vle8_v_u8m1(q6, vl);
-            vuint8m1_t q6_1 = __riscv_vle8_v_u8m1(q6+32, vl);
-
-            vuint8m1_t q6a_0 = __riscv_vand_vx_u8m1(q6_0, 0x0F, vl);
-            vuint8m1_t q6a_1 = __riscv_vand_vx_u8m1(q6_1, 0x0F, vl);
-            vuint8m1_t q6s_0 = __riscv_vsrl_vx_u8m1(q6_0, 0x04, vl);
-            vuint8m1_t q6s_1 = __riscv_vsrl_vx_u8m1(q6_1, 0x04, vl);
-
-            vuint8m1_t qh_0 = __riscv_vand_vx_u8m1(qh_x, 0x03, vl);
-            vuint8m1_t qh_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x2, vl), 0x03 , vl);
-            vuint8m1_t qh_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x4, vl), 0x03 , vl);
-            vuint8m1_t qh_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x6, vl), 0x03 , vl);
-
-            vuint8m1_t qhi_0 = __riscv_vor_vv_u8m1(q6a_0, __riscv_vsll_vx_u8m1(qh_0, 0x04, vl), vl);
-            vuint8m1_t qhi_1 = __riscv_vor_vv_u8m1(q6a_1, __riscv_vsll_vx_u8m1(qh_1, 0x04, vl), vl);
-            vuint8m1_t qhi_2 = __riscv_vor_vv_u8m1(q6s_0, __riscv_vsll_vx_u8m1(qh_2, 0x04, vl), vl);
-            vuint8m1_t qhi_3 = __riscv_vor_vv_u8m1(q6s_1, __riscv_vsll_vx_u8m1(qh_3, 0x04, vl), vl);
-
-            vint8m1_t a_0 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_0), 32, vl);
-            vint8m1_t a_1 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_1), 32, vl);
-            vint8m1_t a_2 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_2), 32, vl);
-            vint8m1_t a_3 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_3), 32, vl);
-
-            // load Q8 and take product
-            vint16m2_t va_q_0 = __riscv_vwmul_vv_i16m2(a_0, __riscv_vle8_v_i8m1(q8, vl), vl);
-            vint16m2_t va_q_1 = __riscv_vwmul_vv_i16m2(a_1, __riscv_vle8_v_i8m1(q8+32, vl), vl);
-            vint16m2_t va_q_2 = __riscv_vwmul_vv_i16m2(a_2, __riscv_vle8_v_i8m1(q8+64, vl), vl);
-            vint16m2_t va_q_3 = __riscv_vwmul_vv_i16m2(a_3, __riscv_vle8_v_i8m1(q8+96, vl), vl);
-
-            vl = 16;
-
-            vint32m2_t vaux_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 0), scale[is+0], vl);
-            vint32m2_t vaux_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 1), scale[is+1], vl);
-            vint32m2_t vaux_2 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 0), scale[is+2], vl);
-            vint32m2_t vaux_3 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 1), scale[is+3], vl);
-            vint32m2_t vaux_4 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 0), scale[is+4], vl);
-            vint32m2_t vaux_5 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 1), scale[is+5], vl);
-            vint32m2_t vaux_6 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 0), scale[is+6], vl);
-            vint32m2_t vaux_7 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 1), scale[is+7], vl);
-
-            vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_0, vaux_1, vl), vzero, vl);
-            vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_2, vaux_3, vl), isum0, vl);
-            vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_4, vaux_5, vl), isum1, vl);
-            vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_6, vaux_7, vl), isum2, vl);
-
-            sum_t += __riscv_vmv_x_s_i32m1_i32(isum3);
-
-            q6 += 64;   qh += 32;   q8 += 128;   is=8;
-
-        }
-
-        sumf += d * sum_t;
-
-    }
-
-    *s = sumf;
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
-    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
-    const vector signed char off = vec_splats((signed char)0x20);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-        vector signed int vsumi4 = v0;
-        vector signed int vsumi5 = v0;
-        vector signed int vsumi6 = v0;
-        vector signed int vsumi7 = v0;
-
-        const uint8_t * restrict q6 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict qs = x[i].scales;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            __builtin_prefetch(q6, 0, 0);
-            __builtin_prefetch(qh, 0, 0);
-            __builtin_prefetch(q8, 0, 0);
-
-            vector signed char qxs0 = (vector signed char)vec_xl( 0, q6);
-            vector signed char qxs1 = (vector signed char)vec_xl(16, q6);
-            vector signed char qxs2 = (vector signed char)vec_xl(32, q6);
-            vector signed char qxs3 = (vector signed char)vec_xl(48, q6);
-            q6 += 64;
-
-            vector signed char qxs00 = vec_and(qxs0, lowMask);
-            vector signed char qxs01 = vec_sr(qxs0, v4);
-            vector signed char qxs10 = vec_and(qxs1, lowMask);
-            vector signed char qxs11 = vec_sr(qxs1, v4);
-            vector signed char qxs20 = vec_and(qxs2, lowMask);
-            vector signed char qxs21 = vec_sr(qxs2, v4);
-            vector signed char qxs30 = vec_and(qxs3, lowMask);
-            vector signed char qxs31 = vec_sr(qxs3, v4);
-
-            vector signed char qxhs0 = (vector signed char)vec_xl( 0, qh);
-            vector signed char qxhs1 = (vector signed char)vec_xl(16, qh);
-            qh += 32;
-
-            vector signed char qxh00 = vec_sl(vec_and((vector signed char)v3, qxhs0), v4);
-            vector signed char qxh01 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v4)), v4);
-            vector signed char qxh10 = vec_sl(vec_and((vector signed char)v3, qxhs1), v4);
-            vector signed char qxh11 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v4)), v4);
-            vector signed char qxh20 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v2)), v4);
-            vector signed char qxh21 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v6)), v4);
-            vector signed char qxh30 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v2)), v4);
-            vector signed char qxh31 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v6)), v4);
-
-            vector signed char q6x00 = vec_sub(vec_or(qxh00, qxs00), off);
-            vector signed char q6x01 = vec_sub(vec_or(qxh01, qxs01), off);
-            vector signed char q6x10 = vec_sub(vec_or(qxh10, qxs10), off);
-            vector signed char q6x11 = vec_sub(vec_or(qxh11, qxs11), off);
-            vector signed char q6x20 = vec_sub(vec_or(qxh20, qxs20), off);
-            vector signed char q6x21 = vec_sub(vec_or(qxh21, qxs21), off);
-            vector signed char q6x30 = vec_sub(vec_or(qxh30, qxs30), off);
-            vector signed char q6x31 = vec_sub(vec_or(qxh31, qxs31), off);
-
-            vector signed char q8y00 = vec_xl(  0, q8);
-            vector signed char q8y10 = vec_xl( 16, q8);
-            vector signed char q8y20 = vec_xl( 32, q8);
-            vector signed char q8y30 = vec_xl( 48, q8);
-            vector signed char q8y01 = vec_xl( 64, q8);
-            vector signed char q8y11 = vec_xl( 80, q8);
-            vector signed char q8y21 = vec_xl( 96, q8);
-            vector signed char q8y31 = vec_xl(112, q8);
-            q8 += 128;
-
-            vector signed short qv00 = vec_add(vec_mule(q6x00, q8y00), vec_mulo(q6x00, q8y00));
-            vector signed short qv10 = vec_add(vec_mule(q6x10, q8y10), vec_mulo(q6x10, q8y10));
-            vector signed short qv20 = vec_add(vec_mule(q6x20, q8y20), vec_mulo(q6x20, q8y20));
-            vector signed short qv30 = vec_add(vec_mule(q6x30, q8y30), vec_mulo(q6x30, q8y30));
-            vector signed short qv01 = vec_add(vec_mule(q6x01, q8y01), vec_mulo(q6x01, q8y01));
-            vector signed short qv11 = vec_add(vec_mule(q6x11, q8y11), vec_mulo(q6x11, q8y11));
-            vector signed short qv21 = vec_add(vec_mule(q6x21, q8y21), vec_mulo(q6x21, q8y21));
-            vector signed short qv31 = vec_add(vec_mule(q6x31, q8y31), vec_mulo(q6x31, q8y31));
-
-            vector signed short vscales = vec_unpackh(vec_xl_len(qs, 8));
-            qs += 8;
-
-            vector signed short vs0 = vec_splat(vscales, 0);
-            vector signed short vs1 = vec_splat(vscales, 1);
-            vector signed short vs2 = vec_splat(vscales, 2);
-            vector signed short vs3 = vec_splat(vscales, 3);
-            vector signed short vs4 = vec_splat(vscales, 4);
-            vector signed short vs5 = vec_splat(vscales, 5);
-            vector signed short vs6 = vec_splat(vscales, 6);
-            vector signed short vs7 = vec_splat(vscales, 7);
-
-            vsumi0 = vec_msum(qv00, vs0, vsumi0);
-            vsumi1 = vec_msum(qv01, vs4, vsumi1);
-            vsumi2 = vec_msum(qv10, vs1, vsumi2);
-            vsumi3 = vec_msum(qv11, vs5, vsumi3);
-            vsumi4 = vec_msum(qv20, vs2, vsumi4);
-            vsumi5 = vec_msum(qv21, vs6, vsumi5);
-            vsumi6 = vec_msum(qv30, vs3, vsumi6);
-            vsumi7 = vec_msum(qv31, vs7, vsumi7);
-        }
-
-        vsumi0 = vec_add(vsumi0, vsumi4);
-        vsumi1 = vec_add(vsumi1, vsumi5);
-        vsumi2 = vec_add(vsumi2, vsumi6);
-        vsumi3 = vec_add(vsumi3, vsumi7);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined __loongarch_asx
-
-    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
-    const __m256i m2 = __lasx_xvreplgr2vr_b(3);
-    const __m256i m32s = __lasx_xvreplgr2vr_b(32);
-
-    __m256 acc = (__m256)__lasx_xvldi(0);
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const __m128i scales = __lsx_vld((const __m128i*)x[i].scales, 0);
-
-        __m256i sumi = __lasx_xvldi(0);
-
-        int is = 0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const __m128i scale_0 = lsx_shuffle_b(scales, get_scale_shuffle(is + 0));
-            const __m128i scale_1 = lsx_shuffle_b(scales, get_scale_shuffle(is + 1));
-            const __m128i scale_2 = lsx_shuffle_b(scales, get_scale_shuffle(is + 2));
-            const __m128i scale_3 = lsx_shuffle_b(scales, get_scale_shuffle(is + 3));
-            is += 4;
-
-            const __m256i q4bits1 = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
-            const __m256i q4bits2 = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
-            const __m256i q4bitsH = __lasx_xvld((const __m256i*)qh, 0); qh += 32;
-
-            const __m256i q4h_0 = __lasx_xvslli_h(__lasx_xvand_v(q4bitsH, m2), 4);
-            const __m256i q4h_1 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 2), m2), 4);
-            const __m256i q4h_2 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 4), m2), 4);
-            const __m256i q4h_3 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 6), m2), 4);
-
-            const __m256i q4_0 = __lasx_xvor_v(__lasx_xvand_v(q4bits1, m4), q4h_0);
-            const __m256i q4_1 = __lasx_xvor_v(__lasx_xvand_v(q4bits2, m4), q4h_1);
-            const __m256i q4_2 = __lasx_xvor_v(__lasx_xvand_v(__lasx_xvsrli_h(q4bits1, 4), m4), q4h_2);
-            const __m256i q4_3 = __lasx_xvor_v(__lasx_xvand_v(__lasx_xvsrli_h(q4bits2, 4), m4), q4h_3);
-
-            const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-
-            __m256i q8s_0 = lasx_maddubs_h(m32s, q8_0);
-            __m256i q8s_1 = lasx_maddubs_h(m32s, q8_1);
-            __m256i q8s_2 = lasx_maddubs_h(m32s, q8_2);
-            __m256i q8s_3 = lasx_maddubs_h(m32s, q8_3);
-
-            __m256i p16_0 = lasx_maddubs_h(q4_0, q8_0);
-            __m256i p16_1 = lasx_maddubs_h(q4_1, q8_1);
-            __m256i p16_2 = lasx_maddubs_h(q4_2, q8_2);
-            __m256i p16_3 = lasx_maddubs_h(q4_3, q8_3);
-
-            p16_0 = __lasx_xvsub_h(p16_0, q8s_0);
-            p16_1 = __lasx_xvsub_h(p16_1, q8s_1);
-            p16_2 = __lasx_xvsub_h(p16_2, q8s_2);
-            p16_3 = __lasx_xvsub_h(p16_3, q8s_3);
-
-            p16_0 = lasx_madd_h(lasx_ext8_16(scale_0), p16_0);
-            p16_1 = lasx_madd_h(lasx_ext8_16(scale_1), p16_1);
-            p16_2 = lasx_madd_h(lasx_ext8_16(scale_2), p16_2);
-            p16_3 = lasx_madd_h(lasx_ext8_16(scale_3), p16_3);
-
-            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1));
-            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_2, p16_3));
-        }
-
-        acc = __lasx_xvfmadd_s((__m256)__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);
-    }
-
-    *s = hsum_float_8(acc);
-
-#else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const  int8_t * restrict q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * restrict a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-#endif
-}
-
-#if defined (__AVX__) || defined (__AVX2__) || defined (__ARM_NEON) || defined (__POWER9_VECTOR__) || defined(__loongarch_asx)
-static const int8_t keven_signs_q2xs[1024] = {
-     1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
-     1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
-     1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1, -1,
-     1,  1, -1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1,  1,
-     1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1, -1,
-     1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1,  1,
-     1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1,  1,
-     1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1,  1,  1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1, -1,
-     1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1, -1,
-     1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1,  1,
-     1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1,  1,
-     1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1, -1,
-     1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1,  1,
-     1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1, -1,
-     1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1, -1,
-     1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1,  1,
-     1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1, -1,
-     1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1,  1,
-     1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1,  1,
-     1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1, -1,
-     1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1,  1,
-     1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1, -1,
-     1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1, -1,
-     1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1,  1,
-     1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1,  1,
-     1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1, -1,
-     1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1, -1,
-     1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1,  1,
-     1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1, -1,
-     1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1,  1,
-     1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1,  1,
-     1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
-};
-#endif
-
-void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(n % QK_K == 0);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_iq2_xxs * restrict x = vx;
-    const block_q8_K    * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_NEON)
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[4];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    ggml_int8x16x4_t q2u;
-    ggml_int8x16x4_t q2s;
-    ggml_int8x16x4_t q8b;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
-        float sumf1 = 0, sumf2 = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
-            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
-            q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 0])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 1])));
-            q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 2])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 3])));
-            q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 8])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 9])));
-            q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[10])), vld1_s8((const void *)(iq2xxs_grid + aux8[11])));
-            q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >>  7) & 127))));
-            q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 21) & 127))));
-            q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[3] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[3] >>  7) & 127))));
-            q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[3] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[3] >> 21) & 127))));
-            q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]);
-            q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]);
-            q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]);
-            q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]);
-            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]), q2u.val[1], q8b.val[1]);
-            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]), q2u.val[3], q8b.val[3]);
-            sumf1 += vaddvq_s32(p1) * (0.5f + (aux32[1] >> 28));
-            sumf2 += vaddvq_s32(p2) * (0.5f + (aux32[3] >> 28));
-        }
-        sumf += d*(sumf1 + sumf2);
-    }
-    *s = 0.25f * sumf;
-
-#elif defined(__AVX2__)
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[4];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    __m256 accumf = _mm256_setzero_ps();
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
-        __m256i sumi1 = _mm256_setzero_si256();
-        __m256i sumi2 = _mm256_setzero_si256();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
-            const __m256i q2_1 = _mm256_set_epi64x(iq2xxs_grid[aux8[ 3]], iq2xxs_grid[aux8[ 2]], iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
-            const __m256i q2_2 = _mm256_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]], iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
-            const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
-                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
-            const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127],
-                                                   signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
-            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
-            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
-            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
-            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
-            const uint16_t ls1 = aux32[1] >> 28;
-            const uint16_t ls2 = aux32[3] >> 28;
-            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
-            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
-            sumi1 = _mm256_add_epi32(sumi1, p1);
-            sumi2 = _mm256_add_epi32(sumi2, p2);
-        }
-
-        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
-
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#elif defined(__AVX__)
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[4];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    __m256 accumf = _mm256_setzero_ps();
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
-        __m128i sumi1_0 = _mm_setzero_si128();
-        __m128i sumi1_1 = _mm_setzero_si128();
-        __m128i sumi2_0 = _mm_setzero_si128();
-        __m128i sumi2_1 = _mm_setzero_si128();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
-            const __m128i q2_1_0 = _mm_set_epi64x(iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
-            const __m128i q2_1_1 = _mm_set_epi64x(iq2xxs_grid[aux8[3]], iq2xxs_grid[aux8[2]]);
-            const __m128i q2_2_0 = _mm_set_epi64x(iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
-            const __m128i q2_2_1 = _mm_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]]);
-            const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
-            const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
-            const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
-            const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127]);
-            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
-            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
-            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
-            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
-            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
-            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
-            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
-            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
-            const uint16_t ls1 = aux32[1] >> 28;
-            const uint16_t ls2 = aux32[3] >> 28;
-            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
-            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
-            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
-            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
-            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
-            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
-            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
-            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
-        }
-
-        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
-
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#elif defined(__POWER9_VECTOR__)
-    const vector int v0 = vec_splats((int32_t)0);
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t  *  restrict q8 = y[i].qs;
-
-        for (int j = 0; j < QK_K/32; j += 2) {
-            __builtin_prefetch(q2, 0, 1);
-            __builtin_prefetch(q8, 0, 1);
-
-            uint32_t aux32[4];
-            const uint8_t * aux8 = (const uint8_t *)aux32;
-
-            memcpy(aux32, q2, 4*sizeof(uint32_t));
-            q2 += 8;
-
-            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2xxs_grid + aux8[ 0]), *(const int64_t *)(iq2xxs_grid + aux8[ 1])};
-            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2xxs_grid + aux8[ 2]), *(const int64_t *)(iq2xxs_grid + aux8[ 3])};
-            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2xxs_grid + aux8[ 8]), *(const int64_t *)(iq2xxs_grid + aux8[ 9])};
-            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2xxs_grid + aux8[10]), *(const int64_t *)(iq2xxs_grid + aux8[11])};
-
-            vector signed long long vsigns0 = {*(const int64_t *)(signs64 + ((aux32[1] >>  0) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >>  7) & 127))};
-            vector signed long long vsigns1 = {*(const int64_t *)(signs64 + ((aux32[1] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >> 21) & 127))};
-            vector signed long long vsigns2 = {*(const int64_t *)(signs64 + ((aux32[3] >>  0) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >>  7) & 127))};
-            vector signed long long vsigns3 = {*(const int64_t *)(signs64 + ((aux32[3] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >> 21) & 127))};
-
-            vector signed char q2x0 = (vector signed char)vec_mul((vector signed char)vsigns0, (vector signed char)aux64x2_0);
-            vector signed char q2x1 = (vector signed char)vec_mul((vector signed char)vsigns1, (vector signed char)aux64x2_1);
-            vector signed char q2x2 = (vector signed char)vec_mul((vector signed char)vsigns2, (vector signed char)aux64x2_2);
-            vector signed char q2x3 = (vector signed char)vec_mul((vector signed char)vsigns3, (vector signed char)aux64x2_3);
-
-            vector signed char q8y0 = vec_xl( 0, q8);
-            vector signed char q8y1 = vec_xl(16, q8);
-            vector signed char q8y2 = vec_xl(32, q8);
-            vector signed char q8y3 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
-            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
-            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
-            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
-
-            const uint16_t ls0 = aux32[1] >> 28;
-            const uint16_t ls1 = aux32[3] >> 28;
-
-            vector signed short vscales01 = vec_splats((int16_t)(2*ls0+1));
-            vector signed short vscales23 = vec_splats((int16_t)(2*ls1+1));
-
-            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
-            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
-            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
-            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = 0.125f * vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[4];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    __m256 accumf = (__m256)__lasx_xvldi(0);
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
-        __m256i sumi1 = __lasx_xvldi(0);
-        __m256i sumi2 = __lasx_xvldi(0);
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
-
-            const __m256i q2_1 = lasx_set_d(iq2xxs_grid[aux8[ 3]], iq2xxs_grid[aux8[ 2]], iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
-            const __m256i q2_2 = lasx_set_d(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]], iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
-            const __m256i s2_1 = lasx_set_d(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
-                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
-            const __m256i s2_2 = lasx_set_d(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127],
-                                                   signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
-            const __m256i q8s_1 = __lasx_xvsigncov_b(s2_1, q8_1);
-            const __m256i q8s_2 = __lasx_xvsigncov_b(s2_2, q8_2);
-            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
-            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
-            const uint16_t ls1 = aux32[1] >> 28;
-            const uint16_t ls2 = aux32[3] >> 28;
-            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
-            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
-            sumi1 = __lasx_xvadd_w(sumi1, p1);
-            sumi2 = __lasx_xvadd_w(sumi2, p2);
-        }
-
-        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#else
-
-    uint32_t aux32[2];
-    const uint8_t * aux8 = (const uint8_t *)aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(aux32, q2, 2*sizeof(uint32_t));
-            q2 += 4;
-            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
-#endif
-}
-
-void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(n % QK_K == 0);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_iq2_xs * restrict x = vx;
-    const block_q8_K   * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_NEON)
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    ggml_int8x16x4_t q2u;
-    ggml_int8x16x4_t q2s;
-    ggml_int8x16x4_t q8b;
-
-    int32x4x4_t scales32;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
-        const uint8x8_t scales8 = vld1_u8(x[i].scales);
-        const uint8x8_t scales_l = vand_u8(scales8, vdup_n_u8(0xf));
-        const uint8x8_t scales_h = vshr_n_u8(scales8, 4);
-        uint8x16_t scales = vcombine_u8(vzip1_u8(scales_l, scales_h), vzip2_u8(scales_l, scales_h));
-        scales = vaddq_u8(vshlq_n_u8(scales, 1), vdupq_n_u8(1));
-        const uint16x8_t scales1 = vmovl_u8(vget_low_u8(scales));
-        const uint16x8_t scales2 = vmovl_u8(vget_high_u8(scales));
-        scales32.val[0] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales1)));
-        scales32.val[1] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales1)));
-        scales32.val[2] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales2)));
-        scales32.val[3] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales2)));
-        int32x4_t sumi = vdupq_n_s32(0);
-        for (int ib64 = 0; ib64 < QK_K/64; ++ib64) {
-            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
-            q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[0] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[1] & 511))));
-            q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[2] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[3] & 511))));
-            q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[4] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[5] & 511))));
-            q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[6] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[7] & 511))));
-            q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[0] >> 9))), vld1_s8((const void *)(signs64 + (q2[1] >> 9))));
-            q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[2] >> 9))), vld1_s8((const void *)(signs64 + (q2[3] >> 9))));
-            q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[4] >> 9))), vld1_s8((const void *)(signs64 + (q2[5] >> 9))));
-            q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[6] >> 9))), vld1_s8((const void *)(signs64 + (q2[7] >> 9))));
-            q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]);
-            q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]);
-            q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]);
-            q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]);
-            const int32x4_t p1 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]);
-            const int32x4_t p2 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[1], q8b.val[1]);
-            const int32x4_t p3 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]);
-            const int32x4_t p4 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[3], q8b.val[3]);
-            const int32x4_t p = vpaddq_s32(vpaddq_s32(p1, p2), vpaddq_s32(p3, p4));
-            sumi = vmlaq_s32(sumi, p, scales32.val[ib64]);
-            q2 += 8;
-        }
-        sumf += d*vaddvq_s32(sumi);
-    }
-    *s = 0.125f * sumf;
-
-#elif defined(__AVX2__)
-
-    const __m256i mone = _mm256_set1_epi8(1);
-    static const char block_sign_shuffle_mask_1[32] = {
-        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
-        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
-    };
-    static const char block_sign_shuffle_mask_2[32] = {
-        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
-        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
-    };
-    static const uint8_t bit_selector_mask_bytes[32] = {
-        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m256i bit_selector_mask = _mm256_loadu_si256((const __m256i*)bit_selector_mask_bytes);
-    const __m256i block_sign_shuffle_1 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_1);
-    const __m256i block_sign_shuffle_2 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_2);
-
-    static const uint8_t k_bit_helper[32] = {
-        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
-        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
-    };
-    const __m256i bit_helper = _mm256_loadu_si256((const __m256i*)k_bit_helper);
-    const __m256i m511 = _mm256_set1_epi16(511);
-    const __m128i m4 = _mm_set1_epi8(0xf);
-    const __m128i m1 = _mm_set1_epi8(1);
-
-    uint64_t aux64;
-
-    // somewhat hacky, but gives a significant boost in performance
-    __m256i aux_gindex;
-    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
-
-    __m256 accumf = _mm256_setzero_ps();
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
-
-        memcpy(&aux64, x[i].scales, 8);
-        __m128i stmp = _mm_set1_epi64x(aux64);
-        stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
-        const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
-
-        __m256i sumi1 = _mm256_setzero_si256();
-        __m256i sumi2 = _mm256_setzero_si256();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
-
-            const __m256i q2_data = _mm256_loadu_si256((const __m256i*)q2);  q2 += 16;
-            aux_gindex = _mm256_and_si256(q2_data, m511);
-
-            const __m256i partial_sign_bits = _mm256_srli_epi16(q2_data, 9);
-            const __m256i partial_sign_bits_upper = _mm256_srli_epi16(q2_data, 13);
-            const __m256i partial_sign_bits_for_counting = _mm256_xor_si256(partial_sign_bits, partial_sign_bits_upper);
-
-            const __m256i odd_bits = _mm256_shuffle_epi8(bit_helper, partial_sign_bits_for_counting);
-            const __m256i full_sign_bits = _mm256_or_si256(partial_sign_bits, odd_bits);
-
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8_3 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8_4 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-
-            const __m256i q2_1 = _mm256_set_epi64x(iq2xs_grid[gindex[ 3]], iq2xs_grid[gindex[ 2]],
-                                                   iq2xs_grid[gindex[ 1]], iq2xs_grid[gindex[ 0]]);
-            const __m256i q2_2 = _mm256_set_epi64x(iq2xs_grid[gindex[ 7]], iq2xs_grid[gindex[ 6]],
-                                                   iq2xs_grid[gindex[ 5]], iq2xs_grid[gindex[ 4]]);
-            const __m256i q2_3 = _mm256_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]],
-                                                   iq2xs_grid[gindex[ 9]], iq2xs_grid[gindex[ 8]]);
-            const __m256i q2_4 = _mm256_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]],
-                                                   iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
-
-            const __m128i full_signs_l = _mm256_castsi256_si128(full_sign_bits);
-            const __m128i full_signs_h = _mm256_extractf128_si256(full_sign_bits, 1);
-            const __m256i full_signs_1 = MM256_SET_M128I(full_signs_l, full_signs_l);
-            const __m256i full_signs_2 = MM256_SET_M128I(full_signs_h, full_signs_h);
-
-            __m256i signs;
-            signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_1);
-            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, _mm256_or_si256(signs, mone));
-
-            signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_2);
-            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, _mm256_or_si256(signs, mone));
-
-            signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_1);
-            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_3 = _mm256_sign_epi8(q8_3, _mm256_or_si256(signs, mone));
-
-            signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_2);
-            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_4 = _mm256_sign_epi8(q8_4, _mm256_or_si256(signs, mone));
-
-            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
-            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
-            const __m256i dot3  = _mm256_maddubs_epi16(q2_3, q8s_3);
-            const __m256i dot4  = _mm256_maddubs_epi16(q2_4, q8s_4);
-
-            const __m256i sc1 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0)));
-            const __m256i sc2 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1)));
-            const __m256i sc3 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2)));
-            const __m256i sc4 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3)));
-
-            sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot1, sc1));
-            sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot2, sc2));
-            sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot3, sc3));
-            sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot4, sc4));
-        }
-
-        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
-
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#elif defined(__AVX__)
-    const __m128i mone = _mm_set1_epi8(1);
-    static const char block_sign_shuffle_mask_1[32] = {
-        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
-        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
-    };
-    static const char block_sign_shuffle_mask_2[32] = {
-        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
-        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
-    };
-    static const uint8_t bit_selector_mask_bytes[32] = {
-        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m128i bit_selector_mask_0 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes);
-    const __m128i bit_selector_mask_1 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes + 1);
-    const __m128i block_sign_shuffle_1_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1);
-    const __m128i block_sign_shuffle_1_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1 + 1);
-    const __m128i block_sign_shuffle_2_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2);
-    const __m128i block_sign_shuffle_2_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2 + 1);
-
-    static const uint8_t k_bit_helper[32] = {
-        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
-        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
-    };
-    const __m128i bit_helper_0 = _mm_loadu_si128((const __m128i*)k_bit_helper);
-    const __m128i bit_helper_1 = _mm_loadu_si128((const __m128i*)k_bit_helper + 1);
-    const __m128i m511 = _mm_set1_epi16(511);
-    const __m128i m4 = _mm_set1_epi8(0xf);
-    const __m128i m1 = _mm_set1_epi8(1);
-
-    uint64_t aux64;
-
-    // somewhat hacky, but gives a significant boost in performance
-    __m256i aux_gindex;
-    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
-
-    __m256 accumf = _mm256_setzero_ps();
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
-
-        memcpy(&aux64, x[i].scales, 8);
-        __m128i stmp = _mm_set1_epi64x(aux64);
-        stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
-        const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
-
-        __m128i sumi1_0 = _mm_setzero_si128();
-        __m128i sumi1_1 = _mm_setzero_si128();
-        __m128i sumi2_0 = _mm_setzero_si128();
-        __m128i sumi2_1 = _mm_setzero_si128();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
-
-            const __m128i q2_data_0 = _mm_loadu_si128((const __m128i*)q2);
-            const __m128i q2_data_1 = _mm_loadu_si128((const __m128i*)q2 + 1);  q2 += 16;
-            aux_gindex = MM256_SET_M128I(_mm_and_si128(q2_data_1, m511), _mm_and_si128(q2_data_0, m511));
-
-            const __m128i partial_sign_bits_0 = _mm_srli_epi16(q2_data_0, 9);
-            const __m128i partial_sign_bits_1 = _mm_srli_epi16(q2_data_1, 9);
-            const __m128i partial_sign_bits_upper_0 = _mm_srli_epi16(q2_data_0, 13);
-            const __m128i partial_sign_bits_upper_1 = _mm_srli_epi16(q2_data_1, 13);
-            const __m128i partial_sign_bits_for_counting_0 = _mm_xor_si128(partial_sign_bits_0, partial_sign_bits_upper_0);
-            const __m128i partial_sign_bits_for_counting_1 = _mm_xor_si128(partial_sign_bits_1, partial_sign_bits_upper_1);
-
-            const __m128i odd_bits_0 = _mm_shuffle_epi8(bit_helper_0, partial_sign_bits_for_counting_0);
-            const __m128i odd_bits_1 = _mm_shuffle_epi8(bit_helper_1, partial_sign_bits_for_counting_1);
-            const __m128i full_sign_bits_0 = _mm_or_si128(partial_sign_bits_0, odd_bits_0);
-            const __m128i full_sign_bits_1 = _mm_or_si128(partial_sign_bits_1, odd_bits_1);
-
-            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_3_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_3_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_4_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_4_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-
-            const __m128i q2_1_0 = _mm_set_epi64x(iq2xs_grid[gindex[1]], iq2xs_grid[gindex[0]]);
-            const __m128i q2_1_1 = _mm_set_epi64x(iq2xs_grid[gindex[3]], iq2xs_grid[gindex[2]]);
-            const __m128i q2_2_0 = _mm_set_epi64x(iq2xs_grid[gindex[5]], iq2xs_grid[gindex[4]]);
-            const __m128i q2_2_1 = _mm_set_epi64x(iq2xs_grid[gindex[7]], iq2xs_grid[gindex[6]]);
-            const __m128i q2_3_0 = _mm_set_epi64x(iq2xs_grid[gindex[9]], iq2xs_grid[gindex[8]]);
-            const __m128i q2_3_1 = _mm_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]]);
-            const __m128i q2_4_0 = _mm_set_epi64x(iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
-            const __m128i q2_4_1 = _mm_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]]);
-
-            // AVX2 full_signs_1 is full_sign_bits_0 here
-            // AVX2 full_signs_2 is full_sign_bits_1 here
-            __m128i signs_0, signs_1;
-            signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_0);
-            signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_1);
-            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
-            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
-            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, _mm_or_si128(signs_0, mone));
-            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, _mm_or_si128(signs_1, mone));
-
-            signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_0);
-            signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_1);
-            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
-            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
-            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, _mm_or_si128(signs_0, mone));
-            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, _mm_or_si128(signs_1, mone));
-
-            signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_0);
-            signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_1);
-            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
-            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
-            const __m128i q8s_3_0 = _mm_sign_epi8(q8_3_0, _mm_or_si128(signs_0, mone));
-            const __m128i q8s_3_1 = _mm_sign_epi8(q8_3_1, _mm_or_si128(signs_1, mone));
-
-            signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_0);
-            signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_1);
-            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
-            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
-            const __m128i q8s_4_0 = _mm_sign_epi8(q8_4_0, _mm_or_si128(signs_0, mone));
-            const __m128i q8s_4_1 = _mm_sign_epi8(q8_4_1, _mm_or_si128(signs_1, mone));
-
-            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
-            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
-            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
-            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
-            const __m128i dot3_0  = _mm_maddubs_epi16(q2_3_0, q8s_3_0);
-            const __m128i dot3_1  = _mm_maddubs_epi16(q2_3_1, q8s_3_1);
-            const __m128i dot4_0  = _mm_maddubs_epi16(q2_4_0, q8s_4_0);
-            const __m128i dot4_1  = _mm_maddubs_epi16(q2_4_1, q8s_4_1);
-
-            __m128i sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0));
-            const __m128i sc1_0 = _mm_cvtepi8_epi16(sc_tmp);
-            const __m128i sc1_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
-            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1));
-            const __m128i sc2_0 = _mm_cvtepi8_epi16(sc_tmp);
-            const __m128i sc2_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
-            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2));
-            const __m128i sc3_0 = _mm_cvtepi8_epi16(sc_tmp);
-            const __m128i sc3_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
-            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3));
-            const __m128i sc4_0 = _mm_cvtepi8_epi16(sc_tmp);
-            const __m128i sc4_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
-
-            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot1_0, sc1_0));
-            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot1_1, sc1_1));
-            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot2_0, sc2_0));
-            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot2_1, sc2_1));
-            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot3_0, sc3_0));
-            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot3_1, sc3_1));
-            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot4_0, sc4_0));
-            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot4_1, sc4_1));
-        }
-
-        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
-
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#elif defined(__loongarch_asx)
-
-    const __m256i mone = __lasx_xvreplgr2vr_b(1);
-    static const char block_sign_shuffle_mask_1[32] = {
-        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
-        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
-    };
-    static const char block_sign_shuffle_mask_2[32] = {
-        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
-        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
-    };
-    static const uint8_t bit_selector_mask_bytes[32] = {
-        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m256i bit_selector_mask = __lasx_xvld((const __m256i*)bit_selector_mask_bytes, 0);
-    const __m256i block_sign_shuffle_1 = __lasx_xvld((const __m256i*)block_sign_shuffle_mask_1, 0);
-    const __m256i block_sign_shuffle_2 = __lasx_xvld((const __m256i*)block_sign_shuffle_mask_2, 0);
-
-    static const uint8_t k_bit_helper[32] = {
-        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
-        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
-    };
-    const __m256i bit_helper = __lasx_xvld((const __m256i*)k_bit_helper, 0);
-    const __m256i m511 = __lasx_xvreplgr2vr_h(511);
-    const __m128i m4 = __lsx_vreplgr2vr_b(0xf);
-    const __m128i m1 = __lsx_vreplgr2vr_b(1);
-
-    uint64_t aux64;
-
-    // somewhat hacky, but gives a significant boost in performance
-    __m256i aux_gindex;
-    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
-
-    __m256 accumf = (__m256)__lasx_xvldi(0);
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const int8_t   * restrict q8 = y[i].qs;
-
-        memcpy(&aux64, x[i].scales, 8);
-        __m128i stmp = __lsx_vreplgr2vr_d(aux64);
-        stmp = __lsx_vilvl_b( __lsx_vand_v(__lsx_vsrli_h(stmp, 4), m4), __lsx_vand_v(stmp, m4));
-        const __m128i scales = __lsx_vadd_b(__lsx_vslli_h(stmp, 1), m1);
-
-        __m256i sumi1 = __lasx_xvldi(0);
-        __m256i sumi2 = __lasx_xvldi(0);
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
-
-            const __m256i q2_data = __lasx_xvld((const __m256i*)q2, 0);  q2 += 16;
-            aux_gindex = __lasx_xvand_v(q2_data, m511);
-
-            const __m256i partial_sign_bits = __lasx_xvsrli_h(q2_data, 9);
-            const __m256i partial_sign_bits_upper = __lasx_xvsrli_h(q2_data, 13);
-            const __m256i partial_sign_bits_for_counting = __lasx_xvxor_v(partial_sign_bits, partial_sign_bits_upper);
-
-            const __m256i odd_bits = lasx_shuffle_b(bit_helper, partial_sign_bits_for_counting);
-            const __m256i full_sign_bits = __lasx_xvor_v(partial_sign_bits, odd_bits);
-
-            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            const __m256i q8_3 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            const __m256i q8_4 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-
-            const __m256i q2_1 = lasx_set_d(iq2xs_grid[gindex[ 3]], iq2xs_grid[gindex[ 2]],
-                                                   iq2xs_grid[gindex[ 1]], iq2xs_grid[gindex[ 0]]);
-            const __m256i q2_2 = lasx_set_d(iq2xs_grid[gindex[ 7]], iq2xs_grid[gindex[ 6]],
-                                                   iq2xs_grid[gindex[ 5]], iq2xs_grid[gindex[ 4]]);
-            const __m256i q2_3 = lasx_set_d(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]],
-                                                   iq2xs_grid[gindex[ 9]], iq2xs_grid[gindex[ 8]]);
-            const __m256i q2_4 = lasx_set_d(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]],
-                                                   iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
-
-            const __m128i full_signs_l = lasx_extracti128(full_sign_bits, 0);
-            const __m128i full_signs_h = lasx_extracti128(full_sign_bits, 1);
-            const __m256i full_signs_1 = lasx_insertf128(full_signs_l, full_signs_l);
-            const __m256i full_signs_2 = lasx_insertf128(full_signs_h, full_signs_h);
-
-            __m256i signs;
-            signs = lasx_shuffle_b(full_signs_1, block_sign_shuffle_1);
-            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_1 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_1);
-
-            signs = lasx_shuffle_b(full_signs_1, block_sign_shuffle_2);
-            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_2 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_2);
-
-            signs = lasx_shuffle_b(full_signs_2, block_sign_shuffle_1);
-            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_3 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_3);
-
-            signs = lasx_shuffle_b(full_signs_2, block_sign_shuffle_2);
-            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
-            const __m256i q8s_4 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_4);
-
-            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
-            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
-            const __m256i dot3  = lasx_maddubs_h(q2_3, q8s_3);
-            const __m256i dot4  = lasx_maddubs_h(q2_4, q8s_4);
-
-            const __m256i sc1 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+0)));
-            const __m256i sc2 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+1)));
-            const __m256i sc3 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+2)));
-            const __m256i sc4 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+3)));
-
-            sumi1 = __lasx_xvadd_w(sumi1, lasx_madd_h(dot1, sc1));
-            sumi2 = __lasx_xvadd_w(sumi2, lasx_madd_h(dot2, sc2));
-            sumi1 = __lasx_xvadd_w(sumi1, lasx_madd_h(dot3, sc3));
-            sumi2 = __lasx_xvadd_w(sumi2, lasx_madd_h(dot4, sc4));
-        }
-
-        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
-
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-#elif defined(__POWER9_VECTOR__)
-    const vector int v0 = vec_splats((int32_t)0);
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        const uint16_t * restrict q2 = x[i].qs;
-        const uint8_t  * restrict sc = x[i].scales;
-        const int8_t  *  restrict q8 = y[i].qs;
-
-        for (int j = 0; j < QK_K/64; ++j) {
-            __builtin_prefetch(q2, 0, 1);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2xs_grid + (q2[0] & 511)), *(const int64_t *)(iq2xs_grid + (q2[1] & 511))};
-            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2xs_grid + (q2[2] & 511)), *(const int64_t *)(iq2xs_grid + (q2[3] & 511))};
-            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2xs_grid + (q2[4] & 511)), *(const int64_t *)(iq2xs_grid + (q2[5] & 511))};
-            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2xs_grid + (q2[6] & 511)), *(const int64_t *)(iq2xs_grid + (q2[7] & 511))};
-
-            vector signed long long vsigns0 = {*(const int64_t *)(signs64 + ((q2[0] >> 9))), *(const int64_t *)(signs64 + ((q2[1] >> 9)))};
-            vector signed long long vsigns1 = {*(const int64_t *)(signs64 + ((q2[2] >> 9))), *(const int64_t *)(signs64 + ((q2[3] >> 9)))};
-            vector signed long long vsigns2 = {*(const int64_t *)(signs64 + ((q2[4] >> 9))), *(const int64_t *)(signs64 + ((q2[5] >> 9)))};
-            vector signed long long vsigns3 = {*(const int64_t *)(signs64 + ((q2[6] >> 9))), *(const int64_t *)(signs64 + ((q2[7] >> 9)))};
-            q2 += 8;
-
-            vector signed char q2x0 = (vector signed char)vec_mul((vector signed char)vsigns0, (vector signed char)aux64x2_0);
-            vector signed char q2x1 = (vector signed char)vec_mul((vector signed char)vsigns1, (vector signed char)aux64x2_1);
-            vector signed char q2x2 = (vector signed char)vec_mul((vector signed char)vsigns2, (vector signed char)aux64x2_2);
-            vector signed char q2x3 = (vector signed char)vec_mul((vector signed char)vsigns3, (vector signed char)aux64x2_3);
-
-            vector signed char q8y0 = vec_xl( 0, q8);
-            vector signed char q8y1 = vec_xl(16, q8);
-            vector signed char q8y2 = vec_xl(32, q8);
-            vector signed char q8y3 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
-            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
-            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
-            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
-
-            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
-            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
-            const uint16_t ls2 = (uint16_t)(sc[1] & 0xf);
-            const uint16_t ls3 = (uint16_t)(sc[1] >>  4);
-            sc += 2;
-
-            vector signed short vscales0 = vec_splats((int16_t)(2*ls0+1));
-            vector signed short vscales1 = vec_splats((int16_t)(2*ls1+1));
-            vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1));
-            vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1));
-
-            vsumi0 = vec_msum(qv0, vscales0, vsumi0);
-            vsumi1 = vec_msum(qv1, vscales1, vsumi1);
-            vsumi2 = vec_msum(qv2, vscales2, vsumi2);
-            vsumi3 = vec_msum(qv3, vscales3, vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = 0.125f * vec_extract(vsumf0, 0);
-#else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint16_t * restrict q2 = x[i].qs;
-        const uint8_t  * restrict sc = x[i].scales;
-        const int8_t   * restrict q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
-            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
-                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
-                for (int j = 0; j < 8; ++j) {
-                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += sumi * ls2;
-            q2 += 4;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.125f * sumf;
-#endif
-}
-
-void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(n % QK_K == 0);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_iq2_s * restrict x = vx;
-    const block_q8_K  * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_NEON)
-
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
-
-    const ggml_uint8x16x2_t mask1 = ggml_vld1q_u8_x2(k_mask1);
-    const uint8x16_t        mask2 = vld1q_u8(k_mask2);
-    const uint8x16_t m1 = vdupq_n_u8(1);
-    const int32x4_t vzero = vdupq_n_s32(0);
-
-    uint8x16x2_t vs;
-    ggml_int8x16x4_t q2s;
-    ggml_int8x16x4_t q8b;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-
-        const uint8_t * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
-        const int8_t  * restrict q8 = y[i].qs;
-
-        int sumi1 = 0, sumi2 = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
-            q2s.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[0] | ((qh[ib32+0] << 8) & 0x300)))),
-                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[1] | ((qh[ib32+0] << 6) & 0x300)))));
-            q2s.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[2] | ((qh[ib32+0] << 4) & 0x300)))),
-                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[3] | ((qh[ib32+0] << 2) & 0x300)))));
-            q2s.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[4] | ((qh[ib32+1] << 8) & 0x300)))),
-                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[5] | ((qh[ib32+1] << 6) & 0x300)))));
-            q2s.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[6] | ((qh[ib32+1] << 4) & 0x300)))),
-                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[7] | ((qh[ib32+1] << 2) & 0x300)))));
-            qs += 8;
-
-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16)));
-            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
-            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
-            vs.val[0] = vceqq_u8(vs.val[0], mask2);
-            vs.val[1] = vceqq_u8(vs.val[1], mask2);
-
-            q2s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[0], m1)), q2s.val[0]);
-            q2s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[1], m1)), q2s.val[1]);
-
-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16)));
-            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
-            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
-            vs.val[0] = vceqq_u8(vs.val[0], mask2);
-            vs.val[1] = vceqq_u8(vs.val[1], mask2);
-
-            signs += 4;
-
-            q2s.val[2] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[0], m1)), q2s.val[2]);
-            q2s.val[3] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[1], m1)), q2s.val[3]);
-
-            const int32x4_t p1 = ggml_vdotq_s32(vzero, q2s.val[0], q8b.val[0]);
-            const int32x4_t p2 = ggml_vdotq_s32(vzero, q2s.val[1], q8b.val[1]);
-            const int32x4_t p3 = ggml_vdotq_s32(vzero, q2s.val[2], q8b.val[2]);
-            const int32x4_t p4 = ggml_vdotq_s32(vzero, q2s.val[3], q8b.val[3]);
-
-            sumi1 += vaddvq_s32(p1) * (1 + 2*(x[i].scales[ib32+0] & 0xf));
-            sumi2 += vaddvq_s32(p2) * (1 + 2*(x[i].scales[ib32+0] >>  4));
-            sumi1 += vaddvq_s32(p3) * (1 + 2*(x[i].scales[ib32+1] & 0xf));
-            sumi2 += vaddvq_s32(p4) * (1 + 2*(x[i].scales[ib32+1] >>  4));
-        }
-        sumf += d*(sumi1 + sumi2);
-    }
-
-    *s = 0.125f * sumf;
-
-#elif defined(__AVX2__)
-
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m128i m4 = _mm_set1_epi8(0xf);
-    const __m128i m1 = _mm_set1_epi8(1);
-
-    const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1);
-    const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2);
-
-    uint64_t aux64;
-
-    __m256 accumf = _mm256_setzero_ps();
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
-        const int8_t  * restrict q8 = y[i].qs;
-
-        memcpy(&aux64, x[i].scales, 8);
-        const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1);
-        const __m256i scales16 = _mm256_cvtepi8_epi16(scales8); // 0 2 4 6 8 10 12 14 1 3 5 7 9 11 13 15
-
-        __m256i sumi1 = _mm256_setzero_si256();
-        __m256i sumi2 = _mm256_setzero_si256();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q2_1 = _mm256_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
-                                                   iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)],
-                                                   iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
-                                                   iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
-            const __m256i q2_2 = _mm256_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
-                                                   iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)],
-                                                   iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
-                                                   iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
-            qs += 8;
-
-            __m256i aux256 = _mm256_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
-            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
-            const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2);
-            const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1);
-
-            aux256 = _mm256_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
-            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
-            const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2);
-            const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2);
-
-            signs += 4;
-
-            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1); // blocks 2*ib32+0, 2*ib32+1
-            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2); // blocks 2*ib32+2, 2*ib32+3
-
-            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_shuffle_epi8(scales16, get_scale_shuffle_k4(ib32+0)));
-            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_shuffle_epi8(scales16, get_scale_shuffle_k4(ib32+1)));
-            sumi1 = _mm256_add_epi32(sumi1, p1);
-            sumi2 = _mm256_add_epi32(sumi2, p2);
-        }
-
-        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
-
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#elif defined(__AVX__)
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m128i m4 = _mm_set1_epi8(0xf);
-    const __m128i m1 = _mm_set1_epi8(1);
-
-    const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
-    const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
-    const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
-    const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
-
-    uint64_t aux64;
-
-    __m256 accumf = _mm256_setzero_ps();
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
-        const int8_t  * restrict q8 = y[i].qs;
-
-        memcpy(&aux64, x[i].scales, 8);
-        const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1);
-        const __m128i scales16_0 = _mm_cvtepi8_epi16(scales8);
-        const __m128i scales16_1 = _mm_cvtepi8_epi16(_mm_srli_si128(scales8, 8));
-
-        __m128i sumi1_0 = _mm_setzero_si128();
-        __m128i sumi1_1 = _mm_setzero_si128();
-        __m128i sumi2_0 = _mm_setzero_si128();
-        __m128i sumi2_1 = _mm_setzero_si128();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q2_1_0 = _mm_set_epi64x(iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
-                                                  iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
-            const __m128i q2_1_1 = _mm_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
-                                                  iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)]);
-            const __m128i q2_2_0 = _mm_set_epi64x(iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
-                                                  iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
-            const __m128i q2_2_1 = _mm_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
-                                                  iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)]);
-            qs += 8;
-
-            __m128i aux128_0 = _mm_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
-            __m128i aux128_1 = aux128_0;
-            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
-            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
-            const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
-            const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
-            const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
-            const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
-
-            aux128_0 = _mm_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
-            aux128_1 = aux128_0;
-            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
-            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
-            const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
-            const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
-            const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
-            const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
-
-            signs += 4;
-
-            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
-            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
-            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
-            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
-
-            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 0)));
-            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 1)));
-            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 0)));
-            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 1)));
-            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
-            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
-            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
-            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
-        }
-
-        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
-
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#elif defined(__POWER9_VECTOR__)
-    static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                        0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-    };
-
-    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
-
-    const vector int v0 = vec_splats((int32_t)0);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    const vector unsigned char mask0 = vec_xl( 0, k_mask1);
-    const vector unsigned char mask1 = vec_xl(16, k_mask1);
-    const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        const uint8_t *  restrict q2 = x[i].qs;
-        const uint8_t *  restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
-        const uint8_t *  restrict sc = x[i].scales;
-        const int8_t  *  restrict q8 = y[i].qs;
-
-        for (int j = 0; j < QK_K/32; j += 2) {
-            __builtin_prefetch(q2, 0, 1);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2s_grid + (q2[0] | ((qh[0] << 8) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[1] | ((qh[0] << 6) & 0x300)))};
-            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2s_grid + (q2[2] | ((qh[0] << 4) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[3] | ((qh[0] << 2) & 0x300)))};
-            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2s_grid + (q2[4] | ((qh[1] << 8) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[5] | ((qh[1] << 6) & 0x300)))};
-            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2s_grid + (q2[6] | ((qh[1] << 4) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[7] | ((qh[1] << 2) & 0x300)))};
-            q2 += 8;
-            qh += 2;
-
-            vector signed char vsigns01 = (vector signed char)vec_splats(*(const uint32_t *)&signs[0]);
-            vector signed char vsigns23 = (vector signed char)vec_splats(*(const uint32_t *)&signs[2]);
-            signs += 4;
-
-            vector signed char vsigns0 = vec_perm(vsigns01, vsigns01, mask0);
-            vector signed char vsigns1 = vec_perm(vsigns01, vsigns01, mask1);
-            vector signed char vsigns2 = vec_perm(vsigns23, vsigns23, mask0);
-            vector signed char vsigns3 = vec_perm(vsigns23, vsigns23, mask1);
-
-            vsigns0 = (vector signed char)vec_cmpeq(vec_and(vsigns0, mask2), mask2);
-            vsigns1 = (vector signed char)vec_cmpeq(vec_and(vsigns1, mask2), mask2);
-            vsigns2 = (vector signed char)vec_cmpeq(vec_and(vsigns2, mask2), mask2);
-            vsigns3 = (vector signed char)vec_cmpeq(vec_and(vsigns3, mask2), mask2);
-
-            vector signed char q2x0 = vec_sub(vec_xor(vsigns0, (vector signed char)aux64x2_0), vsigns0);
-            vector signed char q2x1 = vec_sub(vec_xor(vsigns1, (vector signed char)aux64x2_1), vsigns1);
-            vector signed char q2x2 = vec_sub(vec_xor(vsigns2, (vector signed char)aux64x2_2), vsigns2);
-            vector signed char q2x3 = vec_sub(vec_xor(vsigns3, (vector signed char)aux64x2_3), vsigns3);
-
-            vector signed char q8y0 = vec_xl( 0, q8);
-            vector signed char q8y1 = vec_xl(16, q8);
-            vector signed char q8y2 = vec_xl(32, q8);
-            vector signed char q8y3 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
-            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
-            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
-            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
-
-            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
-            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
-            const uint16_t ls2 = (uint16_t)(sc[1] & 0xf);
-            const uint16_t ls3 = (uint16_t)(sc[1] >>  4);
-            sc += 2;
-
-            vector signed short vscales0 = vec_splats((int16_t)(2*ls0+1));
-            vector signed short vscales1 = vec_splats((int16_t)(2*ls1+1));
-            vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1));
-            vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1));
-
-            vsumi0 = vec_msum(qv0, vscales0, vsumi0);
-            vsumi1 = vec_msum(qv1, vscales1, vsumi1);
-            vsumi2 = vec_msum(qv2, vscales2, vsumi2);
-            vsumi3 = vec_msum(qv3, vscales3, vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = 0.125f * vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-
-    const __m128i m4 = __lsx_vreplgr2vr_b(0xf);
-    const __m128i m1 = __lsx_vreplgr2vr_b(1);
-
-    const __m256i mask1 = __lasx_xvld((const __m256i*)k_mask1, 0);
-    const __m256i mask2 = __lasx_xvld((const __m256i*)k_mask2, 0);
-    uint64_t aux64;
-
-    __m256 accumf = (__m256)__lasx_xvldi(0);
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
-        const int8_t  * restrict q8 = y[i].qs;
-
-        __m128i tmp1;
-        memcpy(&aux64, x[i].scales, 8);
-        tmp1 = __lsx_vinsgr2vr_d(tmp1, aux64, 0);
-        tmp1 = __lsx_vinsgr2vr_d(tmp1, aux64 >> 4, 1);
-        const __m128i scales8 = __lsx_vadd_b(__lsx_vslli_h(__lsx_vand_v(tmp1, m4), 1), m1);
-        const __m256i scales16 = lasx_ext8_16(scales8); // 0 2 4 6 8 10 12 14 1 3 5 7 9 11 13 15
-
-        __m256i sumi1 = __lasx_xvldi(0);
-        __m256i sumi2 = __lasx_xvldi(0);
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            const __m256i q2_1 = lasx_set_d(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
-                                                   iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)],
-                                                   iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
-                                                   iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
-            const __m256i q2_2 = lasx_set_d(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
-                                                   iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)],
-                                                   iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
-                                                   iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
-            qs += 8;
-
-            __m256i aux256 = __lasx_xvreplgr2vr_w(signs[0] | ((uint32_t) signs[1] << 16));
-            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
-            const __m256i s2_1 = __lasx_xvseq_b(aux256, mask2);
-            const __m256i q8s_1 = __lasx_xvsub_b(__lasx_xvxor_v(s2_1, q8_1), s2_1);
-
-            aux256 = __lasx_xvreplgr2vr_w(signs[2] | ((uint32_t) signs[3] << 16));
-            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
-            const __m256i s2_2 = __lasx_xvseq_b(aux256, mask2);
-            const __m256i q8s_2 = __lasx_xvsub_b(__lasx_xvxor_v(s2_2, q8_2), s2_2);
-
-            signs += 4;
-
-            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1); // blocks 2*ib32+0, 2*ib32+1
-            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2); // blocks 2*ib32+2, 2*ib32+3
-
-            const __m256i p1 = lasx_madd_h(dot1, lasx_shuffle_b(scales16, get_scale_shuffle_k4(ib32+0)));
-            const __m256i p2 = lasx_madd_h(dot2, lasx_shuffle_b(scales16, get_scale_shuffle_k4(ib32+1)));
-            sumi1 = __lasx_xvadd_w(sumi1, p1);
-            sumi2 = __lasx_xvadd_w(sumi2, p2);
-        }
-
-        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
-    }
-
-    *s = 0.125f * hsum_float_8(accumf);
-
-#else
-
-    float sumf = 0;
-    for (int i = 0; i < nb; i++) {
-
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const int8_t  * q8 = y[i].qs;
-        const uint8_t * qs = x[i].qs;
-        const uint8_t * qh = x[i].qh;
-        const uint8_t * signs = qs + QK_K/8;
-
-        int bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
-            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
-            int sumi1 = 0, sumi2 = 0;
-            for (int l = 0; l < 2; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            for (int l = 2; l < 4; ++l) {
-                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
-                for (int j = 0; j < 8; ++j) {
-                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            bsum += ls1 * sumi1 + ls2 * sumi2;
-            qs += 4;
-            signs += 4;
-        }
-
-        sumf += d * bsum;
-    }
-
-    *s = 0.125f * sumf;
-
-#endif
-
-}
-
-void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(n % QK_K == 0);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_iq3_xxs * restrict x = vx;
-    const block_q8_K    * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_NEON)
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[2];
-
-    ggml_int8x16x4_t q3s;
-    ggml_int8x16x4_t q8b;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict gas = x[i].qs + QK_K/4;
-        const int8_t   * restrict q8 = y[i].qs;
-        float sumf1 = 0, sumf2 = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
-            memcpy(aux32, gas, 2*sizeof(uint32_t)); gas += 2*sizeof(uint32_t);
-            const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]);
-            const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]);
-            const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]);
-            const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]);
-            q3 += 16;
-            q3s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >>  7) & 127))));
-            q3s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >> 21) & 127))));
-            q3s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >>  7) & 127))));
-            q3s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 21) & 127))));
-            q3s.val[0] = vmulq_s8(q3s.val[0], vreinterpretq_s8_u32(aux32x4_0));
-            q3s.val[1] = vmulq_s8(q3s.val[1], vreinterpretq_s8_u32(aux32x4_1));
-            q3s.val[2] = vmulq_s8(q3s.val[2], vreinterpretq_s8_u32(aux32x4_2));
-            q3s.val[3] = vmulq_s8(q3s.val[3], vreinterpretq_s8_u32(aux32x4_3));
-            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]);
-            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]);
-            sumf1 += vaddvq_s32(p1) * (0.5f + (aux32[0] >> 28));
-            sumf2 += vaddvq_s32(p2) * (0.5f + (aux32[1] >> 28));
-        }
-        sumf += d*(sumf1 + sumf2);
-    }
-    *s = 0.5f * sumf;
-
-#elif defined(__AVX2__)
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[2];
-
-    __m256 accumf = _mm256_setzero_ps();
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict gas = x[i].qs + QK_K/4;
-        const int8_t  * restrict q8 = y[i].qs;
-        __m256i sumi1 = _mm256_setzero_si256();
-        __m256i sumi2 = _mm256_setzero_si256();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q2_1 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
-                                                  iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
-            q3 += 8;
-            const __m256i q2_2 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
-                                                  iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
-            q3 += 8;
-            memcpy(aux32, gas, 8); gas += 8;
-            const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127],
-                                                   signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
-            const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
-                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
-            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
-            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
-            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
-            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
-            const uint16_t ls1 = aux32[0] >> 28;
-            const uint16_t ls2 = aux32[1] >> 28;
-            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
-            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
-            sumi1 = _mm256_add_epi32(sumi1, p1);
-            sumi2 = _mm256_add_epi32(sumi2, p2);
-        }
-
-        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
-
-    }
-
-    *s = 0.25f * hsum_float_8(accumf);
-
-#elif defined(__AVX__)
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[2];
-
-    __m256 accumf = _mm256_setzero_ps();
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict gas = x[i].qs + QK_K/4;
-        const int8_t  * restrict q8 = y[i].qs;
-        __m128i sumi1_0 = _mm_setzero_si128();
-        __m128i sumi1_1 = _mm_setzero_si128();
-        __m128i sumi2_0 = _mm_setzero_si128();
-        __m128i sumi2_1 = _mm_setzero_si128();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q2_1_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
-            const __m128i q2_1_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
-            q3 += 8;
-            const __m128i q2_2_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
-            const __m128i q2_2_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
-            q3 += 8;
-            memcpy(aux32, gas, 8); gas += 8;
-            const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
-            const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127]);
-            const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
-            const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
-            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
-            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
-            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
-            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
-            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
-            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
-            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
-            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
-            const uint16_t ls1 = aux32[0] >> 28;
-            const uint16_t ls2 = aux32[1] >> 28;
-            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
-            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
-            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
-            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
-            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
-            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
-            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
-            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
-        }
-
-        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
-
-    }
-
-    *s = 0.25f * hsum_float_8(accumf);
-
-#elif defined(__POWER9_VECTOR__)
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    const vector int v0 = vec_splats((int32_t)0);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint32_t * restrict signs = (const uint32_t *)(x[i].qs + QK_K/4);
-        const int8_t  * restrict q8 = y[i].qs;
-
-#pragma GCC unroll 1
-        for (int j = 0; j < QK_K/32; j += 2) {
-            __builtin_prefetch(q3, 0, 1);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector unsigned int aux32x4_0 = {iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]};
-            vector unsigned int aux32x4_1 = {iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]};
-            vector unsigned int aux32x4_2 = {iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]};
-            vector unsigned int aux32x4_3 = {iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]};
-            q3 += 16;
-
-            vector unsigned long long aux64x2_0 = {(uint64_t)(signs64[(signs[0] >>  0) & 127]), (uint64_t)(signs64[(signs[0] >>  7) & 127])};
-            vector unsigned long long aux64x2_1 = {(uint64_t)(signs64[(signs[0] >> 14) & 127]), (uint64_t)(signs64[(signs[0] >> 21) & 127])};
-            vector unsigned long long aux64x2_2 = {(uint64_t)(signs64[(signs[1] >>  0) & 127]), (uint64_t)(signs64[(signs[1] >>  7) & 127])};
-            vector unsigned long long aux64x2_3 = {(uint64_t)(signs64[(signs[1] >> 14) & 127]), (uint64_t)(signs64[(signs[1] >> 21) & 127])};
-
-            vector signed char q3x0 = vec_mul((vector signed char)aux64x2_0, (vector signed char)aux32x4_0);
-            vector signed char q3x1 = vec_mul((vector signed char)aux64x2_1, (vector signed char)aux32x4_1);
-            vector signed char q3x2 = vec_mul((vector signed char)aux64x2_2, (vector signed char)aux32x4_2);
-            vector signed char q3x3 = vec_mul((vector signed char)aux64x2_3, (vector signed char)aux32x4_3);
-
-            vector signed char q8y0 = vec_xl( 0, q8);
-            vector signed char q8y1 = vec_xl(16, q8);
-            vector signed char q8y2 = vec_xl(32, q8);
-            vector signed char q8y3 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed short qv0 = vec_add(vec_mule(q3x0, q8y0), vec_mulo(q3x0, q8y0));
-            vector signed short qv1 = vec_add(vec_mule(q3x1, q8y1), vec_mulo(q3x1, q8y1));
-            vector signed short qv2 = vec_add(vec_mule(q3x2, q8y2), vec_mulo(q3x2, q8y2));
-            vector signed short qv3 = vec_add(vec_mule(q3x3, q8y3), vec_mulo(q3x3, q8y3));
-
-            const uint16_t ls0 = (uint16_t)(signs[0] >> 28);
-            const uint16_t ls1 = (uint16_t)(signs[1] >> 28);
-            signs += 2;
-
-            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
-            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
-
-            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
-            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
-            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
-            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = 0.25f * vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-
-    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
-
-    uint32_t aux32[2];
-
-    __m256 accumf = (__m256)__lasx_xvldi(0);
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict gas = x[i].qs + QK_K/4;
-        const int8_t  * restrict q8 = y[i].qs;
-        __m256i sumi1 = __lasx_xvldi(0);
-        __m256i sumi2 = __lasx_xvldi(0);
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            const __m256i q2_1 = lasx_set_w(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
-                                                iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
-            q3 += 8;
-            const __m256i q2_2 = lasx_set_w(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
-                                                iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
-            q3 += 8;
-            memcpy(aux32, gas, 8); gas += 8;
-
-            const __m256i s2_1 = lasx_set_d(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127],
-                                                   signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
-            const __m256i s2_2 = lasx_set_d(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
-                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
-            const __m256i q8s_1 = __lasx_xvsigncov_b(s2_1, q8_1);
-            const __m256i q8s_2 = __lasx_xvsigncov_b(s2_2, q8_2);
-            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
-            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
-            const uint16_t ls1 = aux32[0] >> 28;
-            const uint16_t ls2 = aux32[1] >> 28;
-
-            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
-            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
-            sumi1 = __lasx_xvadd_w(sumi1, p1);
-            sumi2 = __lasx_xvadd_w(sumi2, p2);
-        }
-
-        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
-    }
-
-    *s = 0.25f * hsum_float_8(accumf);
-
-#else
-
-    uint32_t aux32;
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict gas = x[i].qs + QK_K/4;
-        const int8_t  * restrict q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
-            memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
-            const uint32_t ls = 2*(aux32 >> 28) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
-                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
-                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            q3 += 8;
-            bsum += sumi * ls;
-        }
-        sumf += d * bsum;
-    }
-    *s = 0.25f * sumf;
-#endif
-}
-
-void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(n % QK_K == 0);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_iq3_s * restrict x = vx;
-    const block_q8_K  * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined(__ARM_NEON)
-
-    typedef union {
-        uint16x8_t vec_index;
-        uint16_t   index[8];
-    } vec_index_t;
-
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
-
-    static const int16_t k_shift[8] = {8, 7, 6, 5, 4, 3, 2, 1};
-
-    const ggml_uint8x16x2_t mask1 = ggml_vld1q_u8_x2(k_mask1);
-    const uint8x16_t        mask2 = vld1q_u8(k_mask2);
-
-    const int16x8_t  hshift = vld1q_s16(k_shift);
-    const uint16x8_t m256   = vdupq_n_u16(256);
-    const uint8x16_t m1     = vdupq_n_u8(1);
-
-    uint8x16x2_t vs;
-    ggml_int8x16x4_t q3s;
-    ggml_int8x16x4_t q8b;
-    vec_index_t idx;
-
-    uint32_t scales32[2];
-    const uint8_t * scales8 = (const uint8_t *)scales32;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
-        const int8_t   * restrict q8 = y[i].qs;
-
-        memcpy(scales32, x[i].scales, 4);
-        scales32[1] = (((scales32[0] >> 4) & 0x0f0f0f0f) << 1) | 0x01010101;
-        scales32[0] = ((scales32[0] & 0x0f0f0f0f) << 1) | 0x01010101;
-
-        int sumi1 = 0, sumi2 = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            const uint8x16_t idx_l = vld1q_u8(qs); qs += 16;
-            idx.vec_index = vorrq_u16(vmovl_u8(vget_low_u8 (idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+0]), hshift), m256));
-            const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]],
-                                                        iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]);
-            const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]],
-                                                        iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]);
-            idx.vec_index = vorrq_u16(vmovl_u8(vget_high_u8(idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+1]), hshift), m256));
-            const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]],
-                                                        iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]);
-            const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]],
-                                                        iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]);
-
-
-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16)));
-            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
-            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
-            vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1);
-            vs.val[1] = vorrq_u8(vceqq_u8(vs.val[1], mask2), m1);
-
-            q3s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_0));
-            q3s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_1));
-
-            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16)));
-            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
-            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
-            vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1);
-            vs.val[1] = vorrq_u8(vceqq_u8(vs.val[1], mask2), m1);
-
-            signs += 4;
-
-            q3s.val[2] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_2));
-            q3s.val[3] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_3));
-
-            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]);
-            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]);
-
-            sumi1 += vaddvq_s32(p1) * scales8[ib32/2+0];
-            sumi2 += vaddvq_s32(p2) * scales8[ib32/2+4];
-        }
-        sumf += d*(sumi1 + sumi2);
-    }
-    *s = sumf;
-
-#elif defined(__AVX2__)
-
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1);
-    const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2);
-
-    const __m256i idx_shift = _mm256_set_epi32(1, 2, 3, 4, 5, 6, 7, 8);
-    const __m256i idx_mask  = _mm256_set1_epi32(256);
-
-    typedef union {
-        __m256i  vec[2];
-        uint32_t index[16];
-    } index_t;
-
-    index_t idx;
-
-    __m256 accumf = _mm256_setzero_ps();
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
-        const int8_t  * restrict q8 = y[i].qs;
-        __m256i sumi1 = _mm256_setzero_si256();
-        __m256i sumi2 = _mm256_setzero_si256();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i idx_l = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)qs)); qs += 16;
-            idx.vec[0] = _mm256_set1_epi32(qh[ib32+0]);
-            idx.vec[1] = _mm256_set1_epi32(qh[ib32+1]);
-            idx.vec[0] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[0], idx_shift), idx_mask);
-            idx.vec[1] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[1], idx_shift), idx_mask);
-            idx.vec[0] = _mm256_or_si256(idx.vec[0], _mm256_cvtepi16_epi32(_mm256_castsi256_si128(idx_l)));
-            idx.vec[1] = _mm256_or_si256(idx.vec[1], _mm256_cvtepi16_epi32(_mm256_extractf128_si256(idx_l, 1)));
-
-            // At leat on my CPU (Ryzen 7950X), using _mm256_i32gather_epi32 is slower than _mm256_set_epi32. Strange.
-            //const __m256i q2_1 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[0], 4);
-            //const __m256i q2_2 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[1], 4);
-            const __m256i q2_1 = _mm256_set_epi32(
-                    iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]],
-                    iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]
-            );
-            const __m256i q2_2 = _mm256_set_epi32(
-                    iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]],
-                    iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[ 9]], iq3s_grid[idx.index[ 8]]
-            );
-
-            __m256i aux256 = _mm256_set1_epi32(signs[0] | (signs[1] << 16));
-            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
-            const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2);
-            const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1);
-
-            aux256 = _mm256_set1_epi32(signs[2] | (signs[3] << 16));
-            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
-            const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2);
-            const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2);
-
-            signs += 4;
-
-            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
-            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
-            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
-            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
-            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
-            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
-            sumi1 = _mm256_add_epi32(sumi1, p1);
-            sumi2 = _mm256_add_epi32(sumi2, p2);
-        }
-
-        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
-
-    }
-
-    *s = hsum_float_8(accumf);
-
-#elif defined(__AVX__)
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
-    const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
-    const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
-    const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
-
-    const __m128i idx_mul_0 = _mm_set_epi32(32, 64, 128, 256);
-    const __m128i idx_mul_1 = _mm_set_epi32(2, 4, 8, 16);
-    const __m128i idx_mask  = _mm_set1_epi32(256);
-
-    typedef union {
-        __m128i  vec[4];
-        uint32_t index[16];
-    } index_t;
-
-    index_t idx;
-
-    __m256 accumf = _mm256_setzero_ps();
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
-        const int8_t  * restrict q8 = y[i].qs;
-        __m128i sumi1_0 = _mm_setzero_si128();
-        __m128i sumi1_1 = _mm_setzero_si128();
-        __m128i sumi2_0 = _mm_setzero_si128();
-        __m128i sumi2_1 = _mm_setzero_si128();
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i qs_tmp = _mm_loadu_si128((const __m128i *)qs);
-            const __m128i idx_l_0 = _mm_cvtepu8_epi16(qs_tmp);
-            const __m128i idx_l_1 = _mm_cvtepu8_epi16(_mm_srli_si128(qs_tmp, 8)); qs += 16;
-            idx.vec[0] = _mm_set1_epi32(qh[ib32+0]);
-            idx.vec[1] = idx.vec[0];
-            idx.vec[2] = _mm_set1_epi32(qh[ib32+1]);
-            idx.vec[3] = idx.vec[2];
-
-            idx.vec[0] = _mm_and_si128(_mm_mullo_epi32(idx.vec[0], idx_mul_0), idx_mask);
-            idx.vec[1] = _mm_and_si128(_mm_mullo_epi32(idx.vec[1], idx_mul_1), idx_mask);
-            idx.vec[2] = _mm_and_si128(_mm_mullo_epi32(idx.vec[2], idx_mul_0), idx_mask);
-            idx.vec[3] = _mm_and_si128(_mm_mullo_epi32(idx.vec[3], idx_mul_1), idx_mask);
-
-            idx.vec[0] = _mm_or_si128(idx.vec[0], _mm_cvtepi16_epi32(idx_l_0));
-            idx.vec[1] = _mm_or_si128(idx.vec[1], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_0, 8)));
-            idx.vec[2] = _mm_or_si128(idx.vec[2], _mm_cvtepi16_epi32(idx_l_1));
-            idx.vec[3] = _mm_or_si128(idx.vec[3], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_1, 8)));
-
-            const __m128i q2_1_0 = _mm_set_epi32(iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]);
-            const __m128i q2_1_1 = _mm_set_epi32(iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]]);
-            const __m128i q2_2_0 = _mm_set_epi32(iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[9]], iq3s_grid[idx.index[8]]);
-            const __m128i q2_2_1 = _mm_set_epi32(iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]]);
-
-            __m128i aux128_0 = _mm_set1_epi32(signs[0] | (signs[1] << 16));
-            __m128i aux128_1 = aux128_0;
-            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
-            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
-            const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
-            const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
-            const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
-            const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
-
-            aux128_0 = _mm_set1_epi32(signs[2] | (signs[3] << 16));
-            aux128_1 = aux128_0;
-            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
-            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
-            const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
-            const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
-            const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
-            const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
-
-            signs += 4;
-
-            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
-            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
-            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
-            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
-            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
-            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
-            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
-            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
-            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
-            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
-            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
-            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
-            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
-            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
-        }
-
-        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
-
-    }
-
-    *s = hsum_float_8(accumf);
-
-#elif defined(__POWER9_VECTOR__)
-    static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                        0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-    };
-
-    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
-
-    const vector int v0 = vec_splats((int32_t)0);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    const vector unsigned char mask0 = vec_xl( 0, k_mask1);
-    const vector unsigned char mask1 = vec_xl(16, k_mask1);
-    const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        const uint8_t *  restrict q3 = x[i].qs;
-        const uint8_t *  restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)(x[i].signs);
-        const uint8_t *  restrict sc = x[i].scales;
-        const int8_t  *  restrict q8 = y[i].qs;
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        for (int j = 0; j < QK_K/32; j += 2) {
-            __builtin_prefetch(q3, 0, 1);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector unsigned int aux32x4_0 = {iq3s_grid[q3[ 0] | ((qh[0] << 8) & 256)], iq3s_grid[q3[ 1] | ((qh[0] << 7) & 256)],
-                                             iq3s_grid[q3[ 2] | ((qh[0] << 6) & 256)], iq3s_grid[q3[ 3] | ((qh[0] << 5) & 256)]};
-            vector unsigned int aux32x4_1 = {iq3s_grid[q3[ 4] | ((qh[0] << 4) & 256)], iq3s_grid[q3[ 5] | ((qh[0] << 3) & 256)],
-                                             iq3s_grid[q3[ 6] | ((qh[0] << 2) & 256)], iq3s_grid[q3[ 7] | ((qh[0] << 1) & 256)]};
-            vector unsigned int aux32x4_2 = {iq3s_grid[q3[ 8] | ((qh[1] << 8) & 256)], iq3s_grid[q3[ 9] | ((qh[1] << 7) & 256)],
-                                             iq3s_grid[q3[10] | ((qh[1] << 6) & 256)], iq3s_grid[q3[11] | ((qh[1] << 5) & 256)]};
-            vector unsigned int aux32x4_3 = {iq3s_grid[q3[12] | ((qh[1] << 4) & 256)], iq3s_grid[q3[13] | ((qh[1] << 3) & 256)],
-                                             iq3s_grid[q3[14] | ((qh[1] << 2) & 256)], iq3s_grid[q3[15] | ((qh[1] << 1) & 256)]};
-            q3 += 16;
-            qh += 2;
-
-            vector signed char vsigns01 = (vector signed char)vec_splats(*(const uint32_t *)&signs[0]);
-            vector signed char vsigns02 = (vector signed char)vec_splats(*(const uint32_t *)&signs[2]);
-            signs += 4;
-
-            vector signed char vsigns0 = vec_perm(vsigns01, vsigns01, mask0);
-            vector signed char vsigns1 = vec_perm(vsigns01, vsigns01, mask1);
-            vector signed char vsigns2 = vec_perm(vsigns02, vsigns02, mask0);
-            vector signed char vsigns3 = vec_perm(vsigns02, vsigns02, mask1);
-
-            vsigns0 = (vector signed char)vec_cmpeq(vec_and(vsigns0, mask2), mask2);
-            vsigns1 = (vector signed char)vec_cmpeq(vec_and(vsigns1, mask2), mask2);
-            vsigns2 = (vector signed char)vec_cmpeq(vec_and(vsigns2, mask2), mask2);
-            vsigns3 = (vector signed char)vec_cmpeq(vec_and(vsigns3, mask2), mask2);
-
-            vector signed char q3x0 = vec_sub(vec_xor(vsigns0, (vector signed char)aux32x4_0), vsigns0);
-            vector signed char q3x1 = vec_sub(vec_xor(vsigns1, (vector signed char)aux32x4_1), vsigns1);
-            vector signed char q3x2 = vec_sub(vec_xor(vsigns2, (vector signed char)aux32x4_2), vsigns2);
-            vector signed char q3x3 = vec_sub(vec_xor(vsigns3, (vector signed char)aux32x4_3), vsigns3);
-
-            vector signed char q8y0 = vec_xl( 0, q8);
-            vector signed char q8y1 = vec_xl(16, q8);
-            vector signed char q8y2 = vec_xl(32, q8);
-            vector signed char q8y3 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed short qv0 = vec_add(vec_mule(q3x0, q8y0), vec_mulo(q3x0, q8y0));
-            vector signed short qv1 = vec_add(vec_mule(q3x1, q8y1), vec_mulo(q3x1, q8y1));
-            vector signed short qv2 = vec_add(vec_mule(q3x2, q8y2), vec_mulo(q3x2, q8y2));
-            vector signed short qv3 = vec_add(vec_mule(q3x3, q8y3), vec_mulo(q3x3, q8y3));
-
-            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
-            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
-            sc ++;
-
-            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
-            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
-
-            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
-            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
-            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
-            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-
-   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
-                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
-   };
-
-    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
-    };
-
-    const __m256i mask1 = __lasx_xvld((const __m256i*)k_mask1, 0);
-    const __m256i mask2 = __lasx_xvld((const __m256i*)k_mask2, 0);
-
-    __m256i idx_shift = lasx_set_w(1, 2, 3, 4, 5, 6, 7, 8);
-    const __m256i idx_mask  = __lasx_xvreplgr2vr_w(256);
-
-    typedef union {
-        __m256i  vec[2];
-        uint32_t index[16];
-    } index_t;
-
-    index_t idx;
-
-    __m256 accumf = (__m256)__lasx_xvldi(0);
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
-        const int8_t  * restrict q8 = y[i].qs;
-        __m256i sumi1 = __lasx_xvldi(0);
-        __m256i sumi2 = __lasx_xvldi(0);
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            const __m256i idx_l = lasx_extu8_16(__lsx_vld(qs, 0)); qs += 16;
-            idx.vec[0] = __lasx_xvreplgr2vr_w(qh[ib32+0]);
-            idx.vec[1] = __lasx_xvreplgr2vr_w(qh[ib32+1]);
-            idx.vec[0] = __lasx_xvand_v(__lasx_xvsll_w(idx.vec[0], idx_shift), idx_mask);
-            idx.vec[1] = __lasx_xvand_v(__lasx_xvsll_w(idx.vec[1], idx_shift), idx_mask);
-            idx.vec[0] = __lasx_xvor_v(idx.vec[0], lasx_ext16_32(lasx_extracti128(idx_l, 0)));
-            idx.vec[1] = __lasx_xvor_v(idx.vec[1], lasx_ext16_32(lasx_extracti128(idx_l, 1)));
-
-            // At leat on my CPU (Ryzen 7950X), using _mm256_i32gather_epi32 is slower than _mm256_set_epi32. Strange.
-            //const __m256i q2_1 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[0], 4);
-            //const __m256i q2_2 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[1], 4);
-            const __m256i q2_1 = lasx_set_w(
-                    iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]],
-                    iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]
-            );
-            const __m256i q2_2 = lasx_set_w(
-                    iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]],
-                    iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[ 9]], iq3s_grid[idx.index[ 8]]
-            );
-
-            __m256i aux256 = __lasx_xvreplgr2vr_w(signs[0] | (signs[1] << 16));
-            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
-            const __m256i s2_1 = __lasx_xvseq_b(aux256, mask2);
-            const __m256i q8s_1 = __lasx_xvsub_b(__lasx_xvxor_v(s2_1, q8_1), s2_1);
-
-            aux256 = __lasx_xvreplgr2vr_w(signs[2] | (signs[3] << 16));
-            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
-            const __m256i s2_2 = __lasx_xvseq_b(aux256, mask2);
-            const __m256i q8s_2 = __lasx_xvsub_b(__lasx_xvxor_v(s2_2, q8_2), s2_2);
-
-            signs += 4;
-
-            const __m256i dot1 = lasx_maddubs_h(q2_1, q8s_1);
-            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
-            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
-            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
-            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
-            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
-            sumi1 = __lasx_xvadd_w(sumi1, p1);
-            sumi2 = __lasx_xvadd_w(sumi2, p2);
-        }
-
-        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
-    }
-
-    *s = hsum_float_8(accumf);
-
-#else
-
-    float sumf = 0.f;
-    for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const uint8_t * restrict qs = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const uint8_t * restrict signs = x[i].signs;
-        const int8_t  * restrict q8 = y[i].qs;
-        int32_t bsum = 0;
-        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
-            const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
-            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
-            int32_t sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls1;
-            sumi = 0;
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
-                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
-                for (int j = 0; j < 4; ++j) {
-                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
-                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
-                }
-                q8 += 8;
-            }
-            qs += 8;
-            signs += 4;
-            bsum += sumi * ls2;
-        }
-        sumf += d * bsum;
-    }
-    *s = sumf;
-#endif
-}
-
-#if defined(__AVX2__)
-static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
-    const __m256i ax = _mm256_sign_epi8(x, x);
-    const __m256i sy = _mm256_sign_epi8(y, x);
-    return _mm256_maddubs_epi16(ax, sy);
-}
-#elif defined(__loongarch_asx)
-static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
-    const __m256i ax = __lasx_xvsigncov_b(x, x);
-    const __m256i sy = __lasx_xvsigncov_b(x, y);
-    __m256i tmp1, tmp2, tmp3;
-    tmp1 = __lasx_xvmulwev_h_bu_b(ax, sy);
-    tmp2 = __lasx_xvmulwod_h_bu_b(ax, sy);
-    tmp3 = __lasx_xvadd_h(tmp1, tmp2);
-    return __lasx_xvsat_h(tmp3, 15);
-}
-#endif
-
-void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(n % QK_K == 0);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_iq1_s * restrict x = vx;
-    const block_q8_K  * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined __ARM_NEON
-
-    ggml_int8x16x4_t q1b;
-    ggml_int8x16x4_t q8b;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        int sumi1 = 0, sumi2 = 0, sumi3 = 0;
-
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-
-            q1b.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[0] | ((qh[ib+0] << 8) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[1] | ((qh[ib+0] << 5) & 0x700)))));
-            q1b.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[2] | ((qh[ib+0] << 2) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[3] | ((qh[ib+0] >> 1) & 0x700)))));
-            q1b.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[4] | ((qh[ib+1] << 8) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[5] | ((qh[ib+1] << 5) & 0x700)))));
-            q1b.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[6] | ((qh[ib+1] << 2) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[7] | ((qh[ib+1] >> 1) & 0x700)))));
-            qs += 8;
-
-            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q1b.val[0], q8b.val[0]), q1b.val[1], q8b.val[1]);
-            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q1b.val[2], q8b.val[2]), q1b.val[3], q8b.val[3]);
-
-            const int ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
-            const int ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
-            sumi1 += vaddvq_s32(p1) * ls1;
-            sumi2 += vaddvq_s32(p2) * ls2;
-            sumi3 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * ls1 * (qh[ib+0] & 0x8000 ? -1 : 1)
-                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * ls2 * (qh[ib+1] & 0x8000 ? -1 : 1);
-
-        }
-
-        sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3);
-    }
-
-    *s = sumf;
-
-#elif defined __AVX2__
-
-    __m256 accum = _mm256_setzero_ps();
-    float accum1 = 0;
-    for (int i = 0; i < nb; ++i) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        __m256i sumi = _mm256_setzero_si256();
-        int sumi1 = 0;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)],
-                                                    iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]);
-            const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)],
-                                                    iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]);
-            qs += 8;
-            const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
-            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
-            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
-            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
-            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(ls1));
-            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(ls2));
-
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p1, p2));
-            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
-                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
-        }
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        accum = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi), accum);
-        accum1 += d * sumi1;
-
-    }
-
-    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
-
-#elif defined __AVX__
-    __m256 accum = _mm256_setzero_ps();
-    float accum1 = 0;
-    for (int i = 0; i < nb; ++i) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
-
-        __m128i sumi1_0 = _mm_setzero_si128();
-        __m128i sumi1_1 = _mm_setzero_si128();
-        int sumi1 = 0;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const __m128i q1b_1_0 = _mm_set_epi64x(iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]);
-            const __m128i q1b_1_1 = _mm_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)]);
-            const __m128i q1b_2_0 = _mm_set_epi64x(iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]);
-            const __m128i q1b_2_1 = _mm_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)]);
-            qs += 8;
-            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-
-            const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
-            const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
-            const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
-            const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
-            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
-            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
-            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(ls1));
-            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(ls1));
-            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(ls2));
-            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(ls2));
-
-            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
-            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
-            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
-                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
-        }
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum);
-        accum1 += d * sumi1;
-
-    }
-
-    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
-
-#elif defined(__POWER9_VECTOR__)
-    const vector unsigned char v0 = vec_splats((unsigned char)0x0);
-    const vector unsigned short vsign = vec_splats((unsigned short)0x8000);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
-        vector float vyd = vec_splats(y[i].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed int vsumi0 = vec_splats((int32_t)0);
-        vector signed int vsumi1 = vec_splats((int32_t)0);
-        vector signed int vsumi2 = vec_splats((int32_t)0);
-        vector signed int vsumi3 = vec_splats((int32_t)0);
-        vector signed int vsumi8 = vec_splats((int32_t)0);
-
-        const uint8_t  * restrict q1 = x[i].qs;
-        const uint16_t * restrict qh = x[i].qh;
-        const int8_t   * restrict q8 = y[i].qs;
-        const int16_t  * restrict qs = y[i].bsums;
-
-        for (int j = 0; j < QK_K/32; j += 2) {
-            __builtin_prefetch(q1, 0, 1);
-            __builtin_prefetch(qh, 0, 1);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed long long aux64x2_0 = {*(const int64_t *)(iq1s_grid + (q1[0] | ((qh[0] << 8) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[1] | ((qh[0] << 5) & 0x700)))};
-            vector signed long long aux64x2_1 = {*(const int64_t *)(iq1s_grid + (q1[2] | ((qh[0] << 2) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[3] | ((qh[0] >> 1) & 0x700)))};
-            vector signed long long aux64x2_2 = {*(const int64_t *)(iq1s_grid + (q1[4] | ((qh[1] << 8) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[5] | ((qh[1] << 5) & 0x700)))};
-            vector signed long long aux64x2_3 = {*(const int64_t *)(iq1s_grid + (q1[6] | ((qh[1] << 2) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[7] | ((qh[1] >> 1) & 0x700)))};
-            q1 += 8;
-
-            vector signed char q1x0 = (vector signed char)aux64x2_0;
-            vector signed char q1x1 = (vector signed char)aux64x2_1;
-            vector signed char q1x2 = (vector signed char)aux64x2_2;
-            vector signed char q1x3 = (vector signed char)aux64x2_3;
-
-            vector signed char q8y0 = vec_xl( 0, q8);
-            vector signed char q8y1 = vec_xl(16, q8);
-            vector signed char q8y2 = vec_xl(32, q8);
-            vector signed char q8y3 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed short qv0 = vec_add(vec_mule(q1x0, q8y0), vec_mulo(q1x0, q8y0));
-            vector signed short qv1 = vec_add(vec_mule(q1x1, q8y1), vec_mulo(q1x1, q8y1));
-            vector signed short qv2 = vec_add(vec_mule(q1x2, q8y2), vec_mulo(q1x2, q8y2));
-            vector signed short qv3 = vec_add(vec_mule(q1x3, q8y3), vec_mulo(q1x3, q8y3));
-
-            const uint16_t ls0 = (uint16_t)((qh[0] >> 12) & 7);
-            const uint16_t ls1 = (uint16_t)((qh[1] >> 12) & 7);
-
-            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
-            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
-            vector signed short vscales = vec_sld(vscales23, vscales01, 8);
-
-            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
-            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
-            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
-            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
-
-            vector signed short q8ysums = vec_xl_len(qs, 8);
-            qs += 4;
-            q8ysums = vec_mergeh(q8ysums, (vector signed short)v0);
-
-            vector signed short qxh = (vector signed short)vec_sld(vec_splats(qh[1]), vec_splats(qh[0]), 8);
-            qh += 2;
-            vector __bool short vsel = vec_cmpge(qxh, (vector signed short)v0);
-
-            vector signed short q8ysum = vec_sel((vector signed short)vec_xor((vector unsigned short)q8ysums, vsign), q8ysums, vsel);
-
-            vsumi8 = vec_add(vec_mule(q8ysum, vscales), vsumi8);
+            qs += 8;
         }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi8, 0), vec_mul(vd, vec_splats(IQ1S_DELTA)), vsumf0);
     }
+}
 
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-
-    __m256 accum = (__m256)__lasx_xvldi(0);
-    float accum1 = 0;
-    for (int i = 0; i < nb; ++i) {
+// ====================== 3.3125 bpw (de)-quantization
 
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint16_t * qh = x[i].qh;
+void dequantize_row_iq3_s(const block_iq3_s * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
 
-        __m256i sumi = __lasx_xvldi(0);
-        int sumi1 = 0;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            __m256i q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)], 0);
-            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], 1);
-            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)], 2);
-            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], 3);
+    for (int i = 0; i < nb; i++) {
 
-            __m256i q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)], 0);
-            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], 1);
-            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)], 2);
-            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], 3);
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * qh = x[i].qh;
+        const uint8_t * signs = x[i].signs;
 
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const float db1 = d * (1 + 2*(x[i].scales[ib32/2] & 0xf));
+            const float db2 = d * (1 + 2*(x[i].scales[ib32/2] >>  4));
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[0] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[0] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    y[j+0] = db1 * grid1[j] * (signs[l] & kmask_iq2xs[j+0] ? -1.f : 1.f);
+                    y[j+4] = db1 * grid2[j] * (signs[l] & kmask_iq2xs[j+4] ? -1.f : 1.f);
+                }
+                y += 8;
+            }
             qs += 8;
-            const __m256i q8b_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-            const __m256i q8b_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
-
-            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
-            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
-            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
-            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
-
-            __m256i tmp1, tmp5, tmp6;
-            tmp1 = __lasx_xvreplgr2vr_h(ls1);
-            tmp5 = __lasx_xvmulwev_w_h(dot1, tmp1);
-            tmp6 = __lasx_xvmulwod_w_h(dot1, tmp1);
-            const __m256i p1 = __lasx_xvadd_w(tmp5, tmp6);
-
-            tmp1 = __lasx_xvreplgr2vr_h(ls2);
-            tmp5 = __lasx_xvmulwev_w_h(dot2, tmp1);
-            tmp6 = __lasx_xvmulwod_w_h(dot2, tmp1);
-            const __m256i p2 = __lasx_xvadd_w(tmp5, tmp6);
-
-            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p1, p2));
-            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
-                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
+            signs += 4;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[1] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[1] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    y[j+0] = db2 * grid1[j] * (signs[l] & kmask_iq2xs[j+0] ? -1.f : 1.f);
+                    y[j+4] = db2 * grid2[j] * (signs[l] & kmask_iq2xs[j+4] ? -1.f : 1.f);
+                }
+                y += 8;
+            }
+            qh += 2;
+            qs += 8;
+            signs += 4;
         }
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), accum);
-        accum1 += d * sumi1;
     }
+}
 
-    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
+// ====================== 1.5625 bpw (de)-quantization
 
-#else
+void dequantize_row_iq1_s(const block_iq1_s * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
 
-    float sumf = 0;
     for (int i = 0; i < nb; i++) {
 
-        const int8_t   * q8 = y[i].qs;
+        const float d = GGML_FP16_TO_FP32(x[i].d);
         const uint8_t  * qs = x[i].qs;
         const uint16_t * qh = x[i].qh;
 
-        int sumi = 0, sumi1 = 0;
         for (int ib = 0; ib < QK_K/32; ++ib) {
-            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
-            const int delta = qh[ib] & 0x8000 ? -1 : 1;
-            int lsum = 0;
+            const float dl = d * (2*((qh[ib] >> 12) & 7) + 1);
+            const float delta = qh[ib] & 0x8000 ? -IQ1S_DELTA : IQ1S_DELTA;
             for (int l = 0; l < 4; ++l) {
                 const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
                 for (int j = 0; j < 8; ++j) {
-                    lsum += q8[j] * grid[j];
+                    y[j] = dl * (grid[j] + delta);
                 }
-                q8 += 8;
+                y += 8;
             }
-            sumi  += ls * lsum;
-            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
             qs += 4;
         }
-
-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
     }
-
-    *s = sumf;
-
-#endif
 }
 
-void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(n % QK_K == 0);
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-
-    const block_iq1_m * restrict x = vx;
-    const block_q8_K  * restrict y = vy;
+void dequantize_row_iq1_m(const block_iq1_m * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
 
-    const int nb = n / QK_K;
+    float delta[4];
+    uint16_t idx[4];
 
     iq1m_scale_t scale;
 
-#if defined __ARM_NEON
-    const int32x4_t mask  = vdupq_n_s32(0x7);
-    const int32x4_t mone  = vdupq_n_s32(1);
-    const int32x4_t mzero = vdupq_n_s32(0);
-
-    ggml_int8x16x4_t deltas;
-    deltas.val[0] = vcombine_s8(vdup_n_s8(+1), vdup_n_s8(+1));
-    deltas.val[1] = vcombine_s8(vdup_n_s8(-1), vdup_n_s8(+1));
-    deltas.val[2] = vcombine_s8(vdup_n_s8(+1), vdup_n_s8(-1));
-    deltas.val[3] = vcombine_s8(vdup_n_s8(-1), vdup_n_s8(-1));
-
-    ggml_int8x16x4_t q1b;
-    ggml_int8x16x4_t q8b;
-
-    uint32_t aux32;
-    const uint8_t * aux8 = (const uint8_t *)&aux32;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
+    for (int i = 0; i < nb; i++) {
 
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint8_t  * qh = x[i].qh;
         const uint16_t * sc = (const uint16_t *)x[i].scales;
-
         scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+        const float d = GGML_FP16_TO_FP32(scale.f16);
 
-        int32x4_t sumi1 = mzero;
-        int32x4_t sumi2 = mzero;
-
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-
-            q1b.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[0] | ((qh[0] << 8) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[1] | ((qh[0] << 4) & 0x700)))));
-            q1b.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[2] | ((qh[1] << 8) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[3] | ((qh[1] << 4) & 0x700)))));
-            q1b.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[4] | ((qh[2] << 8) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[5] | ((qh[2] << 4) & 0x700)))));
-            q1b.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[6] | ((qh[3] << 8) & 0x700)))),
-                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[7] | ((qh[3] << 4) & 0x700)))));
-
-            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            const int32x4_t p1 = vpaddq_s32(ggml_vdotq_s32(mzero, q1b.val[0], q8b.val[0]), ggml_vdotq_s32(mzero, q1b.val[1], q8b.val[1]));
-            const int32x4_t p2 = vpaddq_s32(ggml_vdotq_s32(mzero, q1b.val[2], q8b.val[2]), ggml_vdotq_s32(mzero, q1b.val[3], q8b.val[3]));
-            const int32x4_t p12 = vpaddq_s32(p1, p2);
-
-            const uint32_t * qh32 = (const uint32_t *)qh; // we are 4-byte aligned, so we can do that
-            aux32 = ((qh32[0] >> 3) & 0x01010101) | ((qh32[0] >> 6) & 0x02020202);
-
-            const int32x4_t p3 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[0]], q8b.val[0]), ggml_vdotq_s32(mzero, deltas.val[aux8[1]], q8b.val[1]));
-            const int32x4_t p4 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[2]], q8b.val[2]), ggml_vdotq_s32(mzero, deltas.val[aux8[3]], q8b.val[3]));
-            const int32x4_t p34 = vpaddq_s32(p3, p4);
-
-            int32x4_t scales_4 = ggml_vld1q_u32(sc[ib/2] >> 0, sc[ib/2] >> 3, sc[ib/2] >> 6, sc[ib/2] >> 9);
-
-            scales_4 = vaddq_s32(vshlq_n_s32(vandq_s32(scales_4, mask), 1), mone);
-
-            sumi1 = vmlaq_s32(sumi1, scales_4, p12);
-            sumi2 = vmlaq_s32(sumi2, scales_4, p34);
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * qh = x[i].qh;
 
-            qs += 8; qh += 4;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const float dl1 = d * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1);
+            const float dl2 = d * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1);
 
+            idx[0] = qs[0] | ((qh[0] << 8) & 0x700);
+            idx[1] = qs[1] | ((qh[0] << 4) & 0x700);
+            idx[2] = qs[2] | ((qh[1] << 8) & 0x700);
+            idx[3] = qs[3] | ((qh[1] << 4) & 0x700);
+            delta[0] = qh[0] & 0x08 ? -IQ1S_DELTA : IQ1S_DELTA;
+            delta[1] = qh[0] & 0x80 ? -IQ1S_DELTA : IQ1S_DELTA;
+            delta[2] = qh[1] & 0x08 ? -IQ1S_DELTA : IQ1S_DELTA;
+            delta[3] = qh[1] & 0x80 ? -IQ1S_DELTA : IQ1S_DELTA;
+            for (int l = 0; l < 2; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + idx[l]);
+                for (int j = 0; j < 8; ++j) {
+                    y[j] = dl1 * (grid[j] + delta[l]);
+                }
+                y += 8;
+            }
+            for (int l = 2; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + idx[l]);
+                for (int j = 0; j < 8; ++j) {
+                    y[j] = dl2 * (grid[j] + delta[l]);
+                }
+                y += 8;
+            }
+            qs += 4;
+            qh += 2;
         }
-
-        sumf += y[i].d * GGML_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2));
     }
+}
 
-    *s = sumf;
-
-#elif defined __AVX2__
-
-    const __m256i mask = _mm256_set1_epi16(0x7);
-    const __m256i mone = _mm256_set1_epi16(1);
-
-    __m256 accum1 = _mm256_setzero_ps();
-    __m256 accum2 = _mm256_setzero_ps();
-    for (int i = 0; i < nb; ++i) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint8_t  * qh = x[i].qh;
-        const uint16_t * sc = (const uint16_t *)x[i].scales;
-
-        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
-
-        __m256i sumi1 = _mm256_setzero_si256();
-        __m256i sumi2 = _mm256_setzero_si256();
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const __m256i q1b_1 = _mm256_set_epi64x(
-                    iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)],
-                    iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)]
-            );
-            const __m256i q1b_2 = _mm256_set_epi64x(
-                    iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)],
-                    iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)]
-            );
-            const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
-            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
-
-            const __m256i delta1 = _mm256_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
-            const __m256i delta2 = _mm256_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
-
-            const __m256i dot3 = mul_add_epi8(delta1, q8b_1);
-            const __m256i dot4 = mul_add_epi8(delta2, q8b_2);
+static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
 
-            __m256i scale1 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 3), _mm_set1_epi16(sc[ib/2] >> 0));
-            __m256i scale2 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 9), _mm_set1_epi16(sc[ib/2] >> 6));
+void dequantize_row_iq4_nl(const block_iq4_nl * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK4_NL == 0);
+    const int64_t nb = k / QK4_NL;
 
-            scale1 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale1, mask), 1), mone);
-            scale2 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale2, mask), 1), mone);
-            const __m256i p1 = _mm256_madd_epi16(dot1, scale1);
-            const __m256i p2 = _mm256_madd_epi16(dot2, scale2);
-            const __m256i p3 = _mm256_madd_epi16(dot3, scale1);
-            const __m256i p4 = _mm256_madd_epi16(dot4, scale2);
+    for (int i = 0; i < nb; i++) {
 
-            sumi1 = _mm256_add_epi32(sumi1, _mm256_add_epi32(p1, p2));
-            sumi2 = _mm256_add_epi32(sumi2, _mm256_add_epi32(p3, p4));
+        const uint8_t * qs = x[i].qs;
 
-            qs += 8; qh += 4;
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        for (int j = 0; j < QK4_NL/2; ++j) {
+            y[j+       0] = d * kvalues_iq4nl[qs[j] & 0xf];
+            y[j+QK4_NL/2] = d * kvalues_iq4nl[qs[j] >>  4];
         }
-
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
-
-        accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1);
-        accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2);
+        y  += QK4_NL;
+        qs += QK4_NL/2;
     }
+}
 
-    *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
-
-#elif defined __AVX__
-    const __m128i mask = _mm_set1_epi16(0x7);
-    const __m128i mone = _mm_set1_epi16(1);
-
-    __m256 accum1 = _mm256_setzero_ps();
-    __m256 accum2 = _mm256_setzero_ps();
-    for (int i = 0; i < nb; ++i) {
-
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint8_t  * qh = x[i].qh;
-        const uint16_t * sc = (const uint16_t *)x[i].scales;
-
-        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
-
-        __m128i sumi1_0 = _mm_setzero_si128();
-        __m128i sumi1_1 = _mm_setzero_si128();
-        __m128i sumi2_0 = _mm_setzero_si128();
-        __m128i sumi2_1 = _mm_setzero_si128();
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const __m128i q1b_1_0 = _mm_set_epi64x(
-                    iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)]);
-            const __m128i q1b_1_1 = _mm_set_epi64x(
-                    iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)]);
-            const __m128i q1b_2_0 = _mm_set_epi64x(
-                    iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)]);
-            const __m128i q1b_2_1 = _mm_set_epi64x(
-                    iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)]);
-            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-
-            const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
-            const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
-            const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
-            const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
-
-            const __m128i delta1_0 = _mm_set_epi64x(qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
-            const __m128i delta1_1 = _mm_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
-            const __m128i delta2_0 = _mm_set_epi64x(qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
-            const __m128i delta2_1 = _mm_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
-                                                     qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
-
-            const __m128i dot3_0 = mul_add_epi8_sse(delta1_0, q8b_1_0);
-            const __m128i dot3_1 = mul_add_epi8_sse(delta1_1, q8b_1_1);
-            const __m128i dot4_0 = mul_add_epi8_sse(delta2_0, q8b_2_0);
-            const __m128i dot4_1 = mul_add_epi8_sse(delta2_1, q8b_2_1);
-
-            __m128i scale1_0 = _mm_set1_epi16(sc[ib/2] >> 0);
-            __m128i scale1_1 = _mm_set1_epi16(sc[ib/2] >> 3);
-            __m128i scale2_0 = _mm_set1_epi16(sc[ib/2] >> 6);
-            __m128i scale2_1 = _mm_set1_epi16(sc[ib/2] >> 9);
-
-            scale1_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_0, mask), 1), mone);
-            scale1_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_1, mask), 1), mone);
-            scale2_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_0, mask), 1), mone);
-            scale2_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_1, mask), 1), mone);
-            const __m128i p1_0 = _mm_madd_epi16(dot1_0, scale1_0);
-            const __m128i p1_1 = _mm_madd_epi16(dot1_1, scale1_1);
-            const __m128i p2_0 = _mm_madd_epi16(dot2_0, scale2_0);
-            const __m128i p2_1 = _mm_madd_epi16(dot2_1, scale2_1);
-            const __m128i p3_0 = _mm_madd_epi16(dot3_0, scale1_0);
-            const __m128i p3_1 = _mm_madd_epi16(dot3_1, scale1_1);
-            const __m128i p4_0 = _mm_madd_epi16(dot4_0, scale2_0);
-            const __m128i p4_1 = _mm_madd_epi16(dot4_1, scale2_1);
-
-            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
-            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
-            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_add_epi32(p3_0, p4_0));
-            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_add_epi32(p3_1, p4_1));
-
-            qs += 8; qh += 4;
-        }
-
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
-
-        accum1 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum1);
-        accum2 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi2_1, sumi2_0))), accum2);
-    }
-
-    *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
-
-#else
-
-    int sum1[2], sum2[2], delta[4];
+void dequantize_row_iq4_xs(const block_iq4_xs * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
 
-    float sumf = 0;
     for (int i = 0; i < nb; i++) {
 
-        const int8_t   * q8 = y[i].qs;
-        const uint8_t  * qs = x[i].qs;
-        const uint8_t  * qh = x[i].qh;
-        const uint16_t * sc = (const uint16_t *)x[i].scales;
+        const uint8_t * qs = x[i].qs;
 
-        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+        const float d = GGML_FP16_TO_FP32(x[i].d);
 
-        int sumi1 = 0, sumi2 = 0;
         for (int ib = 0; ib < QK_K/32; ++ib) {
-            delta[0] = qh[0] & 0x08 ? -1 : 1;
-            delta[1] = qh[0] & 0x80 ? -1 : 1;
-            delta[2] = qh[1] & 0x08 ? -1 : 1;
-            delta[3] = qh[1] & 0x80 ? -1 : 1;
-            sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
-            for (int l = 0; l < 4; ++l) {
-                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
-                int lsum1 = 0, lsum2 = 0;
-                for (int j = 0; j < 8; ++j) {
-                    lsum1 += q8[j] * grid[j];
-                    lsum2 += q8[j];
-                }
-                q8 += 8;
-                sum1[l/2] += lsum1;
-                sum2[l/2] += lsum2*delta[l];
+            const int ls = ((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4);
+            const float dl = d * (ls - 32);
+            for (int j = 0; j < 16; ++j) {
+                y[j+ 0] = dl * kvalues_iq4nl[qs[j] & 0xf];
+                y[j+16] = dl * kvalues_iq4nl[qs[j] >>  4];
             }
-
-            const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
-            const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
-
-            sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
-            sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
-            qs += 4;
-            qh += 2;
+            y  += 32;
+            qs += 16;
         }
-
-        sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
     }
-
-    *s = sumf;
-
-#endif
 }
 
-void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-    assert(n % QK4_NL == 0);
-    static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
-
-    const block_iq4_nl * restrict x = vx;
-    const block_q8_0   * restrict y = vy;
-
-    const int nb = n / QK4_NL;
-
-    int ib = 0;
-    float sumf = 0;
-
-#if defined __ARM_NEON
-    const int8x16_t values = vld1q_s8(kvalues_iq4nl);
-    const uint8x16_t m4b = vdupq_n_u8(0x0f);
-    uint8x16x2_t q4bits;
-    int8x16x4_t q4b;
-    int8x16x4_t q8b;
-    int32x4_t prod_1, prod_2;
-
-    for (; ib + 1 < nb; ib += 2) {
-
-        q4bits.val[0] = vld1q_u8(x[ib + 0].qs);
-        q4bits.val[1] = vld1q_u8(x[ib + 1].qs);
-        q8b.val[0]    = vld1q_s8(y[ib + 0].qs);
-        q8b.val[1]    = vld1q_s8(y[ib + 0].qs + 16);
-        q8b.val[2]    = vld1q_s8(y[ib + 1].qs);
-        q8b.val[3]    = vld1q_s8(y[ib + 1].qs + 16);
-
-        q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[0], m4b));
-        q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
-        q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[1], m4b));
-        q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
-
-        prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
-        prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
-
-        sumf +=
-            GGML_FP16_TO_FP32(x[ib+0].d) * GGML_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
-            GGML_FP16_TO_FP32(x[ib+1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
-    }
-
-#elif defined __AVX2__
-
-    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
-    const __m128i m4b  = _mm_set1_epi8(0x0f);
-    const __m256i mone = _mm256_set1_epi16(1);
-
-    __m256 accum1 = _mm256_setzero_ps();
-    __m256 accum2 = _mm256_setzero_ps();
-    for (; ib + 1 < nb; ib += 2) {
-        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
-        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
-        const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[ib + 0].qs);
-        const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[ib + 1].qs);
-        const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
-                                              _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
-        const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
-                                              _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
-        const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
-        const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
-        const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
-        const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
-        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
-                _mm256_cvtepi32_ps(p_1), accum1);
-        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
-                _mm256_cvtepi32_ps(p_2), accum2);
-    }
-
-    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
-
-#elif defined __AVX__
-    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
-    const __m128i m4b  = _mm_set1_epi8(0x0f);
-    const __m128i mone = _mm_set1_epi16(1);
-
-    __m256 accum1 = _mm256_setzero_ps();
-    __m256 accum2 = _mm256_setzero_ps();
-    for (; ib + 1 < nb; ib += 2) {
-        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
-        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
-        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
-        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
-        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
-        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
-
-        const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
-        const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
-        const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
-        const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
-        const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
-        const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
-        const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
-        const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
-        const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, mone);
-        const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, mone);
-        const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, mone);
-        const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, mone);
-        accum1 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
-                _mm256_cvtepi32_ps(MM256_SET_M128I(p_1_1, p_1_0))), accum1);
-        accum2 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
-                _mm256_cvtepi32_ps(MM256_SET_M128I(p_2_1, p_2_0))), accum2);
-    }
-
-    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector signed int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-
-    const vector signed char values = vec_xl( 0, kvalues_iq4nl);
-
-#pragma GCC unroll 4
-    for (; ib < nb; ++ib) {
-        __builtin_prefetch(x[ib].qs, 0, 1);
-        __builtin_prefetch(y[ib].qs, 0, 1);
-
-
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
-        vector signed char q4x0 = vec_and(qxs, lowMask);
-        vector signed char q4x1 = vec_sr(qxs, v4);
-
-        q4x0 = vec_perm(values, values, (vector unsigned char)q4x0);
-        q4x1 = vec_perm(values, values, (vector unsigned char)q4x1);
-
-        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
-        vector signed char q8y1 = vec_xl(16, y[ib].qs);
-
-        vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
-        vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-
-        vsumi0 = vec_sum4s(qv0, vsumi0);
-        vsumi1 = vec_sum4s(qv1, vsumi1);
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    sumf = vec_extract(vsumf0, 0);
-
-#elif defined (__loongarch_asx)
-
-    const __m128i values128 = __lsx_vld((const __m128i*)kvalues_iq4nl, 0);
-    const __m128i m4b  = __lsx_vreplgr2vr_b(0x0f);
-    const __m256i mone = __lasx_xvreplgr2vr_h(1);
-
-    __m256 accum1 = (__m256)__lasx_xvldi(0);
-    __m256 accum2 = (__m256)__lasx_xvldi(0);
-    for (; ib + 1 < nb; ib += 2) {
-        const __m128i q4bits_1 = __lsx_vld((const __m128i*)x[ib + 0].qs, 0);
-        const __m128i q4bits_2 = __lsx_vld((const __m128i*)x[ib + 1].qs, 0);
-        const __m256i q8b_1 = __lasx_xvld((const __m256i *)y[ib + 0].qs, 0);
-        const __m256i q8b_2 = __lasx_xvld((const __m256i *)y[ib + 1].qs, 0);
-        const __m256i q4b_1 = lasx_insertf128(lsx_shuffle_b(values128, __lsx_vand_v(__lsx_vsrli_h(q4bits_1, 4), m4b)),
-                                              lsx_shuffle_b(values128, __lsx_vand_v(q4bits_1, m4b)));
-        const __m256i q4b_2 = lasx_insertf128(lsx_shuffle_b(values128, __lsx_vand_v(__lsx_vsrli_h(q4bits_2, 4), m4b)),
-                                              lsx_shuffle_b(values128, __lsx_vand_v(q4bits_2, m4b)));
-        const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
-        const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
-        const __m256i p_1 = lasx_madd_h(p16_1, mone);
-        const __m256i p_2 = lasx_madd_h(p16_2, mone);
-        accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
-                __lasx_xvffint_s_w(p_1), accum1);
-        accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
-                __lasx_xvffint_s_w(p_2), accum2);
-    }
+//===================================== Q8_K ==============================================
 
-    sumf = hsum_float_8(__lasx_xvfadd_s(accum1, accum2));
+void quantize_row_q8_K_ref(const float * restrict x, block_q8_K * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
 
-#endif
-    for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
-        int sumi1 = 0, sumi2 = 0;
-        for (int j = 0; j < QK4_NL/2; ++j) {
-            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
-            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
+    for (int i = 0; i < nb; i++) {
+
+        float max = 0;
+        float amax = 0;
+        for (int j = 0; j < QK_K; ++j) {
+            float ax = fabsf(x[j]);
+            if (ax > amax) {
+                amax = ax; max = x[j];
+            }
+        }
+        if (!amax) {
+            y[i].d = 0;
+            memset(y[i].qs, 0, QK_K);
+            x += QK_K;
+            continue;
+        }
+        //const float iscale = -128.f/max;
+        // We need this change for IQ2_XXS, else the AVX implementation becomes very awkward
+        const float iscale = -127.f/max;
+        for (int j = 0; j < QK_K; ++j) {
+            int v = nearest_int(iscale*x[j]);
+            y[i].qs[j] = MIN(127, v);
+        }
+        for (int j = 0; j < QK_K/16; ++j) {
+            int sum = 0;
+            for (int ii = 0; ii < 16; ++ii) {
+                sum += y[i].qs[j*16 + ii];
+            }
+            y[i].bsums[j] = sum;
         }
-        sumf += d * (sumi1 + sumi2);
+        y[i].d = 1/iscale;
+        x += QK_K;
     }
-    *s = sumf;
 }
 
-void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
-    assert(nrc == 1);
-    UNUSED(nrc);
-    UNUSED(bx);
-    UNUSED(by);
-    UNUSED(bs);
-    assert(n % QK_K == 0);
-
-    const block_iq4_xs * restrict x = vx;
-    const block_q8_K   * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#if defined __ARM_NEON
-    const int8x16_t values = vld1q_s8(kvalues_iq4nl);
-    const uint8x16_t m4b = vdupq_n_u8(0x0f);
-    ggml_uint8x16x2_t q4bits;
-    ggml_int8x16x4_t q4b;
-    ggml_int8x16x4_t q8b;
-    int32x4_t prod_1, prod_2;
-
-    float sumf = 0;
-
-    for (int ibl = 0; ibl < nb; ++ibl) {
-
-        const int8_t  * q8 = y[ibl].qs;
-        const uint8_t * q4 = x[ibl].qs;
-        uint16_t h = x[ibl].scales_h;
-
-        int sumi1 = 0, sumi2 = 0;
-        for (int ib = 0; ib < QK_K/64; ++ib) {
-
-            q4bits = ggml_vld1q_u8_x2(q4); q4 += 32;
-            q8b    = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[0], m4b));
-            q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
-            q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[1], m4b));
-            q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
-
-            prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
-            prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
-
-            int ls1 = ((x[ibl].scales_l[ib] & 0xf) | ((h << 4) & 0x30)) - 32;
-            int ls2 = ((x[ibl].scales_l[ib] >>  4) | ((h << 2) & 0x30)) - 32;
-            h >>= 4;
-            sumi1 += vaddvq_s32(prod_1) * ls1;
-            sumi2 += vaddvq_s32(prod_2) * ls2;
-
-        }
-
-        sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
-    }
-
-    *s = sumf;
-
-#elif defined __AVX2__
-
-    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
-    const __m128i m4b  = _mm_set1_epi8(0x0f);
-
-    __m256 accum = _mm256_setzero_ps();
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        uint16_t sh = x[ibl].scales_h;
-        __m256i sumi1 = _mm256_setzero_si256();
-        __m256i sumi2 = _mm256_setzero_si256();
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)qs);  qs += 16;
-            const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)qs);  qs += 16;
-            const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
-            const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
-                                                  _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
-            const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
-                                                  _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
-            const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
-            const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
-            const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;
-            const int16_t ls2 = ((x[ibl].scales_l[ib/2] >>  4) | ((sh << 2) & 0x30)) - 32;
-            sh >>= 4;
-            const __m256i p_1 = _mm256_madd_epi16(p16_1, _mm256_set1_epi16(ls1));
-            const __m256i p_2 = _mm256_madd_epi16(p16_2, _mm256_set1_epi16(ls2));
-            sumi1 = _mm256_add_epi32(p_1, sumi1);
-            sumi2 = _mm256_add_epi32(p_2, sumi2);
-        }
-        accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
-                _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accum);
-    }
-
-    *s = hsum_float_8(accum);
-
-#elif defined __AVX__
-    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
-    const __m128i m4b  = _mm_set1_epi8(0x0f);
-
-    __m256 accum = _mm256_setzero_ps();
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        uint16_t sh = x[ibl].scales_h;
-        __m128i sumi1_0 = _mm_setzero_si128();
-        __m128i sumi1_1 = _mm_setzero_si128();
-        __m128i sumi2_0 = _mm_setzero_si128();
-        __m128i sumi2_1 = _mm_setzero_si128();
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
-            const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
-            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
-            const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
-            const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
-            const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
-            const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
-            const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
-            const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
-            const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
-            const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
-            const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;
-            const int16_t ls2 = ((x[ibl].scales_l[ib/2] >>  4) | ((sh << 2) & 0x30)) - 32;
-            sh >>= 4;
-            const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, _mm_set1_epi16(ls1));
-            const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, _mm_set1_epi16(ls1));
-            const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, _mm_set1_epi16(ls2));
-            const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, _mm_set1_epi16(ls2));
-            sumi1_0 = _mm_add_epi32(p_1_0, sumi1_0);
-            sumi1_1 = _mm_add_epi32(p_1_1, sumi1_1);
-            sumi2_0 = _mm_add_epi32(p_2_0, sumi2_0);
-            sumi2_1 = _mm_add_epi32(p_2_1, sumi2_1);
-        }
-        __m128i sumi12_0 = _mm_add_epi32(sumi1_0, sumi2_0);
-        __m128i sumi12_1 = _mm_add_epi32(sumi1_1, sumi2_1);
-        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
-                _mm256_cvtepi32_ps(MM256_SET_M128I(sumi12_1, sumi12_0))), accum);
-    }
-
-    *s = hsum_float_8(accum);
-
-#elif defined(__POWER9_VECTOR__)
-    const vector signed char lowMask = vec_splats((signed char)0xF);
-    const vector int v0 = vec_splats((int32_t)0);
-    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-
-    vector float vsumf0 = vec_splats(0.0f);
-    vector float vsumf1 = vec_splats(0.0f);
-    vector float vsumf2 = vec_splats(0.0f);
-    vector float vsumf3 = vec_splats(0.0f);
-
-    const vector signed char values = vec_xl( 0, kvalues_iq4nl);
-
-    for (int ibl = 0; ibl < nb; ++ibl) {
-
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ibl].d));
-        vector float vyd = vec_splats(y[ibl].d);
-        vector float vd = vec_mul(vxd, vyd);
-
-        vector signed int vsumi0 = v0;
-        vector signed int vsumi1 = v0;
-        vector signed int vsumi2 = v0;
-        vector signed int vsumi3 = v0;
-
-        uint16_t h = x[ibl].scales_h;
-
-        const uint8_t * restrict q4 = x[ibl].qs;
-        const uint8_t * restrict sc = x[ibl].scales_l;
-        const int8_t  * restrict q8 = y[ibl].qs;
-
-        for (int ib = 0; ib < QK_K/64; ib ++ ) {
-            __builtin_prefetch(q4, 0, 1);
-            __builtin_prefetch(q8, 0, 1);
-
-            vector signed char qxs0 = (vector signed char)vec_xl( 0, q4);
-            vector signed char qxs1 = (vector signed char)vec_xl(16, q4);
-            q4 += 32;
-
-            vector signed char q4x00 = (vector signed char)vec_and(qxs0, lowMask);
-            vector signed char q4x01 = (vector signed char)vec_sr(qxs0, v4);
-            vector signed char q4x10 = (vector signed char)vec_and(qxs1, lowMask);
-            vector signed char q4x11 = (vector signed char)vec_sr(qxs1, v4);
-
-            q4x00 = vec_perm(values, values, (vector unsigned char)q4x00);
-            q4x01 = vec_perm(values, values, (vector unsigned char)q4x01);
-            q4x10 = vec_perm(values, values, (vector unsigned char)q4x10);
-            q4x11 = vec_perm(values, values, (vector unsigned char)q4x11);
-
-            vector signed char q8y0 = vec_xl( 0, q8);
-            vector signed char q8y1 = vec_xl(16, q8);
-            vector signed char q8y2 = vec_xl(32, q8);
-            vector signed char q8y3 = vec_xl(48, q8);
-            q8 += 64;
-
-            vector signed short qv0 = vec_add(vec_mule(q4x00, q8y0), vec_mulo(q4x00, q8y0));
-            vector signed short qv1 = vec_add(vec_mule(q4x01, q8y1), vec_mulo(q4x01, q8y1));
-            vector signed short qv2 = vec_add(vec_mule(q4x10, q8y2), vec_mulo(q4x10, q8y2));
-            vector signed short qv3 = vec_add(vec_mule(q4x11, q8y3), vec_mulo(q4x11, q8y3));
-
-            const uint16_t ls0 = (uint16_t)(((sc[0] & 0xf) | ((h << 4) & 0x30)) - 32);
-            const uint16_t ls1 = (uint16_t)(((sc[0] >>  4) | ((h << 2) & 0x30)) - 32);
-            h >>= 4;
-            sc ++;
-
-            vector signed short vscales01 = vec_splats((int16_t)ls0);
-            vector signed short vscales23 = vec_splats((int16_t)ls1);
-
-            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
-            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
-            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
-            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
-        }
-
-        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
-        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
-        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
-        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
-    }
-
-    vsumf0 = vec_add(vsumf0, vsumf2);
-    vsumf1 = vec_add(vsumf1, vsumf3);
-
-    vsumf0 = vec_add(vsumf0, vsumf1);
-
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
-    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
-
-    *s = vec_extract(vsumf0, 0);
-
-#elif defined(__loongarch_asx)
-
-    const __m128i values128 = __lsx_vld((const __m128i*)kvalues_iq4nl, 0);
-    const __m128i m4b  = __lsx_vreplgr2vr_b(0x0f);
-
-    __m256 accum = (__m256)__lasx_xvldi(0);
-    __m256i tmp1;
-    __m128i tmp0, tmp2, tmp3, tmp4, mask_8f, mask;
-
-    mask_8f = __lsx_vreplgr2vr_b(0x8f);
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        uint16_t sh = x[ibl].scales_h;
-        __m256i sumi1 = __lasx_xvldi(0);
-        __m256i sumi2 = __lasx_xvldi(0);
-        __m128i zero = __lsx_vldi(0);
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const __m128i q4bits_1 = __lsx_vld((const __m128i*)qs, 0);  qs += 16;
-            const __m128i q4bits_2 = __lsx_vld((const __m128i*)qs, 0);  qs += 16;
-            const __m256i q8b_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            const __m256i q8b_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
-            tmp2 = __lsx_vand_v(__lsx_vand_v(__lsx_vsrli_h(q4bits_1, 4), m4b), mask_8f);
-            tmp0 = __lsx_vori_b(tmp2, 0x10);
-            mask = __lsx_vsle_b(zero, tmp2);
-            tmp3 = __lsx_vand_v(tmp0, mask);
-            tmp3 = __lsx_vshuf_b(values128, zero, tmp3);
-
-            tmp2 = __lsx_vand_v(__lsx_vand_v(q4bits_1, m4b), mask_8f);
-            tmp0 = __lsx_vori_b(tmp2, 0x10);
-            mask = __lsx_vsle_b(zero, tmp2);
-            tmp4 = __lsx_vand_v(tmp0, mask);
-            tmp4 = __lsx_vshuf_b(values128, zero, tmp4);
-
-            const __m256i q4b_1 = lasx_insertf128(tmp3, tmp4);
-
-            tmp2 = __lsx_vand_v(__lsx_vand_v(__lsx_vsrli_h(q4bits_2, 4), m4b), mask_8f);
-            tmp0 = __lsx_vori_b(tmp2, 0x10);
-            mask = __lsx_vsle_b(zero, tmp2);
-            tmp3 = __lsx_vand_v(tmp0, mask);
-            tmp3 = __lsx_vshuf_b(values128, zero, tmp3);
-
-            tmp2 = __lsx_vand_v(__lsx_vand_v(q4bits_2, m4b), mask_8f);
-            tmp0 = __lsx_vori_b(tmp2, 0x10);
-            mask = __lsx_vsle_b(zero, tmp2);
-            tmp4 = __lsx_vand_v(tmp0, mask);
-            tmp4 = __lsx_vshuf_b(values128, zero, tmp4);
-
-            const __m256i q4b_2 = lasx_insertf128(tmp3, tmp4);
-
-            const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
-            const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
-            const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;
-            const int16_t ls2 = ((x[ibl].scales_l[ib/2] >>  4) | ((sh << 2) & 0x30)) - 32;
-            sh >>= 4;
-            __m256i tmp5, tmp6;
-            tmp1 = __lasx_xvreplgr2vr_h(ls1);
-            tmp5 = __lasx_xvmulwev_w_h(p16_1, tmp1);
-            tmp6 = __lasx_xvmulwod_w_h(p16_1, tmp1);
-            const __m256i p_1 = __lasx_xvadd_w(tmp5, tmp6);
-            tmp1 = __lasx_xvreplgr2vr_h(ls2);
-            tmp5 = __lasx_xvmulwev_w_h(p16_2, tmp1);
-            tmp6 = __lasx_xvmulwod_w_h(p16_2, tmp1);
-            const __m256i p_2 = __lasx_xvadd_w(tmp5, tmp6);
-            sumi1 = __lasx_xvadd_w(p_1, sumi1);
-            sumi2 = __lasx_xvadd_w(p_2, sumi2);
-        }
-        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
-                __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accum);
-    }
-
-    *s = hsum_float_8(accum);
+void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
 
-#else
-    float sumf = 0;
-    for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
-        uint16_t h = x[ibl].scales_h;
-        const uint8_t * qs = x[ibl].qs;
-        const int8_t  * q8 = y[ibl].qs;
-        for (int ib = 0; ib < QK_K/32; ib += 2) {
-            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
-            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
-            h >>= 4;
-            const float d1 = d4d8*(ls1 - 32);
-            const float d2 = d4d8*(ls2 - 32);
-            int sumi1 = 0, sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d1 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
-            sumi1 = sumi2 = 0;
-            for (int j = 0; j < 16; ++j) {
-                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
-                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
-            }
-            sumf += d2 * (sumi1 + sumi2);
-            qs += 16;
-            q8 += 32;
+    for (int i = 0; i < nb; i++) {
+        for (int j = 0; j < QK_K; ++j) {
+            *y++ = x[i].d * x[i].qs[j];
         }
     }
-    *s = sumf;
-#endif
 }
 
 // ================================ IQ2 quantization =============================================
@@ -14236,12 +3770,6 @@ size_t quantize_iq3_xxs(const float * restrict src, void * restrict dst, int64_t
     return nrow * nblock * sizeof(block_iq3_xxs);
 }
 
-void quantize_row_iq3_xxs(const float * restrict x, void * restrict vy, int64_t k) {
-    assert(k % QK_K == 0);
-    block_iq3_xxs * restrict y = vy;
-    quantize_row_iq3_xxs_ref(x, y, k);
-}
-
 void quantize_row_iq3_xxs_ref(const float * restrict x, block_iq3_xxs * restrict y, int64_t k) {
     assert(k % QK_K == 0);
     quantize_row_iq3_xxs_impl(256, x, y, k, NULL);
@@ -14452,12 +3980,6 @@ size_t quantize_iq3_s(const float * restrict src, void * restrict dst, int64_t n
     return nrow * nblock * sizeof(block_iq3_s);
 }
 
-void quantize_row_iq3_s(const float * restrict x, void * restrict vy, int64_t k) {
-    assert(k % QK_K == 0);
-    block_iq3_s * restrict y = vy;
-    quantize_row_iq3_s_ref(x, y, k);
-}
-
 void quantize_row_iq3_s_ref(const float * restrict x, block_iq3_s * restrict y, int64_t k) {
     assert(k % QK_K == 0);
     quantize_iq3_s(x, y, 1, k, NULL);
@@ -15181,7 +4703,8 @@ size_t quantize_iq4_nl(const float * restrict src, void * restrict dst, int64_t
     return nrow * nblock * sizeof(block_iq4_nl);
 }
 
-void quantize_row_iq4_nl(const float * restrict x, void * restrict vy, int64_t k) {
+//void quantize_row_iq4_nl_ref(const float * restrict x, void * restrict vy, int64_t k) {
+void quantize_row_iq4_nl_ref(const float * restrict x, block_iq4_nl * restrict y, int64_t k) {
     GGML_ASSERT(k%QK4_NL == 0);
     int64_t nblock = k/QK4_NL;
     uint8_t L[QK4_NL];
@@ -15189,18 +4712,13 @@ void quantize_row_iq4_nl(const float * restrict x, void * restrict vy, int64_t k
     uint16_t unused_h;
     uint8_t * unused_l = NULL;
     float scale;
-    block_iq4_nl * iq4 = (block_iq4_nl *)vy;
+    block_iq4_nl * iq4 = y;
     for (int ibl = 0; ibl < nblock; ++ibl) {
         quantize_row_iq4_nl_impl(QK4_NL, 32, x + QK4_NL*ibl, &iq4[ibl].d, iq4[ibl].qs, &unused_h, unused_l,
                 &scale, weight, L, kvalues_iq4nl, NULL, -1);
     }
 }
 
-void quantize_row_iq4_nl_ref(const float * restrict x, block_iq4_nl * restrict y, int64_t k) {
-    assert(k % QK4_NL == 0);
-    quantize_row_iq4_nl(x, y, k);
-}
-
 size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
     GGML_ASSERT(n_per_row%QK_K == 0);
     int64_t nblock = n_per_row/QK_K;
@@ -15221,12 +4739,6 @@ size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int64_t
     return nrow * nblock * sizeof(block_iq4_xs);
 }
 
-void quantize_row_iq4_xs(const float * restrict x, void * restrict vy, int64_t k) {
-    assert(k % QK_K == 0);
-    block_iq4_xs * restrict y = vy;
-    quantize_row_iq4_xs_ref(x, y, k);
-}
-
 void quantize_row_iq4_xs_ref(const float * restrict x, block_iq4_xs * restrict y, int64_t k) {
     assert(k % QK_K == 0);
     quantize_iq4_xs(x, y, 1, k, NULL);
@@ -15419,11 +4931,7 @@ void quantize_row_iq2_s_ref(const float * restrict x, block_iq2_s * restrict y,
     quantize_iq2_s(x, y, 1, k, NULL);
 }
 
-void quantize_row_iq2_s(const float * restrict x, void * restrict vy, int64_t k) {
-    assert(k % QK_K == 0);
-    block_iq2_s * restrict y = vy;
-    quantize_row_iq2_s_ref(x, y, k);
-}
+// =============================== data validation
 
 static bool validate_float(float f, size_t i) {
     if (isinf(f)) {
@@ -15712,15 +5220,6 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
             {
                 VALIDATE_ROW_DATA_D_F16_IMPL(block_iq4_nl, data, nb);
             } break;
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-            {
-                VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x4, data, nbytes / sizeof(block_q4_0x4), 4);
-            } break;
-        case GGML_TYPE_Q4_0_8_8:
-            {
-                VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x8, data, nbytes / sizeof(block_q4_0x8), 8);
-            } break;
 
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
diff --git a/ggml/src/ggml-quants.h b/ggml/src/ggml-quants.h
index df9c4b24ae7..d09173e1116 100644
--- a/ggml/src/ggml-quants.h
+++ b/ggml/src/ggml-quants.h
@@ -11,136 +11,89 @@
 extern "C" {
 #endif
 
+// NOTE: these functions are defined as GGML_API because they used by the CPU backend
+
 // Quantization
-void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1_ref(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0_ref(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1_ref(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K_ref(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K_ref(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K_ref(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K_ref(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_tq1_0_ref(const float * GGML_RESTRICT x, block_tq1_0 * GGML_RESTRICT y, int64_t k);
-void quantize_row_tq2_0_ref(const float * GGML_RESTRICT x, block_tq2_0 * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_iq3_xxs_ref(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_nl_ref (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs_ref (const float * GGML_RESTRICT x, block_iq4_xs  * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq3_s_ref  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq2_s_ref  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq3_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq2_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q4_1_ref(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q5_0_ref(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q5_1_ref(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
+
+GGML_API void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q4_K_ref(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q5_K_ref(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q6_K_ref(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q8_K_ref(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
+
+GGML_API void quantize_row_tq1_0_ref(const float * GGML_RESTRICT x, block_tq1_0 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_tq2_0_ref(const float * GGML_RESTRICT x, block_tq2_0 * GGML_RESTRICT y, int64_t k);
+
+GGML_API void quantize_row_iq3_xxs_ref(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_iq4_nl_ref (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_iq4_xs_ref (const float * GGML_RESTRICT x, block_iq4_xs  * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_iq3_s_ref  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_iq2_s_ref  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_RESTRICT y, int64_t k);
 
 // Dequantization
-void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-
-void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-
-void dequantize_row_tq1_0(const block_tq1_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_tq2_0(const block_tq2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-
-void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq1_m  (const block_iq1_m   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
-
-// Dot product
-void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
-void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+GGML_API void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+//GGML_API void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+GGML_API void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+GGML_API void dequantize_row_tq1_0(const block_tq1_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_tq2_0(const block_tq2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+GGML_API void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq1_m  (const block_iq1_m   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq1_m  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-size_t quantize_tq1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_tq2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
-
-void iq2xs_init_impl(enum ggml_type type);
-void iq2xs_free_impl(enum ggml_type type);
-void iq3xs_init_impl(int grid_size);
-void iq3xs_free_impl(int grid_size);
+GGML_API size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq1_m  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+GGML_API size_t quantize_tq1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_tq2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+GGML_API size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+GGML_API void iq2xs_init_impl(enum ggml_type type);
+GGML_API void iq2xs_free_impl(enum ggml_type type);
+GGML_API void iq3xs_init_impl(int grid_size);
+GGML_API void iq3xs_free_impl(int grid_size);
 
 #ifdef __cplusplus
 }
diff --git a/ggml/src/ggml-rpc/CMakeLists.txt b/ggml/src/ggml-rpc/CMakeLists.txt
new file mode 100644
index 00000000000..f5acb8ec2cb
--- /dev/null
+++ b/ggml/src/ggml-rpc/CMakeLists.txt
@@ -0,0 +1,9 @@
+message(STATUS "Using RPC backend")
+
+ggml_add_backend_library(ggml-rpc
+                         ggml-rpc.cpp
+                        )
+
+if (WIN32)
+    target_link_libraries(ggml-rpc PRIVATE ws2_32)
+endif()
diff --git a/ggml/src/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
similarity index 99%
rename from ggml/src/ggml-rpc.cpp
rename to ggml/src/ggml-rpc/ggml-rpc.cpp
index 8a772f22454..43108242639 100644
--- a/ggml/src/ggml-rpc.cpp
+++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
@@ -671,7 +671,7 @@ static ggml_backend_i ggml_backend_rpc_interface = {
     /* .event_wait              = */ NULL,
 };
 
-GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint) {
+ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint) {
     static std::mutex mutex;
     std::lock_guard<std::mutex> lock(mutex);
     // NOTE: buffer types are allocated and never freed; this is by design
@@ -718,7 +718,7 @@ ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
     return backend;
 }
 
-GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend) {
+bool ggml_backend_is_rpc(ggml_backend_t backend) {
     return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_rpc_guid());
 }
 
@@ -730,7 +730,7 @@ static void get_device_memory(const std::shared_ptr<socket_t> & sock, size_t * f
     *total = response.total_mem;
 }
 
-GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total) {
+void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total) {
     auto sock = get_socket(endpoint);
     if (sock == nullptr) {
         *free = 0;
@@ -1369,8 +1369,9 @@ static const struct ggml_backend_reg_i ggml_backend_rpc_reg_i = {
 
 ggml_backend_reg_t ggml_backend_rpc_reg(void) {
     static struct ggml_backend_reg ggml_backend_rpc_reg = {
-        /* .iface   = */ ggml_backend_rpc_reg_i,
-        /* .context = */ NULL,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_rpc_reg_i,
+        /* .context     = */ NULL,
     };
 
     return &ggml_backend_rpc_reg;
@@ -1401,3 +1402,5 @@ ggml_backend_dev_t ggml_backend_rpc_add_device(const char * endpoint) {
 
     return dev;
 }
+
+GGML_BACKEND_DL_IMPL(ggml_backend_rpc_reg)
diff --git a/ggml/src/ggml-sycl/CMakeLists.txt b/ggml/src/ggml-sycl/CMakeLists.txt
new file mode 100644
index 00000000000..3579a311aac
--- /dev/null
+++ b/ggml/src/ggml-sycl/CMakeLists.txt
@@ -0,0 +1,84 @@
+if (NOT GGML_SYCL_TARGET MATCHES "^(INTEL|NVIDIA|AMD)$")
+    message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL, NVIDIA, or AMD")
+endif()
+
+check_cxx_compiler_flag("-fsycl" SUPPORTS_SYCL)
+
+if (DEFINED ENV{ONEAPI_ROOT})
+    message(STATUS "Using oneAPI Release SYCL compiler (icpx).")
+elseif(SUPPORTS_SYCL)
+    message(WARNING "Using open-source SYCL compiler (clang++). Didn't detect ENV {ONEAPI_ROOT}.
+        If you expected the oneAPI Release compiler, please install oneAPI & source it, like:
+        source /opt/intel/oneapi/setvars.sh")
+else()
+    message(FATAL_ERROR, "C++ compiler lacks SYCL support.")
+endif()
+message(STATUS "SYCL found")
+#todo: AOT
+
+ggml_add_backend_library(ggml-sycl
+                         ggml-sycl.cpp
+                         ../../include/ggml-sycl.h
+                        )
+
+if (GGML_SYCL_F16)
+    if (GGML_SYCL_TARGET STREQUAL "AMD")
+        message(WARNING "AMD target does not entirely support FP16 in the SYCL backend.")
+    endif()
+    add_compile_definitions(GGML_SYCL_F16)
+endif()
+
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
+
+if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
+    add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
+elseif (GGML_SYCL_TARGET STREQUAL "AMD")
+    # INFO: Allowed Sub_group_sizes are not consistent through all
+    # hip targets. For example, 64 is used for certain models, but the backend
+    # does not support it.
+    # Target archs tested working: gfx1030, gfx1031, (Only tested sub_group_size = 32)
+    add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
+else()
+    add_compile_definitions(GGML_SYCL_WARP_SIZE=16)
+endif()
+
+file(GLOB   GGML_HEADERS_SYCL "*.hpp")
+file(GLOB   GGML_SOURCES_SYCL "*.cpp")
+target_sources(ggml-sycl PRIVATE ${GGML_HEADERS_SYCL} ${GGML_SOURCES_SYCL})
+
+find_package(DNNL)
+message("-- DNNL found:" ${DNNL_FOUND})
+
+if (GGML_SYCL_TARGET STREQUAL "INTEL")
+    add_compile_definitions(GGML_SYCL_DNNL=${DNNL_FOUND})
+else()
+    add_compile_definitions(GGML_SYCL_DNNL=0)
+endif()
+
+if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
+    target_link_libraries(ggml-sycl PRIVATE DNNL::dnnl)
+endif()
+
+if (WIN32)
+    find_package(IntelSYCL REQUIRED)
+    find_package(MKL REQUIRED)
+    target_link_libraries(ggml-sycl PRIVATE IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
+else()
+    if (GGML_SYCL_TARGET STREQUAL "INTEL")
+        target_link_libraries(ggml-sycl PRIVATE sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
+    elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
+        add_compile_definitions(GGML_SYCL_NVIDIA)
+        target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl_blas_cublas)
+    elseif (GGML_SYCL_TARGET STREQUAL "AMD")
+        if (NOT GGML_SYCL_DEVICE_ARCH)
+            message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.")
+        endif()
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=amdgcn-amd-amdhsa")
+        target_link_libraries(ggml-sycl PRIVATE sycl pthread m dl onemkl)
+    endif()
+
+    if (GGML_SYCL_DEVICE_ARCH)
+      set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH}")
+  endif()
+endif()
diff --git a/ggml/src/ggml-sycl/common.cpp b/ggml/src/ggml-sycl/common.cpp
index 97ab2003c7f..a9ee4049114 100644
--- a/ggml/src/ggml-sycl/common.cpp
+++ b/ggml/src/ggml-sycl/common.cpp
@@ -11,6 +11,7 @@
 //
 
 #include "common.hpp"
+#include "ggml-impl.h"
 
 int get_current_device_id() {
   return dpct::dev_mgr::instance().current_device_id();
@@ -28,11 +29,7 @@ void* ggml_sycl_host_malloc(size_t size) try {
 
   if (err != 0) {
     // clear the error
-    fprintf(
-        stderr,
-        "WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
-        size / 1024.0 / 1024.0,
-        "syclGetErrorString is not supported");
+    GGML_LOG_ERROR("WARNING: failed to allocate %.2f MB of pinned memory: %s\n", size / 1024.0 / 1024.0,    "syclGetErrorString is not supported");
     return nullptr;
   }
 
@@ -66,18 +63,12 @@ int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block
 void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
                                  const ggml_tensor *src1, ggml_tensor *dst,
                                  const ggml_sycl_op_flatten_t op) try {
-    const int64_t nrows0 = ggml_nrows(src0);
 
     const bool use_src1 = src1 != nullptr;
-    const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1;
 
     GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
     GGML_ASSERT(              dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
 
-    ggml_tensor_extra_gpu * src0_extra =            (ggml_tensor_extra_gpu *) src0->extra;
-    ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
-    ggml_tensor_extra_gpu * dst_extra  =            (ggml_tensor_extra_gpu *)  dst->extra;
-
     // dd = data device
     float * src0_ddf = (float *) src0->data;
     float * src1_ddf = use_src1 ? (float *) src1->data : nullptr;
diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp
index 4549fa5e95a..c1582f610e5 100644
--- a/ggml/src/ggml-sycl/common.hpp
+++ b/ggml/src/ggml-sycl/common.hpp
@@ -626,6 +626,7 @@ struct bin_bcast_sycl {
                     });
             }
         }
+        GGML_UNUSED(ctx);
     }
 };
 
diff --git a/ggml/src/ggml-sycl/concat.cpp b/ggml/src/ggml-sycl/concat.cpp
index c90c452d878..a240968ad2e 100644
--- a/ggml/src/ggml-sycl/concat.cpp
+++ b/ggml/src/ggml-sycl/concat.cpp
@@ -47,7 +47,7 @@ static void concat_f32_dim1(const float *x, const float *y, float *dst,
   // operation
   int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
                    item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
-  if (item_ct1.get_group(1) < ne01) { // src0
+  if (item_ct1.get_group(1) < (size_t) ne01) { // src0
     int offset_src =
         nidx + item_ct1.get_group(1) * ne0 + item_ct1.get_group(0) * ne0 * ne01;
     dst[offset_dst] = x[offset_src];
@@ -70,7 +70,7 @@ static void concat_f32_dim2(const float *x, const float *y, float *dst,
   // operation
   int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
                    item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
-  if (item_ct1.get_group(0) < ne02) { // src0
+  if (item_ct1.get_group(0) < (size_t) ne02) { // src0
     int offset_src = nidx + item_ct1.get_group(1) * ne0 +
                      item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
     dst[offset_dst] = x[offset_src];
diff --git a/ggml/src/ggml-sycl/convert.cpp b/ggml/src/ggml-sycl/convert.cpp
index 5fd15e6cdcc..05b01db2d8b 100644
--- a/ggml/src/ggml-sycl/convert.cpp
+++ b/ggml/src/ggml-sycl/convert.cpp
@@ -424,7 +424,7 @@ static void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y,
     const int64_t global_id = item_ct1.get_local_id(2) + work_group_size * item_ct1.get_group(2);
 
     // make each work-item deal with more elements since sycl global range can not exceed max int
-    const src_t * x = (src_t *) vx;
+    const src_t * x = (const src_t *) vx;
     for (int64_t i = global_id; i < k; i += work_group_size * item_ct1.get_group_range(2)) {
         y[i] = x[i];
     }
diff --git a/ggml/src/ggml-sycl/dmmv.cpp b/ggml/src/ggml-sycl/dmmv.cpp
index 0c3dfaa37eb..0d097357ce7 100644
--- a/ggml/src/ggml-sycl/dmmv.cpp
+++ b/ggml/src/ggml-sycl/dmmv.cpp
@@ -1015,9 +1015,9 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
             break;
     }
 
-    (void) src1;
-    (void) dst;
-    (void) src1_ddq_i;
-    (void) src1_ncols;
-    (void) src1_padded_row_size;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddq_i);
+    GGML_UNUSED(src1_ncols);
+    GGML_UNUSED(src1_padded_row_size);
 }
diff --git a/ggml/src/ggml-sycl/dpct/helper.hpp b/ggml/src/ggml-sycl/dpct/helper.hpp
index fe4a8f744e2..e167948e7a3 100644
--- a/ggml/src/ggml-sycl/dpct/helper.hpp
+++ b/ggml/src/ggml-sycl/dpct/helper.hpp
@@ -15,6 +15,7 @@
 
 #include <sycl/sycl.hpp>
 #include <sycl/half_type.hpp>
+#include <syclcompat/math.hpp>
 #include <oneapi/mkl.hpp>
 #include <map>
 
@@ -1236,7 +1237,7 @@ namespace dpct
 
             std::map<byte_t *, allocation>::iterator get_map_iterator(const void *ptr)
             {
-                auto it = m_map.upper_bound((byte_t *)ptr);
+                auto it = m_map.upper_bound(const_cast<byte_t *>(reinterpret_cast<const byte_t *>(ptr)));
                 if (it == m_map.end())
                 {
                     // Not a virtual pointer.
@@ -1688,9 +1689,14 @@ namespace dpct
             auto data_a = get_memory<const Ta>(a);
             auto data_b = get_memory<const Tb>(b);
             auto data_c = get_memory<Tc>(c);
-            oneapi::mkl::blas::column_major::gemm(
-                q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda,
-                data_b, ldb, beta_value, data_c, ldc);
+#ifdef GGML_SYCL_NVIDIA
+            oneapi::mkl::blas::column_major::gemm(oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q },
+                                                  a_trans, b_trans, m, n, k, alpha_value, data_a, lda, data_b, ldb,
+                                                  beta_value, data_c, ldc);
+#else
+            oneapi::mkl::blas::column_major::gemm(q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda, data_b, ldb,
+                                                  beta_value, data_c, ldc);
+#endif
         }
 
         template <typename VecT, class BinaryOperation, class = void>
@@ -1753,14 +1759,22 @@ namespace dpct
             matrix_info->ld_info[2] = ldc;
             matrix_info->groupsize_info = batch_size;
 
+#ifdef GGML_SYCL_NVIDIA
+            sycl::event e = oneapi::mkl::blas::column_major::gemm_batch(
+                oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q }, matrix_info->transpose_info,
+                matrix_info->transpose_info + 1, matrix_info->size_info, matrix_info->size_info + 1,
+                matrix_info->size_info + 2, matrix_info->value_info, reinterpret_cast<const Ta **>(a),
+                matrix_info->ld_info, reinterpret_cast<const Tb **>(b), matrix_info->ld_info + 1,
+                matrix_info->value_info + 1, reinterpret_cast<Tc **>(c), matrix_info->ld_info + 2, 1,
+                &(matrix_info->groupsize_info));
+#else
             sycl::event e = oneapi::mkl::blas::column_major::gemm_batch(
-                q, matrix_info->transpose_info, matrix_info->transpose_info + 1,
-                matrix_info->size_info, matrix_info->size_info + 1,
-                matrix_info->size_info + 2, matrix_info->value_info,
-                reinterpret_cast<const Ta **>(a), matrix_info->ld_info,
-                reinterpret_cast<const Tb **>(b), matrix_info->ld_info + 1,
-                matrix_info->value_info + 1, reinterpret_cast<Tc **>(c),
+                q, matrix_info->transpose_info, matrix_info->transpose_info + 1, matrix_info->size_info,
+                matrix_info->size_info + 1, matrix_info->size_info + 2, matrix_info->value_info,
+                reinterpret_cast<const Ta **>(a), matrix_info->ld_info, reinterpret_cast<const Tb **>(b),
+                matrix_info->ld_info + 1, matrix_info->value_info + 1, reinterpret_cast<Tc **>(c),
                 matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info));
+#endif
 
             q.submit([&](sycl::handler &cgh)
                      {
@@ -1782,10 +1796,16 @@ namespace dpct
             auto data_a = get_memory<const Ta>(a);
             auto data_b = get_memory<const Tb>(b);
             auto data_c = get_memory<Tc>(c);
+#ifdef GGML_SYCL_NVIDIA
             oneapi::mkl::blas::column_major::gemm_batch(
-                q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda,
-                stride_a, data_b, ldb, stride_b, beta_value,
-                data_c, ldc, stride_c, batch_size);
+                oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q }, a_trans, b_trans, m, n, k,
+                alpha_value, data_a, lda, stride_a, data_b, ldb, stride_b, beta_value, data_c, ldc, stride_c,
+                batch_size);
+#else
+            oneapi::mkl::blas::column_major::gemm_batch(q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda,
+                                                        stride_a, data_b, ldb, stride_b, beta_value, data_c, ldc,
+                                                        stride_c, batch_size);
+#endif
         }
 
     } // namespace detail
@@ -1830,31 +1850,10 @@ namespace dpct
                                            : id);
     }
 
-    template <typename T>
-    sycl::vec<T, 4> extract_and_sign_or_zero_extend4(T val)
-    {
-        return sycl::vec<T, 1>(val)
-            .template as<sycl::vec<
-                std::conditional_t<std::is_signed_v<T>, int8_t, uint8_t>, 4>>()
-            .template convert<T>();
-    }
-
-    template <typename T1, typename T2>
-    using dot_product_acc_t =
-        std::conditional_t<std::is_unsigned_v<T1> && std::is_unsigned_v<T2>,
-                           uint32_t, int32_t>;
-
     template <typename T1, typename T2, typename T3>
     inline auto dp4a(T1 a, T2 b, T3 c)
     {
-        dot_product_acc_t<T1, T2> res = c;
-        auto va = extract_and_sign_or_zero_extend4(a);
-        auto vb = extract_and_sign_or_zero_extend4(b);
-        res += va[0] * vb[0];
-        res += va[1] * vb[1];
-        res += va[2] * vb[2];
-        res += va[3] * vb[3];
-        return res;
+        return syclcompat::dp4a(a, b, c);
     }
 
     struct sub_sat
diff --git a/ggml/src/ggml-sycl/element_wise.cpp b/ggml/src/ggml-sycl/element_wise.cpp
index e5cd736eba9..d05a51f807c 100644
--- a/ggml/src/ggml-sycl/element_wise.cpp
+++ b/ggml/src/ggml-sycl/element_wise.cpp
@@ -237,7 +237,7 @@ void upscale_f32(const float  *x, float *dst, const int nb00, const int nb01,
     int i02 = i12 / sf2;
     int i03 = i13 / sf3;
 
-    dst[index] = *(float *)((char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
+    dst[index] = *(const float *)((const char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
 }
 
 void pad_f32(const float  *x, float *dst, const int ne0, const int ne00, const int ne01, const int ne02,
@@ -251,8 +251,7 @@ void pad_f32(const float  *x, float *dst, const int ne0, const int ne00, const i
     // operation
     int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
                      item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
-    if (nidx < ne00 && item_ct1.get_group(1) < ne01 &&
-        item_ct1.get_group(0) < ne02) {
+    if (nidx < ne00 && item_ct1.get_group(1) < (size_t) ne01 && item_ct1.get_group(0) < (size_t) ne02) {
         int offset_src = nidx + item_ct1.get_group(1) * ne00 +
                          item_ct1.get_group(0) * ne00 * ne01;
             dst[offset_dst] = x[offset_src];
@@ -520,9 +519,10 @@ inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     silu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -535,9 +535,10 @@ inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     gelu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
                                     const ggml_tensor *src1, ggml_tensor *dst,
@@ -550,9 +551,10 @@ inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, const ggml_
 
     gelu_quick_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -564,9 +566,10 @@ inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, const ggml_tensor
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
     tanh_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -579,9 +582,10 @@ inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     relu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -595,9 +599,10 @@ inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, const ggml
 
     hardsigmoid_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -610,9 +615,10 @@ inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, const ggml_t
 
     hardswish_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -625,9 +631,10 @@ inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     exp_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -640,9 +647,10 @@ inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     log_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -655,9 +663,10 @@ inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, const ggml_ten
 
     sigmoid_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -670,9 +679,10 @@ inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     sqrt_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -685,9 +695,10 @@ inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     sin_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -700,9 +711,10 @@ inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     cos_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -715,9 +727,10 @@ inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     step_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -730,9 +743,10 @@ inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     neg_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -749,9 +763,10 @@ inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, const ggml_
 
     leaky_relu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), negative_slope, main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -764,9 +779,10 @@ inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     sqr_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -787,9 +803,10 @@ inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, const ggml_ten
                      dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3,
                      main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -805,9 +822,10 @@ inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, const ggml_tensor
         src0->ne[0], src0->ne[1], src0->ne[2],
         dst->ne[0], dst->ne[1], dst->ne[2], main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -827,7 +845,8 @@ inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     acc_f32_sycl(src0_dd, src1_dd, dst_dd, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, main_stream);
 
-    (void) dst;
+    GGML_UNUSED(dst);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_add(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
diff --git a/ggml/src/ggml-sycl/gemm.hpp b/ggml/src/ggml-sycl/gemm.hpp
index 2ad9b36f419..3f0f34ad603 100644
--- a/ggml/src/ggml-sycl/gemm.hpp
+++ b/ggml/src/ggml-sycl/gemm.hpp
@@ -51,8 +51,8 @@ class DnnlGemmWrapper {
         const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab);
         const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab);
         const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab);
-        auto a_mem = dnnl::memory(a_in_md, eng, (void*)a);
-        auto b_mem = dnnl::memory(b_in_md, eng, (void*)b);
+        auto a_mem = dnnl::memory(a_in_md, eng, const_cast<void*>(a));
+        auto b_mem = dnnl::memory(b_in_md, eng, const_cast<void*>(b));
         auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md);
         auto c_mem = dnnl::memory(matmul_pd.dst_desc(), eng, c);
 
@@ -79,8 +79,8 @@ class DnnlGemmWrapper {
         const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab);
         const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab);
         const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab);
-        auto a_mem = dnnl::memory(a_in_md, eng, (void*)a);
-        auto b_mem = dnnl::memory(b_in_md, eng, (void*)b);
+        auto a_mem = dnnl::memory(a_in_md, eng, const_cast<void*>(a));
+        auto b_mem = dnnl::memory(b_in_md, eng, const_cast<void*>(b));
         auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md);
         auto c_mem = dnnl::memory(matmul_pd.dst_desc(), eng, c);
 
diff --git a/ggml/src/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
similarity index 96%
rename from ggml/src/ggml-sycl.cpp
rename to ggml/src/ggml-sycl/ggml-sycl.cpp
index 2dba15d237e..84f1328e7cf 100644
--- a/ggml/src/ggml-sycl.cpp
+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
@@ -47,7 +47,7 @@ static ggml_sycl_device_info ggml_sycl_init() {
 
     info.device_count = dpct::dev_mgr::instance().device_count();
     if (info.device_count == 0) {
-        fprintf(stderr, "%s: failed to initialize " GGML_SYCL_NAME ": %s\n", __func__);
+        GGML_LOG_ERROR("%s: failed to initialize: %s\n", GGML_SYCL_NAME, __func__);
         return info;
     }
 
@@ -55,16 +55,16 @@ static ggml_sycl_device_info ggml_sycl_init() {
 
     int64_t total_vram = 0;
 #if defined(GGML_SYCL_FORCE_MMQ)
-    fprintf(stderr, "%s: GGML_SYCL_FORCE_MMQ:   yes\n", __func__);
+    GGML_LOG_INFO("%s: GGML_SYCL_FORCE_MMQ:   yes\n", __func__);
 #else
-    fprintf(stderr, "%s: GGML_SYCL_FORCE_MMQ:   no\n", __func__);
+    GGML_LOG_INFO("%s: GGML_SYCL_FORCE_MMQ:   no\n", __func__);
 #endif
 #if defined(SYCL_USE_XMX)
-    fprintf(stderr, "%s: SYCL_USE_XMX: yes\n", __func__);
+    GGML_LOG_INFO("%s: SYCL_USE_XMX: yes\n", __func__);
 #else
-    fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__);
+    GGML_LOG_INFO("%s: SYCL_USE_XMX: no\n", __func__);
 #endif
-    fprintf(stderr, "%s: found %d " GGML_SYCL_NAME " devices:\n", __func__, info.device_count);
+    GGML_LOG_INFO("%s: found %d %s devices:\n", __func__, info.device_count, GGML_SYCL_NAME);
 
     for (int i = 0; i < info.device_count; ++i) {
         info.devices[i].vmm = 0;
@@ -110,7 +110,7 @@ void print_device_detail(int id, sycl::device &device, std::string device_type)
 
     auto global_mem_size = prop.get_global_mem_size()/1000000;
 
-    fprintf(stderr, "|%2d|%19s|%39s|%7s|%7d|%8d|%5d|%6luM|%21s|\n", id, device_type.c_str(),
+    GGML_LOG_INFO("|%2d|%19s|%39s|%7s|%7d|%8d|%5d|%6luM|%21s|\n", id, device_type.c_str(),
             name.c_str(), version.c_str(), prop.get_max_compute_units(),
             prop.get_max_work_group_size(), prop.get_max_sub_group_size(),
             global_mem_size, device.get_info<sycl::info::device::driver_version>().c_str());
@@ -120,19 +120,29 @@ void ggml_backend_sycl_print_sycl_devices() {
     GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_print_sycl_devices\n");
     int device_count = dpct::dev_mgr::instance().device_count();
     std::map<std::string, size_t> DeviceNums;
-    fprintf(stderr, "found %d SYCL devices:\n", device_count);
-    fprintf(stderr, "|  |                   |                                       |       |Max    |        |Max  |Global |                     |\n");
-    fprintf(stderr, "|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |\n");
-    fprintf(stderr, "|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|\n");
-    fprintf(stderr, "|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|\n");
+    GGML_LOG_INFO("Found %d SYCL devices:\n", device_count);
+
+    GGML_LOG_INFO(
+        "|  |                   |                                       |      "
+        " |Max    |        |Max  |Global |                     |\n");
+    GGML_LOG_INFO(
+        "|  |                   |                                       |      "
+        " |compute|Max work|sub  |mem    |                     |\n");
+    GGML_LOG_INFO(
+        "|ID|        Device Type|                                   "
+        "Name|Version|units  |group   |group|size   |       Driver version|\n");
+    GGML_LOG_INFO(
+        "|--|-------------------|---------------------------------------|------"
+        "-|-------|--------|-----|-------|---------------------|\n");
+
     for (int id = 0; id < device_count; ++id) {
-        sycl::device device = dpct::dev_mgr::instance().get_device(id);
-        sycl::backend backend = device.get_backend();
-        std::string backend_type = get_device_backend_and_type(device);
-        int type_id=DeviceNums[backend_type]++;
-        std::stringstream device_type;
-        device_type << "[" <<  backend_type << ":" << std::to_string(type_id) << "]";
-        print_device_detail(id, device, device_type.str());
+      sycl::device device = dpct::dev_mgr::instance().get_device(id);
+      std::string backend_type = get_device_backend_and_type(device);
+      int type_id = DeviceNums[backend_type]++;
+      std::stringstream device_type;
+      device_type << "[" << backend_type << ":" << std::to_string(type_id)
+                  << "]";
+      print_device_detail(id, device, device_type.str());
     }
 }
 
@@ -154,15 +164,14 @@ static void ggml_check_sycl() try {
     static bool initialized = false;
 
     if (!initialized) {
-        fprintf(stderr, "[SYCL] call ggml_check_sycl\n");
+        GGML_LOG_INFO("[SYCL] call ggml_check_sycl\n");
         g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
-
-        fprintf(stderr, "%s: GGML_SYCL_DEBUG: %d\n", __func__, g_ggml_sycl_debug);
+        GGML_LOG_INFO("%s: GGML_SYCL_DEBUG: %d\n", __func__, g_ggml_sycl_debug);
 
 #if defined(GGML_SYCL_F16)
-        fprintf(stderr, "%s: GGML_SYCL_F16: yes\n", __func__);
+        GGML_LOG_INFO("%s: GGML_SYCL_F16: yes\n", __func__);
 #else
-        fprintf(stderr, "%s: GGML_SYCL_F16: no\n", __func__);
+        GGML_LOG_INFO("%s: GGML_SYCL_F16: no\n", __func__);
 #endif
 
 /* NOT REMOVE, keep it for next optimize for XMX.
@@ -180,9 +189,10 @@ static void ggml_check_sycl() try {
             return;
         }
         GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES);
-        ggml_backend_sycl_print_sycl_devices();
+
         initialized = true;
         g_sycl_loaded = true;
+        ggml_backend_sycl_print_sycl_devices();
     }
 }
 catch (sycl::exception const &exc) {
@@ -205,7 +215,7 @@ inline void check_allow_gpu_index(const int device_index) {
         __func__,
         device_index,
         ggml_sycl_info().device_count - 1);
-    fprintf(stderr, "%s\n", error_buf);
+    GGML_LOG_ERROR("%s\n", error_buf);
     assert(false);
   }
 }
@@ -409,14 +419,12 @@ ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
         return true;
     }
     return false;
-}
-catch (sycl::exception const &exc) {
-  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
-            << ", line:" << __LINE__ << std::endl;
-  std::exit(1);
+    GGML_UNUSED(buffer);
+} catch (const sycl::exception & exc) {
+    std::cerr << exc.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl;
+    std::exit(1);
 }
 
-
 static void ggml_backend_sycl_buffer_clear(ggml_backend_buffer_t buffer,
                                            uint8_t value) try {
      ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
@@ -475,8 +483,8 @@ ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
     SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)sycl::malloc_device(
                                     size, *stream)));
     if (!dev_ptr) {
-        fprintf(stderr, "%s: can't malloc %lu Bytes memory on device", __func__, size);
-        return nullptr;
+      GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device\n", __func__, size);
+      return nullptr;
     }
     ggml_backend_sycl_buffer_context * ctx = new  ggml_backend_sycl_buffer_context(buft_ctx->device, dev_ptr, buft_ctx->stream);
     return ggml_backend_buffer_init(buft, ggml_backend_sycl_buffer_interface, ctx, size);
@@ -752,7 +760,7 @@ ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
                                         size, *stream)));
         if (!buf) {
             char err_buf[1024];
-            snprintf(err_buf, 1023, "%s: can't malloc %lu Bytes memory on device", __func__, size);
+            snprintf(err_buf, 1023, "%s: can't allocate %lu Bytes of memory on device\n", __func__, size);
             throw std::runtime_error(err_buf);
         }
         // set padding to 0 to avoid possible NaN values
@@ -1081,10 +1089,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
     ggml_sycl_buffer buffer_pool[MAX_SYCL_BUFFERS] = {};
     size_t pool_size = 0;
 
-    explicit ggml_sycl_pool_leg(queue_ptr qptr_, int device_) :
-        qptr(qptr_),
-        device(device_) {
-    }
+    explicit ggml_sycl_pool_leg(queue_ptr qptr_, int device_) : device(device_), qptr(qptr_) {}
 
     ~ggml_sycl_pool_leg() {
         for (int i = 0; i < MAX_SYCL_BUFFERS; ++i) {
@@ -1142,17 +1147,18 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
             CHECK_TRY_ERROR(ptr = (void *)sycl::malloc_device(
                                 look_ahead_size, *qptr)));
         if (!ptr) {
-            fprintf(stderr, "%s: can't malloc %lu Bytes memory on device", __func__, look_ahead_size);
+            GGML_LOG_ERROR("%s: can't allocate %lu Bytes of memory on device/GPU\n", __func__, look_ahead_size);
             return nullptr;
         }
 
         *actual_size = look_ahead_size;
         pool_size += look_ahead_size;
 
-    #ifdef DEBUG_SYCL_MALLOC
-        fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, id, nnz,
+#ifdef DEBUG_SYCL_MALLOC
+        GGML_LOG_DEBUG("%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, id, nnz,
                 (uint32_t)(max_size/1024/1024), (uint32_t)(g_sycl_pool_size[id]/1024/1024), (uint32_t)(size/1024/1024));
-    #endif
+#endif
+
         // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg look_ahead_size=%lu, return %p\n", look_ahead_size, ptr);
         return ptr;
     }
@@ -1166,7 +1172,7 @@ struct ggml_sycl_pool_leg : public ggml_sycl_pool {
                 return;
             }
         }
-        fprintf(stderr, "WARNING: sycl buffer pool full, increase MAX_sycl_BUFFERS\n");
+        GGML_LOG_WARN("WARNING: sycl buffer pool full, increase MAX_sycl_BUFFERS\n");
         SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, *qptr)));
         pool_size -= size;
     }
@@ -1226,7 +1232,7 @@ static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy,
         zeros[i] = 0.f;
         qzeros[i] = 0;
     }
-    const TC xi = ix < kx ? *(TC *)&x[iy * kx + ix] : zeros;
+    const TC xi = ix < kx ? *(const TC *)&x[iy * kx + ix] : zeros;
     float sum = xi[0];
     float amax = sycl::fabs(xi[0]);
 #pragma unroll
@@ -1787,6 +1793,9 @@ static  void pool2d_nchw_kernel(
         switch (op) {
             case GGML_OP_POOL_AVG: res = 0; break;
             case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
+            default:
+                res      = (To) sycl::nan(uint32_t(0));
+                break;
         }
 
         for (int i = bh; i < eh; i += 1) {
@@ -1805,6 +1814,9 @@ static  void pool2d_nchw_kernel(
                 switch (op) {
                     case GGML_OP_POOL_AVG: res += (cur / (kh * kw)); break;
                     case GGML_OP_POOL_MAX: res = sycl::max(res, (To)cur); break;
+                    default:
+                        res = (To) sycl::nan(uint32_t(0));
+                        break;
                 }
             }
         }
@@ -1843,7 +1855,8 @@ static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *sr
                                  s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
                          });
 
-    (void) dst;
+    GGML_UNUSED(dst);
+    GGML_UNUSED(ctx);
 }
 
 template <typename src0_t>
@@ -1881,10 +1894,10 @@ static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tens
             });
     }
 
-    (void) dst;
+    GGML_UNUSED(dst);
+    GGML_UNUSED(ctx);
 }
 
-
 static void quantize_row_q8_1_sycl(const float *x, void *vy, const int kx,
                                    const int ky, const int kx_padded,
                                    queue_ptr stream) {
@@ -2437,7 +2450,7 @@ static void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, const ggml_te
             break;
         default:
             // TODO: k-quants
-            fprintf(stderr, "%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
+            GGML_LOG_ERROR("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
             GGML_ABORT("fatal error");
             break;
     }
@@ -2452,8 +2465,8 @@ static void ggml_sycl_op_repeat(ggml_backend_sycl_context & ctx, const ggml_tens
 
     ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_repeat>>(ctx, dst, src0, dst, nullptr, src0_d, dst_d, main_stream);
 
-    (void) src1;
-    (void) src1_d;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src1_d);
 }
 
 
@@ -2472,17 +2485,18 @@ inline void ggml_sycl_op_mul_mat_sycl(
     const int64_t ne00 = src0->ne[0];
     const int64_t ne10 = src1->ne[0];
 
-    const int64_t ne0 = dst->ne[0];
 
     const int64_t row_diff = row_high - row_low;
 
     int id;
     SYCL_CHECK(
         CHECK_TRY_ERROR(id = get_current_device_id()));
-
+#if !GGML_SYCL_DNNL
+    const int64_t ne0 = dst->ne[0];
     // the main device has a larger memory buffer to hold the results from all GPUs
     // ldc == nrows of the matrix that cuBLAS writes into
     int ldc = id == ctx.device ? ne0 : row_diff;
+#endif
 
 #ifdef GGML_SYCL_F16
     bool use_fp16 = true;  // TODO(Yu) SYCL capability check
@@ -2519,9 +2533,9 @@ inline void ggml_sycl_op_mul_mat_sycl(
                                          : src1_as_f16.get();
         ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool(), row_diff * src1_ncols);
 
-        const sycl::half alpha_f16 = 1.0f;
-        const sycl::half beta_f16 = 0.0f;
 #if !GGML_SYCL_DNNL
+        const sycl::half alpha_f16 = 1.0f;
+        const sycl::half beta_f16  = 0.0f;
         SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm(
             *stream, oneapi::mkl::transpose::trans,
             oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10,
@@ -2558,24 +2572,29 @@ inline void ggml_sycl_op_mul_mat_sycl(
         const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get();
         const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();
 
-        const float alpha = 1.0f;
-        const float beta = 0.0f;
 #if !GGML_SYCL_DNNL
+        const float alpha = 1.0f;
+        const float beta  = 0.0f;
+#    ifdef GGML_SYCL_NVIDIA
         SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm(
-            *stream, oneapi::mkl::transpose::trans,
-            oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10,
-            dpct::get_value(&alpha, *stream), src0_ddf_i, ne00,
-            src1_ddf1_i, ne10, dpct::get_value(&beta, *stream),
+            oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ *stream }, oneapi::mkl::transpose::trans,
+            oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10, dpct::get_value(&alpha, *stream), src0_ddf_i,
+            ne00, src1_ddf1_i, ne10, dpct::get_value(&beta, *stream), dst_dd_i, ldc)));
+#    else
+        SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm(
+            *stream, oneapi::mkl::transpose::trans, oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10,
+            dpct::get_value(&alpha, *stream), src0_ddf_i, ne00, src1_ddf1_i, ne10, dpct::get_value(&beta, *stream),
             dst_dd_i, ldc)));
+#    endif
 #else
         auto dnnl_stream = ctx.stream_dnnl(stream);
          DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(),
             src0_ddf_i, DnnlGemmWrapper::to_dt<float>(), dst_dd_i, DnnlGemmWrapper::to_dt<float>());
 #endif
     }
-    (void) dst;
-    (void) src1_ddq_i;
-    (void) src1_padded_row_size;
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddq_i);
+    GGML_UNUSED(src1_padded_row_size);
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -2621,8 +2640,9 @@ static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tens
                                item_ct1);
         });
 
-    (void) src1;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_sum(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2637,9 +2657,10 @@ inline void ggml_sycl_op_sum(ggml_backend_sycl_context & ctx, const ggml_tensor
 
     sum_rows_f32_sycl(src0_dd, dst_dd, ne, 1, main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2656,9 +2677,10 @@ inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, const ggml_te
 
     sum_rows_f32_sycl(src0_dd, dst_dd, ncols, nrows, main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2677,9 +2699,10 @@ inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, const ggml_ten
 
     argsort_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, order, main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2696,9 +2719,10 @@ inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, const ggml_tens
 
     argmax_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
@@ -2718,9 +2742,10 @@ inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx, const gg
 
     diag_mask_inf_f32_sycl(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -2741,9 +2766,10 @@ inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, const ggml_tenso
     */
     SYCL_CHECK(0);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
@@ -2766,9 +2792,10 @@ inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, const ggml_tenso
     */
     SYCL_CHECK(0);
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
 
 static void ggml_sycl_set_peer_access(const int n_tokens, int main_device) {
@@ -2845,7 +2872,6 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
 
     ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
     ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
-    ggml_tensor_extra_gpu *  dst_extra = (ggml_tensor_extra_gpu *)  dst->extra;
 
     const bool src0_is_contiguous = ggml_is_contiguous(src0);
     const bool src1_is_contiguous = ggml_is_contiguous(src1);
@@ -3272,7 +3298,6 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx,
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    const int64_t ne_dst = ggml_nelements(dst);
 
     SYCL_CHECK(ggml_sycl_set_device(ctx.device));
     queue_ptr main_stream = ctx.stream();;
@@ -3380,6 +3405,7 @@ catch (sycl::exception const &exc) {
 
 inline bool ggml_sycl_supports_mmq(enum ggml_type type) {
     // TODO: accuracy issues in MMQ
+    GGML_UNUSED(type);
     return false;
 }
 
@@ -3447,8 +3473,15 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
         use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
 
     if (!split && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
-        // KQ single-batch
-        ggml_sycl_mul_mat_vec_p021(ctx, src0, src1, dst);
+        // TODO: Refactor and cleanup of mul mat dispatching.
+        if (src0->ne[3] == 1 && src1->ne[3] == 1) {
+            // KQ single-batch
+            // mmv p021 was specific for these dimensions
+            ggml_sycl_mul_mat_vec_p021(ctx, src0, src1, dst);
+        } else {
+            // The kernel from the if path is faster for that specific case, but does not support all mul mats.
+            ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
+        }
     } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
         // KQV single-batch
         ggml_sycl_mul_mat_vec_nc(ctx, src0, src1, dst);
@@ -3743,12 +3776,12 @@ static void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor *sr
     } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) {
         ggml_cpy_i32_i32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
     } else {
-        fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
+        GGML_LOG_ERROR("%s: unsupported type combination (%s to %s)\n", __func__,
                 ggml_type_name(src0->type), ggml_type_name(src1->type));
         GGML_ABORT("fatal error");
     }
 
-    (void) dst;
+    GGML_UNUSED(dst);
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -3759,7 +3792,7 @@ catch (sycl::exception const &exc) {
 static void ggml_sycl_dup(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     // TODO: why do we pass dst as src1 here?
     ggml_sycl_cpy(ctx, src0, dst, nullptr);
-    (void) src1;
+    GGML_UNUSED(src1);
 }
 
 static void ggml_sycl_diag_mask_inf(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -3804,13 +3837,16 @@ static void ggml_sycl_argmax(ggml_backend_sycl_context & ctx, const ggml_tensor
 }
 
 static void ggml_sycl_nop(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    (void) src0;
-    (void) src1;
-    (void) dst;
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(ctx);
 }
 
 void ggml_sycl_set_main_device(const int main_device) try {
-    if (dpct::get_current_device_id() == main_device) return;
+    if (dpct::get_current_device_id() == static_cast<unsigned int> (main_device)) {
+        return;
+    }
     check_allow_gpu_index(main_device);
     dpct::select_device(main_device);
 
@@ -3818,7 +3854,7 @@ void ggml_sycl_set_main_device(const int main_device) try {
         dpct::device_info prop;
         SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info(
             prop, dpct::dev_mgr::instance().get_device(main_device))));
-        fprintf(stderr, "Using device %d (%s) as main device\n",
+        GGML_LOG_INFO("Using device %d (%s) as main device\n",
                 main_device, prop.get_name());
     }
 }
@@ -4165,7 +4201,7 @@ static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_
 #endif
         bool ok = ggml_sycl_compute_forward(*sycl_ctx, node);
         if (!ok) {
-            fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+            GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
         }
         GGML_ASSERT(ok);
     }
@@ -4178,6 +4214,7 @@ try
 {
     ggml_backend_sycl_context *sycl_ctx =
         (ggml_backend_sycl_context *)backend->context;
+
     sycl::event *sycl_event = static_cast<sycl::event *>(event->context);
 
     const queue_ptr &stream = sycl_ctx->stream(sycl_ctx->device, 0);
@@ -4192,7 +4229,7 @@ catch (sycl::exception const &exc)
 }
 
 static void ggml_backend_sycl_event_wait(ggml_backend_t backend, ggml_backend_event_t event) try {
-    ggml_backend_sycl_context* sycl_ctx = static_cast<ggml_backend_sycl_context*>(backend->context);
+
     sycl::event* sycl_event = static_cast<sycl::event*>(event->context);
 
     if (ggml_backend_is_sycl(backend)) {
@@ -4350,10 +4387,6 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                 if (op->op == GGML_OP_MUL_MAT) {
                     a = op->src[0];
                     b = op->src[1];
-                    if (ggml_is_permuted(a) || ggml_is_permuted(b)) {
-                        // TODO: fix like https://github.com/ggerganov/llama.cpp/pull/10021
-                        return false;
-                    }
                 } else {
                     a = op->src[2];
                     b = op->src[1];
@@ -4455,7 +4488,16 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_SOFT_MAX:
             return true;
         case GGML_OP_ROPE:
-            return ggml_is_contiguous(op->src[0]);
+            {
+                const int mode = ((const int32_t *) op->op_params)[2];
+                if (mode & GGML_ROPE_TYPE_MROPE) {
+                    return false;
+                }
+                if (mode & GGML_ROPE_TYPE_VISION) {
+                    return false;
+                }
+                return ggml_is_contiguous(op->src[0]);
+            }
         case GGML_OP_IM2COL:
             // TODO: add support for the new F32 operations
             return op->src[0]->type == GGML_TYPE_F16;
@@ -4490,7 +4532,7 @@ static bool ggml_backend_sycl_device_supports_buft(ggml_backend_dev_t dev, ggml_
 static int64_t get_op_batch_size(const ggml_tensor * op) {
     switch (op->op) {
         case GGML_OP_GET_ROWS:
-            return op->ne[1]; // this will increse the speed of prefill in test
+            return 0;
         case GGML_OP_MUL_MAT:
             return op->ne[1];
         case GGML_OP_MUL_MAT_ID:
@@ -4604,13 +4646,14 @@ static void *ggml_backend_sycl_reg_get_proc_address(ggml_backend_reg_t reg, cons
     // SYCL doesn't support registering host memory, left here for reference
     // "ggml_backend_register_host_buffer"
     // "ggml_backend_unregister_host_buffer"
+    GGML_UNUSED(name);
     return nullptr;
 }
 
 static const ggml_backend_reg_i ggml_backend_sycl_reg_interface = {
     /* .get_name          = */ ggml_backend_sycl_reg_get_name,
     /* .get_device_count  = */ ggml_backend_sycl_reg_get_device_count,
-    /* .get_device_get    = */ ggml_backend_sycl_reg_get_device,
+    /* .get_device        = */ ggml_backend_sycl_reg_get_device,
     /* .get_proc_address  = */ ggml_backend_sycl_reg_get_proc_address,
 };
 
@@ -4641,16 +4684,17 @@ ggml_backend_reg_t ggml_backend_sycl_reg() {
                 dev_ctx->description = prop.get_name();
 
                 ggml_backend_dev_t dev = new ggml_backend_device {
-                    /* .interface = */ ggml_backend_sycl_device_interface,
-                    /* .reg       = */ &reg,
-                    /* .context   = */ dev_ctx
+                    /* .iface       = */ ggml_backend_sycl_device_interface,
+                    /* .reg         = */ &reg,
+                    /* .context     = */ dev_ctx
                 };
                 ctx->devices.push_back(dev);
             }
 
             reg = ggml_backend_reg {
-                /* .interface = */ ggml_backend_sycl_reg_interface,
-                /* .context   = */ ctx
+                /* .api_version = */ GGML_BACKEND_API_VERSION,
+                /* .iface       = */ ggml_backend_sycl_reg_interface,
+                /* .context     = */ ctx
             };
         }
 
@@ -4668,7 +4712,7 @@ ggml_backend_t ggml_backend_sycl_init(int device) {
 
     ggml_backend_sycl_context * ctx = new ggml_backend_sycl_context(device);
     if (ctx == nullptr) {
-        fprintf(stderr, "%s: error: failed to allocate context\n", __func__);
+        GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
         return nullptr;
     };
 
@@ -4682,3 +4726,4 @@ ggml_backend_t ggml_backend_sycl_init(int device) {
     return sycl_backend;
 }
 
+GGML_BACKEND_DL_IMPL(ggml_backend_sycl_reg)
diff --git a/ggml/src/ggml-sycl/im2col.cpp b/ggml/src/ggml-sycl/im2col.cpp
index 6a0a0fcd08c..6146a99edbe 100644
--- a/ggml/src/ggml-sycl/im2col.cpp
+++ b/ggml/src/ggml-sycl/im2col.cpp
@@ -120,6 +120,7 @@ void ggml_sycl_op_im2col(
         im2col_sycl(src1_dd, (float *)dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
     }
 
-    (void) src0;
-    (void) src0_dd;
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src0_dd);
+    GGML_UNUSED(ctx);
 }
diff --git a/ggml/src/ggml-sycl/mmq.cpp b/ggml/src/ggml-sycl/mmq.cpp
index e952533d310..8ea82c940c7 100644
--- a/ggml/src/ggml-sycl/mmq.cpp
+++ b/ggml/src/ggml-sycl/mmq.cpp
@@ -813,7 +813,7 @@ load_tiles_q4_K(const void *__restrict__ vx, int *__restrict__ x_ql,
         x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
     }
 
-    const int blocks_per_tile_x_row = WARP_SIZE / QI4_K; // == 1 if QK_K == 256
+    constexpr int blocks_per_tile_x_row = QI4_K > WARP_SIZE ? 1 : WARP_SIZE / QI4_K; // == 1 if QK_K == 256
     const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
 
 #pragma unroll
@@ -961,7 +961,7 @@ load_tiles_q5_K(const void *__restrict__ vx, int *__restrict__ x_ql,
         x_ql[i * (2*WARP_SIZE + 1) + kq1] = ql1 | qh1;
     }
 
-    const int blocks_per_tile_x_row = WARP_SIZE / QI5_K; // == 1 if QK_K == 256
+    constexpr int blocks_per_tile_x_row = QI5_K > WARP_SIZE ? 1 : WARP_SIZE / QI5_K; // == 1 if QK_K == 256
     const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
 
 #pragma unroll
@@ -1109,7 +1109,7 @@ load_tiles_q6_K(const void *__restrict__ vx, int *__restrict__ x_ql,
                                                  dpct::sub_sat());
     }
 
-    const int blocks_per_tile_x_row = WARP_SIZE / QI6_K; // == 1 if QK_K == 256
+    constexpr int blocks_per_tile_x_row = QI6_K > WARP_SIZE ? 1 : WARP_SIZE / QI6_K; // == 1 if QK_K == 256
     const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
     float * x_dmf = (float *) x_dm;
 
@@ -3020,9 +3020,9 @@ void ggml_sycl_op_mul_mat_q(
             break;
     }
 
-    (void) src1;
-    (void) dst;
-    (void) src1_ddf_i;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddf_i);
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
diff --git a/ggml/src/ggml-sycl/mmvq.cpp b/ggml/src/ggml-sycl/mmvq.cpp
index 7b10cf68814..221f65c21ea 100644
--- a/ggml/src/ggml-sycl/mmvq.cpp
+++ b/ggml/src/ggml-sycl/mmvq.cpp
@@ -753,11 +753,7 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, QK_WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-            auto iq2xs_grid_ptr_ct1 = &iq2xs_grid[0];
-            auto ksigns64_ptr_ct1 = &ksigns64[0];
-
+        stream->submit([&](sycl::handler & cgh) {
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
@@ -780,9 +776,6 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy,
     {
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq2xs_grid_ptr_ct1 = &iq2xs_grid[0];
-            auto ksigns64_ptr_ct1 = &ksigns64[0];
-
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
@@ -805,9 +798,6 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy,
     {
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3xxs_grid_ptr_ct1 = &iq3xxs_grid[0];
-            auto ksigns64_ptr_ct1 = &ksigns64[0];
-
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
@@ -830,8 +820,6 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy,
     {
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq3s_grid_ptr_ct1 = &iq3s_grid[0];
-
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
@@ -854,9 +842,6 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy,
     {
 
         stream->submit([&](sycl::handler &cgh) {
-            auto iq1s_grid_ptr_ct1 = &iq1s_grid_gpu[0];
-            auto ksigns64_ptr_ct1 = &ksigns64[0];
-
             cgh.parallel_for(
                 sycl::nd_range<3>(block_nums * block_dims, block_dims),
                 [=](sycl::nd_item<3> item_ct1)
@@ -954,7 +939,7 @@ void ggml_sycl_op_mul_mat_vec_q(
     const size_t q8_1_bs = QK8_1;
     // the main device has a larger memory buffer to hold the results from all GPUs
     // nrows_dst == nrows of the matrix that the kernel writes into
-    const int64_t nrows_dst = id == ctx.device ? ne00 : row_diff;
+
     for (int i = 0; i < src1_ncols; i++)
     {
         const size_t src1_ddq_i_offset = i * src1_padded_col_size * q8_1_ts / q8_1_bs;
@@ -1023,7 +1008,8 @@ void ggml_sycl_op_mul_mat_vec_q(
             break;
         }
     }
-    (void) src1;
-    (void) dst;
-    (void) src1_ddf_i;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddf_i);
+    GGML_UNUSED(ctx);
 }
diff --git a/ggml/src/ggml-sycl/norm.cpp b/ggml/src/ggml-sycl/norm.cpp
index 72d8fdb878c..9cf2be15575 100644
--- a/ggml/src/ggml-sycl/norm.cpp
+++ b/ggml/src/ggml-sycl/norm.cpp
@@ -31,7 +31,7 @@ static void norm_f32(const float* x, float* dst, const int ncols, const float ep
         */
         item_ct1.barrier(sycl::access::fence_space::local_space);
         mean_var = 0.f;
-        int nreduce = nwarps / WARP_SIZE;
+        size_t nreduce = nwarps / WARP_SIZE;
         for (size_t i = 0; i < nreduce; i += 1)
         {
             mean_var += s_sum[lane_id + i * WARP_SIZE];
@@ -55,7 +55,7 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con
     const int nthreads = item_ct1.get_local_range(2);
     const int nwarps = nthreads / WARP_SIZE;
     start += item_ct1.get_local_id(2);
-    int nreduce = nwarps / WARP_SIZE;
+    size_t nreduce = nwarps / WARP_SIZE;
 
     if (end >= ne_elements) {
         end = ne_elements;
@@ -163,7 +163,7 @@ static void rms_norm_f32(const float* x, float* dst, const int ncols, const floa
         converged control flow. You may need to adjust the code.
         */
         item_ct1.barrier(sycl::access::fence_space::local_space);
-        int nreduce = nwarps / WARP_SIZE;
+        size_t nreduce = nwarps / WARP_SIZE;
         tmp = 0.f;
         for (size_t i = 0; i < nreduce; i += 1)
         {
@@ -352,6 +352,7 @@ void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor*
     (void)src1;
     (void)dst;
     (void)src1_dd;
+    GGML_UNUSED(ctx);
 }
 
 void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0,
diff --git a/ggml/src/ggml-sycl/outprod.cpp b/ggml/src/ggml-sycl/outprod.cpp
index e61cdc2ca5d..ef9af0b7633 100644
--- a/ggml/src/ggml-sycl/outprod.cpp
+++ b/ggml/src/ggml-sycl/outprod.cpp
@@ -40,14 +40,14 @@ void ggml_sycl_op_out_prod(ggml_backend_sycl_context& ctx, const ggml_tensor* sr
 
     try {
         // Perform matrix multiplication using oneMKL GEMM
-        oneapi::mkl::blas::column_major::gemm(*stream,
-            oneapi::mkl::transpose::nontrans, src1_op,
-            ne0, ne1, ne01,
-            alpha,
-            src0_d, ne00,
-            src1_d, ldb,
-            beta,
-            dst_d, ne0);
+#ifdef GGML_SYCL_NVIDIA
+        oneapi::mkl::blas::column_major::gemm(oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ *stream },
+                                              oneapi::mkl::transpose::nontrans, src1_op, ne0, ne1, ne01, alpha, src0_d,
+                                              ne00, src1_d, ldb, beta, dst_d, ne0);
+#else
+        oneapi::mkl::blas::column_major::gemm(*stream, oneapi::mkl::transpose::nontrans, src1_op, ne0, ne1, ne01, alpha,
+                                              src0_d, ne00, src1_d, ldb, beta, dst_d, ne0);
+#endif
     }
     catch (sycl::exception const& exc) {
         std::cerr << exc.what() << std::endl;
diff --git a/ggml/src/ggml-sycl/rope.cpp b/ggml/src/ggml-sycl/rope.cpp
index 1f06f78fa3d..1244b231af7 100644
--- a/ggml/src/ggml-sycl/rope.cpp
+++ b/ggml/src/ggml-sycl/rope.cpp
@@ -269,7 +269,8 @@ void ggml_sycl_op_rope(
         }
     }
 
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_dd);
+    GGML_UNUSED(ctx);
 }
diff --git a/ggml/src/ggml-sycl/softmax.cpp b/ggml/src/ggml-sycl/softmax.cpp
index 17a542e4903..a9b3fce0dc4 100644
--- a/ggml/src/ggml-sycl/softmax.cpp
+++ b/ggml/src/ggml-sycl/softmax.cpp
@@ -16,7 +16,7 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
     const int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
     const int nthreads = block_size;
     const int nwarps = nthreads / WARP_SIZE;
-    int nreduce = nwarps / WARP_SIZE;
+    size_t nreduce = nwarps / WARP_SIZE;
     float slope = 1.0f;
 
     // ALiBi
@@ -53,8 +53,9 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
     if (block_size > WARP_SIZE) {
         if (warp_id == 0) {
             buf[lane_id] = -INFINITY;
-            for (size_t i = 1; i < nreduce; i += 1)
+            for (size_t i = 1; i < nreduce; i += 1) {
                 buf[lane_id + i * WARP_SIZE] = -INFINITY;
+            }
         }
         item_ct1.barrier(sycl::access::fence_space::local_space);
 
@@ -63,8 +64,7 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
         }
         item_ct1.barrier(sycl::access::fence_space::local_space);
         max_val = buf[lane_id];
-        for (size_t i = 1; i < nreduce; i += 1)
-        {
+        for (size_t i = 1; i < nreduce; i += 1) {
             max_val = std::max(max_val, buf[lane_id + i * WARP_SIZE]);
         }
         max_val = warp_reduce_max(max_val, item_ct1);
@@ -89,8 +89,9 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
         item_ct1.barrier(sycl::access::fence_space::local_space);
         if (warp_id == 0) {
             buf[lane_id] = 0.f;
-            for (size_t i = 1; i < nreduce; i += 1)
+            for (size_t i = 1; i < nreduce; i += 1) {
                 buf[lane_id + i * WARP_SIZE] = 0.f;
+            }
         }
         item_ct1.barrier(sycl::access::fence_space::local_space);
 
@@ -100,8 +101,7 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
         item_ct1.barrier(sycl::access::fence_space::local_space);
 
         tmp = buf[lane_id];
-        for (size_t i = 1; i < nreduce; i += 1)
-        {
+        for (size_t i = 1; i < nreduce; i += 1) {
             tmp += buf[lane_id + i * WARP_SIZE];
         }
         tmp = warp_reduce_sum(tmp, item_ct1);
diff --git a/ggml/src/ggml-sycl/tsembd.cpp b/ggml/src/ggml-sycl/tsembd.cpp
index d5c227cd1ab..2ffe3cca917 100644
--- a/ggml/src/ggml-sycl/tsembd.cpp
+++ b/ggml/src/ggml-sycl/tsembd.cpp
@@ -68,4 +68,5 @@ void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, const ggml
     const int max_period = dst->op_params[1];
 
     timestep_embedding_f32_sycl(src0_d, dst_d, src0->ne[0], dst->nb[1], dim, max_period, stream);
+    GGML_UNUSED(src1);
 }
diff --git a/ggml/src/ggml-sycl/vecdotq.hpp b/ggml/src/ggml-sycl/vecdotq.hpp
index d2dccade20b..c5942008adf 100644
--- a/ggml/src/ggml-sycl/vecdotq.hpp
+++ b/ggml/src/ggml-sycl/vecdotq.hpp
@@ -968,8 +968,8 @@ vec_dot_iq3_xxs_q8_1(const void *__restrict__ vbq,
             grid1[0] ^ signs[0], signs[0], std::minus<>());
         const int grid_h = dpct::vectorized_binary<sycl::uchar4>(
             grid2[0] ^ signs[1], signs[1], std::minus<>());
-        sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi);
-        sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi);
+        sumi = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi);
+        sumi = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi);
         q8 += 8;
         aux32 >>= 7;
     }
@@ -1009,8 +1009,8 @@ vec_dot_iq3_s_q8_1(const void *__restrict__ vbq,
             grid1[0] ^ signs0, signs0, std::minus<>());
         const int grid_h = dpct::vectorized_binary<sycl::uchar4>(
             grid2[0] ^ signs1, signs1, std::minus<>());
-        sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi);
-        sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi);
+        sumi = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi);
+        sumi = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi);
         q8 += 8;
     }
     const float d =
diff --git a/ggml/src/ggml-sycl/wkv6.cpp b/ggml/src/ggml-sycl/wkv6.cpp
index 4c737f4bfce..75ddfb86ac0 100644
--- a/ggml/src/ggml-sycl/wkv6.cpp
+++ b/ggml/src/ggml-sycl/wkv6.cpp
@@ -59,7 +59,7 @@ static void rwkv_wkv_f32_kernel(
         float y = 0;
 
         // Process in chunks of 4 for better vectorization
-        sycl::float4 k4, r4, tf4, td4, s4, kv4;
+        sycl::float4 k4, r4, tf4, td4, s4;
         #pragma unroll
         for (int j = 0; j < head_size; j += 4) {
             // Load data in vec4 chunks
@@ -135,4 +135,7 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, const ggml_tensor* s
                 );
             });
     });
+
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
 }
diff --git a/ggml/src/ggml-threading.cpp b/ggml/src/ggml-threading.cpp
new file mode 100644
index 00000000000..25a19eedb90
--- /dev/null
+++ b/ggml/src/ggml-threading.cpp
@@ -0,0 +1,12 @@
+#include "ggml-threading.h"
+#include <mutex>
+
+std::mutex ggml_critical_section_mutex;
+
+void ggml_critical_section_start() {
+    ggml_critical_section_mutex.lock();
+}
+
+void ggml_critical_section_end(void) {
+    ggml_critical_section_mutex.unlock();
+}
diff --git a/ggml/src/ggml-threading.h b/ggml/src/ggml-threading.h
new file mode 100644
index 00000000000..dec2c8840aa
--- /dev/null
+++ b/ggml/src/ggml-threading.h
@@ -0,0 +1,14 @@
+#pragma once
+
+#include "ggml.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+GGML_API void ggml_critical_section_start(void);
+GGML_API void ggml_critical_section_end(void);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/ggml/src/ggml-vulkan/CMakeLists.txt b/ggml/src/ggml-vulkan/CMakeLists.txt
new file mode 100644
index 00000000000..6d46e5f24c1
--- /dev/null
+++ b/ggml/src/ggml-vulkan/CMakeLists.txt
@@ -0,0 +1,92 @@
+find_package(Vulkan COMPONENTS glslc REQUIRED)
+
+if (Vulkan_FOUND)
+    message(STATUS "Vulkan found")
+
+    ggml_add_backend_library(ggml-vulkan
+                             ggml-vulkan.cpp
+                             ../../include/ggml-vulkan.h
+                            )
+
+    # Compile a test shader to determine whether GL_NV_cooperative_matrix2 is supported.
+    # If it's not, there will be an error to stderr.
+    # If it's supported, set a define to indicate that we should compile those shaders
+    execute_process(COMMAND ${Vulkan_GLSLC_EXECUTABLE} -o - -fshader-stage=compute --target-env=vulkan1.3 "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat2_support.comp"
+                    OUTPUT_VARIABLE glslc_output
+                    ERROR_VARIABLE glslc_error)
+
+    if (${glslc_error} MATCHES ".*extension not supported: GL_NV_cooperative_matrix2.*")
+        message(STATUS "GL_NV_cooperative_matrix2 not supported by glslc")
+    else()
+        message(STATUS "GL_NV_cooperative_matrix2 supported by glslc")
+        add_compile_definitions(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+    endif()
+
+    target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
+    target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
+
+    # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
+    # Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
+    if (MSVC AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+        add_compile_definitions(_ITERATOR_DEBUG_LEVEL=0)
+    endif()
+
+    if (GGML_VULKAN_CHECK_RESULTS)
+        add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
+    endif()
+
+    if (GGML_VULKAN_DEBUG)
+        add_compile_definitions(GGML_VULKAN_DEBUG)
+    endif()
+
+    if (GGML_VULKAN_MEMORY_DEBUG)
+        add_compile_definitions(GGML_VULKAN_MEMORY_DEBUG)
+    endif()
+
+    if (GGML_VULKAN_SHADER_DEBUG_INFO)
+        add_compile_definitions(GGML_VULKAN_SHADER_DEBUG_INFO)
+    endif()
+
+    if (GGML_VULKAN_PERF)
+        add_compile_definitions(GGML_VULKAN_PERF)
+    endif()
+
+    if (GGML_VULKAN_VALIDATE)
+        add_compile_definitions(GGML_VULKAN_VALIDATE)
+    endif()
+
+    if (GGML_VULKAN_RUN_TESTS)
+        add_compile_definitions(GGML_VULKAN_RUN_TESTS)
+    endif()
+
+    add_subdirectory(vulkan-shaders)
+
+    set (_ggml_vk_genshaders_cmd vulkan-shaders-gen)
+    set (_ggml_vk_header     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp)
+    set (_ggml_vk_source     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp)
+    set (_ggml_vk_input_dir  ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders)
+    set (_ggml_vk_output_dir ${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv)
+
+    file(GLOB _ggml_vk_shader_deps "${_ggml_vk_input_dir}/*.comp")
+
+    add_custom_command(
+        OUTPUT ${_ggml_vk_header}
+                ${_ggml_vk_source}
+
+        COMMAND ${_ggml_vk_genshaders_cmd}
+            --glslc      ${Vulkan_GLSLC_EXECUTABLE}
+            --input-dir  ${_ggml_vk_input_dir}
+            --output-dir ${_ggml_vk_output_dir}
+            --target-hpp ${_ggml_vk_header}
+            --target-cpp ${_ggml_vk_source}
+            --no-clean
+
+        DEPENDS ${_ggml_vk_shader_deps} ${_ggml_vk_genshaders_cmd}
+        COMMENT "Generate vulkan shaders"
+    )
+
+    target_sources(ggml-vulkan PRIVATE ${_ggml_vk_source} ${_ggml_vk_header})
+
+else()
+    message(WARNING "Vulkan not found")
+endif()
diff --git a/ggml/src/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
similarity index 73%
rename from ggml/src/ggml-vulkan.cpp
rename to ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 169b5a3b7e9..1696b6e2736 100644
--- a/ggml/src/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -1,7 +1,8 @@
 #include "ggml-vulkan.h"
 #include <vulkan/vulkan_core.h>
-#if defined(GGML_VULKAN_RUN_TESTS) || defined(GGML_VULKAN_PERF)
+#if defined(GGML_VULKAN_RUN_TESTS) || defined(GGML_VULKAN_PERF) || defined(GGML_VULKAN_CHECK_RESULTS)
 #include <chrono>
+#include "ggml-cpu.h"
 #endif
 
 #include <vulkan/vulkan.hpp>
@@ -43,12 +44,6 @@
 
 #define MAX_VK_BUFFERS 256
 
-#ifndef K_QUANTS_PER_ITERATION
-#define K_QUANTS_PER_ITERATION 1
-#else
-static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
-#endif
-
 #define VK_CHECK(err, msg)                                          \
     do {                                                            \
         vk::Result err_ = (err);                                    \
@@ -106,6 +101,15 @@ struct vk_matmul_pipeline_struct {
 
 typedef std::shared_ptr<vk_matmul_pipeline_struct> vk_matmul_pipeline;
 
+struct vk_matmul_pipeline2 {
+    vk_matmul_pipeline2() {
+        f16acc = std::make_shared<vk_matmul_pipeline_struct>();
+        f32acc = std::make_shared<vk_matmul_pipeline_struct>();
+    }
+    vk_matmul_pipeline f32acc;
+    vk_matmul_pipeline f16acc;
+};
+
 struct vk_device_struct;
 typedef std::shared_ptr<vk_device_struct> vk_device;
 typedef std::weak_ptr<vk_device_struct> vk_device_ref;
@@ -149,29 +153,53 @@ struct vk_device_struct {
     std::string name;
     uint64_t max_memory_allocation_size;
     bool fp16;
+    bool pipeline_robustness;
     vk::Device device;
     uint32_t vendor_id;
     vk_queue compute_queue;
     vk_queue transfer_queue;
     bool single_queue;
     uint32_t subgroup_size;
+    uint32_t shader_core_count;
     bool uma;
+    bool float_controls_rte_fp16;
+
+    bool subgroup_size_control;
+    uint32_t subgroup_min_size;
+    uint32_t subgroup_max_size;
+    bool subgroup_require_full_support;
+
+    bool coopmat_support;
+    bool coopmat_acc_f32_support;
+    bool coopmat_acc_f16_support;
+    uint32_t coopmat_m;
+    uint32_t coopmat_n;
+    uint32_t coopmat_k;
+    bool coopmat2;
 
     size_t idx;
 
+    bool mul_mat_l;
+    bool mul_mat_m;
+    bool mul_mat_s;
+    bool mul_mat_id_l;
+    bool mul_mat_id_m;
+    bool mul_mat_id_s;
+
     vk_matmul_pipeline pipeline_matmul_f32;
     vk_matmul_pipeline pipeline_matmul_f32_f16;
-    vk_matmul_pipeline pipeline_matmul_f16;
-    vk_matmul_pipeline pipeline_matmul_f16_f32;
+    vk_matmul_pipeline2 pipeline_matmul_f16;
+    vk_matmul_pipeline2 pipeline_matmul_f16_f32;
     vk_pipeline pipeline_matmul_split_k_reduce;
 
-    vk_matmul_pipeline pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT];
+    vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_COUNT];
+    vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT];
 
     vk_matmul_pipeline pipeline_matmul_id_f32;
-    vk_matmul_pipeline pipeline_matmul_id_f16;
-    vk_matmul_pipeline pipeline_matmul_id_f16_f32;
+    vk_matmul_pipeline2 pipeline_matmul_id_f16;
+    vk_matmul_pipeline2 pipeline_matmul_id_f16_f32;
 
-    vk_matmul_pipeline pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT];
+    vk_matmul_pipeline2 pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT];
 
     vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
     vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT];
@@ -183,9 +211,10 @@ struct vk_device_struct {
     vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
     vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT];
     vk_pipeline pipeline_acc_f32;
-    vk_pipeline pipeline_add_f32, pipeline_add_f16_f32_f16;
-    vk_pipeline pipeline_mul_f32;
-    vk_pipeline pipeline_div_f32;
+    vk_pipeline pipeline_add_f32, pipeline_add_f32_norepeat;
+    vk_pipeline pipeline_add_f16_f32_f16, pipeline_add_f16_f32_f16_norepeat;
+    vk_pipeline pipeline_mul_f32, pipeline_mul_f32_norepeat;
+    vk_pipeline pipeline_div_f32, pipeline_div_f32_norepeat;
     vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
     vk_pipeline pipeline_upscale_f32;
     vk_pipeline pipeline_scale_f32;
@@ -208,6 +237,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_tanh_f32;
     vk_pipeline pipeline_diag_mask_inf_f32;
     vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
+    vk_pipeline pipeline_soft_max_f32_wg512, pipeline_soft_max_f32_f16_wg512;
     vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16;
     vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
     vk_pipeline pipeline_argsort_f32;
@@ -215,6 +245,15 @@ struct vk_device_struct {
     vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
     vk_pipeline pipeline_timestep_embedding_f32;
     vk_pipeline pipeline_pool2d_f32;
+    vk_pipeline pipeline_rwkv_wkv6_f32;
+
+    // [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned}
+    vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D80[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D96[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D112[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D128[GGML_TYPE_COUNT][2][2][2];
+    vk_pipeline pipeline_flash_attn_f32_f16_D256[GGML_TYPE_COUNT][2][2][2];
 
     std::unordered_map<std::string, vk_pipeline_ref> pipelines;
     std::unordered_map<std::string, uint64_t> pipeline_descriptor_set_requirements;
@@ -327,6 +366,40 @@ struct vk_mat_vec_id_push_constants {
     uint32_t nei0; uint32_t ne11;
 };
 
+struct vk_flash_attn_push_constants {
+    uint32_t N;
+    uint32_t KV;
+
+    uint32_t ne1;
+    uint32_t ne2;
+    uint32_t ne3;
+
+    uint32_t neq2;
+    uint32_t neq3;
+    uint32_t nek2;
+    uint32_t nek3;
+    uint32_t nev2;
+    uint32_t nev3;
+    uint32_t nem1;
+
+    uint32_t nb02;
+    uint32_t nb03;
+    uint32_t nb12;
+    uint32_t nb13;
+    uint32_t nb22;
+    uint32_t nb23;
+    uint32_t nb31;
+
+    float scale;
+    float max_bias;
+    float logit_softcap;
+
+    uint32_t mask;
+    uint32_t n_head_log2;
+    float m0;
+    float m1;
+};
+
 struct vk_op_push_constants {
     uint32_t KX;
     uint32_t KY;
@@ -340,7 +413,46 @@ struct vk_op_unary_push_constants {
     uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
     uint32_t d_offset;
     float param1; float param2;
+    uint32_t ne0_012mp; uint32_t ne0_012L;
+    uint32_t ne0_01mp;  uint32_t ne0_01L;
+    uint32_t ne0_0mp;   uint32_t ne0_0L;
+    uint32_t ne1_012mp; uint32_t ne1_012L;
+    uint32_t ne1_01mp;  uint32_t ne1_01L;
+    uint32_t ne1_0mp;   uint32_t ne1_0L;
 };
+static_assert(sizeof(vk_op_unary_push_constants) <= 128, "sizeof(vk_op_unary_push_constants) must be <= 128");
+
+// See https://gmplib.org/~tege/divcnst-pldi94.pdf figure 4.1.
+// Precompute mp (m' in the paper) and L such that division
+// can be computed using a multiply (high 32b of 64b result)
+// and a shift:
+//
+// n/d = (mulhi(n, mp) + n) >> L;
+static void init_fastdiv_values(uint32_t d, uint32_t &mp, uint32_t &L)
+{
+    // compute L = ceil(log2(d));
+    L = 0;
+    while (L < 32 && (uint32_t{1} << L) < d) {
+        L++;
+    }
+
+    mp = (uint32_t)((uint64_t{1} << 32) * ((uint64_t{1} << L) - d) / d + 1);
+}
+
+template <typename T> void init_pushconst_fastdiv(T &p) {
+    GGML_UNUSED(p);
+    static_assert(!std::is_const<T>::value, "unexpected type");
+}
+
+template <> void init_pushconst_fastdiv(vk_op_unary_push_constants &p) {
+    // Compute magic values to divide by these six numbers.
+    init_fastdiv_values(p.ne02*p.ne01*p.ne00,  p.ne0_012mp,    p.ne0_012L);
+    init_fastdiv_values(p.ne01*p.ne00,         p.ne0_01mp,     p.ne0_01L);
+    init_fastdiv_values(p.ne00,                p.ne0_0mp,      p.ne0_0L);
+    init_fastdiv_values(p.ne12*p.ne11*p.ne10,  p.ne1_012mp,    p.ne1_012L);
+    init_fastdiv_values(p.ne11*p.ne10,         p.ne1_01mp,     p.ne1_01L);
+    init_fastdiv_values(p.ne10,                p.ne1_0mp,      p.ne1_0L);
+}
 
 struct vk_op_binary_push_constants {
     uint32_t ne;
@@ -378,6 +490,7 @@ struct vk_op_soft_max_push_constants {
     float m0;
     float m1;
     uint32_t n_head_log2;
+    uint32_t nrows_x;
 };
 
 struct vk_op_argsort_push_constants {
@@ -416,6 +529,13 @@ struct vk_op_pool2d_push_constants {
     int32_t p0; int32_t p1;
 };
 
+struct vk_op_rwkv_wkv6_push_constants {
+    uint32_t B;
+    uint32_t T;
+    uint32_t C;
+    uint32_t H;
+};
+
 // Allow pre-recording command buffers
 struct vk_staging_memcpy {
     vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
@@ -642,8 +762,12 @@ static uint32_t compile_count = 0;
 static std::mutex compile_count_mutex;
 static std::condition_variable compile_count_cond;
 
-static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, const std::string name, size_t spv_size, const void* spv_data, const std::string entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t> specialization_constants, uint32_t align) {
-    VK_LOG_DEBUG("ggml_vk_create_pipeline(" << device->name << ", " << name << ", " << entrypoint << ", " << parameter_count << ", " << push_constant_size << ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " << align << ")");
+static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, const std::string name, size_t spv_size, const void* spv_data, const std::string entrypoint,
+                                         uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t> specialization_constants,
+                                         uint32_t align, bool disable_robustness, bool require_full_subgroups, uint32_t required_subgroup_size) {
+    VK_LOG_DEBUG("ggml_vk_create_pipeline(" << device->name << ", " << name << ", " << entrypoint << ", " << parameter_count << ", " << push_constant_size <<
+                 ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " << align <<
+                 ", " << disable_robustness << ", " << require_full_subgroups << ", " << required_subgroup_size << ")");
     GGML_ASSERT(parameter_count > 0);
     GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT
 
@@ -702,16 +826,39 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
         specialization_constants.data()
     );
 
+    vk::PipelineShaderStageCreateFlags pipeline_shader_stage_create_flags{};
+
+    if (device->subgroup_require_full_support && require_full_subgroups) {
+        pipeline_shader_stage_create_flags |= vk::PipelineShaderStageCreateFlagBits::eRequireFullSubgroupsEXT;
+    }
+
     vk::PipelineShaderStageCreateInfo pipeline_shader_create_info(
-            vk::PipelineShaderStageCreateFlags(),
+            pipeline_shader_stage_create_flags,
             vk::ShaderStageFlagBits::eCompute,
             pipeline->shader_module,
             entrypoint.c_str(),
             &specialization_info);
+
+    vk::PipelineShaderStageRequiredSubgroupSizeCreateInfoEXT pipeline_shader_stage_required_subgroup_size_create_info;
+    pipeline_shader_stage_required_subgroup_size_create_info.requiredSubgroupSize = required_subgroup_size;
+    if (device->subgroup_size_control && required_subgroup_size > 0) {
+        GGML_ASSERT(device->subgroup_min_size <= required_subgroup_size && required_subgroup_size <= device->subgroup_max_size);
+        pipeline_shader_create_info.setPNext(&pipeline_shader_stage_required_subgroup_size_create_info);
+    }
+
     vk::ComputePipelineCreateInfo compute_pipeline_create_info(
-        vk::PipelineCreateFlags(),
+        vk::PipelineCreateFlags{},
         pipeline_shader_create_info,
         pipeline->layout);
+
+    vk::PipelineRobustnessCreateInfoEXT rci;
+
+    if (device->pipeline_robustness && disable_robustness) {
+        rci.storageBuffers = vk::PipelineRobustnessBufferBehaviorEXT::eDisabled;
+        rci.uniformBuffers = vk::PipelineRobustnessBufferBehaviorEXT::eDisabled;
+        compute_pipeline_create_info.setPNext(&rci);
+    }
+
     pipeline->pipeline = device->device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value;
 
     {
@@ -1204,61 +1351,186 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events
     );
 }
 
+// number of rows/cols for flash attention shader
+static constexpr uint32_t flash_attention_num_small_rows = 32;
+static std::array<uint32_t, 2> fa_rows_cols(uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) {
+    GGML_UNUSED(clamp);
+
+    // small rows, large cols
+    if (small_rows) {
+        return {flash_attention_num_small_rows, 128};
+    }
+    // small cols to reduce register count
+    if (ggml_is_quantized(type) || D == 256) {
+        return {64, 32};
+    }
+    return {64, 64};
+};
+
+static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vector<uint32_t>& warptile, bool mul_mat_id) {
+    // Needs to be kept up to date on shader changes
+    const uint32_t bank_conflict_offset = device->coopmat_support ? 8 : 1;
+    const uint32_t type_size = device->fp16 ? sizeof(ggml_fp16_t) : sizeof(float);
+    const uint32_t warps = warptile[0] / warptile[10];
+
+    const uint32_t load_bufs = (warptile[1] + warptile[2]) * (warptile[3] + bank_conflict_offset) * type_size;
+    const uint32_t mmid_row_ids = mul_mat_id ? 3072 * sizeof(uint32_t) : 0;
+    const uint32_t coopmat_stage = device->coopmat_support ? warptile[7] * warptile[8] / warps * sizeof(float) : 0;
+
+    return (load_bufs + mmid_row_ids + coopmat_stage) <= device->properties.limits.maxComputeSharedMemorySize;
+}
+
 static void ggml_vk_load_shaders(vk_device& device) {
     VK_LOG_DEBUG("ggml_vk_load_shaders(" << device->name << ")");
 
     std::cerr << "ggml_vulkan: Compiling shaders";
 
-    // mulmat
-    std::initializer_list<uint32_t> warptile_l = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
-    std::initializer_list<uint32_t> warptile_m = { 128,  64,  64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
-    std::initializer_list<uint32_t> warptile_s = { std::max(device->subgroup_size, 16u),  32,  32, 16, 32, 32, 2, 2, 2, device->subgroup_size };
-
-    std::initializer_list<uint32_t> warptile_mmq_l = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
-    std::initializer_list<uint32_t> warptile_mmq_m = { 128,  64,  64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
-    std::initializer_list<uint32_t> warptile_mmq_s = { std::max(device->subgroup_size, 16u),  32,  32, 32, 32, 32, 2, 2, 2, device->subgroup_size };
-
-    std::array<uint32_t, 3> l_wg_denoms = {128, 128, 1 };
-    std::array<uint32_t, 3> m_wg_denoms = { 64,  64, 1 };
-    std::array<uint32_t, 3> s_wg_denoms = { 32,  32, 1 };
+    // some shaders have a minimum subgroup size
+    const uint32_t subgroup_size_16 = std::max(device->subgroup_size, 16u);
+    const uint32_t subgroup_size_32 = std::max(device->subgroup_size, 32u);
 
-    uint32_t l_align = 128;
-    uint32_t m_align =  64;
-    uint32_t s_align =  32;
+    // mulmat
+    std::vector<uint32_t> l_warptile, m_warptile, s_warptile,
+                          l_warptile_mmq, m_warptile_mmq, s_warptile_mmq,
+                          l_warptile_mmq_k, m_warptile_mmq_k, s_warptile_mmq_k,
+                          l_warptile_mmqid, m_warptile_mmqid, s_warptile_mmqid;
+    std::array<uint32_t, 3> l_wg_denoms, m_wg_denoms, s_wg_denoms,
+                            l_mmq_wg_denoms, m_mmq_wg_denoms, s_mmq_wg_denoms,
+                            l_mmq_wg_denoms_k, m_mmq_wg_denoms_k, s_mmq_wg_denoms_k,
+                            l_mmqid_wg_denoms, m_mmqid_wg_denoms, s_mmqid_wg_denoms;
+
+    uint32_t l_align, m_align, s_align;
+    if (device->coopmat2) {
+        // spec constants and tile sizes for non-quant matmul/matmul_id
+        l_warptile = { 256, 128, 256, 64 };
+        m_warptile = { 256, 128, 128, 64 };
+        s_warptile = { 128,  32,  16, 64 };
+        l_wg_denoms = {128, 256, 1 };
+        m_wg_denoms = {128, 128, 1 };
+        s_wg_denoms = { 32,  16, 1 };
+
+        // spec constants and tile sizes for quant matmul (non-Qi_K)
+        l_warptile_mmq = { 256, 128, 256, 64 };
+        m_warptile_mmq = { 256, 128, 128, 64 };
+        s_warptile_mmq = { 256, 128, 128, 64 };
+        l_mmq_wg_denoms = { 128, 256, 1 };
+        m_mmq_wg_denoms = { 128, 128, 1 };
+        s_mmq_wg_denoms = { 128, 128, 1 };
+
+        // spec constants and tile sizes for quant matmul (Qi_K)
+        l_warptile_mmq_k = { 256, 128, 512, 16 };
+        m_warptile_mmq_k = { 256, 128, 256, 16 };
+        s_warptile_mmq_k = { 256, 32, 128, 64 };
+        l_mmq_wg_denoms_k = { 128, 512, 1 };
+        m_mmq_wg_denoms_k = { 128, 256, 1 };
+        s_mmq_wg_denoms_k = { 32, 128, 1 };
+
+        // spec constants and tile sizes for quant matmul_id
+        l_warptile_mmqid = { 256, 128, 128, 16 };
+        m_warptile_mmqid = { 256, 128, 64, 16 };
+        s_warptile_mmqid = { 256, 64, 64, 16 };
+        l_mmqid_wg_denoms = { 128, 128, 1 };
+        m_mmqid_wg_denoms = { 128, 64, 1 };
+        s_mmqid_wg_denoms = { 64, 64, 1 };
+
+        l_align = 128;
+        m_align =  64;
+        s_align =  32;
+    } else {
+        // Matrix cores require different warp group sizes
+        const uint32_t tm_l = device->coopmat_support ? device->coopmat_m : 4;
+        const uint32_t tm_m = device->coopmat_support ? device->coopmat_m : 4;
+        const uint32_t tm_s = device->coopmat_support ? device->coopmat_m : 2;
+        const uint32_t tn_l = device->coopmat_support ? device->coopmat_n : 4;
+        const uint32_t tn_m = device->coopmat_support ? device->coopmat_n : 2;
+        const uint32_t tn_s = device->coopmat_support ? device->coopmat_n : 2;
+        const uint32_t tk_l = device->coopmat_support ? device->coopmat_k : 1;
+        const uint32_t tk_m = device->coopmat_support ? device->coopmat_k : 1;
+        const uint32_t tk_s = device->coopmat_support ? device->coopmat_k : 1;
+
+        l_warptile = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, tm_l, tn_l, tk_l, device->subgroup_size };
+        m_warptile = { 128,  64,  64, 16, device->subgroup_size, 32, 2, tm_m, tn_m, tk_m, device->subgroup_size };
+        s_warptile = { subgroup_size_16, 32, 32, 16, 32, 32, 2, tm_s, tn_s, tk_s, device->subgroup_size };
+
+        l_warptile_mmq = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, tm_l, tn_l, tk_l, device->subgroup_size };
+        m_warptile_mmq = { 128,  64,  64, 32, device->subgroup_size, 32, 2, tm_m, tn_m, tk_m, device->subgroup_size };
+        s_warptile_mmq = { subgroup_size_32, 32, 32, 32, 32, 32, 2, tm_s, tn_s, tk_s, device->subgroup_size };
+
+        l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 };
+        m_mmq_wg_denoms = m_wg_denoms = { 64,  64, 1 };
+        s_mmq_wg_denoms = s_wg_denoms = { 32,  32, 1 };
+        l_align = 128;
+        m_align =  64;
+        s_align =  32;
+
+        // Fallback to smaller sizes if there's not enough shared memory. Given the current shaders
+        // and tile sizes, this should handle 16KB, 32KB, and 48KB+.
+        // This logic doesn't explicitly account for the 12KB row_ids in the mul_mat_mat_id shaders.
+        // But the numbers happen to work out for 32KB shared memory size that when using the medium
+        // size there's enough room for everything, and we assert for this.
+        uint32_t shmem_needed = (l_warptile[1] + l_warptile[2]) * (l_warptile[3] + 1) * sizeof(float);
+        if (shmem_needed > device->properties.limits.maxComputeSharedMemorySize) {
+            l_warptile = m_warptile;
+            l_wg_denoms = m_wg_denoms;
+            shmem_needed = (l_warptile[1] + l_warptile[2]) * (l_warptile[3] + 1) * sizeof(float);
+            GGML_ASSERT(shmem_needed <= device->properties.limits.maxComputeSharedMemorySize);
+        }
+        if (device->properties.limits.maxComputeSharedMemorySize >= 32768) {
+            // assert mul_mat_mat_id shaders will fit.
+            GGML_ASSERT(shmem_needed + 3072*4 <= device->properties.limits.maxComputeSharedMemorySize);
+        }
+
+        shmem_needed = (l_warptile_mmq[1] + l_warptile_mmq[2]) * (l_warptile_mmq[3] + 1) * sizeof(float);
+        if (shmem_needed > device->properties.limits.maxComputeSharedMemorySize) {
+            if (device->properties.limits.maxComputeSharedMemorySize == 32768) {
+                l_warptile_mmq = m_warptile_mmq;
+                l_mmq_wg_denoms = m_mmq_wg_denoms;
+            } else {
+                l_warptile_mmq = s_warptile_mmq;
+                l_mmq_wg_denoms = s_mmq_wg_denoms;
+            }
+            shmem_needed = (l_warptile_mmq[1] + l_warptile_mmq[2]) * (l_warptile_mmq[3] + 1) * sizeof(float);
+            GGML_ASSERT(shmem_needed <= device->properties.limits.maxComputeSharedMemorySize);
+        }
+        if (device->properties.limits.maxComputeSharedMemorySize >= 32768) {
+            // assert mul_mat_mat_id shaders will fit.
+            GGML_ASSERT(shmem_needed + 3072*4 <= device->properties.limits.maxComputeSharedMemorySize);
+        }
+        // Disable medium and large matrix multiplication if not enough shared memory is available
+        // Check mmq warptiles as the largest configuration
+        // Throw an error if not enough for any matrix multiplication is available
+        if (!ggml_vk_matmul_shmem_support(device, s_warptile_mmq, false)) {
+            std::cerr << "ggml_vulkan: Error: Shared memory size too small for matrix multiplication." << std::endl;
+            throw std::runtime_error("Shared memory size too small for matrix multiplication.");
+        } else if (!ggml_vk_matmul_shmem_support(device, m_warptile_mmq, false)) {
+            device->mul_mat_m = false;
+            device->mul_mat_l = false;
+        } else if (!ggml_vk_matmul_shmem_support(device, l_warptile_mmq, false)) {
+            device->mul_mat_l = false;
+        }
+
+        // Disable mul_mat_id if not enough shared memory is available
+        if (!ggml_vk_matmul_shmem_support(device, s_warptile_mmq, true)) {
+            device->mul_mat_id_s = false;
+            device->mul_mat_id_m = false;
+            device->mul_mat_id_l = false;
+        } else if (!ggml_vk_matmul_shmem_support(device, m_warptile_mmq, true)) {
+            device->mul_mat_id_m = false;
+            device->mul_mat_id_l = false;
+        } else if (!ggml_vk_matmul_shmem_support(device, l_warptile_mmq, true)) {
+            device->mul_mat_id_l = false;
+        }
+    }
 
     device->pipeline_matmul_f32 = std::make_shared<vk_matmul_pipeline_struct>();
     device->pipeline_matmul_f32_f16 = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_matmul_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_matmul_f16 = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>();
 
     device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_matmul_id_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_matmul_id_f16 = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K] = std::make_shared<vk_matmul_pipeline_struct>();
-    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>();
 
     std::vector<std::future<void>> compiles;
-    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t>&& specialization_constants, uint32_t align) {
+    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint,
+                                              uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants,
+                                              uint32_t align, bool disable_robustness = false, bool require_full_subgroups = false, uint32_t required_subgroup_size = 0) {
         {
             // wait until fewer than N compiles are in progress
             uint32_t N = std::max(1u, std::thread::hardware_concurrency());
@@ -1268,459 +1540,368 @@ static void ggml_vk_load_shaders(vk_device& device) {
             }
             compile_count++;
         }
-        compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint, parameter_count, push_constant_size, wg_denoms, specialization_constants, align));
+        compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint,
+                                      parameter_count, push_constant_size, wg_denoms, specialization_constants, align, disable_robustness, require_full_subgroups, required_subgroup_size));
     };
 
-    if (device->fp16) {
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->l, "matmul_f32_f16_l", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->m, "matmul_f32_f16_m", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->s, "matmul_f32_f16_s", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_l, "matmul_f32_f16_aligned_l", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_m, "matmul_f32_f16_aligned_m", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_s, "matmul_f32_f16_aligned_s", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_m, "matmul_q6_k_f32_aligned_m", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_s, "matmul_q6_k_f32_aligned_s", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->l, "matmul_iq4_nl_f32_l", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->m, "matmul_iq4_nl_f32_m", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->s, "matmul_iq4_nl_f32_s", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_l, "matmul_iq4_nl_f32_aligned_l", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_m, "matmul_iq4_nl_f32_aligned_m", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_s, "matmul_iq4_nl_f32_aligned_s", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_1_f32_l", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_1_f32_m", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_1_f32_s", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_1_f32_aligned_l", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_1_f32_aligned_m", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_1_f32_aligned_s", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->l, "matmul_id_iq4_nl_f32_l", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->m, "matmul_id_iq4_nl_f32_m", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->s, "matmul_id_iq4_nl_f32_s", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_l, "matmul_id_iq4_nl_f32_aligned_l", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_m, "matmul_id_iq4_nl_f32_aligned_m", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_s, "matmul_id_iq4_nl_f32_aligned_s", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+    if (device->coopmat2) {
+
+        auto const &fa_wg_denoms = [&](uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::array<uint32_t, 3> {
+            return {fa_rows_cols(D, clamp, type, small_rows)[0], 1, 1};
+        };
+
+        auto const &fa_spec_constants = [&](uint32_t D, uint32_t clamp, ggml_type type, bool small_rows) -> std::vector<uint32_t> {
+            // For large number of rows, 128 invocations seems to work best.
+            // For small number of rows (e.g. N==1), 256 works better. But matrix granularity for 256 is 32, so we
+            // can't use 256 for D==80.
+            uint32_t wg_size = (small_rows && (D % 32) == 0) ? 256 : 128;
+            auto rows_cols = fa_rows_cols(D, clamp, type, small_rows);
+            return {wg_size, rows_cols[0], rows_cols[1], (D), clamp};
+        };
+
+#define CREATE_FA2(TYPE, NAMELC, D) \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][0][0][0], "flash_attn_f32_f16_D" #D "_f16acc"         #NAMELC,           flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,1,TYPE,false), fa_spec_constants(D,1,TYPE,false), 1);     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][0][0][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc" #NAMELC,           flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,0,TYPE,false), fa_spec_constants(D,0,TYPE,false), fa_rows_cols(D,0,TYPE,false)[1]);     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][1][0][0], "flash_attn_f32_f16_D" #D "_f32acc"         #NAMELC,           flash_attn_f32_f16_ ## NAMELC ## _cm2_len,         flash_attn_f32_f16_ ## NAMELC ## _cm2_data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,1,TYPE,false), fa_spec_constants(D,1,TYPE,false), 1);     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][1][0][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc" #NAMELC,           flash_attn_f32_f16_ ## NAMELC ## _cm2_len,         flash_attn_f32_f16_ ## NAMELC ## _cm2_data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,0,TYPE,false), fa_spec_constants(D,0,TYPE,false), fa_rows_cols(D,0,TYPE,false)[1]);     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][0][1][0], "flash_attn_f32_f16_D" #D "_f16acc_smallrows"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,1,TYPE,true), fa_spec_constants(D,1,TYPE,true), 1);     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][0][1][1], "flash_attn_f32_f16_D" #D "_aligned_f16acc_smallrows" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_len,  flash_attn_f32_f16_ ## NAMELC ## _f16acc_cm2_data,  "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,0,TYPE,true), fa_spec_constants(D,0,TYPE,true), fa_rows_cols(D,0,TYPE,true)[1]);     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][1][1][0], "flash_attn_f32_f16_D" #D "_f32acc_smallrows"         #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _cm2_len,         flash_attn_f32_f16_ ## NAMELC ## _cm2_data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,1,TYPE,true), fa_spec_constants(D,1,TYPE,true), 1);     \
+        ggml_vk_create_pipeline(device, device->pipeline_flash_attn_f32_f16_D ## D[TYPE][1][1][1], "flash_attn_f32_f16_D" #D "_aligned_f32acc_smallrows" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _cm2_len,         flash_attn_f32_f16_ ## NAMELC ## _cm2_data,         "main", 5, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(D,0,TYPE,true), fa_spec_constants(D,0,TYPE,true), fa_rows_cols(D,0,TYPE,true)[1]);     \
+
+#define CREATE_FA(TYPE, NAMELC) \
+        CREATE_FA2(TYPE, NAMELC, 64) \
+        CREATE_FA2(TYPE, NAMELC, 80) \
+        CREATE_FA2(TYPE, NAMELC, 96) \
+        CREATE_FA2(TYPE, NAMELC, 112) \
+        CREATE_FA2(TYPE, NAMELC, 128) \
+        CREATE_FA2(TYPE, NAMELC, 256)
+
+        CREATE_FA(GGML_TYPE_F16, f16)
+        CREATE_FA(GGML_TYPE_Q4_0, q4_0)
+        CREATE_FA(GGML_TYPE_Q4_1, q4_1)
+        CREATE_FA(GGML_TYPE_Q5_0, q5_0)
+        CREATE_FA(GGML_TYPE_Q5_1, q5_1)
+        CREATE_FA(GGML_TYPE_Q8_0, q8_0)
+        // K dequants currently disabled because D dimension is rounded up to 256 and runs inefficiently
+        //CREATE_FA(GGML_TYPE_Q2_K, q2_k)
+        //CREATE_FA(GGML_TYPE_Q3_K, q3_k)
+        //CREATE_FA(GGML_TYPE_Q4_K, q4_k)
+        //CREATE_FA(GGML_TYPE_Q5_K, q5_k)
+        //CREATE_FA(GGML_TYPE_Q6_K, q6_k)
+        CREATE_FA(GGML_TYPE_IQ4_NL, iq4_nl)
+#undef CREATE_FA
+
+        // Create 6 variants, {s,m,l}x{unaligned,aligned}
+#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _cm2_len, NAMELC ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align);   \
+        ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _cm2_len, NAMELC ## _aligned ## F16ACC ## _cm2_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align);   \
+
+        // Create 2 variants, {f16,f32} accumulator
+#define CREATE_MM2(PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT) \
+        CREATE_MM(PIPELINE_NAME . f16acc, NAMELC, _f16acc, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT)   \
+        CREATE_MM(PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT)   \
+
+        CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3)
+
+        CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3)
+        CREATE_MM2(pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f16, _f16acc, mmq_wg_denoms_k, warptile_mmq_k, vk_mat_mat_push_constants, 3)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_f16[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f16, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3)
+
+        CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
+
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+#undef CREATE_MM
+#undef CREATE_MM2
+    } else
+#endif  // defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+    if (device->coopmat_support) {
+        // Create 6 variants, {s,m,l}x{unaligned,aligned}
+#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+        if (device->mul_mat ## ID ## _l) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1, false, true);   \
+        if (device->mul_mat ## ID ## _m) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1, false, true);   \
+        if (device->mul_mat ## ID ## _s) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _coopmat_len, NAMELC ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1, false, true);   \
+        if (device->mul_mat ## ID ## _l) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align, false, true);   \
+        if (device->mul_mat ## ID ## _m) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align, false, true);   \
+        if (device->mul_mat ## ID ## _s) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _coopmat_len, NAMELC ## _aligned ## F16ACC ## _coopmat_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align, false, true);   \
+
+        // Create 2 variants, {f16,f32} accumulator
+#define CREATE_MM2(PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+        if (device->coopmat_acc_f16_support) { \
+            CREATE_MM(PIPELINE_NAME . f16acc, NAMELC, _f16acc, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+        } \
+        if (device->coopmat_acc_f32_support) { \
+            CREATE_MM(PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+        } \
+
+        CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+
+        if (device->coopmat_acc_f16_support) {
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        } else {
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+            CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        }
+
+        // If there's not enough shared memory for row_ids and the result tile, don't create these pipelines.
+        if (device->mul_mat_id_s || device->mul_mat_id_m || device->mul_mat_id_l) {
+            CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+            CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+            CREATE_MM2(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+
+            if (device->coopmat_acc_f16_support) {
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            } else {
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+                CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, , wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            }
+        }
+#undef CREATE_MM2
+#undef CREATE_MM
+    } else if (device->fp16) {
+        // Create 6 variants, {s,m,l}x{unaligned,aligned}
+#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+        if (device->mul_mat ## ID ## _l) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
+        if (device->mul_mat ## ID ## _m) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
+        if (device->mul_mat ## ID ## _s) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _len, NAMELC ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \
+        if (device->mul_mat ## ID ## _l) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align);   \
+        if (device->mul_mat ## ID ## _m) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align);   \
+        if (device->mul_mat ## ID ## _s) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _len, NAMELC ## _aligned ## F16ACC ## _data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align);   \
+
+        // Create 2 variants, {f16,f32} accumulator
+#define CREATE_MM2(PIPELINE_NAME, NAMELC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+        CREATE_MM(PIPELINE_NAME . f16acc, NAMELC, _f16acc, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+        CREATE_MM(PIPELINE_NAME . f32acc, NAMELC, , WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+
+        CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(pipeline_matmul_f16, matmul_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+        CREATE_MM2(pipeline_matmul_f16_f32, matmul_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f16acc, matmul_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f16acc, matmul_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f16acc, matmul_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f16acc, matmul_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f16acc, matmul_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f16acc, matmul_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f16acc, matmul_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f16acc, matmul_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f16acc, matmul_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f16acc, matmul_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f16acc, matmul_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+
+        // If there's not enough shared memory for row_ids and the result tile, don't create these pipelines.
+        if (device->mul_mat_id_s || device->mul_mat_id_m || device->mul_mat_id_l) {
+            CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+            CREATE_MM2(pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+            CREATE_MM2(pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc, matmul_id_iq4_nl_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        }
+#undef CREATE_MM2
+#undef CREATE_MM
     } else {
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->l, "matmul_f32_f16_l", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->m, "matmul_f32_f16_m", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->s, "matmul_f32_f16_s", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_l, "matmul_f32_f16_aligned_l", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_m, "matmul_f32_f16_aligned_m", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_s, "matmul_f32_f16_aligned_s", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_m, "matmul_q6_k_f32_aligned_m", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_s, "matmul_q6_k_f32_aligned_s", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->l, "matmul_iq4_nl_f32_l", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->m, "matmul_iq4_nl_f32_m", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->s, "matmul_iq4_nl_f32_s", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_l, "matmul_iq4_nl_f32_aligned_l", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_m, "matmul_iq4_nl_f32_aligned_m", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_s, "matmul_iq4_nl_f32_aligned_s", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_1_f32_l", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_1_f32_m", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_1_f32_s", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_1_f32_aligned_l", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_1_f32_aligned_m", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_1_f32_aligned_s", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->l, "matmul_id_iq4_nl_f32_l", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->m, "matmul_id_iq4_nl_f32_m", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->s, "matmul_id_iq4_nl_f32_s", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_l, "matmul_id_iq4_nl_f32_aligned_l", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_m, "matmul_id_iq4_nl_f32_aligned_m", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
-        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_s, "matmul_id_iq4_nl_f32_aligned_s", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        // Create 6 variants, {s,m,l}x{unaligned,aligned}
+#define CREATE_MM(PIPELINE_NAME, NAMELC, F16ACC, WG_DENOMS, WARPTILE, PUSHCONST, PARAMCOUNT, ID) \
+        if (device->mul_mat ## ID ## _l) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->l, #NAMELC #F16ACC "_l", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, 1);   \
+        if (device->mul_mat ## ID ## _m) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->m, #NAMELC #F16ACC "_m", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, 1);   \
+        if (device->mul_mat ## ID ## _s) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->s, #NAMELC #F16ACC "_s", NAMELC ## F16ACC ## _fp32_len, NAMELC ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, 1);   \
+        if (device->mul_mat ## ID ## _l) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_l, #NAMELC #F16ACC "_aligned_l", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), l_ ## WG_DENOMS, l_ ## WARPTILE, l_align);   \
+        if (device->mul_mat ## ID ## _m) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_m, #NAMELC #F16ACC "_aligned_m", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), m_ ## WG_DENOMS, m_ ## WARPTILE, m_align);   \
+        if (device->mul_mat ## ID ## _s) \
+            ggml_vk_create_pipeline(device, device-> PIPELINE_NAME ->a_s, #NAMELC #F16ACC "_aligned_s", NAMELC ## _aligned ## F16ACC ## _fp32_len, NAMELC ## _aligned ## F16ACC ## _fp32_data, "main", PARAMCOUNT, sizeof(PUSHCONST), s_ ## WG_DENOMS, s_ ## WARPTILE, s_align);   \
+
+        CREATE_MM(pipeline_matmul_f32, matmul_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_matmul_f32_f16, matmul_f32_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_matmul_f16.f32acc, matmul_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_matmul_f16_f32.f32acc, matmul_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, );
+
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0].f32acc, matmul_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1].f32acc, matmul_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0].f32acc, matmul_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1].f32acc, matmul_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0].f32acc, matmul_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K].f32acc, matmul_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K].f32acc, matmul_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K].f32acc, matmul_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K].f32acc, matmul_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K].f32acc, matmul_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+        CREATE_MM(pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL].f32acc, matmul_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_push_constants, 3, );
+
+        // If there's not enough shared memory for row_ids and the result tile, don't create these pipelines.
+        if (device->mul_mat_id_s || device->mul_mat_id_m || device->mul_mat_id_l) {
+            CREATE_MM(pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+            CREATE_MM(pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+            CREATE_MM(pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, 4, _id);
+
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f32acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f32acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f32acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f32acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f32acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f32acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f32acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f32acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f32acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+            CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc, matmul_id_iq4_nl_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        }
+#undef CREATE_MM
     }
 
     // mul mat vec
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32",  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32",  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32",  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32",  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+
+    // AMD GCN and Intel graphics cards perform best when the number of rows per shader is doubled
+    uint32_t rm = 1;
+    if ((device->vendor_id == VK_VENDOR_ID_AMD && device->subgroup_min_size == 64 && device->subgroup_max_size == 64) || device->vendor_id == VK_VENDOR_ID_INTEL)
+        rm = 2;
+
+    // computing additional rows per workgroup is a benefit for Q4_0 -> Q5_1, but not for Q8_0.
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32",  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32",  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
+
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32",  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32",  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
+
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1*rm, 1, 1}, {device->subgroup_size, 1*rm}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
 
     // dequant shaders
     ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16",   dequant_f32_len,  dequant_f32_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
@@ -1755,7 +1936,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q8_0], "get_rows_q8_0_f32", get_rows_q8_0_f32_len, get_rows_q8_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ4_NL], "get_rows_iq4_nl_f32", get_rows_iq4_nl_f32_len, get_rows_iq4_nl_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32, "mul_mat_vec_p021_f16_f32", mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
@@ -1772,13 +1953,17 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16, "add_f16_f32_f16", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_add_f32_norepeat, "add_f32_norepeat", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16, "add_f16_f32_f16", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16_norepeat, "add_f16_f32_f16_norepeat", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_div_f32, "div_f32", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_mul_f32_norepeat, "mul_f32_norepeat", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_div_f32, "div_f32", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {0}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_div_f32_norepeat, "div_f32_norepeat", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {1}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_concat_f32, "concat_f32", concat_f32_len, concat_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_concat_f16, "concat_f16", concat_f16_len, concat_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
@@ -1807,26 +1992,39 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_wg512, "soft_max_f32_wg512", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16_wg512, "soft_max_f32_f16_wg512", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, { 512 }, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-
     ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+
+    if (device->float_controls_rte_fp16) {
+        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_rte_len, rope_norm_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_rte_len, rope_neox_f16_rte_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+    } else {
+        ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+    }
 
     ggml_vk_create_pipeline(device, device->pipeline_argsort_f32, "argsort_f32", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1024, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32_len, im2col_f32_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
-    ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+    if (device->float_controls_rte_fp16) {
+        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_rte_len, im2col_f32_f16_rte_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+    } else {
+        ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+    }
 
     ggml_vk_create_pipeline(device, device->pipeline_timestep_embedding_f32, "timestep_embedding_f32", timestep_embedding_f32_len, timestep_embedding_f32_data, "main", 2, sizeof(vk_op_timestep_embedding_push_constants), {256, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_pool2d_f32, "pool2d_f32", pool2d_f32_len, pool2d_f32_data, "main", 2, sizeof(vk_op_pool2d_push_constants), {512, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv6_f32, "rwkv_wkv6_f32", rwkv_wkv6_f32_len, rwkv_wkv6_f32_data, "main", 7, sizeof(vk_op_rwkv_wkv6_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
+
     for (auto &c : compiles) {
         c.wait();
     }
@@ -1860,12 +2058,40 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->physical_device = physical_devices[dev_num];
         const std::vector<vk::ExtensionProperties> ext_props = device->physical_device.enumerateDeviceExtensionProperties();
 
+        bool fp16_storage = false;
+        bool fp16_compute = false;
         bool maintenance4_support = false;
+        bool sm_builtins = false;
+        bool amd_shader_core_properties2 = false;
+        bool pipeline_robustness = false;
+        bool coopmat2_support = false;
+        device->coopmat_support = false;
 
         // Check if maintenance4 is supported
         for (const auto& properties : ext_props) {
             if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
                 maintenance4_support = true;
+            } else if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) {
+                fp16_storage = true;
+            } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) {
+                fp16_compute = true;
+            } else if (strcmp("VK_NV_shader_sm_builtins", properties.extensionName) == 0) {
+                sm_builtins = true;
+            } else if (strcmp("VK_AMD_shader_core_properties2", properties.extensionName) == 0) {
+                amd_shader_core_properties2 = true;
+            } else if (strcmp("VK_EXT_pipeline_robustness", properties.extensionName) == 0) {
+                pipeline_robustness = true;
+            } else if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) {
+                device->subgroup_size_control = true;
+            } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 &&
+                       !getenv("GGML_VK_DISABLE_COOPMAT")) {
+                device->coopmat_support = true;
+                device->coopmat_m = 0;
+                device->coopmat_n = 0;
+                device->coopmat_k = 0;
+            } else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 &&
+                       !getenv("GGML_VK_DISABLE_COOPMAT2")) {
+                coopmat2_support = true;
             }
         }
 
@@ -1873,18 +2099,51 @@ static vk_device ggml_vk_get_device(size_t idx) {
         vk::PhysicalDeviceMaintenance3Properties props3;
         vk::PhysicalDeviceMaintenance4Properties props4;
         vk::PhysicalDeviceSubgroupProperties subgroup_props;
+        vk::PhysicalDeviceDriverProperties driver_props;
+        vk::PhysicalDeviceShaderSMBuiltinsPropertiesNV sm_props;
+        vk::PhysicalDeviceShaderCoreProperties2AMD amd_shader_core_properties2_props;
+        vk::PhysicalDeviceVulkan12Properties vk12_props;
+        vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;
+
         props2.pNext = &props3;
         props3.pNext = &subgroup_props;
+        subgroup_props.pNext = &driver_props;
+        driver_props.pNext = &vk12_props;
+
+        VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&vk12_props;
+
         if (maintenance4_support) {
-            subgroup_props.pNext = &props4;
+            last_struct->pNext = (VkBaseOutStructure *)&props4;
+            last_struct = (VkBaseOutStructure *)&props4;
+        }
+        if (sm_builtins) {
+            last_struct->pNext = (VkBaseOutStructure *)&sm_props;
+            last_struct = (VkBaseOutStructure *)&sm_props;
         }
+        if (amd_shader_core_properties2) {
+            last_struct->pNext = (VkBaseOutStructure *)&amd_shader_core_properties2_props;
+            last_struct = (VkBaseOutStructure *)&amd_shader_core_properties2_props;
+        }
+        if (device->subgroup_size_control) {
+            last_struct->pNext = (VkBaseOutStructure *)&subgroup_size_control_props;
+            last_struct = (VkBaseOutStructure *)&subgroup_size_control_props;
+        }
+
+#if defined(VK_NV_cooperative_matrix2)
+        vk::PhysicalDeviceCooperativeMatrix2PropertiesNV coopmat2_props;
+        if (coopmat2_support) {
+            last_struct->pNext = (VkBaseOutStructure *)&coopmat2_props;
+            last_struct = (VkBaseOutStructure *)&coopmat2_props;
+        }
+#endif
+
         device->physical_device.getProperties2(&props2);
         device->properties = props2.properties;
 
         const char* GGML_VK_FORCE_MAX_ALLOCATION_SIZE = getenv("GGML_VK_FORCE_MAX_ALLOCATION_SIZE");
 
         if (GGML_VK_FORCE_MAX_ALLOCATION_SIZE != nullptr) {
-            device->max_memory_allocation_size = std::stoi(GGML_VK_FORCE_MAX_ALLOCATION_SIZE);
+            device->max_memory_allocation_size = std::stoul(GGML_VK_FORCE_MAX_ALLOCATION_SIZE);
         } else if (maintenance4_support) {
             device->max_memory_allocation_size = std::min(props3.maxMemoryAllocationSize, props4.maxBufferSize);
         } else {
@@ -1894,23 +2153,25 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->vendor_id = device->properties.vendorID;
         device->subgroup_size = subgroup_props.subgroupSize;
         device->uma = device->properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu;
-
-        bool fp16_storage = false;
-        bool fp16_compute = false;
-
-        for (const auto& properties : ext_props) {
-            if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) {
-                fp16_storage = true;
-            } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) {
-                fp16_compute = true;
-            }
+        if (sm_builtins) {
+            device->shader_core_count = sm_props.shaderSMCount;
+        } else if (amd_shader_core_properties2) {
+            device->shader_core_count = amd_shader_core_properties2_props.activeComputeUnitCount;
+        } else {
+            device->shader_core_count = 0;
         }
+        device->float_controls_rte_fp16 = vk12_props.shaderRoundingModeRTEFloat16;
 
-        const char* GGML_VK_DISABLE_F16 = getenv("GGML_VK_DISABLE_F16");
-        const bool force_disable_f16 = GGML_VK_DISABLE_F16 != nullptr;
+        const bool force_disable_f16 = getenv("GGML_VK_DISABLE_F16") != nullptr;
 
         device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute;
 
+        if (device->vendor_id == VK_VENDOR_ID_INTEL || (device->vendor_id == VK_VENDOR_ID_AMD && (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource))) {
+            // Intel drivers don't support coopmat properly yet
+            // Only RADV supports coopmat properly on AMD
+            device->coopmat_support = false;
+        }
+
         std::vector<vk::QueueFamilyProperties> queue_family_props = device->physical_device.getQueueFamilyProperties();
 
         // Try to find a non-graphics compute queue and transfer-focused queues
@@ -1948,10 +2209,149 @@ static vk_device ggml_vk_get_device(size_t idx) {
         vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES;
         vk11_features.pNext = &vk12_features;
 
+        last_struct = (VkBaseOutStructure *)&vk12_features;
+
+        VkPhysicalDevicePipelineRobustnessFeaturesEXT pl_robustness_features;
+        pl_robustness_features.pNext = nullptr;
+        pl_robustness_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PIPELINE_ROBUSTNESS_FEATURES_EXT;
+        pl_robustness_features.pipelineRobustness = VK_FALSE;
+
+        if (pipeline_robustness) {
+            last_struct->pNext = (VkBaseOutStructure *)&pl_robustness_features;
+            last_struct = (VkBaseOutStructure *)&pl_robustness_features;
+            device_extensions.push_back("VK_EXT_pipeline_robustness");
+        }
+
+        VkPhysicalDeviceSubgroupSizeControlFeaturesEXT subgroup_size_control_features;
+        subgroup_size_control_features.pNext = nullptr;
+        subgroup_size_control_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SUBGROUP_SIZE_CONTROL_FEATURES_EXT;
+        subgroup_size_control_features.computeFullSubgroups = false;
+        subgroup_size_control_features.subgroupSizeControl = false;
+
+        if (device->subgroup_size_control) {
+            last_struct->pNext = (VkBaseOutStructure *)&subgroup_size_control_features;
+            last_struct = (VkBaseOutStructure *)&subgroup_size_control_features;
+        }
+
+        VkPhysicalDeviceCooperativeMatrixFeaturesKHR coopmat_features;
+        coopmat_features.pNext = nullptr;
+        coopmat_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR;
+        coopmat_features.cooperativeMatrix = VK_FALSE;
+
+        if (device->coopmat_support) {
+            last_struct->pNext = (VkBaseOutStructure *)&coopmat_features;
+            last_struct = (VkBaseOutStructure *)&coopmat_features;
+        }
+
+#if defined(VK_NV_cooperative_matrix2)
+        VkPhysicalDeviceCooperativeMatrix2FeaturesNV coopmat2_features {};
+        coopmat2_features.pNext = nullptr;
+        coopmat2_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_2_FEATURES_NV;
+        if (coopmat2_support) {
+            last_struct->pNext = (VkBaseOutStructure *)&coopmat2_features;
+            last_struct = (VkBaseOutStructure *)&coopmat2_features;
+            device_extensions.push_back("VK_NV_cooperative_matrix2");
+        }
+#endif
+
         vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2);
 
         device->fp16 = device->fp16 && vk12_features.shaderFloat16;
 
+        device->pipeline_robustness = pl_robustness_features.pipelineRobustness;
+
+        if (device->subgroup_size_control) {
+            device->subgroup_min_size = subgroup_size_control_props.minSubgroupSize;
+            device->subgroup_max_size = subgroup_size_control_props.maxSubgroupSize;
+        }
+
+        device->subgroup_size_control = device->subgroup_size_control &&
+                (subgroup_size_control_props.requiredSubgroupSizeStages & vk::ShaderStageFlagBits::eCompute) &&
+                subgroup_size_control_features.subgroupSizeControl;
+
+        if (device->subgroup_size_control) {
+            device->subgroup_require_full_support = subgroup_size_control_features.computeFullSubgroups;
+            device_extensions.push_back("VK_EXT_subgroup_size_control");
+        }
+
+        device->coopmat_support = device->coopmat_support && coopmat_features.cooperativeMatrix;
+
+        if (coopmat2_support) {
+#if defined(VK_NV_cooperative_matrix2) && defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+            if (coopmat2_features.cooperativeMatrixWorkgroupScope &&
+                coopmat2_features.cooperativeMatrixFlexibleDimensions &&
+                coopmat2_features.cooperativeMatrixReductions &&
+                coopmat2_features.cooperativeMatrixConversions &&
+                coopmat2_features.cooperativeMatrixPerElementOperations &&
+                coopmat2_features.cooperativeMatrixTensorAddressing &&
+                coopmat2_features.cooperativeMatrixBlockLoads &&
+                vk12_features.bufferDeviceAddress) {
+
+                std::vector<VkCooperativeMatrixFlexibleDimensionsPropertiesNV> flexible_dimensions;
+                uint32_t count = 0;
+
+                PFN_vkGetPhysicalDeviceCooperativeMatrixFlexibleDimensionsPropertiesNV
+                    _vkGetPhysicalDeviceCooperativeMatrixFlexibleDimensionsPropertiesNV =
+                        (PFN_vkGetPhysicalDeviceCooperativeMatrixFlexibleDimensionsPropertiesNV)
+                        vk_instance.instance.getProcAddr("vkGetPhysicalDeviceCooperativeMatrixFlexibleDimensionsPropertiesNV");
+
+                _vkGetPhysicalDeviceCooperativeMatrixFlexibleDimensionsPropertiesNV(device->physical_device, &count, nullptr);
+
+                VkCooperativeMatrixFlexibleDimensionsPropertiesNV empty_prop {};
+                empty_prop.sType = VK_STRUCTURE_TYPE_COOPERATIVE_MATRIX_FLEXIBLE_DIMENSIONS_PROPERTIES_NV;
+                flexible_dimensions.resize(count, empty_prop);
+
+                _vkGetPhysicalDeviceCooperativeMatrixFlexibleDimensionsPropertiesNV(device->physical_device, &count, flexible_dimensions.data());
+
+                bool found_fp16_128 = false,
+                     found_fp16_256 = false,
+                     found_fp32_128 = false,
+                     found_fp32_256 = false;
+                // need to support fp16*fp16 with fp16/fp32 accumulator, for workgroupsize 128
+                // with 32x16x16 and 256 with 32x32x16.
+                for (auto &prop : flexible_dimensions) {
+                    if (prop.saturatingAccumulation == VK_FALSE &&
+                        prop.scope == VK_SCOPE_WORKGROUP_KHR &&
+                        prop.AType == VK_COMPONENT_TYPE_FLOAT16_KHR &&
+                        prop.BType == VK_COMPONENT_TYPE_FLOAT16_KHR) {
+
+                        if (prop.workgroupInvocations == 128 &&
+                            prop.MGranularity <= 32 &&
+                            prop.NGranularity <= 16 &&
+                            prop.KGranularity <= 16) {
+                            if (prop.CType == VK_COMPONENT_TYPE_FLOAT16_KHR &&
+                                prop.ResultType == VK_COMPONENT_TYPE_FLOAT16_KHR) {
+                                found_fp16_128 = true;
+                            }
+                            if (prop.CType == VK_COMPONENT_TYPE_FLOAT32_KHR &&
+                                prop.ResultType == VK_COMPONENT_TYPE_FLOAT32_KHR) {
+                                found_fp32_128 = true;
+                            }
+                        }
+                        if (prop.workgroupInvocations == 256 &&
+                            prop.MGranularity <= 32 &&
+                            prop.NGranularity <= 32 &&
+                            prop.KGranularity <= 16) {
+                            if (prop.CType == VK_COMPONENT_TYPE_FLOAT16_KHR &&
+                                prop.ResultType == VK_COMPONENT_TYPE_FLOAT16_KHR) {
+                                found_fp16_256 = true;
+                            }
+                            if (prop.CType == VK_COMPONENT_TYPE_FLOAT32_KHR &&
+                                prop.ResultType == VK_COMPONENT_TYPE_FLOAT32_KHR) {
+                                found_fp32_256 = true;
+                            }
+                        }
+                    }
+                }
+                if (found_fp16_128 && found_fp16_256 &&
+                    found_fp32_128 && found_fp32_256 &&
+                    coopmat2_props.cooperativeMatrixFlexibleDimensionsMaxDimension >= 512) {
+                    device->coopmat2 = true;
+                }
+            }
+#endif
+        }
+
         if (!vk11_features.storageBuffer16BitAccess) {
             std::cerr << "ggml_vulkan: device " << GGML_VK_NAME << idx << " does not support 16-bit storage." << std::endl;
             throw std::runtime_error("Unsupported device");
@@ -1966,6 +2366,74 @@ static vk_device ggml_vk_get_device(size_t idx) {
         if (device->fp16) {
             device_extensions.push_back("VK_KHR_shader_float16_int8");
         }
+
+        if (device->coopmat_support) {
+            // Query supported shapes
+            std::vector<VkCooperativeMatrixPropertiesKHR> cm_props;
+
+            PFN_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR pfn_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR =
+                (PFN_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR)vkGetInstanceProcAddr(vk_instance.instance, "vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR");
+
+            uint32_t cm_props_num;
+
+            pfn_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR(device->physical_device, &cm_props_num, nullptr);
+
+            cm_props.resize(cm_props_num);
+
+            for (auto& prop : cm_props) {
+                prop.sType = VK_STRUCTURE_TYPE_COOPERATIVE_MATRIX_PROPERTIES_KHR;
+            }
+
+            pfn_vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR(device->physical_device, &cm_props_num, cm_props.data());
+
+            VK_LOG_DEBUG("ggml_vulkan: Cooperative Matrix Shapes: " << cm_props.size());
+
+            for (auto& prop : cm_props) {
+                VK_LOG_DEBUG("ggml_vulkan: M: " << prop.MSize << " N: " << prop.NSize << " K: " << prop.KSize << " A: " << vk::to_string((vk::ComponentTypeKHR)prop.AType) << " B: " << vk::to_string((vk::ComponentTypeKHR)prop.BType) << " C: " << vk::to_string((vk::ComponentTypeKHR)prop.CType) << " Result: " << vk::to_string((vk::ComponentTypeKHR)prop.ResultType) << " saturatingAccumulation: " << prop.saturatingAccumulation << " scope: " << vk::to_string((vk::ScopeKHR)prop.scope));
+
+                if ((vk::ComponentTypeKHR)prop.AType == vk::ComponentTypeKHR::eFloat16 &&
+                    (vk::ComponentTypeKHR)prop.BType == vk::ComponentTypeKHR::eFloat16 &&
+                    (vk::ScopeKHR)prop.scope == vk::ScopeKHR::eSubgroup
+                ) {
+                    if ((vk::ComponentTypeKHR)prop.CType == vk::ComponentTypeKHR::eFloat32 &&
+                        (vk::ComponentTypeKHR)prop.ResultType == vk::ComponentTypeKHR::eFloat32) {
+                        // coopmat sizes not set yet
+                        if (device->coopmat_m == 0) {
+                            device->coopmat_acc_f32_support = true;
+                            device->coopmat_m = prop.MSize;
+                            device->coopmat_n = prop.NSize;
+                            device->coopmat_k = prop.KSize;
+                        } else if (device->coopmat_m == prop.MSize && device->coopmat_n == prop.NSize && device->coopmat_k == prop.KSize) {
+                            // Only enable if shape is identical
+                            device->coopmat_acc_f32_support = true;
+                        }
+                    } else if ((vk::ComponentTypeKHR)prop.CType == vk::ComponentTypeKHR::eFloat16 &&
+                               (vk::ComponentTypeKHR)prop.ResultType == vk::ComponentTypeKHR::eFloat16) {
+                        // coopmat sizes not set yet
+                        if (device->coopmat_m == 0) {
+                            device->coopmat_acc_f16_support = true;
+                            device->coopmat_m = prop.MSize;
+                            device->coopmat_n = prop.NSize;
+                            device->coopmat_k = prop.KSize;
+                        } else if (device->coopmat_m == prop.MSize && device->coopmat_n == prop.NSize && device->coopmat_k == prop.KSize) {
+                            // Only enable if shape is identical
+                            device->coopmat_acc_f16_support = true;
+                        }
+                    }
+                }
+            }
+
+            if (device->coopmat_m == 0 || !device->coopmat_acc_f32_support) {
+                // No suitable matmul mode found
+                GGML_LOG_DEBUG("ggml_vulkan: WARNING: No suitable matrix core mode found. Disabling matrix cores.\n");
+                device->coopmat_support = false;
+            }
+        }
+
+        if (device->coopmat_support) {
+            device_extensions.push_back("VK_KHR_cooperative_matrix");
+        }
+
         device->name = GGML_VK_NAME + std::to_string(idx);
 
         device_create_info = {
@@ -1981,6 +2449,37 @@ static vk_device ggml_vk_get_device(size_t idx) {
         ggml_vk_create_queue(device, device->compute_queue, compute_queue_family_index, 0, { vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer }, false);
 
         // Shaders
+        // Disable matmul tile sizes early if performance low or not supported
+        switch (device->vendor_id) {
+#ifndef GGML_VULKAN_RUN_TESTS
+        case VK_VENDOR_ID_AMD:
+        case VK_VENDOR_ID_INTEL:
+            device->mul_mat_l = false;
+            device->mul_mat_m = true;
+            device->mul_mat_s = true;
+            device->mul_mat_id_l = false;
+            device->mul_mat_id_m = true;
+            device->mul_mat_id_s = true;
+            break;
+        case VK_VENDOR_ID_APPLE:
+            device->mul_mat_l = false;
+            device->mul_mat_m = true;
+            device->mul_mat_s = false;
+            device->mul_mat_id_l = false;
+            device->mul_mat_id_m = true;
+            device->mul_mat_id_s = false;
+            break;
+#endif
+        default:
+            device->mul_mat_l = true;
+            device->mul_mat_m = true;
+            device->mul_mat_s = true;
+            device->mul_mat_id_l = true;
+            device->mul_mat_id_m = true;
+            device->mul_mat_id_s = true;
+            break;
+        }
+
         ggml_vk_load_shaders(device);
 
         if (!device->single_queue) {
@@ -2038,15 +2537,31 @@ static void ggml_vk_print_gpu_info(size_t idx) {
 
     bool fp16_storage = false;
     bool fp16_compute = false;
+    bool coopmat_support = false;
+    bool coopmat2_support = false;
 
     for (auto properties : ext_props) {
         if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) {
             fp16_storage = true;
         } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) {
             fp16_compute = true;
+        } else if (strcmp("VK_KHR_cooperative_matrix", properties.extensionName) == 0 &&
+                   !getenv("GGML_VK_DISABLE_COOPMAT")) {
+            coopmat_support = true;
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+        } else if (strcmp("VK_NV_cooperative_matrix2", properties.extensionName) == 0 &&
+                   !getenv("GGML_VK_DISABLE_COOPMAT2")) {
+            coopmat2_support = true;
+#endif
         }
     }
 
+    if (props2.properties.vendorID == VK_VENDOR_ID_INTEL || (props2.properties.vendorID == VK_VENDOR_ID_AMD && (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource))) {
+        // Intel drivers don't support coopmat properly yet
+        // Only RADV supports coopmat properly on AMD
+        coopmat_support = false;
+    }
+
     const char* GGML_VK_DISABLE_F16 = getenv("GGML_VK_DISABLE_F16");
     bool force_disable_f16 = GGML_VK_DISABLE_F16 != nullptr;
 
@@ -2069,16 +2584,33 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES;
     vk11_features.pNext = &vk12_features;
 
+    // Pointer to the last chain element
+    VkBaseOutStructure * last_struct = (VkBaseOutStructure *)&vk12_features;
+
+    VkPhysicalDeviceCooperativeMatrixFeaturesKHR coopmat_features;
+    coopmat_features.pNext = nullptr;
+    coopmat_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_COOPERATIVE_MATRIX_FEATURES_KHR;
+    coopmat_features.cooperativeMatrix = VK_FALSE;
+
+    if (coopmat_support) {
+        last_struct->pNext = (VkBaseOutStructure *)&coopmat_features;
+        last_struct = (VkBaseOutStructure *)&coopmat_features;
+    }
+
     vkGetPhysicalDeviceFeatures2(physical_device, &device_features2);
 
     fp16 = fp16 && vk12_features.shaderFloat16;
 
+    coopmat_support = coopmat_support && coopmat_features.cooperativeMatrix;
+
+    std::string matrix_cores = coopmat2_support ? "NV_coopmat2" : coopmat_support ? "KHR_coopmat" : "none";
+
     std::string device_name = props2.properties.deviceName.data();
-    GGML_LOG_DEBUG("ggml_vulkan: %d = %s (%s) | uma: %d | fp16: %d | warp size: %d\n",
-              idx, device_name.c_str(), driver_props.driverName, uma, fp16, subgroup_size);
+    GGML_LOG_DEBUG("ggml_vulkan: %zu = %s (%s) | uma: %d | fp16: %d | warp size: %zu | matrix cores: %s\n",
+              idx, device_name.c_str(), driver_props.driverName.data(), uma, fp16, subgroup_size, matrix_cores.c_str());
 
     if (props2.properties.deviceType == vk::PhysicalDeviceType::eCpu) {
-        std::cerr << "ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want." << std::endl;
+        GGML_LOG_DEBUG("ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want.\n");
     }
 }
 
@@ -2247,8 +2779,7 @@ void ggml_vk_instance_init() {
             vk_instance.device_indices.push_back(0);
         }
     }
-    GGML_LOG_DEBUG("ggml_vulkan: Found %d Vulkan devices:\n", vk_instance.device_indices.size());
-
+    GGML_LOG_DEBUG("ggml_vulkan: Found %zu Vulkan devices:\n", vk_instance.device_indices.size());
 
     for (size_t i = 0; i < vk_instance.device_indices.size(); i++) {
         ggml_vk_print_gpu_info(i);
@@ -2304,7 +2835,7 @@ static vk_pipeline ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type
     return ctx->device->pipeline_dequant[type];
 }
 
-static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type) {
+static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type, ggml_prec prec) {
     VK_LOG_DEBUG("ggml_vk_get_mul_mat_mat_pipeline(" << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ")");
     if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
         return ctx->device->pipeline_matmul_f32;
@@ -2312,14 +2843,23 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
     if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
         return ctx->device->pipeline_matmul_f32_f16;
     }
-    if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
-        return ctx->device->pipeline_matmul_f16_f32;
-    }
-    if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
-        return ctx->device->pipeline_matmul_f16;
+    if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) {
+        if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_matmul_f16_f32.f16acc;
+        }
+        if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
+            return ctx->device->pipeline_matmul_f16.f16acc;
+        }
+    } else {
+        if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_matmul_f16_f32.f32acc;
+        }
+        if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
+            return ctx->device->pipeline_matmul_f16.f32acc;
+        }
     }
 
-    if (src1_type != GGML_TYPE_F32) {
+    if (src1_type != GGML_TYPE_F32 && !ctx->device->coopmat2) {
         return nullptr;
     }
 
@@ -2340,7 +2880,11 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_conte
             return nullptr;
     }
 
-    return ctx->device->pipeline_dequant_mul_mat_mat[src0_type];
+    if (ctx->device->coopmat2) {
+        assert(src1_type == GGML_TYPE_F16);
+        return ctx->device->pipeline_dequant_mul_mat_mat_f16[src0_type].f16acc;
+    }
+    return ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat[src0_type].f32acc;
 }
 
 static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
@@ -2369,16 +2913,25 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
     return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type];
 }
 
-static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type) {
+static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type, ggml_prec prec) {
     VK_LOG_DEBUG("ggml_vk_get_mul_mat_mat_id_pipeline()");
     if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
         return ctx->device->pipeline_matmul_id_f32;
     }
-    if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
-        return ctx->device->pipeline_matmul_id_f16_f32;
-    }
-    if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
-        return ctx->device->pipeline_matmul_id_f16;
+    if (prec == GGML_PREC_DEFAULT && ctx->device->fp16 && !(ctx->device->coopmat_support && !ctx->device->coopmat_acc_f16_support)) {
+        if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_matmul_id_f16_f32.f16acc;
+        }
+        if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
+            return ctx->device->pipeline_matmul_id_f16.f16acc;
+        }
+    } else {
+        if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_matmul_id_f16_f32.f32acc;
+        }
+        if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
+            return ctx->device->pipeline_matmul_id_f16.f32acc;
+        }
     }
 
     GGML_ASSERT(src1_type == GGML_TYPE_F32);
@@ -2400,7 +2953,7 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
             return nullptr;
     }
 
-    return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type];
+    return ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc;
 }
 
 static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
@@ -2969,55 +3522,44 @@ static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, siz
     dst->device->device.resetFences({ dst->device->fence });
 }
 
-static uint32_t ggml_vk_guess_split_k(int m, int n, int k) {
+static uint32_t ggml_vk_guess_split_k(ggml_backend_vk_context * ctx, int m, int n, int k, const vk_pipeline& pipeline) {
     VK_LOG_DEBUG("ggml_vk_guess_split_k(" << m << ", " << n << ", " << k << ")");
-    // if (k > 128 && (m < 128 || n < 128) && m > 2 && n > 2) {
-    //     return 4;
-    // }
-
-    return 1;
 
-    GGML_UNUSED(m); GGML_UNUSED(n); GGML_UNUSED(k);
-}
-
-static vk_pipeline ggml_vk_guess_matmul_pipeline_amd(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) {
-    if (m <= 32 || n <= 32) {
-        return aligned ? mmp->a_s : mmp->s;
+    uint32_t split_k = 1;
+    if (ctx->device->shader_core_count != 0 && m >= (int)pipeline->wg_denoms[0] && n >= (int)pipeline->wg_denoms[1]) {
+        // If k is 'large' and the SMs will fill less than halfway, use split_k.
+        uint32_t m_tiles = CEIL_DIV(m, pipeline->wg_denoms[0]);
+        uint32_t n_tiles = CEIL_DIV(n, pipeline->wg_denoms[1]);
+        if (k >= 2048 && m_tiles * n_tiles < ctx->device->shader_core_count / 2) {
+            split_k = ctx->device->shader_core_count / (m_tiles * n_tiles);
+            // Clamp to 2 or 4
+            split_k = std::min(split_k, 4u);
+            if (split_k == 3) {
+                split_k = 2;
+            }
+        }
     }
-    return aligned ? mmp->a_m : mmp->m;
 
-    GGML_UNUSED(ctx);
-}
-
-static vk_pipeline ggml_vk_guess_matmul_pipeline_apple(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, bool aligned) {
-    return aligned ? mmp->a_m : mmp->m;
-
-    GGML_UNUSED(ctx);
-}
-
-static vk_pipeline ggml_vk_guess_matmul_pipeline_intel(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, bool aligned) {
-    return aligned ? mmp->a_s : mmp->s;
-
-    GGML_UNUSED(ctx);
+    return split_k;
 }
 
 static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) {
     VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ")");
-    switch (ctx->device->vendor_id) {
-    case VK_VENDOR_ID_AMD:
-        return ggml_vk_guess_matmul_pipeline_amd(ctx, mmp, m, n, aligned);
-    case VK_VENDOR_ID_APPLE:
-        return ggml_vk_guess_matmul_pipeline_apple(ctx, mmp, aligned);
-    case VK_VENDOR_ID_INTEL:
-        return ggml_vk_guess_matmul_pipeline_intel(ctx, mmp, aligned);
-    default:
-        break;
+
+    if (ctx->device->coopmat2) {
+        if ((ctx->device->mul_mat_l && (m % mmp->l->wg_denoms[0]) == 0 && (n % mmp->l->wg_denoms[1]) == 0) || (!ctx->device->mul_mat_m && !ctx->device->mul_mat_s)) {
+            return aligned ? mmp->a_l : mmp->l;
+        }
+        if ((ctx->device->mul_mat_m && (m % mmp->m->wg_denoms[0]) == 0 && (n % mmp->m->wg_denoms[1]) == 0) || !ctx->device->mul_mat_s) {
+            return aligned ? mmp->a_m : mmp->m;
+        }
+        return aligned ? mmp->a_s : mmp->s;
     }
 
-    if (m <= 32 || n <= 32) {
+    if ((ctx->device->mul_mat_s && (m <= 32 || n <= 32)) || (!ctx->device->mul_mat_m && !ctx->device->mul_mat_l)) {
         return aligned ? mmp->a_s : mmp->s;
     }
-    if (m <= 64 || n <= 64) {
+    if ((ctx->device->mul_mat_m && (m <= 64 || n <= 64)) || !ctx->device->mul_mat_l) {
         return aligned ? mmp->a_m : mmp->m;
     }
     return aligned ? mmp->a_l : mmp->l;
@@ -3052,6 +3594,33 @@ static void ggml_vk_matmul(
     ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2.size() * sizeof(uint32_t), pc2.data(), { m * n * batch, 1, 1 });
 }
 
+static vk_pipeline ggml_vk_guess_matmul_id_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) {
+    VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ")");
+
+    if (ctx->device->coopmat2) {
+        if ((ctx->device->mul_mat_id_l && (m % mmp->l->wg_denoms[0]) == 0 && (n % mmp->l->wg_denoms[1]) == 0) || (!ctx->device->mul_mat_id_m && !ctx->device->mul_mat_id_s)) {
+            return aligned ? mmp->a_l : mmp->l;
+        }
+        if ((ctx->device->mul_mat_id_m && (m % mmp->m->wg_denoms[0]) == 0 && (n % mmp->m->wg_denoms[1]) == 0) || !ctx->device->mul_mat_id_s) {
+            return aligned ? mmp->a_m : mmp->m;
+        }
+        return aligned ? mmp->a_s : mmp->s;
+    }
+
+    if ((ctx->device->mul_mat_id_s && (m <= 32 || n <= 32)) || (!ctx->device->mul_mat_id_m && !ctx->device->mul_mat_id_l)) {
+        return aligned ? mmp->a_s : mmp->s;
+    }
+    if ((ctx->device->mul_mat_id_m && (m <= 64 || n <= 64)) || !ctx->device->mul_mat_id_l) {
+        return aligned ? mmp->a_m : mmp->m;
+    }
+    return aligned ? mmp->a_l : mmp->l;
+}
+
+static uint32_t ggml_vk_guess_matmul_id_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n) {
+    VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline_align(" << m << ", " << n << ")");
+    return ggml_vk_guess_matmul_id_pipeline(ctx, mmp, m, n, true)->align;
+}
+
 static void ggml_vk_matmul_id(
         ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline& pipeline,
         vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& ids,
@@ -3122,13 +3691,15 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
         elements = { ne, 1, 1 };
     }
 
-    const vk_op_unary_push_constants pc = {
+    vk_op_unary_push_constants pc = {
         (uint32_t)ne,
         (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3], (uint32_t)tensor->nb[0] / tensor_type_size, (uint32_t)tensor->nb[1] / tensor_type_size, (uint32_t)tensor->nb[2] / tensor_type_size, (uint32_t)tensor->nb[3] / tensor_type_size,
         (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3],                       1                   , (uint32_t)tensor->ne[0]                   , (uint32_t)(tensor->ne[0] * tensor->ne[1]) , (uint32_t)(tensor->ne[0] * tensor->ne[1] * tensor->ne[2]),
         0,
         0.0f, 0.0f,
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     };
+    init_pushconst_fastdiv(pc);
     ggml_vk_sync_buffers(subctx);
     ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, elements);
 }
@@ -3177,18 +3748,20 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     }
 
     const bool x_non_contig = !ggml_vk_dim01_contiguous(src0);
-    const bool y_non_contig = !ggml_vk_dim01_contiguous(src1);
+    // Reformat and convert to fp16 if src1 is non-contiguous, or for coopmat2 for better perf
+    const bool y_non_contig = (ctx->device->coopmat2 && src1->type == GGML_TYPE_F32) ||
+                              !ggml_vk_dim01_contiguous(src1);
 
     const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
 
-    vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type);
+    vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]);
 
     const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
     const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
 
     if (qx_needs_dequant) {
         // Fall back to dequant + f16 mulmat
-        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16);
+        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16, (ggml_prec)dst->op_params[0]);
     }
 
     // Not implemented
@@ -3201,10 +3774,10 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11));
     const bool aligned = ne10 == kpad && ne01 > 8 && ne11 > 8;
 
-    const uint32_t split_k = ggml_vk_guess_split_k(ne01, ne11, ne10);
-
     vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned);
 
+    const uint32_t split_k = ggml_vk_guess_split_k(ctx, ne01, ne11, ne10, pipeline);
+
     const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
     const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
     const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne;
@@ -3230,7 +3803,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
     if (dryrun) {
         const uint64_t x_sz_upd = x_sz * ne02 * ne03;
         const uint64_t y_sz_upd = y_sz * ne12 * ne13;
-        const uint64_t split_k_size = split_k > 1 ? d_sz * ne12 * ne13 * 4 : 0;
+        const uint64_t split_k_size = split_k > 1 ? d_sz * ne12 * ne13 * split_k : 0;
         if (
                 (qx_needs_dequant && x_sz_upd > ctx->device->max_memory_allocation_size) ||
                 (qy_needs_dequant && y_sz_upd > ctx->device->max_memory_allocation_size) ||
@@ -3497,7 +4070,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
 
     if (ne01 > max_groups_x) {
         groups_z = 64;
-        groups_x /= groups_z;
+        groups_x = CEIL_DIV(groups_x, groups_z);
     }
 
     // compute
@@ -3752,7 +4325,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
 
     const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
 
-    vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type);
+    vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type, (ggml_prec)dst->op_params[0]);
 
     const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
     const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
@@ -3768,10 +4341,10 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
     const uint64_t y_ne = ne11 * ne10;
     const uint64_t d_ne = ne21 * ne20;
 
-    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, nei1));
+    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_id_pipeline_align(ctx, mmp, ne01, nei1));
     const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8;
 
-    vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, nei1, aligned);
+    vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned);
 
     const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
     const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
@@ -4074,7 +4647,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
 
     if (ne01 > max_groups_x) {
         groups_z = 64;
-        groups_x /= groups_z;
+        groups_x = CEIL_DIV(groups_x, groups_z);
     }
 
     // compute
@@ -4099,6 +4672,167 @@ static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx
     }
 }
 
+static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * q, const ggml_tensor * k, const ggml_tensor * v, const ggml_tensor * mask, ggml_tensor * dst, bool dryrun = false) {
+    VK_LOG_DEBUG("ggml_vk_flash_attn((" << q << ", name=" << q->name << ", type=" << q->type << ", ne0=" << q->ne[0] << ", ne1=" << q->ne[1] << ", ne2=" << q->ne[2] << ", ne3=" << q->ne[3] << ", nb0=" << q->nb[0] << ", nb1=" << q->nb[1] << ", nb2=" << q->nb[2] << ", nb3=" << q->nb[3];
+    std::cerr << "), (" << k << ", name=" << k->name << ", type=" << k->type << ", ne0=" << k->ne[0] << ", ne1=" << k->ne[1] << ", ne2=" << k->ne[2] << ", ne3=" << k->ne[3] << ", nb0=" << k->nb[0] << ", nb1=" << k->nb[1] << ", nb2=" << k->nb[2] << ", nb3=" << k->nb[3];
+    std::cerr << "), (" << v << ", name=" << v->name << ", type=" << v->type << ", ne0=" << v->ne[0] << ", ne1=" << v->ne[1] << ", ne2=" << v->ne[2] << ", ne3=" << v->ne[3] << ", nb0=" << v->nb[0] << ", nb1=" << v->nb[1] << ", nb2=" << v->nb[2] << ", nb3=" << v->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
+    std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
+
+    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
+
+    const uint32_t nem1 = mask ? mask->ne[1] : 0;
+    const uint32_t nbm1 = mask ? mask->nb[1] : 0;
+
+    const uint32_t D = neq0;
+    const uint32_t N = neq1;
+    const uint32_t KV = nek1;
+
+    GGML_ASSERT(ne0 == D);
+    GGML_ASSERT(ne2 == N);
+
+    // input tensor rows must be contiguous
+    GGML_ASSERT(nbq0 == ggml_type_size(q->type));
+    GGML_ASSERT(nbk0 == ggml_type_size(k->type));
+    GGML_ASSERT(nbv0 == ggml_type_size(v->type));
+
+    GGML_ASSERT(neq0 == D);
+    GGML_ASSERT(nek0 == D);
+    GGML_ASSERT(nev0 == D);
+
+    GGML_ASSERT(neq1 == N);
+    GGML_ASSERT(nev0 == D);
+
+    GGML_ASSERT(nev1 == nek1);
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    assert(dst->type == GGML_TYPE_F32);
+    assert(q->type == GGML_TYPE_F32);
+    assert(k->type == v->type);
+
+    vk_pipeline *pipelines;
+    // XXX TODO other backends may be changing accumulator precision to default to f32 soon
+    bool f32acc = dst->op_params[3] == GGML_PREC_F32;
+    bool small_rows = N <= flash_attention_num_small_rows;
+    switch (D) {
+    case 64: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D64[k->type][f32acc][small_rows][0]; break;
+    case 80: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D80[k->type][f32acc][small_rows][0]; break;
+    case 96: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D96[k->type][f32acc][small_rows][0]; break;
+    case 112: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D112[k->type][f32acc][small_rows][0]; break;
+    case 128: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D128[k->type][f32acc][small_rows][0]; break;
+    case 256: pipelines = &ctx->device->pipeline_flash_attn_f32_f16_D256[k->type][f32acc][small_rows][0]; break;
+    default:
+        assert(!"unsupported D value");
+        return;
+    }
+    assert(pipelines);
+
+    bool aligned = (KV % pipelines[1]->align) == 0;
+    vk_pipeline pipeline = pipelines[aligned];
+    assert(pipeline);
+
+    if (dryrun) {
+        // Request descriptor sets
+        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        return;
+    }
+
+    float scale         = 1.0f;
+    float max_bias      = 0.0f;
+    float logit_softcap = 0.0f;
+
+    memcpy(&scale,         (const float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias,      (const float *) dst->op_params + 1, sizeof(float));
+    memcpy(&logit_softcap, (const float *) dst->op_params + 2, sizeof(float));
+
+    if (logit_softcap != 0) {
+        scale /= logit_softcap;
+    }
+
+    const uint32_t n_head_kv   = neq2;
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    ggml_vk_sync_buffers(subctx);
+
+    vk_buffer d_Q, d_K, d_V, d_D, d_M;
+    uint64_t q_buf_offset, k_buf_offset, v_buf_offset, d_buf_offset, m_buf_offset;
+
+    bool Q_uma = false, K_uma = false, V_uma = false, D_uma = false, M_uma = false;
+
+    if (ctx->device->uma) {
+        ggml_vk_host_get(ctx->device, q->data, d_Q, q_buf_offset);
+        ggml_vk_host_get(ctx->device, k->data, d_K, q_buf_offset);
+        ggml_vk_host_get(ctx->device, v->data, d_V, q_buf_offset);
+        ggml_vk_host_get(ctx->device, dst->data, d_D, q_buf_offset);
+        Q_uma = d_Q != nullptr;
+        K_uma = d_K != nullptr;
+        V_uma = d_V != nullptr;
+        D_uma = d_D != nullptr;
+        if (mask) {
+            ggml_vk_host_get(ctx->device, mask->data, d_M, q_buf_offset);
+            M_uma = d_M != nullptr;
+        }
+    }
+
+
+    ggml_backend_vk_buffer_context * d_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
+    ggml_backend_vk_buffer_context * q_buf_ctx = (ggml_backend_vk_buffer_context *)q->buffer->context;
+    ggml_backend_vk_buffer_context * k_buf_ctx = (ggml_backend_vk_buffer_context *)k->buffer->context;
+    ggml_backend_vk_buffer_context * v_buf_ctx = (ggml_backend_vk_buffer_context *)v->buffer->context;
+
+    if (!Q_uma) {
+        d_Q = q_buf_ctx->dev_buffer;
+        q_buf_offset = vk_tensor_offset(q) + q->view_offs;
+    }
+    if (!K_uma) {
+        d_K = k_buf_ctx->dev_buffer;
+        k_buf_offset = vk_tensor_offset(k) + k->view_offs;
+    }
+    if (!V_uma) {
+        d_V = v_buf_ctx->dev_buffer;
+        v_buf_offset = vk_tensor_offset(v) + v->view_offs;
+    }
+    if (!D_uma) {
+        d_D = d_buf_ctx->dev_buffer;
+        d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
+    }
+
+    if (!M_uma) {
+        d_M = d_Q;
+        m_buf_offset = q_buf_offset;
+        if (mask) {
+            ggml_backend_vk_buffer_context * m_buf_ctx = (ggml_backend_vk_buffer_context*)mask->buffer->context;
+            d_M = m_buf_ctx->dev_buffer;
+            m_buf_offset = vk_tensor_offset(mask) + mask->view_offs;
+        }
+    }
+
+    const vk_flash_attn_push_constants pc = { N, KV, (uint32_t)ne1, (uint32_t)ne2, (uint32_t)ne3, (uint32_t)neq2, (uint32_t)neq3, (uint32_t)nek2, (uint32_t)nek3, (uint32_t)nev2, (uint32_t)nev3, nem1, (uint32_t)nbq2, (uint32_t)nbq3, (uint32_t)nbk2, (uint32_t)nbk3, (uint32_t)nbv2, (uint32_t)nbv3, nbm1, scale, max_bias, logit_softcap, mask != nullptr, n_head_log2, m0, m1 };
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
+                                {
+                                    vk_subbuffer{d_Q, q_buf_offset, VK_WHOLE_SIZE},
+                                    vk_subbuffer{d_K, k_buf_offset, VK_WHOLE_SIZE},
+                                    vk_subbuffer{d_V, v_buf_offset, VK_WHOLE_SIZE},
+                                    vk_subbuffer{d_M, m_buf_offset, VK_WHOLE_SIZE},
+                                    vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE},
+                                },
+                                sizeof(vk_flash_attn_push_constants), &pc, { (uint32_t)neq1, (uint32_t)neq2, (uint32_t)neq3 });
+}
+
 static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op) {
     switch (op) {
     case GGML_OP_GET_ROWS:
@@ -4117,20 +4851,20 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         return nullptr;
     case GGML_OP_ADD:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_add_f32;
+            return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_f32_norepeat : ctx->device->pipeline_add_f32;
         }
         if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
-            return ctx->device->pipeline_add_f16_f32_f16;
+            return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_f16_f32_f16_norepeat : ctx->device->pipeline_add_f16_f32_f16;
         }
         return nullptr;
     case GGML_OP_MUL:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_mul_f32;
+            return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_mul_f32_norepeat : ctx->device->pipeline_mul_f32;
         }
         return nullptr;
     case GGML_OP_DIV:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_div_f32;
+            return ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_div_f32_norepeat : ctx->device->pipeline_div_f32;
         }
         return nullptr;
     case GGML_OP_CONCAT:
@@ -4243,10 +4977,10 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);
 
         if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_soft_max_f32;
+            return src0->ne[0] > 1024 ? ctx->device->pipeline_soft_max_f32_wg512 : ctx->device->pipeline_soft_max_f32;
         }
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
-            return ctx->device->pipeline_soft_max_f32_f16;
+            return src0->ne[0] > 1024 ? ctx->device->pipeline_soft_max_f32_f16_wg512 : ctx->device->pipeline_soft_max_f32_f16;
         }
         return nullptr;
     case GGML_OP_ROPE:
@@ -4299,6 +5033,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_pool2d_f32;
         }
         return nullptr;
+    case GGML_OP_RWKV_WKV6:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_rwkv_wkv6_f32;
+        }
+        return nullptr;
     case GGML_OP_LEAKY_RELU:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_leaky_relu_f32;
@@ -4333,7 +5072,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
 }
 
 template<typename PC>
-static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, const PC&& pc, bool dryrun = false) {
+static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, PC&& pc, bool dryrun = false) {
     VK_LOG_DEBUG("ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     if (src1 != nullptr) {
         std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
@@ -4373,6 +5112,8 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
     const uint64_t ned3 = dst->ne[3];
     const uint64_t ned = ned0 * ned1;
 
+    init_pushconst_fastdiv(pc);
+
     vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, dst, op);
 
     if (pipeline == nullptr) {
@@ -4699,6 +5440,134 @@ static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const
     }, dryrun);
 }
 
+static void ggml_vk_op_f32_rwkv6(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, const vk_op_rwkv_wkv6_push_constants&& pc, bool dryrun = false) {
+    const ggml_tensor * k = dst->src[0];
+    const ggml_tensor * v = dst->src[1];
+    const ggml_tensor * r = dst->src[2];
+    const ggml_tensor * tf = dst->src[3];
+    const ggml_tensor * td = dst->src[4];
+    const ggml_tensor * state = dst->src[5];
+
+    GGML_ASSERT(!ggml_is_quantized(k->type));
+    GGML_ASSERT(!ggml_is_quantized(v->type));
+    GGML_ASSERT(!ggml_is_quantized(r->type));
+    GGML_ASSERT(!ggml_is_quantized(tf->type));
+    GGML_ASSERT(!ggml_is_quantized(td->type));
+    GGML_ASSERT(!ggml_is_quantized(state->type));
+    GGML_ASSERT(dst->buffer != nullptr);
+
+    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, k, v, r, dst, GGML_OP_RWKV_WKV6);
+    GGML_ASSERT(pipeline != nullptr);
+
+    if (dryrun) {
+        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        return;
+    }
+
+    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
+    ggml_backend_vk_buffer_context * k_buf_ctx = (ggml_backend_vk_buffer_context *)k->buffer->context;
+    ggml_backend_vk_buffer_context * v_buf_ctx = (ggml_backend_vk_buffer_context *)v->buffer->context;
+    ggml_backend_vk_buffer_context * r_buf_ctx = (ggml_backend_vk_buffer_context *)r->buffer->context;
+    ggml_backend_vk_buffer_context * tf_buf_ctx = (ggml_backend_vk_buffer_context *)tf->buffer->context;
+    ggml_backend_vk_buffer_context * td_buf_ctx = (ggml_backend_vk_buffer_context *)td->buffer->context;
+    ggml_backend_vk_buffer_context * state_buf_ctx = (ggml_backend_vk_buffer_context *)state->buffer->context;
+
+    ggml_vk_sync_buffers(subctx);
+
+    vk_buffer d_D, d_K, d_V, d_R, d_TF, d_TD, d_State;
+    uint64_t k_offset, v_offset, r_offset, tf_offset, td_offset, state_offset, dst_offset;
+    bool K_uma = false, V_uma = false, R_uma = false, TF_uma = false, TD_uma = false, STATE_uma = false, DST_uma = false;
+
+    if (ctx->device->uma) {
+        ggml_vk_host_get(ctx->device, k->data, d_K, k_offset);
+        ggml_vk_host_get(ctx->device, v->data, d_V, v_offset);
+        ggml_vk_host_get(ctx->device, r->data, d_R, r_offset);
+        ggml_vk_host_get(ctx->device, tf->data, d_TF, tf_offset);
+        ggml_vk_host_get(ctx->device, td->data, d_TD, td_offset);
+        ggml_vk_host_get(ctx->device, state->data, d_State, state_offset);
+        ggml_vk_host_get(ctx->device, dst->data, d_D, dst_offset);
+
+        K_uma = d_K != nullptr;
+        V_uma = d_V != nullptr;
+        R_uma = d_R != nullptr;
+        TF_uma = d_TF != nullptr;
+        TD_uma = d_TD != nullptr;
+        STATE_uma = d_State != nullptr;
+        DST_uma = d_D != nullptr;
+    }
+
+    if (!K_uma) {
+        d_K = k_buf_ctx->dev_buffer;
+        k_offset = vk_tensor_offset(k) + k->view_offs;
+    }
+    if (!V_uma) {
+        d_V = v_buf_ctx->dev_buffer;
+        v_offset = vk_tensor_offset(v) + v->view_offs;
+    }
+    if (!R_uma) {
+        d_R = r_buf_ctx->dev_buffer;
+        r_offset = vk_tensor_offset(r) + r->view_offs;
+    }
+    if (!TF_uma) {
+        d_TF = tf_buf_ctx->dev_buffer;
+        tf_offset = vk_tensor_offset(tf) + tf->view_offs;
+    }
+    if (!TD_uma) {
+        d_TD = td_buf_ctx->dev_buffer;
+        td_offset = vk_tensor_offset(td) + td->view_offs;
+    }
+    if (!STATE_uma) {
+        d_State = state_buf_ctx->dev_buffer;
+        state_offset = vk_tensor_offset(state) + state->view_offs;
+    }
+    if (!DST_uma) {
+        d_D = dst_buf_ctx->dev_buffer;
+        dst_offset = vk_tensor_offset(dst) + dst->view_offs;
+    }
+
+    const uint64_t k_size = ggml_nbytes(k);
+    const uint64_t v_size = ggml_nbytes(v);
+    const uint64_t r_size = ggml_nbytes(r);
+    const uint64_t tf_size = ggml_nbytes(tf);
+    const uint64_t td_size = ggml_nbytes(td);
+    const uint64_t state_size = ggml_nbytes(state);
+    const uint64_t dst_size = ggml_nbytes(dst);
+
+    std::array<uint32_t, 3> elements = {
+        (uint32_t)(pc.B * pc.H),
+        1,
+        1
+    };
+
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, {
+        vk_subbuffer{ d_K, k_offset, k_size },
+        vk_subbuffer{ d_V, v_offset, v_size },
+        vk_subbuffer{ d_R, r_offset, r_size },
+        vk_subbuffer{ d_TF, tf_offset, tf_size },
+        vk_subbuffer{ d_TD, td_offset, td_size },
+        vk_subbuffer{ d_State, state_offset, state_size },
+        vk_subbuffer{ d_D, dst_offset, dst_size }
+    }, sizeof(vk_op_rwkv_wkv6_push_constants), &pc, elements);
+}
+
+static void ggml_vk_rwkv_wkv6(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) {
+    const size_t seq_length = dst->src[0]->ne[3];
+    const size_t n_embed = dst->ne[0];
+    const size_t n_heads = dst->src[0]->ne[2];
+    const size_t n_seqs = dst->src[5]->ne[1];
+
+    ggml_vk_op_f32_rwkv6(
+        ctx, subctx, dst,
+        {
+            (uint32_t)n_seqs,
+            (uint32_t)seq_length,
+            (uint32_t)n_embed,
+            (uint32_t)n_heads,
+        },
+        dryrun
+    );
+}
+
 static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
     int * op_params = (int *)dst->op_params;
 
@@ -4742,7 +5611,8 @@ static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, con
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
-        op_params[0], 0.0f
+        op_params[0], 0.0f,
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 
@@ -4756,6 +5626,7 @@ static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f,
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 
@@ -4769,6 +5640,7 @@ static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f,
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 
@@ -4782,6 +5654,7 @@ static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f,
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 
@@ -4796,6 +5669,7 @@ static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, con
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         op_params[0], op_params[1],
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 
@@ -4809,6 +5683,7 @@ static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f,
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 
@@ -4822,6 +5697,7 @@ static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, co
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         0,
         0.0f, 0.0f,
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 
@@ -4836,6 +5712,7 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
         d_offset,
         0.0f, 0.0f,
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
     }, dryrun);
 }
 
@@ -4892,6 +5769,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx,
         scale, max_bias,
         m0, m1,
         n_head_log2,
+        nrows_x,
     }, dryrun);
 }
 
@@ -5070,10 +5948,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             p = ctx->device->pipeline_matmul_f32_f16->a_s;
             shname = "F32_F16_ALIGNED_S";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16_f32->a_s;
+            p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_s;
             shname = "F16_F32_ALIGNED_S";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16->a_s;
+            p = ctx->device->pipeline_matmul_f16.f32acc->a_s;
             shname = "F16_ALIGNED_S";
         } else {
             GGML_ABORT("fatal error");
@@ -5086,10 +5964,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             p = ctx->device->pipeline_matmul_f32_f16->a_m;
             shname = "F32_F16_ALIGNED_M";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16_f32->a_m;
+            p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_m;
             shname = "F16_F32_ALIGNED_M";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16->a_m;
+            p = ctx->device->pipeline_matmul_f16.f32acc->a_m;
             shname = "F16_ALIGNED_M";
         } else {
             GGML_ABORT("fatal error");
@@ -5102,10 +5980,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
             p = ctx->device->pipeline_matmul_f32_f16->a_l;
             shname = "F32_F16_ALIGNED_L";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16_f32->a_l;
+            p = ctx->device->pipeline_matmul_f16_f32.f32acc->a_l;
             shname = "F16_F32_ALIGNED_L";
         } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-            p = ctx->device->pipeline_matmul_f16->a_l;
+            p = ctx->device->pipeline_matmul_f16.f32acc->a_l;
             shname = "F16_ALIGNED_L";
         } else {
             GGML_ABORT("fatal error");
@@ -5125,10 +6003,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
                 p = ctx->device->pipeline_matmul_f32_f16->s;
                 shname = "F32_F16_S";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16_f32->s;
+                p = ctx->device->pipeline_matmul_f16_f32.f32acc->s;
                 shname = "F16_F32_S";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16->s;
+                p = ctx->device->pipeline_matmul_f16.f32acc->s;
                 shname = "F16_S";
             }
         } else if (shader_size == 1) {
@@ -5139,10 +6017,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
                 p = ctx->device->pipeline_matmul_f32_f16->m;
                 shname = "F32_F16_M";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16_f32->m;
+                p = ctx->device->pipeline_matmul_f16_f32.f32acc->m;
                 shname = "F16_F32_M";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16->m;
+                p = ctx->device->pipeline_matmul_f16.f32acc->m;
                 shname = "F16_M";
             }
         } else if (shader_size == 2) {
@@ -5153,10 +6031,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
                 p = ctx->device->pipeline_matmul_f32_f16->l;
                 shname = "F32_F16_L";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16_f32->l;
+                p = ctx->device->pipeline_matmul_f16_f32.f32acc->l;
                 shname = "F16_F32_L";
             } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
-                p = ctx->device->pipeline_matmul_f16->l;
+                p = ctx->device->pipeline_matmul_f16.f32acc->l;
                 shname = "F16_L";
             }
         }
@@ -5188,19 +6066,27 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
     for (size_t i = 0; i < x_ne; i++) {
         if (std::is_same<float, X_TYPE>()) {
             x[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f;
+            // x[i] = 1.0f;
+            // x[i] = i + 1;
+            // x[i] = (i % k == i / k) ? 1.0f : 0.0f;
         } else if (std::is_same<ggml_fp16_t, X_TYPE>()) {
             x[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f);
+            // x[i] = ggml_fp32_to_fp16(1.0f);
+            // x[i] = ggml_fp32_to_fp16(i + 1);
+            // x[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f);
         } else {
             GGML_ABORT("fatal error");
         }
     }
     for (size_t i = 0; i < y_ne; i++) {
         if (std::is_same<float, Y_TYPE>()) {
-            // y[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f;
-            y[i] = (i % k == i / k) ? 1.0f : 0.0f;
+            y[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f;
+            // y[i] = (i % k == i / k) ? 1.0f : 0.0f;
+            // y[i] = i + 1;
         } else if (std::is_same<ggml_fp16_t, Y_TYPE>()) {
-            // y[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f);
-            y[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f);
+            y[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f);
+            // y[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f);
+            // y[i] = ggml_fp32_to_fp16(i + 1);
         } else {
             GGML_ABORT("fatal error");
         }
@@ -5210,16 +6096,16 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
     ggml_vk_buffer_write(d_Y, 0, y, sizeof(Y_TYPE) * k * n * batch);
 
     vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+    ggml_vk_ctx_begin(ctx->device, subctx);
     for (size_t i = 0; i < num_it; i++) {
-        ggml_vk_ctx_begin(ctx->device, subctx);
         ggml_vk_matmul(
             ctx, subctx, p, ggml_vk_subbuffer(d_X), ggml_vk_subbuffer(d_Y), ggml_vk_subbuffer(d_D), ggml_vk_subbuffer(ctx->prealloc_split_k),
             m, n, k,
             k, k, m, k*m, k*n, m*n,
             split_k, batch, batch, batch, 1, 1
         );
-        ggml_vk_ctx_end(subctx);
     }
+    ggml_vk_ctx_end(subctx);
 
     auto begin = std::chrono::high_resolution_clock::now();
     ggml_vk_submit(subctx, ctx->fence);
@@ -5284,7 +6170,7 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
         double err = std::fabs(d[i] - d_chk[i]);
         avg_err += err;
 
-        if (err > 0.05f && first_err_n == -1) {
+        if ((err > 0.05f || std::isnan(err)) && first_err_n == -1) {
             first_err_b = i / (m * n);
             first_err_n = (i % (m * n)) / m;
             first_err_m = (i % (m * n)) % m;
@@ -5297,12 +6183,10 @@ static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t
 
     std::cerr << "TEST " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time / num_it << "ms " << tflops << " TFLOPS avg_err=" << avg_err << std::endl;
 
-    if (avg_err > 0.1) {
+    if (avg_err > 0.1 || std::isnan(avg_err)) {
         std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl;
         std::cerr << "Actual result: " << std::endl << std::endl;
         ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
-        std::cerr << std::endl;
-        ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n + 15, first_err_b);
         std::cerr << "Expected result: " << std::endl << std::endl;
         ggml_vk_print_matrix_area(d_chk, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
 
@@ -5485,13 +6369,13 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
     vk_pipeline p;
     std::string shname;
     if (shader_size == 0) {
-        p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_s;
+        p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->a_s : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_s;
         shname = std::string(ggml_type_name(quant)) + "_ALIGNED_S";
     } else if (shader_size == 1) {
-        p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_m;
+        p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->a_m : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_m;
         shname = std::string(ggml_type_name(quant)) + "_ALIGNED_M";
     } else if (shader_size == 2) {
-        p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_l;
+        p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->a_l : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->a_l;
         shname = std::string(ggml_type_name(quant)) + "_ALIGNED_L";
     } else {
         GGML_ASSERT(0);
@@ -5501,13 +6385,13 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
 
     if (k != kpad) {
         if (shader_size == 0) {
-            p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->s;
+            p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->s : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->s;
             shname = std::string(ggml_type_name(quant)) + "_S";
         } else if (shader_size == 1) {
-            p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->m;
+            p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->m : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->m;
             shname = std::string(ggml_type_name(quant)) + "_M";
         } else if (shader_size == 2) {
-            p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->l;
+            p = ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat[quant].f16acc->l : ctx->device->pipeline_dequant_mul_mat_mat[quant].f32acc->l;
             shname = std::string(ggml_type_name(quant)) + "_L";
         } else {
             GGML_ASSERT(0);
@@ -5557,16 +6441,16 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
     ggml_vk_buffer_write(y_buf, 0, y, y_sz);
 
     vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+    ggml_vk_ctx_begin(ctx->device, subctx);
     for (size_t i = 0; i < num_it; i++) {
-        ggml_vk_ctx_begin(ctx->device, subctx);
         ggml_vk_matmul(
             ctx, subctx, p, ggml_vk_subbuffer(qx_buf), ggml_vk_subbuffer(y_buf), ggml_vk_subbuffer(d_buf), ggml_vk_subbuffer(ctx->prealloc_split_k),
             m, n, k,
             k, k, m, k*m, k*n, m*n,
             split_k, batch, batch, batch, 1, 1
         );
-        ggml_vk_ctx_end(subctx);
     }
+    ggml_vk_ctx_end(subctx);
 
     auto begin = std::chrono::high_resolution_clock::now();
 
@@ -5666,105 +6550,13 @@ static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m,
 
 static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
 #if defined(GGML_VULKAN_RUN_TESTS)
-    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_F32);
-    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_0);
-    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_1);
-    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_0);
-    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_1);
-    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q8_0);
-    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q2_K);
-    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q3_K);
-    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_K);
-    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_K);
-    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q6_K);
-    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_IQ4_NL);
-
-    ggml_vk_test_matmul<ggml_fp16_t, ggml_fp16_t>(ctx, 512, 512, 100, 32, 100, 1, 2);
-
-    ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 0);
-    ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 1);
-    ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 2);
-    // ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 0);
-    // ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 1);
-    // ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 2);
-
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_0);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_0);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_0);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_0);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_0);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_0);
-
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_1);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_1);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_1);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_1);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_1);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_1);
-
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_0);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_0);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_0);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_0);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_0);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_0);
-
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_1);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_1);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_1);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_1);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_1);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_1);
-
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q8_0);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q8_0);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q8_0);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q8_0);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q8_0);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q8_0);
-
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q2_K);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q2_K);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q2_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q2_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q2_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q2_K);
-
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q3_K);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q3_K);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q3_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q3_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q3_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q3_K);
-
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_K);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_K);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_K);
-
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_K);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_K);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_K);
-
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q6_K);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q6_K);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q6_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q6_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q6_K);
-    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q6_K);
-
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_IQ4_NL);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_IQ4_NL);
-    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_IQ4_NL);
-
-    std::cerr << std::endl;
-
     const std::vector<size_t> vals {
+        512, 512, 128,
+        128, 512, 512,
+        4096, 512, 4096,
+        11008, 512, 4096,
+        4096, 512, 11008,
+        32000, 512, 4096,
         8, 8, 8,
         100, 46, 576,
         623, 111, 128,
@@ -5777,25 +6569,52 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
         49, 49, 128,
         128, 49, 49,
         4096, 49, 4096,
-        11008, 49, 4096,
-        4096, 49, 11008,
-        32000, 49, 4096,
-        512, 512, 128,
-        128, 512, 512,
-        4096, 512, 4096,
-        11008, 512, 4096,
-        4096, 512, 11008,
-        32000, 512, 4096,
     };
-    const size_t num_it = 1;
+    const size_t num_it = 100;
+
     for (size_t i = 0; i < vals.size(); i += 3) {
         ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0);
         ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1);
         ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 2);
-        // ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0);
-        // ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1);
-        // ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2);
-        std::cerr << std::endl;
+        std::cerr << '\n';
+        ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 0);
+        ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 1);
+        ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 2);
+        std::cerr << '\n';
+        ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0);
+        ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1);
+        ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2);
+        std::cerr << '\n' << std::endl;
+
+        if (vals[i + 2] % 32 == 0) {
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0, GGML_TYPE_Q4_0);
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1, GGML_TYPE_Q4_0);
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 2, GGML_TYPE_Q4_0);
+            std::cerr << '\n';
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 0, GGML_TYPE_Q4_0);
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 1, GGML_TYPE_Q4_0);
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 2, GGML_TYPE_Q4_0);
+            std::cerr << '\n';
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0, GGML_TYPE_Q4_0);
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1, GGML_TYPE_Q4_0);
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2, GGML_TYPE_Q4_0);
+            std::cerr << '\n' << std::endl;
+        }
+
+        if (vals[i + 2] % 256 == 0) {
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0, GGML_TYPE_Q4_K);
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1, GGML_TYPE_Q4_K);
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 2, GGML_TYPE_Q4_K);
+            std::cerr << '\n';
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 0, GGML_TYPE_Q4_K);
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 1, GGML_TYPE_Q4_K);
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 2, 2, GGML_TYPE_Q4_K);
+            std::cerr << '\n';
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0, GGML_TYPE_Q4_K);
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1, GGML_TYPE_Q4_K);
+            ggml_vk_test_dequant_matmul(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2, GGML_TYPE_Q4_K);
+            std::cerr << '\n' << std::endl;
+        }
     }
 
     GGML_ABORT("fatal error");
@@ -5842,6 +6661,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     const ggml_tensor * src0 = node->src[0];
     const ggml_tensor * src1 = node->src[1];
     const ggml_tensor * src2 = node->src[2];
+    const ggml_tensor * src3 = node->src[3];
 
     switch (node->op) {
     // Return on empty ops to avoid generating a compute_ctx and setting exit_tensor
@@ -5893,7 +6713,9 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_IM2COL:
     case GGML_OP_TIMESTEP_EMBEDDING:
     case GGML_OP_POOL_2D:
+    case GGML_OP_RWKV_WKV6:
     case GGML_OP_LEAKY_RELU:
+    case GGML_OP_FLASH_ATTN_EXT:
         break;
     default:
         std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl;
@@ -5911,6 +6733,48 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
         } else {
             compute_ctx = ctx->compute_ctx.lock();
         }
+    } else {
+        switch (node->op) {
+        case GGML_OP_REPEAT:
+        case GGML_OP_ACC:
+        case GGML_OP_GET_ROWS:
+        case GGML_OP_ADD:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_CONCAT:
+        case GGML_OP_UPSCALE:
+        case GGML_OP_SCALE:
+        case GGML_OP_SQR:
+        case GGML_OP_SIN:
+        case GGML_OP_COS:
+        case GGML_OP_CLAMP:
+        case GGML_OP_PAD:
+        case GGML_OP_CPY:
+        case GGML_OP_CONT:
+        case GGML_OP_DUP:
+        case GGML_OP_NORM:
+        case GGML_OP_GROUP_NORM:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_UNARY:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_ROPE:
+        case GGML_OP_ARGSORT:
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_IM2COL:
+        case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_POOL_2D:
+        case GGML_OP_LEAKY_RELU:
+            {
+                // These operations all go through ggml_vk_op_f32, so short-circuit and
+                // do the only thing needed for the dryrun.
+                vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, node, node->op);
+                ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+                return false;
+            }
+        default:
+            break;
+        }
     }
 
     switch (node->op) {
@@ -6044,6 +6908,16 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
     case GGML_OP_MUL_MAT_ID:
         ggml_vk_mul_mat_id(ctx, compute_ctx, src0, src1, src2, node, dryrun);
 
+        break;
+
+    case GGML_OP_FLASH_ATTN_EXT:
+        ggml_vk_flash_attn(ctx, compute_ctx, src0, src1, src2, src3, node, dryrun);
+
+        break;
+
+    case GGML_OP_RWKV_WKV6:
+        ggml_vk_rwkv_wkv6(ctx, compute_ctx, node, dryrun);
+
         break;
     default:
         return false;
@@ -6124,6 +6998,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
     case GGML_OP_IM2COL:
     case GGML_OP_TIMESTEP_EMBEDDING:
     case GGML_OP_POOL_2D:
+    case GGML_OP_RWKV_WKV6:
     case GGML_OP_LEAKY_RELU:
     case GGML_OP_REPEAT:
         buf = tensor->buffer;
@@ -6144,6 +7019,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
         break;
     case GGML_OP_MUL_MAT:
     case GGML_OP_MUL_MAT_ID:
+    case GGML_OP_FLASH_ATTN_EXT:
         buf = tensor->buffer;
 
         break;
@@ -6640,16 +7516,17 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
     bool first_node_in_batch = true; // true if next node will be first node in a batch
     int submit_node_idx = 0; // index to first node in a batch
 
-    // submit work every submit_count node to overlap CPU cmdbuffer generation with GPU execution
-    constexpr int submit_count = 100;
+    // Submit work every nodes_per_submit nodes to overlap CPU cmdbuffer generation with GPU execution.
+    // Start with a smaller count to get work submitted right away, and increase it after each submit.
+    int nodes_per_submit = 20;
     int submitted_nodes = 0;
+    int submit_count = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
         if (first_node_in_batch) {
             submit_node_idx = i;
         }
 
-        bool submit = (submitted_nodes >= submit_count) || (i == last_node);
-
+        bool submit = (submitted_nodes >= nodes_per_submit) || (i == last_node);
 
         bool enqueued = ggml_vk_build_graph(ctx, cgraph->nodes[i], i, cgraph->nodes[submit_node_idx], submit_node_idx, false, i == last_node, submit);
 
@@ -6666,6 +7543,15 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
         if (submit) {
             first_node_in_batch = true;
             submitted_nodes = 0;
+            switch (submit_count) {
+            case 0:
+                nodes_per_submit = 50;
+                break;
+            default:
+                nodes_per_submit = 100;
+                break;
+            }
+            submit_count++;
         }
     }
 
@@ -6822,6 +7708,12 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             {
+                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+                const vk_device& device = ggml_vk_get_device(ctx->device);
+                if (op->op == GGML_OP_MUL_MAT_ID && !device->mul_mat_id_s && !device->mul_mat_id_m && !device->mul_mat_id_l) {
+                    // If there's not enough shared memory for row_ids and the result tile, fallback to CPU
+                    return false;
+                }
                 switch (op->src[0]->type) {
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:
@@ -6859,6 +7751,57 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
 
                 return true;
             } break;
+        case GGML_OP_FLASH_ATTN_EXT:
+            {
+                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+                if (!ggml_vk_get_device(ctx->device)->coopmat2) {
+                    return false;
+                }
+                switch (op->src[0]->ne[0]) {
+                case 64:
+                case 80:
+                case 96:
+                case 112:
+                case 128:
+                case 256:
+                    break;
+                default:
+                    return false;
+                }
+                if (op->src[0]->type != GGML_TYPE_F32) {
+                    return false;
+                }
+                if (op->type != GGML_TYPE_F32) {
+                    return false;
+                }
+                if (op->src[3] && op->src[3]->type != GGML_TYPE_F16) {
+                    return false;
+                }
+                // It's straightforward to support different K/V dequant, but would
+                // significantly increase the number of pipelines
+                if (op->src[1]->type != op->src[2]->type) {
+                    return false;
+                }
+                switch (op->src[1]->type) {
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q5_0:
+                case GGML_TYPE_Q5_1:
+                case GGML_TYPE_Q8_0:
+                // K dequants currently disabled because D dimension is rounded up to 256 and runs inefficiently
+                //case GGML_TYPE_Q2_K:
+                //case GGML_TYPE_Q3_K:
+                //case GGML_TYPE_Q4_K:
+                //case GGML_TYPE_Q5_K:
+                //case GGML_TYPE_Q6_K:
+                case GGML_TYPE_IQ4_NL:
+                    break;
+                default:
+                    return false;
+                }
+                return true;
+            }
         case GGML_OP_GET_ROWS:
             {
                 switch (op->src[0]->type) {
@@ -6895,7 +7838,16 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_REPEAT:
             return ggml_type_size(op->type) == sizeof(float) && ggml_type_size(op->src[0]->type) == sizeof(float);
         case GGML_OP_ROPE:
-            return ggml_is_contiguous(op->src[0]);
+            {
+                const int mode = ((const int32_t *) op->op_params)[2];
+                if (mode & GGML_ROPE_TYPE_MROPE) {
+                    return false;
+                }
+                if (mode & GGML_ROPE_TYPE_VISION) {
+                    return false;
+                }
+                return ggml_is_contiguous(op->src[0]);
+            }
         case GGML_OP_NONE:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
@@ -6923,6 +7875,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_IM2COL:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_POOL_2D:
+        case GGML_OP_RWKV_WKV6:
         case GGML_OP_LEAKY_RELU:
             return true;
         default:
@@ -7019,8 +7972,9 @@ static const struct ggml_backend_reg_i ggml_backend_vk_reg_i = {
 
 ggml_backend_reg_t ggml_backend_vk_reg() {
     static ggml_backend_reg reg = {
-        /* .iface   = */ ggml_backend_vk_reg_i,
-        /* .context = */ nullptr,
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_vk_reg_i,
+        /* .context     = */ nullptr,
     };
 
     return &reg;
@@ -7172,6 +8126,7 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
     ggml_tensor * src0 = tensor->src[0];
     ggml_tensor * src1 = tensor->src[1];
     ggml_tensor * src2 = tensor->src[2];
+    ggml_tensor * src3 = tensor->src[3];
 
     struct ggml_init_params iparams = {
         /*.mem_size   =*/ 2ul*1024ul*1024ul*1024ul,
@@ -7184,15 +8139,18 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
     struct ggml_tensor * src0_clone = nullptr;
     struct ggml_tensor * src1_clone = nullptr;
     struct ggml_tensor * src2_clone = nullptr;
+    struct ggml_tensor * src3_clone = nullptr;
     struct ggml_tensor * tensor_clone = nullptr;
 
     size_t src0_size;
     size_t src1_size;
     size_t src2_size;
+    size_t src3_size;
 
     void * src0_buffer = nullptr;
     void * src1_buffer = nullptr;
     void * src2_buffer = nullptr;
+    void * src3_buffer = nullptr;
 
     if (src0 != nullptr) {
         src0_clone = ggml_dup_tensor(ggml_ctx, src0);
@@ -7320,8 +8278,53 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
             ggml_vk_print_tensor(src2, "src2");
         }
     }
+    if (src3 != nullptr) {
+        src3_clone = ggml_dup_tensor(ggml_ctx, src3);
 
-    if (tensor->op == GGML_OP_MUL_MAT) {
+        src3_size = ggml_nbytes(src3);
+
+        src3_buffer = malloc(src3_size);
+        src3_clone->data = src3_buffer;
+        if (ggml_backend_buffer_is_host(src3->buffer)) {
+            memcpy(src3_clone->data, src3->data, src3_size);
+            memcpy(src3_clone->nb, src3->nb, sizeof(size_t) * GGML_MAX_DIMS);
+        } else if (ggml_backend_buffer_is_vk(src3->buffer)) {
+            ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)src3->buffer->context;
+            vk_buffer& buffer_gpu = buf_ctx->dev_buffer;
+            uint64_t offset = vk_tensor_offset(src3) + src3->view_offs;
+            if (!ggml_is_contiguous(src3) && ggml_vk_dim01_contiguous(src3)) {
+                for (int i3 = 0; i3 < src3->ne[3]; i3++) {
+                    for (int i2 = 0; i2 < src3->ne[2]; i2++) {
+                        const int idx = i3*src3->ne[2] + i2;
+                        ggml_vk_buffer_read(buffer_gpu, offset + idx * src3->nb[2], ((char *)src3_clone->data + idx * src3_clone->nb[2]), src3->ne[1] * src3->nb[1]);
+                    }
+                }
+
+                src3_clone->nb[0] = src3->nb[0];
+                src3_clone->nb[1] = src3->nb[1];
+                for (int i = 2; i < GGML_MAX_DIMS; i++) {
+                    src3_clone->nb[i] = src3_clone->nb[i - 1]*src3_clone->ne[i - 1];
+                }
+            } else {
+                if (offset + src3_size >= buffer_gpu->size) {
+                    src3_size = buffer_gpu->size - offset;
+                }
+                ggml_vk_buffer_read(buffer_gpu, offset, src3_clone->data, src3_size);
+                memcpy(src3_clone->nb, src3->nb, sizeof(size_t) * GGML_MAX_DIMS);
+            }
+        } else {
+            GGML_ABORT("fatal error");
+        }
+
+        if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
+            ggml_vk_print_tensor(src3, "src3");
+        }
+    }
+
+    if (tensor->op == GGML_OP_FLASH_ATTN_EXT) {
+        const float *params = (const float *)tensor->op_params;
+        tensor_clone = ggml_flash_attn_ext(ggml_ctx, src0_clone, src1_clone, src2_clone, src3_clone, params[0], params[1], params[2]);
+    } else if (tensor->op == GGML_OP_MUL_MAT) {
         tensor_clone = ggml_mul_mat(ggml_ctx, src0_clone, src1_clone);
     } else if (tensor->op == GGML_OP_MUL_MAT_ID) {
         tensor_clone = ggml_mul_mat_id(ggml_ctx, src0_clone, src1_clone, src2_clone);
@@ -7437,7 +8440,7 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
         const int32_t max_period = tensor->op_params[1];
         tensor_clone = ggml_timestep_embedding(ggml_ctx, src0_clone, dim, max_period);
     } else if (tensor->op == GGML_OP_POOL_2D) {
-        enum ggml_op_pool op = static_cast<ggml_op_pool>(dst->op_params[0]);
+        enum ggml_op_pool op = static_cast<ggml_op_pool>(tensor->op_params[0]);
         const int32_t k0 = tensor->op_params[1];
         const int32_t k1 = tensor->op_params[2];
         const int32_t s0 = tensor->op_params[3];
@@ -7449,7 +8452,11 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
     } else if (tensor->op == GGML_OP_LEAKY_RELU) {
         const float * op_params = (const float *)tensor->op_params;
         tensor_clone = ggml_leaky_relu(ggml_ctx, src0_clone, op_params[0], false);
-    } else {
+    } else if (tensor->op == GGML_OP_RWKV_WKV6) {
+        tensor_clone = ggml_rwkv_wkv6(ggml_ctx, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3],
+        tensor->src[4], tensor->src[5]);
+    }
+    else {
         std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
         GGML_ABORT("fatal error");
     }
@@ -7646,3 +8653,5 @@ static void ggml_vk_check_results_1(ggml_tensor * tensor) {
     VK_LOG_DEBUG("END ggml_vk_check_results_1(" << tensor->name << ")");
 }
 #endif
+
+GGML_BACKEND_DL_IMPL(ggml_backend_vk_reg)
diff --git a/ggml/src/vulkan-shaders/CMakeLists.txt b/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt
similarity index 56%
rename from ggml/src/vulkan-shaders/CMakeLists.txt
rename to ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt
index 10075db3377..bd0c74cb1c7 100644
--- a/ggml/src/vulkan-shaders/CMakeLists.txt
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt
@@ -1,7 +1,9 @@
 find_package (Threads REQUIRED)
+find_package(Vulkan COMPONENTS glslc REQUIRED)
 
 set(TARGET vulkan-shaders-gen)
 add_executable(${TARGET} vulkan-shaders-gen.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_compile_features(${TARGET} PRIVATE cxx_std_17)
 target_link_libraries(vulkan-shaders-gen PUBLIC Threads::Threads)
+target_link_libraries(vulkan-shaders-gen PRIVATE Vulkan::Vulkan)
diff --git a/ggml/src/vulkan-shaders/acc.comp b/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp
similarity index 51%
rename from ggml/src/vulkan-shaders/acc.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/acc.comp
index 4c8739efee2..4f5a04e71ce 100644
--- a/ggml/src/vulkan-shaders/acc.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/acc.comp
@@ -3,6 +3,8 @@
 #include "types.comp"
 #include "generic_binary_head.comp"
 
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
 void main() {
     const uint idx = gl_GlobalInvocationID.x;
     if (idx >= p.ne) {
@@ -15,10 +17,13 @@ void main() {
     const uint oy = (src1_i - (oz * p.nb02)) / p.nb01;
     const uint ox = src1_i % p.nb01;
 
+    uint i00, i01, i02, i03;
+    get_indices(idx, i00, i01, i02, i03);
+
     if (ox < p.ne10 && oy < p.ne11 && oz < p.ne12) {
-        data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) + FLOAT_TYPE(data_b[ox + oy * p.ne10 + oz * p.ne10 * p.ne11]));
+        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[ox + oy * p.ne10 + oz * p.ne10 * p.ne11]));
     } else {
-        data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]));
+        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]));
     }
 }
 
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/add.comp b/ggml/src/ggml-vulkan/vulkan-shaders/add.comp
new file mode 100644
index 00000000000..da61b76dfc8
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/add.comp
@@ -0,0 +1,29 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+const uint num_threads = 256;
+
+layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    uint idx = get_idx();
+
+    // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
+    const uint num_iter = 2;
+
+    [[unroll]] for (uint i = 0; i < num_iter; ++i) {
+        if (idx >= p.ne) {
+            continue;
+        }
+        uint i00, i01, i02, i03;
+        get_indices(idx, i00, i01, i02, i03);
+
+        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) + FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
+
+        idx += num_threads;
+    }
+}
diff --git a/ggml/src/vulkan-shaders/argsort.comp b/ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/argsort.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/argsort.comp
diff --git a/ggml/src/vulkan-shaders/clamp.comp b/ggml/src/ggml-vulkan/vulkan-shaders/clamp.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/clamp.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/clamp.comp
diff --git a/ggml/src/vulkan-shaders/concat.comp b/ggml/src/ggml-vulkan/vulkan-shaders/concat.comp
similarity index 95%
rename from ggml/src/vulkan-shaders/concat.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/concat.comp
index c23b6eb1b0c..683f9ac3c1d 100644
--- a/ggml/src/vulkan-shaders/concat.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/concat.comp
@@ -3,6 +3,8 @@
 #include "types.comp"
 #include "generic_binary_head.comp"
 
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
 void main() {
     const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
     const int dim = p.param3;
diff --git a/ggml/src/vulkan-shaders/contig_copy.comp b/ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/contig_copy.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/contig_copy.comp
diff --git a/ggml/src/vulkan-shaders/copy.comp b/ggml/src/ggml-vulkan/vulkan-shaders/copy.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/copy.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/copy.comp
diff --git a/ggml/src/vulkan-shaders/cos.comp b/ggml/src/ggml-vulkan/vulkan-shaders/cos.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/cos.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/cos.comp
diff --git a/ggml/src/vulkan-shaders/dequant_f32.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_f32.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/dequant_f32.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/dequant_f32.comp
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp
new file mode 100644
index 00000000000..91bb8f8db61
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp
@@ -0,0 +1,118 @@
+#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
+#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
+#endif
+
+#include "types.comp"
+
+#if defined(A_TYPE_PACKED16)
+layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
+#endif
+#if defined(A_TYPE_PACKED32)
+layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
+#endif
+
+#if defined(DATA_A_F32)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
+}
+#endif
+
+#if defined(DATA_A_F16)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
+}
+#endif
+
+#if defined(DATA_A_Q4_0)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return (vec2(vui & 0xF, vui >> 4) - 8.0f);
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return (vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, vui >> 12) - 8.0f);
+}
+#endif
+
+#if defined(DATA_A_Q4_1)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return vec2(vui & 0xF, vui >> 4);
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return vec4(vui & 0xF, (vui >> 4) & 0xF, (vui >> 8) & 0xF, vui >> 12);
+}
+#endif
+
+#if defined(DATA_A_Q5_0)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint uint_qh = uint(data_a[a_offset + ib].qh[1]) << 16 | data_a[a_offset + ib].qh[0];
+    const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f);
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const uint uint_qh = uint(data_a_packed16[a_offset + ib].qh[1]) << 16 | data_a_packed16[a_offset + ib].qh[0];
+    const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+    const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10);
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return (vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y) - 16.0f);
+}
+#endif
+
+#if defined(DATA_A_Q5_1)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint uint_qh = data_a[a_offset + ib].qh;
+    const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y);
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const uint uint_qh = data_a_packed16[a_offset + ib].qh;
+    const ivec2 qh0 = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+    const ivec2 qh1 = ivec2(((uint_qh >> (iqs + 1)) << 4) & 0x10, (uint_qh >> (iqs + 13)) & 0x10);
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y);
+}
+#endif
+
+#if defined(DATA_A_Q8_0)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1]));
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    uint32_t v0 = data_a_packed16[a_offset + ib].qs[iqs/2];
+    uint32_t v1 = data_a_packed16[a_offset + ib].qs[iqs/2 + 1];
+    return vec4(int8_t(v0 & 0xFF), int8_t(v0 >> 8), int8_t(v1 & 0xFF), int8_t(v1 >> 8));
+}
+#endif
+
+#if defined(DATA_A_IQ4_NL)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]);
+}
+vec4 dequantize4(uint ib, uint iqs, uint a_offset) {
+    const uint vui = uint(data_a_packed16[a_offset + ib].qs[iqs/2]);
+    return vec4(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[(vui >> 4) & 0xF], kvalues_iq4nl[(vui >> 8) & 0xF], kvalues_iq4nl[vui >> 12]);
+}
+#endif
+
+#if defined(DATA_A_F32) || defined(DATA_A_F16)
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(0, 0);
+}
+#endif
+
+#if defined(DATA_A_Q4_0) || defined(DATA_A_Q5_0) || defined(DATA_A_Q8_0) || defined(DATA_A_IQ4_NL)
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(float(data_a[a_offset + ib].d), 0);
+}
+#endif
+
+#if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
+vec2 get_dm(uint ib, uint a_offset) {
+    return vec2(float(data_a[a_offset + ib].d), float(data_a[a_offset + ib].m));
+}
+#endif
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp
new file mode 100644
index 00000000000..a8707b621e7
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp
@@ -0,0 +1,305 @@
+
+#include "types.comp"
+
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ4_0 {
+   block_q4_0_packed16 block;
+};
+
+float16_t dequantFuncQ4_0(const in decodeBufQ4_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const float16_t d = bl.block.d;
+    const uint idx = coordInBlock[1];
+    const uint shift = (idx & 0x10) >> 2;
+    uint32_t qs = unpack8(uint32_t(bl.block.qs[(idx & 0xE) >> 1]))[idx & 1];
+    qs >>= shift;
+    qs &= 0xF;
+    float16_t ret = (float16_t(qs) - float16_t(8)) * d;
+    return ret;
+}
+
+layout(buffer_reference, std430, buffer_reference_align = 4) buffer decodeBufQ4_1 {
+   block_q4_1 block;
+};
+
+float16_t dequantFuncQ4_1(const in decodeBufQ4_1 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const float16_t d = bl.block.d;
+    const float16_t m = bl.block.m;
+    const uint idx = coordInBlock[1];
+    const uint iqs = idx & 0xF;
+    const uint shift = (idx & 0x10) >> 2;
+    uint32_t qs = bl.block.qs[iqs];
+    qs >>= shift;
+    qs &= 0xF;
+    float16_t ret = float16_t(qs) * d + m;
+    return ret;
+}
+
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ5_0 {
+   block_q5_0 block;
+};
+
+float16_t dequantFuncQ5_0(const in decodeBufQ5_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const float16_t d = bl.block.d;
+    const uint idx = coordInBlock[1];
+    const uint iqs = idx & 0xF;
+
+    const uint uint_qh = uint(bl.block.qh[1]) << 16 | bl.block.qh[0];
+    const uint qh = ((uint_qh >> idx) << 4) & 0x10;
+
+    const uint shift = (idx & 0x10) >> 2;
+    uint32_t qs = bl.block.qs[iqs];
+    qs >>= shift;
+    qs &= 0xF;
+
+    float16_t ret = (float16_t(qs | qh) - float16_t(16)) * d;
+    return ret;
+}
+
+layout(buffer_reference, std430, buffer_reference_align = 8) buffer decodeBufQ5_1 {
+   block_q5_1 block;
+};
+
+float16_t dequantFuncQ5_1(const in decodeBufQ5_1 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const float16_t d = bl.block.d;
+    const float16_t m = bl.block.m;
+    const uint idx = coordInBlock[1];
+    const uint iqs = idx & 0xF;
+
+    const uint uint_qh = bl.block.qh;
+    const uint qh = ((uint_qh >> idx) << 4) & 0x10;
+
+    const uint shift = (idx & 0x10) >> 2;
+    uint32_t qs = bl.block.qs[iqs];
+    qs >>= shift;
+    qs &= 0xF;
+
+    float16_t ret = float16_t(qs | qh) * d + m;
+    return ret;
+}
+
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ8_0 {
+   block_q8_0_packed16 block;
+};
+
+float16_t dequantFuncQ8_0(const in decodeBufQ8_0 bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const float16_t d = bl.block.d;
+    const uint idx = coordInBlock[1];
+    const uint iqs = idx;
+
+    // Load 16b and select the byte for this element
+    int32_t qs = unpack8(int32_t(bl.block.qs[(iqs & 0x1E) >> 1]))[iqs & 1];
+    float16_t ret = float16_t(qs) * d;
+    return ret;
+}
+
+layout(buffer_reference, std430, buffer_reference_align = 4) buffer decodeBufQ2_K {
+   block_q2_K block;
+};
+
+float16_t dequantFuncQ2_K(const in decodeBufQ2_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const f16vec2 d = bl.block.d;
+    const uint idx = coordInBlock[1];
+    const uint iqs = idx;
+
+    const uint qsi = (iqs / 128) * 32 + (iqs % 32);     // 0..31
+    const uint scalesi = iqs / 16;                      // 0..15
+    const uint qsshift = ((iqs % 128) / 32) * 2;        // 0,2,4,6
+
+    uint32_t qs = bl.block.qs[qsi];
+    const uint scales = bl.block.scales[scalesi];
+    float16_t ret = d.x * float16_t(scales & 0xF) * float16_t((qs >> qsshift) & 3) - d.y * float16_t(scales >> 4);
+    return ret;
+}
+
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ3_K {
+   block_q3_K block;
+};
+
+float16_t dequantFuncQ3_K(const in decodeBufQ3_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const uint idx = coordInBlock[1];
+    const uint iqs = idx;
+
+    const uint n = iqs / 128;                    // 0,1
+    const uint qsi = n * 32 + (iqs % 32);        // 0..63
+    const uint hmi =          (iqs % 32);        // 0..31
+    const uint j = (iqs % 128) / 8;              // 0..15
+    const uint is = iqs / 16;                    // 0..15
+    const uint halfsplit = ((iqs % 128) / 32);   // 0,1,2,3
+    const uint qsshift = halfsplit * 2;          // 0,2,4,6
+    const uint m = 1 << (4 * n + halfsplit);     // 1,2,4,8,16,32,64,128
+
+    uint32_t scaleidx0 = (is < 8) ? is : (is-8);
+    uint32_t scaleidx0shift = (is < 8) ? 0 : 4;
+    uint32_t scaleidx1 = is + 8 - (is/4)*4;
+    uint32_t scaleidx1shift = (is/4)*2;
+
+    const int8_t us = int8_t(((bl.block.scales[scaleidx0] >> scaleidx0shift) & 0xF) | (((bl.block.scales[scaleidx1] >> scaleidx1shift) & 3) << 4));
+
+    const float16_t dl = bl.block.d * float16_t(us - 32);
+
+    float16_t ret = dl * float16_t(int8_t((bl.block.qs[qsi    ] >> qsshift) & 3) - (((bl.block.hmask[hmi    ] & m) != 0) ? 0 : 4));
+
+    return ret;
+}
+
+layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ4_K {
+   block_q4_K block;
+};
+
+float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const uint idx = coordInBlock[1];
+    const uint iqs = idx;
+
+    const uint n = iqs / 64;                   // 0,1,2,3
+    const uint b = (iqs % 64) / 32;            // 0,1
+    const uint is = (idx & 0xE0) >> 5;         // 0..7
+    const uint qsi = n * 32 + (iqs % 32);      // 0..127
+
+    const f16vec2 loadd = bl.block.d;
+
+    uint32_t sc;
+    uint32_t mbyte;
+
+    uint32_t scidx0 = (is < 4) ? is : (is + 4);
+    uint32_t scidx1 = (is < 4) ? is : (is - 4);
+    uint32_t scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+    uint32_t scidxshift1 = (is < 4) ? 0 : 2;
+    uint32_t mbidx0 = is + 4;
+    uint32_t mbidx1 = (is < 4) ? is + 4 : is;
+    uint32_t mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+    uint32_t mbidxshift0 = (is < 4) ? 0 : 4;
+    uint32_t mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+    uint32_t mbidxshift1 = (is < 4) ? 0 : 2;
+
+    sc    = uint8_t((bl.block.scales[scidx0] & 0xF)                         | ((bl.block.scales[scidx1] & scidxmask1) >> scidxshift1));
+    mbyte = uint8_t(((bl.block.scales[mbidx0] & mbidxmask0) >> mbidxshift0) | ((bl.block.scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
+    const float16_t d = loadd.x * float16_t(sc);
+    const float16_t m = loadd.y * float16_t(mbyte);
+
+    uint32_t dmask = 0xF << (b * 4);
+
+    float16_t ret = d * float16_t((bl.block.qs[qsi    ] & dmask) >> (b * 4)) - m;
+
+    return ret;
+}
+
+layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K {
+   block_q5_K block;
+};
+
+float16_t dequantFuncQ5_K(const in decodeBufQ5_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const uint idx = coordInBlock[1];
+    const uint iqs = idx;
+
+    const uint n = iqs / 64;                   // 0,1,2,3
+    const uint b = (iqs % 64) / 32;            // 0,1
+    const uint is = (idx & 0xE0) >> 5;         // 0..7
+    const uint qsi = n * 32 + (iqs % 32);      // 0..127
+    const uint qhi = (iqs % 32);               // 0..31
+
+    const uint8_t hm = uint8_t(1 << (iqs / 32));
+
+    const f16vec2 loadd = bl.block.d;
+
+    uint32_t sc;
+    uint32_t mbyte;
+
+    uint32_t scidx0 = (is < 4) ? is : (is + 4);
+    uint32_t scidx1 = (is < 4) ? is : (is - 4);
+    uint32_t scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+    uint32_t scidxshift1 = (is < 4) ? 0 : 2;
+    uint32_t mbidx0 = is + 4;
+    uint32_t mbidx1 = (is < 4) ? is + 4 : is;
+    uint32_t mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+    uint32_t mbidxshift0 = (is < 4) ? 0 : 4;
+    uint32_t mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+    uint32_t mbidxshift1 = (is < 4) ? 0 : 2;
+
+    sc    = uint8_t((bl.block.scales[scidx0] & 0xF)                         | ((bl.block.scales[scidx1] & scidxmask1) >> scidxshift1));
+    mbyte = uint8_t(((bl.block.scales[mbidx0] & mbidxmask0) >> mbidxshift0) | ((bl.block.scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
+    const float16_t d = loadd.x * float16_t(sc);
+    const float16_t m = loadd.y * float16_t(mbyte);
+
+    uint32_t dmask = 0xF << (b * 4);
+
+    float16_t ret = d * (float16_t((bl.block.qs[qsi    ] & dmask) >> (b * 4)) + float16_t((bl.block.qh[qhi    ] & hm) != 0 ? 16 : 0)) - m;
+
+    return ret;
+}
+
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufQ6_K {
+   block_q6_K block;
+};
+
+float16_t dequantFuncQ6_K(const in decodeBufQ6_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const uint idx = coordInBlock[1];
+    const uint iqs = idx;
+
+    const uint n = iqs / 128;                   // 0,1
+    const uint b = (iqs % 128) / 64;            // 0,1
+    const uint is_b = (iqs % 32) / 16;          // 0,1
+    const uint qhshift = ((iqs % 128) / 32) * 2;// 0,2,4,6
+    const uint is = 8 * n + qhshift + is_b;     // 0..15
+    const uint qsi = n * 64 + (iqs % 64);       // 0..127
+    const uint qhi = n * 32 + (iqs % 32);       // 0..63
+
+    const float16_t dscale = bl.block.d * float16_t(bl.block.scales[is]);
+
+    float16_t ret = dscale * float16_t(int8_t(((bl.block.ql[qsi    ] >> (b * 4)) & 0xF) | (((bl.block.qh[qhi    ] >> qhshift) & 3) << 4)) - 32);
+
+    return ret;
+}
+
+#if defined(DATA_A_IQ4_NL)
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ4_NL {
+   block_iq4_nl block;
+};
+
+float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const float16_t d = bl.block.d;
+    const uint idx = coordInBlock[1];
+    const uint iqs = idx & 0xF;
+    const uint shift = (idx & 0x10) >> 2;
+    uint32_t qs = bl.block.qs[iqs];
+    qs >>= shift;
+    qs &= 0xF;
+    float16_t ret = float16_t(kvalues_iq4nl[qs]) * d;
+    return ret;
+}
+#endif
+
+#if defined(DATA_A_Q4_0)
+#define dequantFuncA dequantFuncQ4_0
+#elif defined(DATA_A_Q4_1)
+#define dequantFuncA dequantFuncQ4_1
+#elif defined(DATA_A_Q5_0)
+#define dequantFuncA dequantFuncQ5_0
+#elif defined(DATA_A_Q5_1)
+#define dequantFuncA dequantFuncQ5_1
+#elif defined(DATA_A_Q8_0)
+#define dequantFuncA dequantFuncQ8_0
+#elif defined(DATA_A_Q2_K)
+#define dequantFuncA dequantFuncQ2_K
+#elif defined(DATA_A_Q3_K)
+#define dequantFuncA dequantFuncQ3_K
+#elif defined(DATA_A_Q4_K)
+#define dequantFuncA dequantFuncQ4_K
+#elif defined(DATA_A_Q5_K)
+#define dequantFuncA dequantFuncQ5_K
+#elif defined(DATA_A_Q6_K)
+#define dequantFuncA dequantFuncQ6_K
+#elif defined(DATA_A_IQ4_NL)
+#define dequantFuncA dequantFuncIQ4_NL
+#endif
diff --git a/ggml/src/vulkan-shaders/dequant_head.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_head.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/dequant_head.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/dequant_head.comp
diff --git a/ggml/src/vulkan-shaders/dequant_iq4_nl.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp
similarity index 97%
rename from ggml/src/vulkan-shaders/dequant_iq4_nl.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp
index 34ef3da30b8..8de14fc03f1 100644
--- a/ggml/src/vulkan-shaders/dequant_iq4_nl.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp
@@ -10,6 +10,8 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
 
+    init_iq4nl_shmem();
+
     const uint tid = gl_LocalInvocationID.x % 64;
     const uint il  = tid/32;
     const uint ir  = tid%32;
diff --git a/ggml/src/vulkan-shaders/dequant_q2_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/dequant_q2_k.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp
diff --git a/ggml/src/vulkan-shaders/dequant_q3_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/dequant_q3_k.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp
diff --git a/ggml/src/vulkan-shaders/dequant_q4_0.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_0.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/dequant_q4_0.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_0.comp
diff --git a/ggml/src/vulkan-shaders/dequant_q4_1.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_1.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/dequant_q4_1.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_1.comp
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp
new file mode 100644
index 00000000000..987f113a35a
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp
@@ -0,0 +1,68 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
+        const uint ib = gl_WorkGroupID.x * 256 + wgy;
+        if (ib >= p.M * p.K / QUANT_K) {
+            return;
+        }
+
+        const uint tid = gl_LocalInvocationID.x;
+        const uint il = tid / 8;
+        const uint ir = tid % 8;
+        const uint is = 2 * il;
+        const uint n = 4;
+
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);
+
+        const uint y_idx = ib * QUANT_K + 64 * il + n * ir;
+        const uint qs_idx = 32*il + n * ir;
+
+        uint scidx0 = (is < 4) ? is : (is + 4);
+        uint scidx1 = (is < 4) ? is : (is - 4);
+        uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        uint scidxshift1 = (is < 4) ? 0 : 2;
+        uint mbidx0 = is + 4;
+        uint mbidx1 = (is < 4) ? is + 4 : is;
+        uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+        uint mbidxshift0 = (is < 4) ? 0 : 4;
+        uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        uint mbidxshift1 = (is < 4) ? 0 : 2;
+
+        uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
+        uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
+        const FLOAT_TYPE d1 = dall * sc;
+        const FLOAT_TYPE m1 = dmin * mbyte;
+
+        scidx0 = (is < 4) ? is + 1 : (is + 5);
+        scidx1 = (is < 4) ? is + 1 : (is - 3);
+        scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        scidxshift1 = (is < 4) ? 0 : 2;
+        mbidx0 = is + 5;
+        mbidx1 = (is < 4) ? is + 5 : is + 1;
+        mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+        mbidxshift0 = (is < 4) ? 0 : 4;
+        mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        mbidxshift1 = (is < 4) ? 0 : 2;
+
+        sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
+        mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
+        const FLOAT_TYPE d2 = dall * sc;
+        const FLOAT_TYPE m2 = dmin * mbyte;
+
+        [[unroll]] for (uint l = 0; l < n; ++l) {
+            data_b[y_idx + l     ] = D_TYPE(d1 * FLOAT_TYPE(data_a[ib].qs[qs_idx + l] & 0xF) - m1);
+            data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[ib].qs[qs_idx + l] >>  4) - m2);
+        }
+    }
+}
diff --git a/ggml/src/vulkan-shaders/dequant_q5_0.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_0.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/dequant_q5_0.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_0.comp
diff --git a/ggml/src/vulkan-shaders/dequant_q5_1.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_1.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/dequant_q5_1.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_1.comp
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp
new file mode 100644
index 00000000000..6db5403b661
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp
@@ -0,0 +1,70 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
+        const uint ib = gl_WorkGroupID.x * 256 + wgy;
+        if (ib >= p.M * p.K / QUANT_K) {
+            return;
+        }
+
+        const uint tid = gl_LocalInvocationID.x;
+        const uint il = tid / 16;
+        const uint ir = tid % 16;
+        const uint is = 2 * il;
+
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib].d.y);
+
+        const uint y_idx = ib * QUANT_K + 64 * il + 2 * ir;
+        const uint qs_idx = 32*il + 2 * ir;
+        const uint qh_idx = 2 * ir;
+
+        uint scidx0 = (is < 4) ? is : (is + 4);
+        uint scidx1 = (is < 4) ? is : (is - 4);
+        uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        uint scidxshift1 = (is < 4) ? 0 : 2;
+        uint mbidx0 = is + 4;
+        uint mbidx1 = (is < 4) ? is + 4 : is;
+        uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+        uint mbidxshift0 = (is < 4) ? 0 : 4;
+        uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        uint mbidxshift1 = (is < 4) ? 0 : 2;
+
+        uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
+        uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
+        const FLOAT_TYPE d1 = dall * sc;
+        const FLOAT_TYPE m1 = dmin * mbyte;
+
+        scidx0 = (is < 4) ? is + 1 : (is + 5);
+        scidx1 = (is < 4) ? is + 1 : (is - 3);
+        scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        scidxshift1 = (is < 4) ? 0 : 2;
+        mbidx0 = is + 5;
+        mbidx1 = (is < 4) ? is + 5 : is + 1;
+        mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+        mbidxshift0 = (is < 4) ? 0 : 4;
+        mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+        mbidxshift1 = (is < 4) ? 0 : 2;
+
+        sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
+        mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
+        const FLOAT_TYPE d2 = dall * sc;
+        const FLOAT_TYPE m2 = dmin * mbyte;
+
+        const uint8_t hm1 = uint8_t(1 << (2 * il    ));
+        const uint8_t hm2 = uint8_t(1 << (2 * il + 1));
+        data_b[y_idx     ] = D_TYPE(d1 * FLOAT_TYPE((data_a[ib].qs[qs_idx    ] & 0xF) + (((data_a[ib].qh[qh_idx    ] & hm1) != 0) ? 16 : 0)) - m1);
+        data_b[y_idx +  1] = D_TYPE(d1 * FLOAT_TYPE((data_a[ib].qs[qs_idx + 1] & 0xF) + (((data_a[ib].qh[qh_idx + 1] & hm1) != 0) ? 16 : 0)) - m1);
+        data_b[y_idx + 32] = D_TYPE(d2 * FLOAT_TYPE((data_a[ib].qs[qs_idx    ]  >> 4) + (((data_a[ib].qh[qh_idx    ] & hm2) != 0) ? 16 : 0)) - m2);
+        data_b[y_idx + 33] = D_TYPE(d2 * FLOAT_TYPE((data_a[ib].qs[qs_idx + 1]  >> 4) + (((data_a[ib].qh[qh_idx + 1] & hm2) != 0) ? 16 : 0)) - m2);
+    }
+}
diff --git a/ggml/src/vulkan-shaders/dequant_q6_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/dequant_q6_k.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp
diff --git a/ggml/src/vulkan-shaders/dequant_q8_0.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q8_0.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/dequant_q8_0.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/dequant_q8_0.comp
diff --git a/ggml/src/vulkan-shaders/diag_mask_inf.comp b/ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/diag_mask_inf.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/diag_mask_inf.comp
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/div.comp b/ggml/src/ggml-vulkan/vulkan-shaders/div.comp
new file mode 100644
index 00000000000..e581905b3f0
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/div.comp
@@ -0,0 +1,27 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+const uint num_threads = 256;
+
+layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    uint idx = get_idx();
+
+    // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
+    const uint num_iter = 2;
+
+    [[unroll]] for (uint i = 0; i < num_iter; ++i) {
+        if (idx >= p.ne) {
+            continue;
+        }
+        uint i00, i01, i02, i03;
+        get_indices(idx, i00, i01, i02, i03);
+
+        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) / FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
+
+        idx += num_threads;
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
new file mode 100644
index 00000000000..c5be8131b30
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
@@ -0,0 +1,289 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : require
+
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
+
+#extension GL_KHR_memory_scope_semantics : enable
+#extension GL_KHR_cooperative_matrix : enable
+#extension GL_NV_cooperative_matrix2 : enable
+#extension GL_EXT_buffer_reference : enable
+#extension GL_KHR_shader_subgroup_ballot : enable
+#extension GL_KHR_shader_subgroup_vote : enable
+#extension GL_EXT_null_initializer : enable
+
+#include "types.comp"
+#include "dequant_funcs_cm2.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (constant_id = 1) const uint32_t Br = 32;
+layout (constant_id = 2) const uint32_t Bc = 32;
+layout (constant_id = 3) const uint32_t D = 32;
+layout (constant_id = 4) const uint32_t Clamp = gl_CooperativeMatrixClampModeConstantNV;
+
+layout (push_constant) uniform parameter {
+    uint32_t N;
+    uint32_t KV;
+
+    uint32_t ne1;
+    uint32_t ne2;
+    uint32_t ne3;
+
+    uint32_t neq2;
+    uint32_t neq3;
+    uint32_t nek2;
+    uint32_t nek3;
+    uint32_t nev2;
+    uint32_t nev3;
+    uint32_t nem1;
+
+    uint32_t nb02;
+    uint32_t nb03;
+    uint32_t nb12;
+    uint32_t nb13;
+    uint32_t nb22;
+    uint32_t nb23;
+    uint32_t nb31;
+
+    float scale;
+    float max_bias;
+    float logit_softcap;
+
+    uint32_t mask;
+    uint32_t n_head_log2;
+    float m0;
+    float m1;
+} p;
+
+layout (binding = 0) readonly buffer Q {uint8_t data_q[];};
+layout (binding = 1) readonly buffer K {uint8_t data_k[];};
+layout (binding = 2) readonly buffer V {uint8_t data_v[];};
+layout (binding = 3) readonly buffer M {uint8_t data_m[];};
+layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
+
+#define CEIL_DIV(a, b) (((a) + (b) - 1) / (b))
+
+ACC_TYPE maxReduce(const in ACC_TYPE x, const in ACC_TYPE y) {
+    return max(x, y);
+}
+
+ACC_TYPE smearReduce(const in ACC_TYPE x, const in ACC_TYPE y) {
+    return x;
+}
+
+// Replace matrix elements >= numRows or numCols with 'replace'
+ACC_TYPE replacePadding(const in uint32_t row, const in uint32_t col, const in ACC_TYPE elem, const in ACC_TYPE replace, const in uint32_t numRows, const in uint32_t numCols) {
+    if (row >= numRows || col >= numCols) {
+        return replace;
+    }
+    return elem;
+}
+
+ACC_TYPE Exp(const in uint32_t row, const in uint32_t col, const in ACC_TYPE elem)
+{
+    return exp(elem);
+}
+
+ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE elem0, const in ACC_TYPE elem1)
+{
+    return max(elem0, elem1);
+}
+
+#if defined(BLOCK_SIZE)
+#define DECODEFUNC , DEQUANTFUNC
+#else
+#define DECODEFUNC
+#endif
+
+void main() {
+#if defined(DATA_A_IQ4_NL)
+    init_iq4nl_shmem();
+#endif
+
+    const uint32_t N = p.N;
+    const uint32_t KV = p.KV;
+
+    const uint32_t Tr = CEIL_DIV(N, Br);
+    const uint32_t Tc = CEIL_DIV(KV, Bc);
+
+    const uint32_t i = gl_WorkGroupID.x;
+
+    const uint32_t iq2 = gl_WorkGroupID.y;
+    const uint32_t iq3 = gl_WorkGroupID.z;
+
+    // broadcast factors
+    const uint32_t rk2 = p.neq2/p.nek2;
+    const uint32_t rk3 = p.neq3/p.nek3;
+
+    const uint32_t rv2 = p.neq2/p.nev2;
+    const uint32_t rv3 = p.neq3/p.nev3;
+
+    // k indices
+    const uint32_t ik3 = iq3 / rk3;
+    const uint32_t ik2 = iq2 / rk2;
+
+    // v indices
+    const uint32_t iv3 = iq3 / rv3;
+    const uint32_t iv2 = iq2 / rv2;
+
+    tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutQ = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
+    tensorLayoutNV<2, Clamp> tensorLayoutK = createTensorLayoutNV(2, Clamp);
+    tensorLayoutNV<2, Clamp> tensorLayoutV = createTensorLayoutNV(2, Clamp);
+
+    tensorViewNV<2, false, 1, 0> tensorViewTranspose = createTensorViewNV(2, false, 1, 0);
+
+#if defined(BLOCK_SIZE)
+    tensorLayoutK = setTensorLayoutBlockSizeNV(tensorLayoutK, 1, BLOCK_SIZE);
+    tensorLayoutV = setTensorLayoutBlockSizeNV(tensorLayoutV, 1, BLOCK_SIZE);
+#endif
+
+    tensorLayoutQ = setTensorLayoutDimensionNV(tensorLayoutQ, N, D);
+    tensorLayoutK = setTensorLayoutDimensionNV(tensorLayoutK, KV, D);
+    tensorLayoutV = setTensorLayoutDimensionNV(tensorLayoutV, KV, D);
+
+    coopmat<Q_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA> Q;
+    coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA> Qf16;
+
+    uint32_t q_offset = iq2*p.nb02+iq3*p.nb03;
+    coopMatLoadTensorNV(Q, data_q, q_offset, sliceTensorLayoutNV(tensorLayoutQ, i * Br, Br, 0, D));
+
+    Qf16 = coopmat<float16_t, gl_ScopeWorkgroup, Br, D, gl_MatrixUseA>(Q);
+    Qf16 *= float16_t(p.scale);
+
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(0);
+
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> L, M;
+
+    L = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
+    M = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(-1.0/0.0);
+
+    ACC_TYPE slope = ACC_TYPE(1.0);
+
+    // ALiBi
+    if (p.max_bias > 0.0f) {
+        const uint32_t h = iq2;
+
+        const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
+        const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
+
+        slope = pow(base, ACC_TYPE(exph));
+    }
+
+    [[dont_unroll]]
+    for (uint32_t j = 0; j < Tc; ++j) {
+
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> S = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0);
+
+        coopmat<float16_t, gl_ScopeWorkgroup, D, Bc, gl_MatrixUseB> K_T;
+
+        uint32_t k_offset = ik2*p.nb12 + ik3*p.nb13;
+        coopMatLoadTensorNV(K_T, data_k, k_offset, sliceTensorLayoutNV(tensorLayoutK, j * Bc, Bc, 0, D), tensorViewTranspose DECODEFUNC);
+        S = coopMatMulAdd(Qf16, K_T, S);
+
+        if (p.logit_softcap != 0.0f) {
+            [[unroll]]
+            for (int k = 0; k < S.length(); ++k) {
+                S[k] = ACC_TYPE(p.logit_softcap)*tanh(S[k]);
+            }
+        }
+
+        if (p.mask != 0) {
+            tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutM = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
+            tensorLayoutM = setTensorLayoutDimensionNV(tensorLayoutM, p.nem1, KV);
+
+            coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> mv;
+
+            coopMatLoadTensorNV(mv, data_m, 0, sliceTensorLayoutNV(tensorLayoutM, i * Br, Br, j * Bc, Bc));
+
+            S += slope*coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(mv);
+        }
+
+        // Clear padding elements to -inf, so they don't contribute to rowmax
+        if (Clamp != 0 &&
+            ((j + 1) * Bc > KV ||
+             (i + 1) * Br > N)) {
+
+            uint R = ((i + 1) * Br >  N) ?  (N % Br) : Br;
+            uint C = ((j + 1) * Bc > KV) ? (KV % Bc) : Bc;
+
+            coopMatPerElementNV(S, S, replacePadding, ACC_TYPE(-1.0/0.0), R, C);
+        }
+
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> rowmax, P, rowsum, eM;
+
+        coopMatReduceNV(rowmax, S, gl_CooperativeMatrixReduceRowNV, maxReduce);
+
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator> Mold = M;
+
+        // M = max(rowmax, Mold)
+        // P = e^(S - M)
+        // eM = e^(Mold - M)
+        coopMatPerElementNV(M, rowmax, Max, Mold);
+        coopMatPerElementNV(P, S - M, Exp);
+        coopMatPerElementNV(eM, Mold - M, Exp);
+
+        // Clear padding elements to 0, so they don't contribute to rowsum
+        if (Clamp != 0 &&
+            ((j + 1) * Bc > KV ||
+             (i + 1) * Br > N)) {
+
+            uint R = ((i + 1) * Br >  N) ?  (N % Br) : Br;
+            uint C = ((j + 1) * Bc > KV) ? (KV % Bc) : Bc;
+
+            coopMatPerElementNV(P, P, replacePadding, ACC_TYPE(0.0), R, C);
+        }
+
+        coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseA> P_A = coopmat<float16_t, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseA>(P);
+
+        // compute rowsum by multiplying by matrix of all ones.
+        coopmat<float16_t, gl_ScopeWorkgroup, Bc, Bc, gl_MatrixUseB> One = coopmat<float16_t, gl_ScopeWorkgroup, Bc, Bc, gl_MatrixUseB>(1.0);
+
+        rowsum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, Bc, gl_MatrixUseAccumulator>(0.0);
+        rowsum = coopMatMulAdd(P_A, One, rowsum);
+
+        coopmat<float16_t, gl_ScopeWorkgroup, Bc, D, gl_MatrixUseB> V;
+        uint32_t v_offset = iv2*p.nb22 + iv3*p.nb23;
+        coopMatLoadTensorNV(V,  data_v, v_offset, sliceTensorLayoutNV(tensorLayoutV, j * Bc, Bc, 0, D) DECODEFUNC);
+
+        L = eM*L + rowsum;
+
+        // This is the "diagonal" matrix in the paper, but since we do componentwise
+        // multiply rather than matrix multiply it has the diagonal element smeared
+        // across the row
+        coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> eMdiag;
+
+        // resize eM by using smear/reduce
+        coopMatReduceNV(eMdiag, eM, gl_CooperativeMatrixReduceRowNV, smearReduce);
+
+        O = eMdiag * O;
+
+        O = coopMatMulAdd(P_A, V, O);
+    }
+
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> Ldiag;
+
+    // resize L by using smear/reduce
+    coopMatReduceNV(Ldiag, L, gl_CooperativeMatrixReduceRowNV, smearReduce);
+
+    [[unroll]]
+    for (int k = 0; k < Ldiag.length(); ++k) {
+        Ldiag[k] = ACC_TYPE(1.0) / Ldiag[k];
+    }
+
+    O = Ldiag*O;
+
+    tensorLayoutNV<3, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(3, gl_CooperativeMatrixClampModeConstantNV);
+    tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.ne2, p.ne1, D);
+
+    // permute dimensions
+    tensorViewNV<3, false, 1, 0, 2> tensorViewPermute = createTensorViewNV(3, false, 1, 0, 2);
+    uint32_t o_offset = iq3*p.ne2*p.ne1;
+
+    coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, D, gl_MatrixUseAccumulator>(O);
+    coopMatStoreTensorNV(O_D, data_o, o_offset, sliceTensorLayoutNV(tensorLayoutD, i * Br, Br, iq2, 1, 0, D), tensorViewPermute);
+}
diff --git a/ggml/src/vulkan-shaders/gelu.comp b/ggml/src/ggml-vulkan/vulkan-shaders/gelu.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/gelu.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/gelu.comp
diff --git a/ggml/src/vulkan-shaders/gelu_quick.comp b/ggml/src/ggml-vulkan/vulkan-shaders/gelu_quick.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/gelu_quick.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/gelu_quick.comp
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp b/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp
new file mode 100644
index 00000000000..a6555fa2706
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp
@@ -0,0 +1,60 @@
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_control_flow_attributes : require
+
+layout (push_constant) uniform parameter
+{
+    uint ne;
+    uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
+    uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
+    uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
+    uint d_offset;
+    float param1; float param2; int param3;
+} p;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+// true if src0/src1 are the same shape and the indices can be reused without additional modulus
+layout(constant_id = 0) const bool norepeat = false;
+
+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
+// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
+uint fastmod(uint a, uint b) {
+    if ((b & (b-1)) == 0) {
+        return a & (b-1);
+    }
+    return a % b;
+}
+
+uint fastdiv(uint a, uint b) {
+    return (a < b) ? 0 : (a / b);
+}
+
+void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03) {
+    i03 = fastdiv(idx, (p.ne02*p.ne01*p.ne00));
+    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
+    i02 = fastdiv((idx - i03_offset), (p.ne01*p.ne00));
+    const uint i02_offset = i02*p.ne01*p.ne00;
+    i01 = (idx - i03_offset - i02_offset) / p.ne00;
+    i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+}
+
+uint src0_idx(uint i00, uint i01, uint i02, uint i03) {
+    return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00;
+}
+
+uint src1_idx(uint i00, uint i01, uint i02, uint i03) {
+    if (norepeat) {
+        return i03*p.nb13 + i02*p.nb12 + i01*p.nb11 + i00*p.nb10;
+    } else {
+        return fastmod(i03, p.ne13)*p.nb13 + fastmod(i02, p.ne12)*p.nb12 + fastmod(i01, p.ne11)*p.nb11 + fastmod(i00, p.ne10)*p.nb10;
+    }
+}
+
+uint dst_idx(uint i00, uint i01, uint i02, uint i03) {
+    return i03*p.nb23 + i02*p.nb22 + i01*p.nb21 + i00*p.nb20;
+}
diff --git a/ggml/src/vulkan-shaders/generic_head.comp b/ggml/src/ggml-vulkan/vulkan-shaders/generic_head.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/generic_head.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/generic_head.comp
diff --git a/ggml/src/vulkan-shaders/generic_unary_head.comp b/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.comp
similarity index 58%
rename from ggml/src/vulkan-shaders/generic_unary_head.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.comp
index 4e1fa3af3ad..ab7c9d7eb1b 100644
--- a/ggml/src/vulkan-shaders/generic_unary_head.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/generic_unary_head.comp
@@ -8,6 +8,13 @@ layout (push_constant) uniform parameter
     uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
     uint d_offset;
     float param1; float param2;
+
+    uint ne0_012mp; uint ne0_012L;
+    uint ne0_01mp;  uint ne0_01L;
+    uint ne0_0mp;   uint ne0_0L;
+    uint ne1_012mp; uint ne1_012L;
+    uint ne1_01mp;  uint ne1_01L;
+    uint ne1_0mp;   uint ne1_0L;
 } p;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
@@ -17,22 +24,30 @@ uint get_idx() {
     return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
 }
 
+// see init_fastdiv_values in ggml-vulkan.cpp
+uint fastdiv(uint n, uint mp, uint L) {
+    uint msbs, lsbs;
+    // msbs = mulhi(n, mp)
+    umulExtended(n, mp, msbs, lsbs);
+    return (msbs + n) >> L;
+}
+
 uint src0_idx(uint idx) {
-    const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
+    const uint i03 = fastdiv(idx, p.ne0_012mp, p.ne0_012L);
     const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
-    const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
+    const uint i02 = fastdiv(idx - i03_offset, p.ne0_01mp, p.ne0_01L);
     const uint i02_offset = i02*p.ne01*p.ne00;
-    const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
+    const uint i01 = fastdiv(idx - i03_offset - i02_offset, p.ne0_0mp, p.ne0_0L);
     const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
     return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00;
 }
 
 uint dst_idx(uint idx) {
-    const uint i13 = idx / (p.ne12*p.ne11*p.ne10);
+    const uint i13 = fastdiv(idx, p.ne1_012mp, p.ne1_012L);
     const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
-    const uint i12 = (idx - i13_offset) / (p.ne11*p.ne10);
+    const uint i12 = fastdiv(idx - i13_offset, p.ne1_01mp, p.ne1_01L);
     const uint i12_offset = i12*p.ne11*p.ne10;
-    const uint i11 = (idx - i13_offset - i12_offset) / p.ne10;
+    const uint i11 = fastdiv(idx - i13_offset - i12_offset, p.ne1_0mp, p.ne1_0L);
     const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
     return i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + i10*p.nb10;
 }
diff --git a/ggml/src/vulkan-shaders/get_rows.comp b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp
similarity index 91%
rename from ggml/src/vulkan-shaders/get_rows.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp
index e9ff22efa9c..a7b81e52ce5 100644
--- a/ggml/src/vulkan-shaders/get_rows.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows.comp
@@ -3,6 +3,8 @@
 #include "types.comp"
 #include "generic_binary_head.comp"
 
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
 void main() {
     const uint i00 = gl_GlobalInvocationID.x;
     const uint i10 = gl_GlobalInvocationID.y;
diff --git a/ggml/src/vulkan-shaders/get_rows_quant.comp b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp
similarity index 83%
rename from ggml/src/vulkan-shaders/get_rows_quant.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp
index 53a9a96f236..1426fde6597 100644
--- a/ggml/src/vulkan-shaders/get_rows_quant.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp
@@ -4,12 +4,18 @@
 #include "generic_binary_head.comp"
 #include "dequant_funcs.comp"
 
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
 void main() {
     const uint i00 = (gl_GlobalInvocationID.x)*2;
     const uint i10 = gl_GlobalInvocationID.y;
     const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
     const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
 
+#if defined(DATA_A_IQ4_NL)
+    init_iq4nl_shmem();
+#endif
+
     if (i00 >= p.ne00) {
         return;
     }
@@ -25,6 +31,8 @@ void main() {
     const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
 
     vec2 v = dequantize(ib, iqs, 0);
+    const vec2 dm = get_dm(ib, 0);
+    v = v * dm.x + dm.y;
 
     data_d[d_offset + iybs + iqs           ] = D_TYPE(v.x);
     data_d[d_offset + iybs + iqs + y_offset] = D_TYPE(v.y);
diff --git a/ggml/src/vulkan-shaders/group_norm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp
similarity index 96%
rename from ggml/src/vulkan-shaders/group_norm.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp
index 5ad9b28daff..b6a0d564549 100644
--- a/ggml/src/vulkan-shaders/group_norm.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/group_norm.comp
@@ -19,7 +19,7 @@ void main() {
 
     const uint tid = gl_LocalInvocationID.x;
     const uint start = gl_WorkGroupID.x * group_size + tid;
-    const uint end = start + group_size;
+    const uint end = (gl_WorkGroupID.x + 1) * group_size;
 
     tmp[tid] = 0.0f;
 
diff --git a/ggml/src/vulkan-shaders/im2col.comp b/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp
similarity index 91%
rename from ggml/src/vulkan-shaders/im2col.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp
index 4d48610a3ad..966fedf8fa7 100644
--- a/ggml/src/vulkan-shaders/im2col.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/im2col.comp
@@ -1,6 +1,11 @@
 #version 450
 
 #extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_spirv_intrinsics: enable
+
+#if RTE16
+spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
+#endif
 
 layout (push_constant) uniform parameter
 {
diff --git a/ggml/src/vulkan-shaders/leaky_relu.comp b/ggml/src/ggml-vulkan/vulkan-shaders/leaky_relu.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/leaky_relu.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/leaky_relu.comp
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul.comp
new file mode 100644
index 00000000000..5ce57cbcfc7
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul.comp
@@ -0,0 +1,27 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+const uint num_threads = 256;
+
+layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    uint idx = get_idx();
+
+    // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
+    const uint num_iter = 2;
+
+    [[unroll]] for (uint i = 0; i < num_iter; ++i) {
+        if (idx >= p.ne) {
+            continue;
+        }
+        uint i00, i01, i02, i03;
+        get_indices(idx, i00, i01, i02, i03);
+
+        data_d[p.d_offset + dst_idx(i00, i01, i02, i03)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(i00, i01, i02, i03)]) * FLOAT_TYPE(data_b[src1_idx(i00, i01, i02, i03)]));
+
+        idx += num_threads;
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp
new file mode 100644
index 00000000000..4c64fd47af7
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_split_k_reduce.comp
@@ -0,0 +1,48 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {float data_a[];};
+layout (binding = 0) readonly buffer A4 {vec4 data_a4[];};
+layout (binding = 1) writeonly buffer D {float data_d[];};
+layout (binding = 1) writeonly buffer D4 {vec4 data_d4[];};
+
+layout (push_constant) uniform parameter {
+    uint ne;
+    uint k_num;
+} p;
+
+void main() {
+    // Each invocation handles four consecutive components
+    const uint idx = gl_GlobalInvocationID.x * 4;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    // Check if all four components are in bounds and aligned,
+    // then use vector loads
+    if (idx + 3 < p.ne && (p.ne % 4) == 0) {
+        vec4 result = vec4(0.0f);
+
+        [[unroll]] for (uint i = 0; i < p.k_num; i++) {
+            result += data_a4[(i * p.ne + idx) / 4];
+        }
+
+        data_d4[idx / 4] = result;
+    } else {
+        [[unroll]] for (uint j = 0; j < 4; ++j) {
+            if (idx + j < p.ne) {
+                float result = 0.0f;
+
+                [[unroll]] for (uint i = 0; i < p.k_num; i++) {
+                    result += data_a[i * p.ne + idx + j];
+                }
+
+                data_d[idx + j] = result;
+            }
+        }
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
new file mode 100644
index 00000000000..187c31916d1
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
@@ -0,0 +1,170 @@
+#version 450
+
+#ifdef FLOAT16
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#endif
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+layout (constant_id = 1) const uint NUM_ROWS = 1;
+
+#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
+#define K_PER_ITER 8
+#else
+#define K_PER_ITER 2
+#endif
+
+
+uint a_offset, b_offset, d_offset, y_offset;
+
+shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];
+
+void iter(inout FLOAT_TYPE temp[NUM_ROWS], const uint first_row, const uint num_rows, const uint tid, const uint i, bool lastiter)
+{
+    const uint col = i*BLOCK_SIZE + K_PER_ITER*tid;
+    const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
+    const uint iybs = col - col%QUANT_K; // y block start index
+
+#if K_PER_ITER == 8
+#if QUANT_R == 2
+    const B_TYPE_VEC4 bv02 = data_b_v4[(b_offset + iybs + iqs) / 4];
+    const B_TYPE_VEC4 bv13 = data_b_v4[(b_offset + iybs + iqs + y_offset) / 4];
+    const vec4 bv0 = vec4(bv02.x, bv13.x, bv02.y, bv13.y);
+    const vec4 bv1 = vec4(bv02.z, bv13.z, bv02.w, bv13.w);
+#else
+    const vec4 bv0 = vec4(data_b_v4[(b_offset + iybs + iqs) / 4]);
+    const vec4 bv1 = vec4(data_b_v4[(b_offset + iybs + iqs) / 4 + 1]);
+#endif
+#else
+    // Check if the second of the pair of elements is OOB, and don't fetch B or
+    // accumulate it. We still fetch a pair of elements for A, which is fine for
+    // quantized formats since they'll be within the same block. We should
+    // probably skip fetching the second element for F16/F32, but as of now we
+    // still do.
+    const bool OOB = lastiter && (iybs + iqs + y_offset >= p.ncols);
+
+    FLOAT_TYPE b0 = 0, b1 = 0;
+    b0 = FLOAT_TYPE(data_b[b_offset + iybs + iqs]);
+    if (!OOB) {
+        b1 = FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]);
+    }
+#endif
+    uint ibi = first_row*p.ncols;
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        const uint ib = (ibi + col)/QUANT_K; // block index
+        ibi += p.ncols;
+
+#if K_PER_ITER == 8
+        vec4 v = dequantize4(ib, iqs, a_offset);
+        vec4 v2 = dequantize4(ib, iqs+(4/QUANT_R), a_offset);
+
+        const vec2 dm = get_dm(ib, a_offset);
+        if (dm.y != 0) { // quant has min component
+            v = v * dm.x + dm.y;
+            v2 = v2 * dm.x + dm.y;
+        }
+
+        // matrix multiplication
+        FLOAT_TYPE rowtmp = dot(bv0, v);
+        rowtmp += dot(bv1, v2);
+
+        if (dm.y == 0)
+            rowtmp *= dm.x;
+
+        temp[n] += rowtmp;
+#else
+        const vec2 v = dequantize(ib, iqs, a_offset);
+
+        // matrix multiplication
+        temp[n] = fma(FLOAT_TYPE(v.x), b0, temp[n]);
+        if (!OOB) {
+            temp[n] = fma(FLOAT_TYPE(v.y), b1, temp[n]);
+        }
+#endif
+    }
+}
+
+void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
+    const uint tid = gl_LocalInvocationID.x;
+
+    get_offsets(a_offset, b_offset, d_offset);
+    a_offset /= QUANT_K;
+
+    y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
+
+    FLOAT_TYPE temp[NUM_ROWS];
+
+    for (uint i = 0; i < NUM_ROWS; ++i) {
+        temp[i] = FLOAT_TYPE(0);
+    }
+
+    uint num_iters = p.ncols / (K_PER_ITER * BLOCK_SIZE);
+    if (num_iters * K_PER_ITER * BLOCK_SIZE + K_PER_ITER*tid < p.ncols) {
+        num_iters++;
+    }
+    int unroll_count = 4;
+    uint unrolled_iters = num_iters & ~(unroll_count - 1);
+
+    uint i = 0;
+    while (i < unrolled_iters) {
+        // Manually partially unroll the loop
+        [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
+            iter(temp, first_row, num_rows, tid, i*K_PER_ITER, false);
+            i++;
+        }
+    }
+    unroll_count = 2;
+    unrolled_iters = num_iters & ~(unroll_count - 1);
+    while (i < unrolled_iters) {
+        // Manually partially unroll the loop
+        [[unroll]] for (uint k = 0; k < unroll_count; ++k) {
+            iter(temp, first_row, num_rows, tid, i*K_PER_ITER, false);
+            i++;
+        }
+    }
+    while (i < num_iters) {
+        iter(temp, first_row, num_rows, tid, i*K_PER_ITER, true);
+        i++;
+    }
+
+    // sum up partial sums and write back result
+    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+        tmpsh[n][tid] = temp[n];
+    }
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
+        if (tid < s) {
+            [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+                tmpsh[n][tid] += tmpsh[n][tid + s];
+            }
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        [[unroll]] for (uint n = 0; n < num_rows; ++n) {
+            data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
+        }
+    }
+}
+
+void main() {
+    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
+
+#if defined(DATA_A_IQ4_NL)
+    init_iq4nl_shmem();
+#endif
+
+    // do NUM_ROWS at a time, unless there aren't enough remaining rows
+    if (first_row + NUM_ROWS <= p.stride_d) {
+        compute_outputs(first_row, NUM_ROWS);
+    } else {
+        if (first_row >= p.stride_d) {
+            return;
+        }
+        compute_outputs(first_row, p.stride_d - first_row);
+    }
+}
diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_base.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
similarity index 78%
rename from ggml/src/vulkan-shaders/mul_mat_vec_base.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
index 5920bc93641..3894fca8260 100644
--- a/ggml/src/vulkan-shaders/mul_mat_vec_base.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.comp
@@ -2,8 +2,6 @@
 #extension GL_EXT_shader_16bit_storage : require
 #extension GL_EXT_shader_8bit_storage : require
 
-#define K_QUANTS_PER_ITERATION 2
-
 #ifdef MUL_MAT_ID
 #define EXPERT_COUNT 8
 #endif
@@ -12,6 +10,9 @@
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 1) readonly buffer BV2 {B_TYPE_VEC2 data_b_v2[];};
+layout (binding = 1) readonly buffer BV4 {B_TYPE_VEC4 data_b_v4[];};
+
 layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
 #ifdef MUL_MAT_ID
 layout (binding = 3) readonly buffer IDS {int data_ids[];};
@@ -49,13 +50,16 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 #endif
 
 #ifndef MUL_MAT_ID
-    const uint i13 = batch_idx / p.ne12;
-    const uint i12 = batch_idx % p.ne12;
+    uint batch_idx_a = 0;
+    if (batch_idx != 0) {
+        const uint i13 = batch_idx / p.ne12;
+        const uint i12 = batch_idx % p.ne12;
 
-    const uint i03 = i13 / p.broadcast3;
-    const uint i02 = i12 / p.broadcast2;
+        const uint i03 = i13 / p.broadcast3;
+        const uint i02 = i12 / p.broadcast2;
 
-    const uint batch_idx_a = i03 * p.ne02 + i02;
+        batch_idx_a = i03 * p.ne02 + i02;
+    }
 #else
     const uint expert_id = data_ids[expert_idx];
 #endif
diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_nc.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/mul_mat_vec_nc.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp
diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_p021.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/mul_mat_vec_p021.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_p021.comp
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
new file mode 100644
index 00000000000..1a5350d99ea
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
@@ -0,0 +1,113 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+
+shared FLOAT_TYPE tmp[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
+
+    if (row >= p.stride_d) {
+        return;
+    }
+
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
+
+    // 16 threads are used to process each block
+    const uint it_size = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid%16;  // 0...16
+    const uint ix  = tid/16;
+
+    const uint step = 8;
+
+    const uint v_im = itid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = itid - step*v_im;                      // 0...15 or 0...7
+
+    const uint l0 = 2*v_in;                                  // 0...15
+    const uint q_offset = 32*v_im + l0;
+    const uint s_offset = 8*v_im;
+    const uint y_offset = 128*v_im + l0;
+
+    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
+        const uint y_idx = i * QUANT_K + y_offset;
+
+        f16vec2 d = data_a[ib0 + i].d;
+        const FLOAT_TYPE dall = d.x;
+        const FLOAT_TYPE dmin = d.y;
+
+        B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
+        B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
+        B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
+        B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
+        B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
+        B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
+        B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
+        B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
+
+        uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
+        uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];
+
+        uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F;
+        uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F;
+        uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F;
+        uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F;
+
+        uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32));
+        uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32));
+        uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32));
+        uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32));
+
+        uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0];
+        uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8];
+        uvec2 qs0 =  uvec2(unpack8(qs0_u16));
+        uvec2 qs16 = uvec2(unpack8(qs16_u16));
+
+        FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
+        FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
+        [[unroll]] for (int l = 0; l < 2; ++l) {
+            sum1 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 0) & 3),
+                   fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
+                   fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 2) & 3),
+                   fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
+                   fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l]  >> 4) & 3),
+                   fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
+                   fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l]  >> 6) & 3),
+                   fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
+            sum2 = fma(FLOAT_TYPE(b0[l]),   FLOAT_TYPE(s0_hi4[0]),
+                   fma(FLOAT_TYPE(b16[l]),  FLOAT_TYPE(s0_hi4[1]),
+                   fma(FLOAT_TYPE(b32[l]),  FLOAT_TYPE(s0_hi4[2]),
+                   fma(FLOAT_TYPE(b48[l]),  FLOAT_TYPE(s0_hi4[3]),
+                   fma(FLOAT_TYPE(b64[l]),  FLOAT_TYPE(s4_hi4[0]),
+                   fma(FLOAT_TYPE(b80[l]),  FLOAT_TYPE(s4_hi4[1]),
+                   fma(FLOAT_TYPE(b96[l]),  FLOAT_TYPE(s4_hi4[2]),
+                   fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
+        }
+        temp = fma(dall, sum1, fma(-dmin, sum2, temp));
+    }
+
+    tmp[gl_LocalInvocationID.x] = temp;
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
new file mode 100644
index 00000000000..b19c3811136
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
@@ -0,0 +1,100 @@
+#version 450
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+
+shared FLOAT_TYPE tmp[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
+
+    if (row >= p.stride_d) {
+        return;
+    }
+
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
+
+    // 16 threads are used to process each block
+    const uint it_size = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid%16;  // 0...16
+    const uint ix  = tid/16;
+
+    const uint step = 8;
+
+    const uint v_im = itid/step;                            // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = itid - step*v_im;                     // 0...15 or 0...7
+
+    const uint8_t m = uint8_t(1 << (4 * v_im));
+
+    const uint l0 = 2*v_in;                                // 0...15
+    const uint q_offset = 32*v_im + l0;
+    const uint y_offset = 128*v_im + l0;
+
+    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    const uint s_shift = 4 * v_im;
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
+        const uint y_idx = i * QUANT_K + y_offset;
+
+        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
+
+        B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
+        B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
+        B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
+        B_TYPE_VEC2 b48 = data_b_v2[(b_offset + y_idx) / 2 + 24];
+        B_TYPE_VEC2 b64 = data_b_v2[(b_offset + y_idx) / 2 + 32];
+        B_TYPE_VEC2 b80 = data_b_v2[(b_offset + y_idx) / 2 + 40];
+        B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
+        B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];
+
+        uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
+        uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
+        uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
+        uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
+        uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
+        uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
+        u8vec2 s0 = unpack8(s0_16);
+        u8vec2 s2 = unpack8(s2_16);
+        u8vec2 s4 = unpack8(s4_16);
+        u8vec2 s6 = unpack8(s6_16);
+        u8vec2 s8 = unpack8(s8_16);
+        u8vec2 s10 = unpack8(s10_16);
+
+        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+        [[unroll]] for (int l = 0; l < 2; ++l) {
+            sum = fma(FLOAT_TYPE(b0[l])   * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b32[l])  * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b64[l])  * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b96[l])  * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b16[l])  * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b48[l])  * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b80[l])  * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1]  >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
+        }
+        temp = fma(d, sum, temp);
+    }
+
+    tmp[gl_LocalInvocationID.x] = temp;
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp
new file mode 100644
index 00000000000..b86d28589c6
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp
@@ -0,0 +1,131 @@
+#version 450
+
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+
+shared FLOAT_TYPE tmp[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
+
+    if (row >= p.stride_d) {
+        return;
+    }
+
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
+
+    // 16 threads are used to process each block
+    const uint it_size = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid%16;  // 0...16
+    const uint ix  = tid/16;
+
+    const uint step = 4;
+
+    const uint il = itid/step;                               // 0...3
+    const uint ir = itid - step*il;                          // 0...7 or 0...3
+    const uint n =  4;
+
+    const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const uint v_in = il % 2;
+
+    const uint l0 = n * (2 * ir + v_in);            // 0...15
+    const uint q_offset = 32*v_im + l0;
+    const uint y_offset = 64*v_im + l0;
+
+    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
+        const uint y1_idx = i * QUANT_K + y_offset;
+        const uint y2_idx = y1_idx + 128;
+
+        f16vec2 d = data_a[ib0 + i].d;
+        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+
+        uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
+        uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
+        uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
+        uvec4 scale0 = uvec4(unpack8(scale0_u32));
+        uvec4 scale4 = uvec4(unpack8(scale4_u32));
+        uvec4 scale8 = uvec4(unpack8(scale8_u32));
+
+        const uint32_t sc0 = (  scale0.x       & 0x3f);
+        const uint32_t sc1 = (  scale0.y       & 0x3f);
+        const uint32_t sc2 = (  scale4.x       & 0x3f);
+        const uint32_t sc3 = (  scale4.y       & 0x3f);
+        const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
+        const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
+        const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
+        const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
+
+        uint32_t qs0_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4];
+        uint32_t qs64_u32 = data_a_packed32[ib0 + i].qs[q_offset / 4 + 16];
+
+        uint32_t qs0_u32_lo4 = qs0_u32 & 0x0F0F0F0F;
+        uint32_t qs0_u32_hi4 = (qs0_u32 >> 4) & 0x0F0F0F0F;
+        uint32_t qs64_u32_lo4 = qs64_u32 & 0x0F0F0F0F;
+        uint32_t qs64_u32_hi4 = (qs64_u32 >> 4) & 0x0F0F0F0F;
+
+        uvec4 qs0_lo4 = uvec4(unpack8(qs0_u32_lo4));
+        uvec4 qs64_lo4 = uvec4(unpack8(qs64_u32_lo4));
+        uvec4 qs0_hi4 = uvec4(unpack8(qs0_u32_hi4));
+        uvec4 qs64_hi4 = uvec4(unpack8(qs64_u32_hi4));
+
+        const uint32_t q4_0  = qs0_lo4.x;
+        const uint32_t q4_1  = qs0_lo4.y;
+        const uint32_t q4_2  = qs0_lo4.z;
+        const uint32_t q4_3  = qs0_lo4.w;
+        const uint32_t q4_4  = qs0_hi4.x;
+        const uint32_t q4_5  = qs0_hi4.y;
+        const uint32_t q4_6  = qs0_hi4.z;
+        const uint32_t q4_7  = qs0_hi4.w;
+        const uint32_t q4_8  = qs64_lo4.x;
+        const uint32_t q4_9  = qs64_lo4.y;
+        const uint32_t q4_10 = qs64_lo4.z;
+        const uint32_t q4_11 = qs64_lo4.w;
+        const uint32_t q4_12 = qs64_hi4.x;
+        const uint32_t q4_13 = qs64_hi4.y;
+        const uint32_t q4_14 = qs64_hi4.z;
+        const uint32_t q4_15 = qs64_hi4.w;
+
+        B_TYPE_VEC4 by10 =  data_b_v4[(b_offset + y1_idx) / 4];
+        B_TYPE_VEC4 by132 = data_b_v4[(b_offset + y1_idx) / 4 + 8];
+        B_TYPE_VEC4 by20 =  data_b_v4[(b_offset + y2_idx) / 4];
+        B_TYPE_VEC4 by232 = data_b_v4[(b_offset + y2_idx) / 4 + 8];
+
+        const FLOAT_TYPE sx = fma(FLOAT_TYPE(by10.x),      q4_0,  fma(FLOAT_TYPE(by10.y),  q4_1,  fma(FLOAT_TYPE(by10.z),  q4_2,  FLOAT_TYPE(by10.w) *  q4_3)));
+        const FLOAT_TYPE sy = fma(FLOAT_TYPE(by132.x),     q4_4,  fma(FLOAT_TYPE(by132.y), q4_5,  fma(FLOAT_TYPE(by132.z), q4_6,  FLOAT_TYPE(by132.w) * q4_7)));
+        const FLOAT_TYPE sz = fma(FLOAT_TYPE(by20.x),      q4_8,  fma(FLOAT_TYPE(by20.y),  q4_9,  fma(FLOAT_TYPE(by20.z),  q4_10, FLOAT_TYPE(by20.w) *  q4_11)));
+        const FLOAT_TYPE sw = fma(FLOAT_TYPE(by232.x),     q4_12, fma(FLOAT_TYPE(by232.y), q4_13, fma(FLOAT_TYPE(by232.z), q4_14, FLOAT_TYPE(by232.w) * q4_15)));
+        const FLOAT_TYPE smin =
+            fma(FLOAT_TYPE(by10.x), sc2, fma(FLOAT_TYPE(by132.x), sc3, fma(FLOAT_TYPE(by20.x), sc6, fma(FLOAT_TYPE(by232.x), sc7,
+            fma(FLOAT_TYPE(by10.y), sc2, fma(FLOAT_TYPE(by132.y), sc3, fma(FLOAT_TYPE(by20.y), sc6, fma(FLOAT_TYPE(by232.y), sc7,
+            fma(FLOAT_TYPE(by10.z), sc2, fma(FLOAT_TYPE(by132.z), sc3, fma(FLOAT_TYPE(by20.z), sc6, fma(FLOAT_TYPE(by232.z), sc7,
+            fma(FLOAT_TYPE(by10.w), sc2, fma(FLOAT_TYPE(by132.w), sc3, fma(FLOAT_TYPE(by20.w), sc6,     FLOAT_TYPE(by232.w) * sc7)))))))))))))));
+        temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp));
+    }
+
+    tmp[gl_LocalInvocationID.x] = temp;
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
new file mode 100644
index 00000000000..fd243cf9161
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp
@@ -0,0 +1,160 @@
+#version 450
+
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+
+shared FLOAT_TYPE tmp[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
+
+    if (row >= p.stride_d) {
+        return;
+    }
+
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
+
+    // 16 threads are used to process each block
+    const uint it_size = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid%16;  // 0...16
+    const uint ix  = tid/16;
+
+    const uint il = itid/4;                          // 0...3
+    const uint ir = itid - 4*il;                     // 0...7 or 0...3
+
+    const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const uint v_in = il % 2;
+
+    const uint l0 = 4*ir + 2*v_in;                   // 0...15
+    const uint q_offset = 32*v_im + l0;
+    const uint y_offset = 64*v_im + l0;
+
+    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
+        const uint y1_idx = i * QUANT_K + y_offset;
+        const uint y2_idx = y1_idx + 128;
+
+        f16vec2 d = data_a[ib0 + i].d;
+        const FLOAT_TYPE dall = FLOAT_TYPE(d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(d.y);
+
+        uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];
+        uint32_t scale4_u32 = data_a_packed16[ib0 + i].scales[v_im + 2];
+        uint32_t scale8_u32 = data_a_packed16[ib0 + i].scales[v_im + 4];
+        uvec4 scale0 = uvec4(unpack8(scale0_u32));
+        uvec4 scale4 = uvec4(unpack8(scale4_u32));
+        uvec4 scale8 = uvec4(unpack8(scale8_u32));
+
+        const uint32_t sc0 = (  scale0.x       & 0x3f);
+        const uint32_t sc1 = (  scale0.y       & 0x3f);
+        const uint32_t sc2 = (  scale4.x       & 0x3f);
+        const uint32_t sc3 = (  scale4.y       & 0x3f);
+        const uint32_t sc4 = (( scale8.x       & 0x0f) | ((scale0.x & 0xc0) >> 2));
+        const uint32_t sc5 = (( scale8.y       & 0x0f) | ((scale0.y & 0xc0) >> 2));
+        const uint32_t sc6 = (((scale8.x >> 4) & 0x0f) | ((scale4.x & 0xc0) >> 2));
+        const uint32_t sc7 = (((scale8.y >> 4) & 0x0f) | ((scale4.y & 0xc0) >> 2));
+
+        uint32_t qs0_16_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 8]) << 16);
+        uint32_t qs64_80_u32 = uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 32]) | (uint32_t(data_a_packed16[ib0 + i].qs[q_offset / 2 + 40]) << 16);
+
+        uint32_t qs0_16_u32_lo4 = qs0_16_u32 & 0x0F0F0F0F;
+        uint32_t qs0_16_u32_hi4 = (qs0_16_u32 >> 4) & 0x0F0F0F0F;
+        uint32_t qs64_80_u32_lo4 = qs64_80_u32 & 0x0F0F0F0F;
+        uint32_t qs64_80_u32_hi4 = (qs64_80_u32 >> 4) & 0x0F0F0F0F;
+
+        uint32_t qh = pack32(u16vec2(data_a_packed16[ib0 + i].qh[l0 / 2], data_a_packed16[ib0 + i].qh[l0 / 2 + 8]));
+
+        uint32_t qs0_16_lo4_offset16 = ((qh >> (2*v_im)) & 0x01010101) << 4;
+        uint32_t qs0_16_hi4_offset16 = ((qh >> (2*v_im)) & 0x02020202) << 3;
+        uint32_t qs64_80_lo4_offset16 = ((qh >> (2*v_im)) & 0x10101010) << 0;
+        uint32_t qs64_80_hi4_offset16 = ((qh >> (2*v_im)) & 0x20202020) >> 1;
+
+        qs0_16_u32_lo4 += qs0_16_lo4_offset16;
+        qs0_16_u32_hi4 += qs0_16_hi4_offset16;
+        qs64_80_u32_lo4 += qs64_80_lo4_offset16;
+        qs64_80_u32_hi4 += qs64_80_hi4_offset16;
+
+        uvec4 qs0_16_lo4 = uvec4(unpack8(qs0_16_u32_lo4));
+        uvec4 qs64_80_lo4 = uvec4(unpack8(qs64_80_u32_lo4));
+        uvec4 qs0_16_hi4 = uvec4(unpack8(qs0_16_u32_hi4));
+        uvec4 qs64_80_hi4 = uvec4(unpack8(qs64_80_u32_hi4));
+
+        const uint32_t q4_0  = qs0_16_lo4.x;
+        const uint32_t q4_1  = qs0_16_lo4.y;
+        const uint32_t q4_2  = qs0_16_lo4.z;
+        const uint32_t q4_3  = qs0_16_lo4.w;
+        const uint32_t q4_4  = qs0_16_hi4.x;
+        const uint32_t q4_5  = qs0_16_hi4.y;
+        const uint32_t q4_6  = qs0_16_hi4.z;
+        const uint32_t q4_7  = qs0_16_hi4.w;
+        const uint32_t q4_8  = qs64_80_lo4.x;
+        const uint32_t q4_9  = qs64_80_lo4.y;
+        const uint32_t q4_10 = qs64_80_lo4.z;
+        const uint32_t q4_11 = qs64_80_lo4.w;
+        const uint32_t q4_12 = qs64_80_hi4.x;
+        const uint32_t q4_13 = qs64_80_hi4.y;
+        const uint32_t q4_14 = qs64_80_hi4.z;
+        const uint32_t q4_15 = qs64_80_hi4.w;
+
+        B_TYPE_VEC2 by10 =  data_b_v2[(b_offset + y1_idx) / 2];
+        B_TYPE_VEC2 by116 = data_b_v2[(b_offset + y1_idx) / 2 + 8];
+        B_TYPE_VEC2 by132 = data_b_v2[(b_offset + y1_idx) / 2 + 16];
+        B_TYPE_VEC2 by148 = data_b_v2[(b_offset + y1_idx) / 2 + 24];
+        B_TYPE_VEC2 by20 =  data_b_v2[(b_offset + y2_idx) / 2];
+        B_TYPE_VEC2 by216 = data_b_v2[(b_offset + y2_idx) / 2 + 8];
+        B_TYPE_VEC2 by232 = data_b_v2[(b_offset + y2_idx) / 2 + 16];
+        B_TYPE_VEC2 by248 = data_b_v2[(b_offset + y2_idx) / 2 + 24];
+
+        const FLOAT_TYPE sx =
+          fma(FLOAT_TYPE(by10.x), q4_0,
+          fma(FLOAT_TYPE(by10.y), q4_1,
+          fma(FLOAT_TYPE(by116.x), q4_2,
+             FLOAT_TYPE(by116.y) * q4_3)));
+        const FLOAT_TYPE sy =
+          fma(FLOAT_TYPE(by132.x), q4_4,
+          fma(FLOAT_TYPE(by132.y), q4_5,
+          fma(FLOAT_TYPE(by148.x), q4_6,
+             FLOAT_TYPE(by148.y) * q4_7)));
+        const FLOAT_TYPE sz =
+          fma(FLOAT_TYPE(by20.x), q4_8,
+          fma(FLOAT_TYPE(by20.y), q4_9,
+          fma(FLOAT_TYPE(by216.x), q4_10,
+             FLOAT_TYPE(by216.y) * q4_11)));
+        const FLOAT_TYPE sw =
+          fma(FLOAT_TYPE(by232.x), q4_12,
+          fma(FLOAT_TYPE(by232.y), q4_13,
+          fma(FLOAT_TYPE(by248.x), q4_14,
+             FLOAT_TYPE(by248.y) * q4_15)));
+        const FLOAT_TYPE smin =
+          fma(FLOAT_TYPE(by10.x) + FLOAT_TYPE(by10.y) + FLOAT_TYPE(by116.x) + FLOAT_TYPE(by116.y), sc2,
+          fma(FLOAT_TYPE(by132.x) + FLOAT_TYPE(by132.y) + FLOAT_TYPE(by148.x) + FLOAT_TYPE(by148.y), sc3,
+          fma(FLOAT_TYPE(by20.x) + FLOAT_TYPE(by20.y) + FLOAT_TYPE(by216.x) + FLOAT_TYPE(by216.y), sc6,
+              (FLOAT_TYPE(by232.x) + FLOAT_TYPE(by232.y) + FLOAT_TYPE(by248.x) + FLOAT_TYPE(by248.y)) * sc7)));
+        temp = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, temp));
+    }
+
+    tmp[gl_LocalInvocationID.x] = temp;
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp
new file mode 100644
index 00000000000..760aff85499
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp
@@ -0,0 +1,110 @@
+#version 450
+
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+
+shared FLOAT_TYPE tmp[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
+
+    if (row >= p.stride_d) {
+        return;
+    }
+
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
+
+    // 16 threads are used to process each block
+    const uint it_size = gl_WorkGroupSize.x/16;
+    const uint tid = gl_LocalInvocationID.x;
+    const uint itid = tid%16;  // 0...16
+    const uint ix  = tid/16;
+
+    const uint step = 8;
+
+    const uint v_im = itid/step;                            // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = itid - step*v_im;                     // 0...15 or 0...7
+
+    const uint l0 = 4 * v_in;                               // 0, 4, 8, ..., 28
+    const uint is = v_in / 4;
+
+    const uint ql_offset = 64*v_im + l0;
+    const uint qh_offset = 32*v_im + l0;
+    const uint s_offset  =  8*v_im + is;
+    const uint y_offset = 128*v_im + l0;
+
+    FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
+        const uint y_idx   = i * QUANT_K + y_offset;
+
+        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
+
+        FLOAT_TYPE scales[4];
+        scales[0] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]);
+        scales[1] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]);
+        scales[2] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]);
+        scales[3] = FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]);
+
+        uint32_t ql0_u32 =  uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 1]) << 16);
+        uint32_t ql32_u32 = uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 16]) | (uint32_t(data_a_packed16[ib0 + i].ql[ql_offset / 2 + 17]) << 16);
+
+        uint32_t ql0_u32_lo4 = ql0_u32 & 0x0F0F0F0F;
+        uint32_t ql0_u32_hi4 = (ql0_u32 >> 4) & 0x0F0F0F0F;
+        uint32_t ql32_u32_lo4 = ql32_u32 & 0x0F0F0F0F;
+        uint32_t ql32_u32_hi4 = (ql32_u32 >> 4) & 0x0F0F0F0F;
+
+        uint32_t qh_u32 = uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2]) | (uint32_t(data_a_packed16[ib0 + i].qh[qh_offset / 2 + 1]) << 16);
+        uint32_t qh0_u32 = (qh_u32 & 0x03030303) << 4;
+        uint32_t qh2_u32 = (qh_u32 & 0x0C0C0C0C) << 2;
+        uint32_t qh4_u32 = (qh_u32 & 0x30303030) << 0;
+        uint32_t qh6_u32 = (qh_u32 & 0xC0C0C0C0) >> 2;
+
+        uint32_t q0_u32 = ql0_u32_lo4  | qh0_u32;
+        uint32_t q1_u32 = ql32_u32_lo4 | qh2_u32;
+        uint32_t q2_u32 = ql0_u32_hi4  | qh4_u32;
+        uint32_t q3_u32 = ql32_u32_hi4 | qh6_u32;
+
+        uvec4 q0 = uvec4(unpack8(q0_u32));
+        uvec4 q1 = uvec4(unpack8(q1_u32));
+        uvec4 q2 = uvec4(unpack8(q2_u32));
+        uvec4 q3 = uvec4(unpack8(q3_u32));
+
+        B_TYPE_VEC4 by0  = data_b_v4[(b_offset + y_idx) / 4];
+        B_TYPE_VEC4 by32 = data_b_v4[(b_offset + y_idx) / 4 + 8];
+        B_TYPE_VEC4 by64 = data_b_v4[(b_offset + y_idx) / 4 + 16];
+        B_TYPE_VEC4 by96 = data_b_v4[(b_offset + y_idx) / 4 + 24];
+
+        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+        [[unroll]] for (int l = 0; l < 4; ++l) {
+            sum = fma(FLOAT_TYPE(by0[l])  * scales[0], FLOAT_TYPE(int8_t(q0[l]) - 32),
+                  fma(FLOAT_TYPE(by32[l]) * scales[1], FLOAT_TYPE(int8_t(q1[l]) - 32),
+                  fma(FLOAT_TYPE(by64[l]) * scales[2], FLOAT_TYPE(int8_t(q2[l]) - 32),
+                  fma(FLOAT_TYPE(by96[l]) * scales[3], FLOAT_TYPE(int8_t(q3[l]) - 32), sum))));
+        }
+        temp += sum * d;
+    }
+
+    tmp[gl_LocalInvocationID.x] = temp;
+    // sum up partial sums and write back result
+
+    barrier();
+    [[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}
diff --git a/ggml/src/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
similarity index 66%
rename from ggml/src/vulkan-shaders/mul_mm.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
index fffdd18189d..48122cbef90 100644
--- a/ggml/src/vulkan-shaders/mul_mm.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
@@ -7,6 +7,12 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
 #endif
 
+#ifdef COOPMAT
+#extension GL_KHR_cooperative_matrix : enable
+#extension GL_KHR_memory_scope_semantics : enable
+#extension GL_KHR_shader_subgroup_basic : enable
+#endif
+
 #ifdef MUL_MAT_ID
 #extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
 #endif
@@ -57,6 +63,7 @@ layout (push_constant) uniform parameter
 #endif
 } p;
 
+layout (constant_id = 0) const uint BLOCK_SIZE = 64;
 layout (constant_id = 1) const uint BM = 64;
 layout (constant_id = 2) const uint BN = 64;
 layout (constant_id = 3) const uint BK = 16;  // Assumed to be 32 if working with a quant
@@ -65,16 +72,33 @@ layout (constant_id = 5) const uint WN = 32;
 layout (constant_id = 6) const uint WMITER = 2;
 layout (constant_id = 7) const uint TM = 4;
 layout (constant_id = 8) const uint TN = 2;
-layout (constant_id = 9) const uint WARP = 32;
+layout (constant_id = 9) const uint TK = 1;  // Only needed for coopmat
+layout (constant_id = 10) const uint WARP = 32;
+
+#ifdef COOPMAT
+#define SHMEM_STRIDE (BK + 8)
+#else
+#define SHMEM_STRIDE (BK + 1)
+#endif
 
-shared FLOAT_TYPE buf_a[BM * (BK+1)];
-shared FLOAT_TYPE buf_b[BN * (BK+1)];
+shared FLOAT_TYPE buf_a[BM * SHMEM_STRIDE];
+shared FLOAT_TYPE buf_b[BN * SHMEM_STRIDE];
 
 #ifdef MUL_MAT_ID
 shared u16vec2 row_ids[3072];
+#endif // MUL_MAT_ID
+
+#define NUM_WARPS (BLOCK_SIZE / WARP)
+
+#ifdef COOPMAT
+shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
 #endif
 
 void main() {
+#if defined(DATA_A_IQ4_NL)
+    init_iq4nl_shmem();
+#endif
+
 #ifdef MUL_MAT_ID
     const uint expert_idx = gl_GlobalInvocationID.z;
 #else
@@ -94,17 +118,32 @@ void main() {
     const uint ik = gl_WorkGroupID.x / blocks_m;
     const uint ic = gl_WorkGroupID.y;
 
-    const uint warp_i = gl_LocalInvocationID.x / WARP;
-    const uint warp_r = warp_i % (BM / WM);
-    const uint warp_c = warp_i / (BM / WM);
-
     const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER);
     const uint WSUBM = WM / WMITER;
     const uint WSUBN = WN / WNITER;
 
+#ifdef COOPMAT
+    const uint warp_i = gl_SubgroupID;
+
+    const uint tiw = gl_SubgroupInvocationID;
+
+    const uint cms_per_row = WM / TM;
+    const uint cms_per_col = WN / TN;
+
+    const uint storestride = WARP / TM;
+    const uint store_r = tiw % TM;
+    const uint store_c = tiw / TM;
+#else
+    const uint warp_i = gl_LocalInvocationID.x / WARP;
+
     const uint tiw = gl_LocalInvocationID.x % WARP;
+
     const uint tiwr = tiw % (WSUBM / TM);
     const uint tiwc = tiw / (WSUBM / TM);
+#endif
+
+    const uint warp_r = warp_i % (BM / WM);
+    const uint warp_c = warp_i / (BM / WM);
 
     const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A);
     const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A);
@@ -152,21 +191,31 @@ void main() {
     uint pos_b = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / LOAD_VEC_B;
 #endif
 
-    float sums[WMITER * TM * WNITER * TN];
+#ifdef COOPMAT
+    coopmat<float16_t, gl_ScopeSubgroup, TM, TK, gl_MatrixUseA> cache_a;
+    coopmat<float16_t, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
+    coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> sums[cms_per_row * cms_per_col];
+
+    [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) {
+        sums[i] = coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0.0f);
+    }
+#else
+    ACC_TYPE sums[WMITER * TM * WNITER * TN];
     FLOAT_TYPE cache_a[WMITER * TM];
     FLOAT_TYPE cache_b[WNITER * TN];
 
     [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
-        sums[i] = 0.0f;
+        sums[i] = ACC_TYPE(0.0f);
     }
+#endif
 
-    [[unroll]] for (uint block = start_k; block < end_k; block += BK) {
+    for (uint block = start_k; block < end_k; block += BK) {
         [[unroll]] for (uint l = 0; l < BM; l += loadstride_a) {
 
 #if defined(DATA_A_F32) || defined(DATA_A_F16)
 #if LOAD_VEC_A == 8
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
-            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
             buf_a[buf_idx    ] = FLOAT_TYPE(data_a[idx][0].x);
             buf_a[buf_idx + 1] = FLOAT_TYPE(data_a[idx][0].y);
             buf_a[buf_idx + 2] = FLOAT_TYPE(data_a[idx][0].z);
@@ -177,21 +226,21 @@ void main() {
             buf_a[buf_idx + 7] = FLOAT_TYPE(data_a[idx][1].w);
 #elif LOAD_VEC_A == 4
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
-            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
             buf_a[buf_idx    ] = FLOAT_TYPE(data_a[idx].x);
             buf_a[buf_idx + 1] = FLOAT_TYPE(data_a[idx].y);
             buf_a[buf_idx + 2] = FLOAT_TYPE(data_a[idx].z);
             buf_a[buf_idx + 3] = FLOAT_TYPE(data_a[idx].w);
 #else
             if (ir * BM + loadc_a + l < p.M && block + loadr_a < end_k) {
-                buf_a[(loadc_a + l) * (BK+1) + loadr_a] = FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]);
+                buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]);
             } else {
-                buf_a[(loadc_a + l) * (BK+1) + loadr_a] = FLOAT_TYPE(0.0f);
+                buf_a[(loadc_a + l) * SHMEM_STRIDE + loadr_a] = FLOAT_TYPE(0.0f);
             }
 #endif
 #elif defined(DATA_A_Q4_0)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
-            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a;
 
             const uint ib = idx / 16;
             const uint iqs = idx & 0xF;
@@ -204,7 +253,7 @@ void main() {
             buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
 #elif defined(DATA_A_Q4_1)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
-            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a;
 
             const uint ib = idx / 16;
             const uint iqs = idx & 0xF;
@@ -218,7 +267,7 @@ void main() {
             buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
 #elif defined(DATA_A_Q5_0)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
-            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a;
 
             const uint ib = idx / 16;
             const uint iqs = idx & 0xF;
@@ -233,7 +282,7 @@ void main() {
             buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
 #elif defined(DATA_A_Q5_1)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
-            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a;
 
             const uint ib = idx / 16;
             const uint iqs = idx & 0xF;
@@ -249,7 +298,7 @@ void main() {
             buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
 #elif defined(DATA_A_Q8_0)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
-            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
 
             const uint ib = idx / 16;
             const uint iqs = (idx & 0xF) * 2;
@@ -261,7 +310,7 @@ void main() {
             buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
 #elif defined(DATA_A_Q2_K)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
-            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
 
             const uint ib = idx / 128;                         // 2 values per idx
             const uint iqs = idx % 128;                        // 0..127
@@ -280,7 +329,7 @@ void main() {
             buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
 #elif defined(DATA_A_Q3_K)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
-            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
 
             const uint ib = idx / 128;                   // 2 values per idx
             const uint iqs = idx % 128;                  // 0..127
@@ -304,7 +353,7 @@ void main() {
             buf_a[buf_idx + 1] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[ib].hmask[hmi + 1] & m) != 0) ? 0 : 4)));
 #elif defined(DATA_A_Q4_K)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
-            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
 
             const uint ib = idx / 128;                 // 2 values per idx
             const uint iqs = idx % 128;                // 0..127
@@ -316,15 +365,20 @@ void main() {
 
             const vec2 loadd = vec2(data_a[ib].d);
 
-            uint8_t sc;
-            uint8_t mbyte;
-            if (is < 4) {
-                sc    = uint8_t(data_a[ib].scales[is    ] & 63);
-                mbyte = uint8_t(data_a[ib].scales[is + 4] & 63);
-            } else {
-                sc    = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4));
-                mbyte = uint8_t((data_a[ib].scales[is + 4] >>  4) | ((data_a[ib].scales[is    ] >> 6) << 4));
-            }
+            const uint scidx0 = (is < 4) ? is : (is + 4);
+            const uint scidx1 = (is < 4) ? is : (is - 4);
+            const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+            const uint scidxshift1 = (is < 4) ? 0 : 2;
+            const uint mbidx0 = is + 4;
+            const uint mbidx1 = (is < 4) ? is + 4 : is;
+            const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+            const uint mbidxshift0 = (is < 4) ? 0 : 4;
+            const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+            const uint mbidxshift1 = (is < 4) ? 0 : 2;
+
+            const uint8_t sc = uint8_t((data_a[ib].scales[scidx0] & 0xF) | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
+            const uint8_t mbyte = uint8_t((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0 | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
             const float d = loadd.x * sc;
             const float m = -loadd.y * mbyte;
 
@@ -332,7 +386,7 @@ void main() {
             buf_a[buf_idx + 1] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF), m));
 #elif defined(DATA_A_Q5_K)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
-            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
 
             const uint ib = idx / 128;                 // 2 values per idx
             const uint iqs = idx % 128;                // 0..127
@@ -347,15 +401,20 @@ void main() {
 
             const vec2 loadd = vec2(data_a[ib].d);
 
-            uint8_t sc;
-            uint8_t mbyte;
-            if (is < 4) {
-                sc    = uint8_t(data_a[ib].scales[is    ] & 63);
-                mbyte = uint8_t(data_a[ib].scales[is + 4] & 63);
-            } else {
-                sc    = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4));
-                mbyte = uint8_t((data_a[ib].scales[is + 4] >>  4) | ((data_a[ib].scales[is    ] >> 6) << 4));
-            }
+            const uint scidx0 = (is < 4) ? is : (is + 4);
+            const uint scidx1 = (is < 4) ? is : (is - 4);
+            const uint scidxmask1 = (is < 4) ? 0x30 : 0xC0;
+            const uint scidxshift1 = (is < 4) ? 0 : 2;
+            const uint mbidx0 = is + 4;
+            const uint mbidx1 = (is < 4) ? is + 4 : is;
+            const uint mbidxmask0 = (is < 4) ? 0xF : 0xF0;
+            const uint mbidxshift0 = (is < 4) ? 0 : 4;
+            const uint mbidxmask1 = (is < 4) ? 0x30 : 0xC0;
+            const uint mbidxshift1 = (is < 4) ? 0 : 2;
+
+            const uint8_t sc    = uint8_t((data_a[ib].scales[scidx0] & 0xF)                         | ((data_a[ib].scales[scidx1] & scidxmask1) >> scidxshift1));
+            const uint8_t mbyte = uint8_t(((data_a[ib].scales[mbidx0] & mbidxmask0) >> mbidxshift0) | ((data_a[ib].scales[mbidx1] & mbidxmask1) >> mbidxshift1));
+
             const float d = loadd.x * sc;
             const float m = -loadd.y * mbyte;
 
@@ -363,7 +422,7 @@ void main() {
             buf_a[buf_idx + 1] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0), m));
 #elif defined(DATA_A_Q6_K)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
-            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
 
             const uint ib = idx / 128;                  // 2 values per idx
             const uint iqs = idx % 128;                 // 0..127
@@ -382,7 +441,7 @@ void main() {
             buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
 #elif defined(DATA_A_IQ4_NL)
             const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
-            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+            const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a;
 
             const uint ib = idx / 16;
             const uint iqs = idx & 0xF;
@@ -403,7 +462,7 @@ void main() {
 #else
             const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
 #endif
-            const uint buf_idx = (loadc_b + l) * (BK+1) + loadr_b * LOAD_VEC_B;
+            const uint buf_idx = (loadc_b + l) * SHMEM_STRIDE + loadr_b * LOAD_VEC_B;
             buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx][0].x);
             buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx][0].y);
             buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx][0].z);
@@ -419,24 +478,24 @@ void main() {
 #else
             const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
 #endif
-            const uint buf_idx = (loadc_b + l) * (BK+1) + loadr_b * LOAD_VEC_B;
+            const uint buf_idx = (loadc_b + l) * SHMEM_STRIDE + loadr_b * LOAD_VEC_B;
             buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx].x);
             buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx].y);
             buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx].z);
             buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx].w);
 #elif !MUL_MAT_ID
             if (ic * BN + loadc_b + l < p.N && block + loadr_b < end_k) {
-                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]);
+                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]);
             } else {
-                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(0.0f);
+                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f);
             }
 #else
             const uint row_i = ic * BN + loadc_b + l;
             if (row_i < _ne1) {
                 const u16vec2 row_idx = row_ids[row_i];
-                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
+                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
             } else {
-                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(0.0f);
+                buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = FLOAT_TYPE(0.0f);
             }
 #endif
         }
@@ -446,16 +505,30 @@ void main() {
         pos_a += BK / LOAD_VEC_A;
         pos_b += BK / LOAD_VEC_B;
 
-        for (uint i = 0; i < BK; i++) {
+#ifdef COOPMAT
+        [[unroll]] for (uint i = 0; i < BK; i += TK) {
+            [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
+                // Load from shared into cache
+                coopMatLoad(cache_a, buf_a, (warp_r * WM + cm_row * TM) * SHMEM_STRIDE + i, SHMEM_STRIDE, gl_CooperativeMatrixLayoutRowMajor);
+
+                [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
+                    coopMatLoad(cache_b, buf_b, (warp_c * WN + cm_col * TN) * SHMEM_STRIDE + i, SHMEM_STRIDE, gl_CooperativeMatrixLayoutColumnMajor);
+
+                    sums[cm_col * cms_per_row + cm_row] = coopMatMulAdd(cache_a, cache_b, sums[cm_col * cms_per_row + cm_row]);
+                }
+            }
+        }
+#else
+        [[unroll]] for (uint i = 0; i < BK; i++) {
             // Load from shared into cache
             [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
                 [[unroll]] for (uint j = 0; j < TM; j++) {
-                    cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * (BK+1) + i];
+                    cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * SHMEM_STRIDE + i];
                 }
             }
             [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
                 [[unroll]] for (uint j = 0; j < TN; j++) {
-                    cache_b[wsic * TN + j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * (BK+1) + i];
+                    cache_b[wsic * TN + j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * SHMEM_STRIDE + i];
                 }
             }
 
@@ -464,12 +537,13 @@ void main() {
                     [[unroll]] for (uint cc = 0; cc < TN; cc++) {
                         [[unroll]] for (uint cr = 0; cr < TM; cr++) {
                             const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;
-                            sums[sums_idx] = fma(float(cache_a[wsir * TM + cr]), float(cache_b[wsic * TN + cc]), sums[sums_idx]);
+                            sums[sums_idx] = fma(ACC_TYPE(cache_a[wsir * TM + cr]), ACC_TYPE(cache_b[wsic * TN + cc]), sums[sums_idx]);
                         }
                     }
                 }
             }
         }
+#endif
 
         barrier();
     }
@@ -481,6 +555,54 @@ void main() {
     const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
 #endif
 
+#ifdef COOPMAT
+#ifdef MUL_MAT_ID
+    [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
+        [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
+            coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
+
+            [[unroll]] for (uint col = 0; col < BN; col += storestride) {
+                const uint row_i = dc + cm_col * TN + col + store_c;
+                if (row_i >= _ne1) break;
+
+                const u16vec2 row_idx = row_ids[row_i];
+
+                data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
+            }
+        }
+    }
+#else
+    const bool is_aligned = p.stride_d % 4 == 0;  // Assumption: D_TYPE == float
+
+    [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
+        [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
+            const bool is_in_bounds = dr + (cm_row + 1) * TM <= p.M && dc + (cm_col + 1) * TN <= p.N;
+
+            if (is_aligned && is_in_bounds) {
+                // Full coopMat is within bounds and stride_d is aligned with 16B
+                coopmat<D_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> cm_dtype = coopmat<D_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(sums[cm_col * cms_per_row + cm_row]);
+                coopMatStore(cm_dtype, data_d, offsets + (dc + cm_col * TN) * p.stride_d + dr + cm_row * TM, p.stride_d, gl_CooperativeMatrixLayoutColumnMajor);
+            } else if (is_in_bounds) {
+                // Full coopMat is within bounds, but stride_d is not aligned
+                coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
+
+                [[unroll]] for (uint col = 0; col < TN; col += storestride) {
+                    data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
+                }
+            } else if (dr + cm_row * TM < p.M && dc + cm_col * TN < p.N) {
+                // Partial coopMat is within bounds
+                coopMatStore(sums[cm_col * cms_per_row + cm_row], coopmat_stage, warp_i * TM * TN, TM, gl_CooperativeMatrixLayoutColumnMajor);
+
+                [[unroll]] for (uint col = 0; col < TN; col += storestride) {
+                    if (dr + cm_row * TM + store_r < p.M && dc + cm_col * TN + col + store_c < p.N) {
+                        data_d[offsets + (dc + cm_col * TN + col + store_c) * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
+                    }
+                }
+            }
+        }
+    }
+#endif // MUL_MAT_ID
+#else
     [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
         [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
 
@@ -492,7 +614,7 @@ void main() {
                 if (row_i >= _ne1) break;
 
                 const u16vec2 row_idx = row_ids[row_i];
-#endif
+#endif // MUL_MAT_ID
                 [[unroll]] for (uint cr = 0; cr < TM; cr++) {
 #ifdef MUL_MAT_ID
                     data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
@@ -500,9 +622,10 @@ void main() {
                     if (dr_warp + cr < p.M && dc_warp + cc < p.N) {
                         data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
                     }
-#endif
+#endif // MUL_MAT_ID
                 }
             }
         }
     }
+#endif // COOPMAT
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
new file mode 100644
index 00000000000..cbfa5dce1b0
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
@@ -0,0 +1,328 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : require
+
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
+
+#extension GL_KHR_memory_scope_semantics : enable
+#extension GL_KHR_cooperative_matrix : enable
+#extension GL_NV_cooperative_matrix2 : enable
+#extension GL_EXT_buffer_reference : enable
+#extension GL_KHR_shader_subgroup_ballot : enable
+#extension GL_KHR_shader_subgroup_vote : enable
+
+#include "types.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (constant_id = 1) const uint BM = 64;
+layout (constant_id = 2) const uint BN = 64;
+layout (constant_id = 3) const uint BK = 16;  // Assumed to be 32 if working with a quant
+
+layout (push_constant) uniform parameter
+{
+    uint M;
+    uint N;
+    uint K;
+    uint stride_a;
+    uint stride_b;
+    uint stride_d;
+
+    uint batch_stride_a;
+    uint batch_stride_b;
+    uint batch_stride_d;
+
+#ifdef MUL_MAT_ID
+    uint nei0;
+    uint nei1;
+    uint nbi1;
+    uint ne11;
+#else
+    uint k_split;
+    uint ne02;
+    uint ne12;
+    uint broadcast2;
+    uint broadcast3;
+#endif
+} p;
+
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+#if QUANT_K > 1
+#define DECODEFUNCA , dequantFuncA
+#define MAT_A_TYPE float16_t
+
+#include "dequant_funcs_cm2.comp"
+
+#else
+#define DECODEFUNCA
+#define MAT_A_TYPE A_TYPE
+#endif
+
+#define MAT_B_TYPE B_TYPE
+
+#ifdef MUL_MAT_ID
+layout (binding = 3) readonly buffer IDS {int data_ids[];};
+
+shared u16vec4 row_ids[3072];
+
+layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufB {
+   B_TYPE b[];
+};
+
+uint _ne1;
+shared uint _ne1_sh;
+
+B_TYPE decodeFuncB(const in decodeBufB bl, const in uint blockCoords[2], const in uint coordInBlock[2])
+{
+    const uint row_i = blockCoords[0];
+
+    if (row_i >= _ne1) {
+        return B_TYPE(0.0);
+    }
+
+    const u16vec4 row_idx = row_ids[row_i];
+    B_TYPE ret = data_b[row_idx.y * p.batch_stride_b + row_idx.x * p.stride_b + blockCoords[1]];
+
+    return ret;
+}
+
+D_TYPE perElemOpD(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t ir, const in uint32_t ic)
+{
+    uint dr = ir * BM + r;
+    uint dc = ic * BN + c;
+
+    if (dr < p.M && dc < _ne1) {
+        uint row_i = dc;
+        const u16vec4 row_idx = row_ids[row_i];
+        data_d[row_idx.y * p.batch_stride_d + row_idx.z * p.stride_d + dr] = elem;
+    }
+    return elem;
+}
+
+#endif
+
+void main() {
+#if defined(DATA_A_IQ4_NL)
+    init_iq4nl_shmem();
+#endif
+
+#ifdef MUL_MAT_ID
+    const uint expert_idx = gl_GlobalInvocationID.z;
+#else
+    const uint batch_idx = gl_GlobalInvocationID.z;
+
+    const uint i13 = batch_idx / p.ne12;
+    const uint i12 = batch_idx % p.ne12;
+
+    const uint i03 = i13 / p.broadcast3;
+    const uint i02 = i12 / p.broadcast2;
+
+    const uint batch_idx_a = i03 * p.ne02 + i02;
+#endif
+
+    const uint blocks_m = (p.M + BM - 1) / BM;
+    const uint ir = gl_WorkGroupID.x % blocks_m;
+    const uint ik = gl_WorkGroupID.x / blocks_m;
+    const uint ic = gl_WorkGroupID.y;
+
+#ifdef MUL_MAT_ID
+    // Spread the search across all elements in the first subgroup
+    if (gl_SubgroupID == 0) {
+        _ne1 = 0;
+        uint num_elements = p.nei1 * p.nei0;
+
+        for (uint i = gl_SubgroupInvocationID; subgroupAny(i < num_elements); i += gl_SubgroupSize) {
+            bool in_range = i < num_elements;
+            uint ii0 = i % p.nei0;
+            uint ii1 = i / p.nei0;
+            uint id = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
+            uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
+            uint idx = subgroupBallotExclusiveBitCount(ballot);
+            if (in_range && id == expert_idx) {
+                row_ids[_ne1 + idx] = u16vec4(ii0 % p.ne11, ii1, ii0, 0);
+            }
+            _ne1 += subgroupBallotBitCount(ballot);
+        }
+        _ne1_sh = _ne1;
+    }
+
+    barrier();
+
+    _ne1 = _ne1_sh;
+
+    // Workgroup has no work
+    if (ic * BN >= _ne1) return;
+#endif
+
+#ifdef MUL_MAT_ID
+    uint start_k = 0;
+    const uint end_k = p.K;
+#else
+    uint start_k = ik * p.k_split;
+    const uint end_k = min(p.K, (ik + 1) * p.k_split);
+#endif
+
+    coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator> sum;
+    sum = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator>(0.0);
+
+#ifdef MUL_MAT_ID
+    uint pos_a = (expert_idx * p.batch_stride_a) / QUANT_K;
+    uint pos_b = 0;
+#else
+    uint pos_a = (batch_idx_a * p.batch_stride_a) / QUANT_K;
+    uint pos_b = batch_idx * p.batch_stride_b;
+#endif
+
+    uint stride_a = p.stride_a / QUANT_K;
+    uint stride_b = p.stride_b;
+
+    // Hint to the compiler that values are aligned (want 16B alignment).
+    // Quants are always block-aligned, no alignment needed.
+#if ALIGNED
+#if QUANT_K == 1
+    stride_a &= ~7;
+#endif
+    stride_b &= ~7;
+#endif
+
+    // Create layouts for both clamped and unclamped accesses
+    tensorLayoutNV<2> tensorLayoutA = createTensorLayoutNV(2);
+    tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutAClamp = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
+    tensorLayoutNV<2> tensorLayoutB = createTensorLayoutNV(2);
+    tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutBClamp = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
+    tensorLayoutNV<2, gl_CooperativeMatrixClampModeConstantNV> tensorLayoutD = createTensorLayoutNV(2, gl_CooperativeMatrixClampModeConstantNV);
+
+#if QUANT_K > 1
+    tensorLayoutA = setTensorLayoutBlockSizeNV(tensorLayoutA, 1, QUANT_K);
+    tensorLayoutAClamp = setTensorLayoutBlockSizeNV(tensorLayoutAClamp, 1, QUANT_K);
+#endif
+
+    // Use end_k rather than p.K as the dimension because that's what
+    // we need to bound check against when using split_k
+    tensorLayoutA = setTensorLayoutDimensionNV(tensorLayoutA, p.M, end_k);
+    tensorLayoutB = setTensorLayoutDimensionNV(tensorLayoutB, p.N, end_k);
+    tensorLayoutD = setTensorLayoutDimensionNV(tensorLayoutD, p.N, p.M);
+    tensorLayoutAClamp = setTensorLayoutDimensionNV(tensorLayoutAClamp, p.M, end_k);
+    tensorLayoutBClamp = setTensorLayoutDimensionNV(tensorLayoutBClamp, p.N, end_k);
+
+    tensorViewNV<2, false, 1, 0> tensorViewTranspose = createTensorViewNV(2, false, 1, 0);
+
+#if !defined(MUL_MAT_ID)
+    // Detect a fast path where all loads are entirely in bounds and no clamping is required
+    if ((ir + 1) * BM <= p.M && (ic + 1) * BN <= p.N && (start_k % BK) == 0 && (end_k % BK) == 0 &&
+#if QUANT_K == 1
+        (stride_a % 8) == 0 &&
+#endif
+        (stride_b % 8) == 0 && (start_k % 8) == 0) {
+        // Hint to the compiler that values are aligned (want 16B alignment)
+        start_k &= ~7;
+        stride_b &= ~7;
+#if QUANT_K == 1
+        stride_a &= ~7;
+#endif
+
+        tensorLayoutA = setTensorLayoutStrideNV(tensorLayoutA, stride_a, 1);
+        tensorLayoutB = setTensorLayoutStrideNV(tensorLayoutB, stride_b, 1);
+
+        uint k_iters = (end_k - start_k + BK - 1) / BK;
+
+        for (uint block_k = start_k, i = 0; i < k_iters; block_k += BK, ++i) {
+
+            coopmat<MAT_A_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+            coopmat<MAT_B_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+
+            coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, block_k, BK) DECODEFUNCA);
+            coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a_ft = coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA>(mat_a);
+
+            coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose);
+            coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b_ft = coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB>(mat_b);
+
+            sum = coopMatMulAdd(mat_a_ft, mat_b_ft, sum);
+        }
+    } else
+#endif // !defined(MUL_MAT_ID)
+    {
+        tensorLayoutA = setTensorLayoutStrideNV(tensorLayoutA, stride_a, 1);
+
+        tensorLayoutAClamp = setTensorLayoutStrideNV(tensorLayoutAClamp, stride_a, 1);
+
+        tensorLayoutB = setTensorLayoutStrideNV(tensorLayoutB, stride_b, 1);
+
+        tensorLayoutBClamp = setTensorLayoutStrideNV(tensorLayoutBClamp, stride_b, 1);
+
+        [[dont_unroll]]
+        for (uint block_k = start_k; block_k < end_k; block_k += BK) {
+
+            coopmat<MAT_A_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a;
+            coopmat<MAT_B_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b;
+            coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA> mat_a_ft;
+            coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB> mat_b_ft;
+
+            // Clamping is expensive, so detect different code paths for each combination
+            // of A and B needing clamping.
+            bool unclampedA = (ir + 1) * BM <= p.M && block_k + BK <= end_k && (block_k % 8) == 0;
+#ifdef MUL_MAT_ID
+            bool unclampedB = true;
+#else
+            bool unclampedB = (ic + 1) * BN <= p.N && block_k + BK <= end_k && (block_k % 8) == 0;
+#endif
+            if (unclampedA && unclampedB) {
+                coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, (block_k & ~7), BK) DECODEFUNCA);
+#ifdef MUL_MAT_ID
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
+#else
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, (block_k & ~7), BK), tensorViewTranspose);
+#endif
+                mat_a_ft = coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA>(mat_a);
+                mat_b_ft = coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB>(mat_b);
+                sum = coopMatMulAdd(mat_a_ft, mat_b_ft, sum);
+            } else if (unclampedA && !unclampedB) {
+                coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutA, ir * BM, BM, (block_k & ~7), BK) DECODEFUNCA);
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
+
+                mat_a_ft = coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA>(mat_a);
+                mat_b_ft = coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB>(mat_b);
+                sum = coopMatMulAdd(mat_a_ft, mat_b_ft, sum);
+            } else if (!unclampedA && unclampedB) {
+                coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
+#ifdef MUL_MAT_ID
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, block_k, BK), tensorViewTranspose, decodeFuncB);
+#else
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutB, ic * BN, BN, (block_k & ~7), BK), tensorViewTranspose);
+#endif
+                mat_a_ft = coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA>(mat_a);
+                mat_b_ft = coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB>(mat_b);
+                sum = coopMatMulAdd(mat_a_ft, mat_b_ft, sum);
+            } else if (!unclampedA && !unclampedB) {
+                coopMatLoadTensorNV(mat_a, data_a, pos_a, sliceTensorLayoutNV(tensorLayoutAClamp, ir * BM, BM, block_k, BK) DECODEFUNCA);
+                coopMatLoadTensorNV(mat_b, data_b, pos_b, sliceTensorLayoutNV(tensorLayoutBClamp, ic * BN, BN, block_k, BK), tensorViewTranspose);
+
+                mat_a_ft = coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BM, BK, gl_MatrixUseA>(mat_a);
+                mat_b_ft = coopmat<FLOAT_TYPE, gl_ScopeWorkgroup, BK, BN, gl_MatrixUseB>(mat_b);
+                sum = coopMatMulAdd(mat_a_ft, mat_b_ft, sum);
+            }
+        }
+    }
+
+    // Convert from ACC_TYPE to D_TYPE
+    coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator> mat_d;
+    mat_d = coopmat<D_TYPE, gl_ScopeWorkgroup, BM, BN, gl_MatrixUseAccumulator>(sum);
+
+#ifdef MUL_MAT_ID
+    // Call callback to store each element, remapping row through shared memory
+    coopMatPerElementNV(mat_d, mat_d, perElemOpD, ir, ic);
+#else
+    tensorLayoutD = setTensorLayoutStrideNV(tensorLayoutD, p.stride_d, 1);
+
+    uint pos_d = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
+    coopMatStoreTensorNV(mat_d, data_d, pos_d, sliceTensorLayoutNV(tensorLayoutD, ic * BN, BN, ir * BM, BM), tensorViewTranspose);
+#endif
+}
diff --git a/ggml/src/vulkan-shaders/norm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/norm.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/norm.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/norm.comp
diff --git a/ggml/src/vulkan-shaders/pad.comp b/ggml/src/ggml-vulkan/vulkan-shaders/pad.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/pad.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/pad.comp
diff --git a/ggml/src/vulkan-shaders/pool2d.comp b/ggml/src/ggml-vulkan/vulkan-shaders/pool2d.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/pool2d.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/pool2d.comp
diff --git a/ggml/src/vulkan-shaders/relu.comp b/ggml/src/ggml-vulkan/vulkan-shaders/relu.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/relu.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/relu.comp
diff --git a/ggml/src/vulkan-shaders/repeat.comp b/ggml/src/ggml-vulkan/vulkan-shaders/repeat.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/repeat.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/repeat.comp
diff --git a/ggml/src/vulkan-shaders/rms_norm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/rms_norm.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/rms_norm.comp
diff --git a/ggml/src/vulkan-shaders/rope_head.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp
similarity index 91%
rename from ggml/src/vulkan-shaders/rope_head.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp
index ea89542261c..574b51ca553 100644
--- a/ggml/src/vulkan-shaders/rope_head.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp
@@ -1,6 +1,11 @@
 #include "types.comp"
 
 #extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_spirv_intrinsics: enable
+
+#if RTE16
+spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
+#endif
 
 layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in;
 
diff --git a/ggml/src/vulkan-shaders/rope_neox.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/rope_neox.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/rope_neox.comp
diff --git a/ggml/src/vulkan-shaders/rope_norm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/rope_norm.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/rope_norm.comp
diff --git a/ggml/src/vulkan-shaders/scale.comp b/ggml/src/ggml-vulkan/vulkan-shaders/scale.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/scale.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/scale.comp
diff --git a/ggml/src/vulkan-shaders/silu.comp b/ggml/src/ggml-vulkan/vulkan-shaders/silu.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/silu.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/silu.comp
diff --git a/ggml/src/vulkan-shaders/sin.comp b/ggml/src/ggml-vulkan/vulkan-shaders/sin.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/sin.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/sin.comp
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp b/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
new file mode 100644
index 00000000000..a25808e1656
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/soft_max.comp
@@ -0,0 +1,174 @@
+#version 450
+
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#extension GL_EXT_control_flow_attributes : enable
+
+layout (push_constant) uniform parameter
+{
+    uint KX;
+    uint KY;
+    float scale;
+    float max_bias;
+    float m0;
+    float m1;
+    uint n_head_log2;
+    uint nrows_x;
+} p;
+
+#include "types.comp"
+
+layout(constant_id = 0) const uint BLOCK_SIZE = 32;
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
+layout (binding = 2) buffer D {D_TYPE data_d[];};
+
+shared FLOAT_TYPE vals[BLOCK_SIZE];
+
+// num_iters is the number of BLOCK_SIZE loop iterations we need to iterate
+// over all the columns. The main function tries to pass a constant here,
+// as if it were a template function, to allow unrolling.
+void soft_max(uint num_iters) {
+    const uint tid = gl_LocalInvocationID.x;
+    const uint rowx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint rowy = (p.KY > 0) ? (rowx % p.KY) : 0;
+
+    if (rowx >= p.nrows_x) {
+        return;
+    }
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (p.max_bias > 0.0f) {
+        const uint h = rowx/p.KY; // head index
+
+        const float base = h < p.n_head_log2 ? p.m0 : p.m1;
+        const uint   exp  = h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // Find max
+    FLOAT_TYPE max_val = uintBitsToFloat(0xFF800000);
+
+    // Cache values while we compute the max, so we don't need to read them
+    // again when we're ready to compute exp(x-max).
+    const uint DATA_CACHE_SIZE = 16;
+    FLOAT_TYPE data_cache[DATA_CACHE_SIZE];
+
+    [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {
+        const uint col = col0 + tid;
+
+        FLOAT_TYPE a = FLOAT_TYPE(0);
+        if (col < p.KX) {
+            a = data_a[rowx * p.KX + col];
+        }
+
+        FLOAT_TYPE b = FLOAT_TYPE(0);
+        if (p.KY > 0 && col < p.KX) {
+            b = data_b[rowy * p.KX + col];
+        }
+
+        FLOAT_TYPE v = a * p.scale + slope * b;
+
+        if (col < p.KX) {
+            max_val = max(max_val, v);
+        }
+
+        if (idx < DATA_CACHE_SIZE) {
+            data_cache[idx] = v;
+        }
+    }
+
+    // reduce across the workgroup
+    vals[tid] = max_val;
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            vals[tid] = max(vals[tid], vals[tid + s]);
+        }
+        barrier();
+    }
+
+    max_val = vals[0];
+    barrier();
+
+    FLOAT_TYPE sum = FLOAT_TYPE(0.0f);
+
+    // Compute sum{exp(x - max)}
+    [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {
+        const uint col = col0 + tid;
+
+        if (col >= p.KX) {
+            break;
+        }
+
+        // compute exp(a*scale+b*slope), add it to sum, and cache the new value
+        // in data_cache if possible.
+        const uint i = rowx * p.KX + col;
+        FLOAT_TYPE val;
+        if (idx < DATA_CACHE_SIZE) {
+            val = exp(data_cache[idx] - max_val);
+        } else {
+            val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
+        }
+        sum += val;
+        if (idx < DATA_CACHE_SIZE) {
+            data_cache[idx] = val;
+        } else {
+            data_d[i] = D_TYPE(val);
+        }
+    }
+
+    // reduce across the workgroup
+    vals[tid] = sum;
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            vals[tid] += vals[tid + s];
+        }
+        barrier();
+    }
+    sum = vals[0];
+
+    FLOAT_TYPE rcpdivisor = 1.0/sum;
+
+    [[unroll]] for (uint col0 = 0, idx = 0; idx < num_iters; col0 += BLOCK_SIZE, ++idx) {
+        const uint col = col0 + tid;
+
+        if (col >= p.KX) {
+            continue;
+        }
+
+        if (idx < DATA_CACHE_SIZE) {
+            data_d[rowx*p.KX + col] = D_TYPE(data_cache[idx] * rcpdivisor);
+        } else {
+            data_d[rowx*p.KX + col] *= D_TYPE(rcpdivisor);
+        }
+    }
+}
+
+void main() {
+    // instantiate the soft_max function for several different
+    // dimensions, to allow loop unrolling
+    uint num_blocks = (p.KX + BLOCK_SIZE - 1) / BLOCK_SIZE;
+    if (num_blocks > 32) {
+        soft_max(num_blocks);
+    } else if (num_blocks > 16) {
+        soft_max(32);
+    } else if (num_blocks > 8) {
+        soft_max(16);
+    } else if (num_blocks > 4) {
+        soft_max(8);
+    } else if (num_blocks == 4) {
+        soft_max(4);
+    } else if (num_blocks == 3) {
+        soft_max(3);
+    } else if (num_blocks == 2) {
+        soft_max(2);
+    } else if (num_blocks == 1) {
+        soft_max(1);
+    }
+}
diff --git a/ggml/src/vulkan-shaders/square.comp b/ggml/src/ggml-vulkan/vulkan-shaders/square.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/square.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/square.comp
diff --git a/ggml/src/vulkan-shaders/sum_rows.comp b/ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/sum_rows.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/sum_rows.comp
diff --git a/ggml/src/vulkan-shaders/tanh.comp b/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp
similarity index 88%
rename from ggml/src/vulkan-shaders/tanh.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp
index 74630dc7fef..495f966bdc6 100644
--- a/ggml/src/vulkan-shaders/tanh.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/tanh.comp
@@ -16,6 +16,5 @@ void main() {
     if (i >= p.KX) {
         return;
     }
-
-    data_d[i] = D_TYPE(tanh(data_a[i]));
+    data_d[i] = D_TYPE(1. - 2. / (exp(2.*data_a[i]) + 1.));
 }
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat2_support.comp b/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat2_support.comp
new file mode 100644
index 00000000000..28eb24e11f8
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/test_coopmat2_support.comp
@@ -0,0 +1,7 @@
+#version 460
+
+#extension GL_NV_cooperative_matrix2 : require
+
+void main()
+{
+}
diff --git a/ggml/src/vulkan-shaders/timestep_embedding.comp b/ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/timestep_embedding.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/timestep_embedding.comp
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/types.comp b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
new file mode 100644
index 00000000000..eecc47f3a97
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/types.comp
@@ -0,0 +1,323 @@
+
+#if !defined(GGML_TYPES_COMP)
+#define GGML_TYPES_COMP
+
+#extension GL_EXT_shader_explicit_arithmetic_types : require
+
+#if defined(DATA_A_F32)
+#define QUANT_K 1
+#define QUANT_R 1
+
+#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
+#define A_TYPE float
+#elif LOAD_VEC_A == 4
+#define A_TYPE vec4
+#elif LOAD_VEC_A == 8
+#define A_TYPE mat2x4
+#endif
+#endif
+
+#if defined(DATA_A_F16)
+#define QUANT_K 1
+#define QUANT_R 1
+
+#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
+#define A_TYPE float16_t
+#elif LOAD_VEC_A == 4
+#define A_TYPE f16vec4
+#elif LOAD_VEC_A == 8
+#define A_TYPE f16mat2x4
+#endif
+#endif
+
+#define QUANT_K_Q4_0 32
+#define QUANT_R_Q4_0 2
+
+struct block_q4_0
+{
+    float16_t d;
+    uint8_t qs[16];
+};
+struct block_q4_0_packed16
+{
+    float16_t d;
+    uint16_t qs[16/2];
+};
+
+#if defined(DATA_A_Q4_0)
+#define QUANT_K QUANT_K_Q4_0
+#define QUANT_R QUANT_R_Q4_0
+#define A_TYPE block_q4_0
+#define A_TYPE_PACKED16 block_q4_0_packed16
+#endif
+
+#define QUANT_K_Q4_1 32
+#define QUANT_R_Q4_1 2
+
+struct block_q4_1
+{
+    float16_t d;
+    float16_t m;
+    uint8_t qs[16];
+};
+
+struct block_q4_1_packed16
+{
+    float16_t d;
+    float16_t m;
+    uint16_t qs[16/2];
+};
+
+#if defined(DATA_A_Q4_1)
+#define QUANT_K QUANT_K_Q4_1
+#define QUANT_R QUANT_R_Q4_1
+#define A_TYPE block_q4_1
+#define A_TYPE_PACKED16 block_q4_1_packed16
+#endif
+
+#define QUANT_K_Q5_0 32
+#define QUANT_R_Q5_0 2
+
+struct block_q5_0
+{
+    float16_t d;
+    uint16_t qh[2];
+    uint8_t qs[16];
+};
+
+struct block_q5_0_packed16
+{
+    float16_t d;
+    uint16_t qh[2];
+    uint16_t qs[16/2];
+};
+
+#if defined(DATA_A_Q5_0)
+#define QUANT_K QUANT_K_Q5_0
+#define QUANT_R QUANT_R_Q5_0
+#define A_TYPE block_q5_0
+#define A_TYPE_PACKED16 block_q5_0_packed16
+#endif
+
+#define QUANT_K_Q5_1 32
+#define QUANT_R_Q5_1 2
+
+struct block_q5_1
+{
+    float16_t d;
+    float16_t m;
+    uint qh;
+    uint8_t qs[16];
+};
+
+struct block_q5_1_packed16
+{
+    float16_t d;
+    float16_t m;
+    uint qh;
+    uint16_t qs[16/2];
+};
+
+#if defined(DATA_A_Q5_1)
+#define QUANT_K QUANT_K_Q5_1
+#define QUANT_R QUANT_R_Q5_1
+#define A_TYPE block_q5_1
+#define A_TYPE_PACKED16 block_q5_1_packed16
+#endif
+
+#define QUANT_K_Q8_0 32
+#define QUANT_R_Q8_0 1
+
+struct block_q8_0
+{
+    float16_t d;
+    int8_t qs[32];
+};
+struct block_q8_0_packed16
+{
+    float16_t d;
+    uint16_t qs[32/2];
+};
+
+#if defined(DATA_A_Q8_0)
+#define QUANT_K QUANT_K_Q8_0
+#define QUANT_R QUANT_R_Q8_0
+#define A_TYPE block_q8_0
+#define A_TYPE_PACKED16 block_q8_0_packed16
+#endif
+
+// K-quants
+#define QUANT_K_Q2_K 256
+
+struct block_q2_K
+{
+    uint8_t scales[QUANT_K_Q2_K/16];
+    uint8_t qs[QUANT_K_Q2_K/4];
+    f16vec2 d;
+};
+
+struct block_q2_K_packed16
+{
+    uint16_t scales[QUANT_K_Q2_K/16/2];
+    uint16_t qs[QUANT_K_Q2_K/4/2];
+    f16vec2 d;
+};
+
+struct block_q2_K_packed32
+{
+    uint32_t scales[QUANT_K_Q2_K/16/4];
+    uint32_t qs[QUANT_K_Q2_K/4/4];
+    f16vec2 d;
+};
+
+#if defined(DATA_A_Q2_K)
+#define QUANT_K QUANT_K_Q2_K
+#define A_TYPE block_q2_K
+#define A_TYPE_PACKED16 block_q2_K_packed16
+#define A_TYPE_PACKED32 block_q2_K_packed32
+#endif
+
+#define QUANT_K_Q3_K 256
+
+struct block_q3_K
+{
+    uint8_t hmask[QUANT_K_Q3_K/8];
+    uint8_t qs[QUANT_K_Q3_K/4];
+    uint8_t scales[12];
+    float16_t d;
+};
+
+struct block_q3_K_packed16
+{
+    uint16_t hmask[QUANT_K_Q3_K/8/2];
+    uint16_t qs[QUANT_K_Q3_K/4/2];
+    uint16_t scales[12/2];
+    float16_t d;
+};
+
+#if defined(DATA_A_Q3_K)
+#define QUANT_K QUANT_K_Q3_K
+#define A_TYPE block_q3_K
+#define A_TYPE_PACKED16 block_q3_K_packed16
+#endif
+
+#define QUANT_K_Q4_K 256
+
+struct block_q4_K
+{
+    f16vec2 d;
+    uint8_t scales[3*QUANT_K_Q4_K/64];
+    uint8_t qs[QUANT_K_Q4_K/2];
+};
+
+struct block_q4_K_packed16
+{
+    f16vec2 d;
+    uint16_t scales[3*QUANT_K_Q4_K/64/2];
+    uint16_t qs[QUANT_K_Q4_K/2/2];
+};
+
+struct block_q4_K_packed32
+{
+    f16vec2 d;
+    uint32_t scales[3*QUANT_K_Q4_K/64/4];
+    uint32_t qs[QUANT_K_Q4_K/2/4];
+};
+
+#if defined(DATA_A_Q4_K)
+#define QUANT_K QUANT_K_Q4_K
+#define A_TYPE block_q4_K
+#define A_TYPE_PACKED16 block_q4_K_packed16
+#define A_TYPE_PACKED32 block_q4_K_packed32
+#endif
+
+#define QUANT_K_Q5_K 256
+
+struct block_q5_K
+{
+    f16vec2 d;
+    uint8_t scales[12];
+    uint8_t qh[QUANT_K_Q5_K/8];
+    uint8_t qs[QUANT_K_Q5_K/2];
+};
+
+struct block_q5_K_packed16
+{
+    f16vec2 d;
+    uint16_t scales[12/2];
+    uint16_t qh[QUANT_K_Q5_K/8/2];
+    uint16_t qs[QUANT_K_Q5_K/2/2];
+};
+
+#if defined(DATA_A_Q5_K)
+#define QUANT_K QUANT_K_Q5_K
+#define A_TYPE block_q5_K
+#define A_TYPE_PACKED16 block_q5_K_packed16
+#endif
+
+#define QUANT_K_Q6_K 256
+
+struct block_q6_K
+{
+    uint8_t ql[QUANT_K_Q6_K/2];
+    uint8_t qh[QUANT_K_Q6_K/4];
+    int8_t scales[QUANT_K_Q6_K/16];
+    float16_t d;
+};
+
+struct block_q6_K_packed16
+{
+    uint16_t ql[QUANT_K_Q6_K/2/2];
+    uint16_t qh[QUANT_K_Q6_K/4/2];
+    int8_t scales[QUANT_K_Q6_K/16];
+    float16_t d;
+};
+
+#if defined(DATA_A_Q6_K)
+#define QUANT_K QUANT_K_Q6_K
+#define A_TYPE block_q6_K
+#define A_TYPE_PACKED16 block_q6_K_packed16
+#endif
+
+// IQuants
+
+#define QUANT_K_IQ4_NL 32
+#define QUANT_R_IQ4_NL 2
+
+struct block_iq4_nl
+{
+    float16_t d;
+    uint8_t qs[QUANT_K_IQ4_NL/2];
+};
+
+struct block_iq4_nl_packed16
+{
+    float16_t d;
+    uint16_t qs[QUANT_K_IQ4_NL/2/2];
+};
+
+#if defined(DATA_A_IQ4_NL)
+
+const int8_t kvalues_iq4nl_const[16] = {
+    int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10),
+    int8_t(1), int8_t(13), int8_t(25), int8_t(38), int8_t(53), int8_t(69), int8_t(89), int8_t(113)
+};
+
+shared FLOAT_TYPE kvalues_iq4nl[16];
+
+void init_iq4nl_shmem()
+{
+    // copy the table into shared memory and sync
+    if (gl_LocalInvocationIndex.x < 16) {
+        kvalues_iq4nl[gl_LocalInvocationIndex.x] = FLOAT_TYPE(kvalues_iq4nl_const[gl_LocalInvocationIndex.x]);
+    }
+    barrier();
+}
+
+#define QUANT_K QUANT_K_IQ4_NL
+#define QUANT_R QUANT_R_IQ4_NL
+#define A_TYPE block_iq4_nl
+#define A_TYPE_PACKED16 block_iq4_nl_packed16
+#endif
+
+#endif // !defined(GGML_TYPES_COMP)
diff --git a/ggml/src/vulkan-shaders/upscale.comp b/ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp
similarity index 100%
rename from ggml/src/vulkan-shaders/upscale.comp
rename to ggml/src/ggml-vulkan/vulkan-shaders/upscale.comp
diff --git a/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
similarity index 72%
rename from ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
rename to ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index 5c84f473fc0..7a0d7285dcb 100644
--- a/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -30,6 +30,8 @@
     #include <fcntl.h>
 #endif
 
+#include <vulkan/vulkan_core.h>
+
 #define ASYNCIO_CONCURRENCY 64
 
 std::mutex lock;
@@ -58,6 +60,7 @@ const std::vector<std::string> type_names = {
     "iq4_nl"
 };
 
+namespace {
 void execute_command(const std::string& command, std::string& stdout_str, std::string& stderr_str) {
 #ifdef _WIN32
     HANDLE stdout_read, stdout_write;
@@ -196,15 +199,20 @@ static uint32_t compile_count = 0;
 static std::mutex compile_count_mutex;
 static std::condition_variable compile_count_cond;
 
-void string_to_spv_func(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true) {
-    std::string name = _name + (fp16 ? "" : "_fp32");
+void string_to_spv_func(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true, bool coopmat = false, bool coopmat2 = false, bool f16acc = false) {
+    std::string name = _name + (f16acc ? "_f16acc" : "") + (coopmat ? "_coopmat" : "") + (coopmat2 ? "_cm2" : (fp16 ? "" : "_fp32"));
     std::string out_fname = join_paths(output_dir, name + ".spv");
     std::string in_path = join_paths(input_dir, in_fname);
 
+    std::string target_env = (name.find("_cm2") != std::string::npos) ? "--target-env=vulkan1.3" : "--target-env=vulkan1.2";
+
+    // disable spirv-opt for coopmat shaders for https://github.com/ggerganov/llama.cpp/issues/10734
+    std::string opt_level = coopmat ? "" : "-O";
+
     #ifdef _WIN32
-        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", "\"" + in_path + "\"", "-o", "\"" + out_fname + "\""};
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, "\"" + in_path + "\"", "-o", "\"" + out_fname + "\""};
     #else
-        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o",  out_fname};
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", target_env, opt_level, in_path, "-o",  out_fname};
     #endif
 
     #ifdef GGML_VULKAN_SHADER_DEBUG_INFO
@@ -254,7 +262,7 @@ std::map<std::string, std::string> merge_maps(const std::map<std::string, std::s
 }
 
 static std::vector<std::future<void>> compiles;
-void string_to_spv(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true) {
+void string_to_spv(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true, bool coopmat = false, bool coopmat2 = false, bool f16acc = false) {
     {
         // wait until fewer than N compiles are in progress.
         // 16 is an arbitrary limit, the goal is to avoid "failed to create pipe" errors.
@@ -265,15 +273,15 @@ void string_to_spv(const std::string& _name, const std::string& in_fname, const
         }
         compile_count++;
     }
-    compiles.push_back(std::async(string_to_spv_func, _name, in_fname, defines, fp16));
+    compiles.push_back(std::async(string_to_spv_func, _name, in_fname, defines, fp16, coopmat, coopmat2, f16acc));
 }
 
-void matmul_shaders(bool fp16, bool matmul_id) {
-    std::string load_vec = fp16 ? "8" : "4";
-    std::string aligned_b_type_f32 = fp16 ? "mat2x4" : "vec4";
-    std::string aligned_b_type_f16 = fp16 ? "f16mat2x4" : "f16vec4";
+void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool f16acc) {
+    std::string load_vec = coopmat2 ? "1" : fp16 ? "8" : "4";
+    std::string aligned_b_type_f32 = coopmat2 ? "float" : fp16 ? "mat2x4" : "vec4";
+    std::string aligned_b_type_f16 = coopmat2 ? "float16_t" : fp16 ? "f16mat2x4" : "f16vec4";
 
-    std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", fp16 ? "float16_t" : "float"}};
+    std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", (coopmat2 || fp16) ? "float16_t" : "float"}};
     std::string shader_name = "matmul";
 
     if (matmul_id) {
@@ -285,21 +293,37 @@ void matmul_shaders(bool fp16, bool matmul_id) {
         base_dict["FLOAT16"] = "1";
     }
 
+    base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float";
+
+    if (coopmat) {
+        base_dict["COOPMAT"] = "1";
+    }
+
+    base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float";
+
+    std::string source_name = coopmat2 ? "mul_mm_cm2.comp" : "mul_mm.comp";
+
     // Shaders with f16 B_TYPE
-    string_to_spv(shader_name + "_f32_f16", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16);
-    string_to_spv(shader_name + "_f32_f16_aligned", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}}), fp16);
+    string_to_spv(shader_name + "_f32_f16", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, }), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f32_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
 
-    string_to_spv(shader_name + "_f16", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16);
-    string_to_spv(shader_name + "_f16_aligned", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}}), fp16);
+    string_to_spv(shader_name + "_f16_aligned", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+    string_to_spv(shader_name + "_f16", source_name, merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16, coopmat, coopmat2, f16acc);
 
     for (const auto& tname : type_names) {
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);
         // For unaligned, load one at a time for f32/f16, or two at a time for quants
-        std::string load_vec_a_unaligned = (tname == "f32" || tname == "f16") ? "1" : "2";
+        std::string load_vec_a_unaligned = (coopmat2 || tname == "f32" || tname == "f16") ? "1" : "2";
         // For aligned matmul loads
-        std::string load_vec_a = (tname == "f32" || tname == "f16") ? load_vec : "2";
-        string_to_spv(shader_name + "_" + tname + "_f32", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16);
-        string_to_spv(shader_name + "_" + tname + "_f32_aligned", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}}), fp16);
+        std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16") ? load_vec : "2";
+
+        string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc);
+        string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+
+        if (tname != "f16" && tname != "f32") {
+            string_to_spv(shader_name + "_" + tname + "_f16", source_name,          merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned},                           {"B_TYPE", "float16_t"},        {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc);
+            string_to_spv(shader_name + "_" + tname + "_f16_aligned", source_name,  merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a},           {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
+        }
     }
 }
 
@@ -307,20 +331,58 @@ void process_shaders() {
     std::cout << "ggml_vulkan: Generating and compiling shaders to SPIR-V" << std::endl;
     std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}};
 
-    for (const auto& fp16 : {false, true}) {
-        matmul_shaders(fp16, false);
-        matmul_shaders(fp16, true);
+    // matmul
+    for (const auto& matmul_id : {false, true}) {
+        // No coopmats
+        // fp32
+        matmul_shaders(false, matmul_id, false, false, false);
+
+        // fp16, fp32acc and fp16acc
+        matmul_shaders(true, matmul_id, false, false, false);
+        matmul_shaders(true, matmul_id, false, false, true);
+
+        // Coopmat, fp32acc and fp16acc
+        matmul_shaders(true, matmul_id, true, false, false);
+        matmul_shaders(true, matmul_id, true, false, true);
+
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+        // Coopmat2, fp32acc and fp16acc
+        matmul_shaders(true, matmul_id, false, true, false);
+        matmul_shaders(true, matmul_id, false, true, true);
+#endif
     }
 
+#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
+    // flash attention
+    for (const auto& f16acc : {false, true}) {
+        std::string acctype = f16acc ? "float16_t" : "float";
+
+        for (const auto& tname : type_names) {
+            if (tname == "f32") {
+                continue;
+            }
+
+            if (tname == "f16") {
+                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
+                    merge_maps(base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}}), true, false, true, f16acc);
+            } else {
+                std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+                string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
+                    merge_maps(base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"ACC_TYPE", acctype}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, true, f16acc);
+            }
+        }
+    }
+#endif
+
     for (const auto& tname : type_names) {
         // mul mat vec
         std::string data_a_key = "DATA_A_" + to_uppercase(tname);
         std::string shader = (string_ends_with(tname, "_k")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
 
-        string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
-        string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
+        string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
+        string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"B_TYPE_VEC2", "f16vec2"}, {"B_TYPE_VEC4", "f16vec4"}, {"D_TYPE", "float"}}));
 
-        string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+        string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"B_TYPE_VEC2", "vec2"}, {"B_TYPE_VEC4", "vec4"}, {"D_TYPE", "float"}}));
 
         // Dequant shaders
         if (tname != "f16") {
@@ -331,11 +393,11 @@ void process_shaders() {
             shader = (tname == "f32" || tname == "f16") ? "get_rows.comp" : "get_rows_quant.comp";
 
             if (tname == "f16") {
-                string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+                string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}}));
             } else {
-                string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}});
+                string_to_spv("get_rows_" + tname, shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}}));
             }
-            string_to_spv("get_rows_" + tname + "_f32", shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}});
+            string_to_spv("get_rows_" + tname + "_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}}));
         }
     }
 
@@ -399,9 +461,11 @@ void process_shaders() {
 
     string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
+    string_to_spv("rope_norm_f16_rte", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
     string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
     string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
+    string_to_spv("rope_neox_f16_rte", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}});
 
     string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}});
 
@@ -409,11 +473,14 @@ void process_shaders() {
 
     string_to_spv("im2col_f32", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
     string_to_spv("im2col_f32_f16", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}));
+    string_to_spv("im2col_f32_f16_rte", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"RTE16", "1"}}));
 
     string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
     string_to_spv("pool2d_f32", "pool2d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
 
+    string_to_spv("rwkv_wkv6_f32", "wkv6.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
+
     for (auto &c : compiles) {
         c.wait();
     }
@@ -471,12 +538,19 @@ void write_output_files() {
     fclose(hdr);
     fclose(src);
 }
+}
 
 int main(int argc, char** argv) {
     std::map<std::string, std::string> args;
-    for (int i = 1; i < argc; i += 2) {
-        if (i + 1 < argc) {
-            args[argv[i]] = argv[i + 1];
+    for (int i = 1; i < argc; ++i) {
+        std::string arg = argv[i];
+        if (arg.rfind("--", 0) == 0) {
+            if (i + 1 < argc && argv[i + 1][0] != '-') {
+                args[arg] = argv[i + 1];
+                ++i;
+            } else {
+                args[arg] = "";
+            }
         }
     }
 
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/wkv6.comp b/ggml/src/ggml-vulkan/vulkan-shaders/wkv6.comp
new file mode 100644
index 00000000000..35cc6c45f90
--- /dev/null
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/wkv6.comp
@@ -0,0 +1,87 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : require
+
+#define BLOCK_SIZE 64
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout(push_constant) uniform Parameters {
+    uint B;
+    uint T;
+    uint C;
+    uint H;
+};
+
+layout(binding = 0) readonly buffer KBuf { A_TYPE k[]; };
+layout(binding = 1) readonly buffer VBuf { A_TYPE v[]; };
+layout(binding = 2) readonly buffer RBuf { A_TYPE r[]; };
+layout(binding = 3) readonly buffer TimeFBuf { A_TYPE tf[]; };
+layout(binding = 4) readonly buffer TimeDBuf { A_TYPE td[]; };
+layout(binding = 5) readonly buffer StateBuf { A_TYPE state_in[]; };
+layout(binding = 6) buffer DstBuf { A_TYPE dst[]; };
+
+shared A_TYPE _k[BLOCK_SIZE], _r[BLOCK_SIZE], _tf[BLOCK_SIZE], _td[BLOCK_SIZE];
+
+void main() {
+    const uint head_size = BLOCK_SIZE;
+    const uint batch_id = gl_WorkGroupID.x / H;
+    const uint head_id = gl_WorkGroupID.x % H;
+    const uint tid = gl_LocalInvocationID.x;
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    A_TYPE state[BLOCK_SIZE];
+    [[unroll]] for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + i * head_size + tid];
+    }
+
+    barrier();
+    _tf[tid] = tf[head_id * head_size + tid];
+    barrier();
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        barrier();
+        _k[tid] = k[t];
+        _r[tid] = r[t];
+        _td[tid] = td[t];
+        barrier();
+
+        const A_TYPE v_val = v[t];
+        A_TYPE y = 0.0;
+
+        [[unroll]] for (uint j = 0; j < head_size; j += 4) {
+            vec4 k_vec = vec4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
+            vec4 r_vec = vec4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
+            vec4 tf_vec = vec4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]);
+            vec4 td_vec = vec4(_td[j], _td[j+1], _td[j+2], _td[j+3]);
+            vec4 s_vec = vec4(state[j], state[j+1], state[j+2], state[j+3]);
+
+            vec4 kv = k_vec * v_val;
+
+            vec4 temp = tf_vec * kv + s_vec;
+            y += dot(r_vec, temp);
+
+            s_vec = s_vec * td_vec + kv;
+            state[j] = s_vec.x;
+            state[j+1] = s_vec.y;
+            state[j+2] = s_vec.z;
+            state[j+3] = s_vec.w;
+        }
+
+        dst[t] = y;
+    }
+
+    [[unroll]] for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + i * head_size + tid] = state[i];
+    }
+}
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index cd26a361b84..030d93a5177 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -3,10 +3,15 @@
 
 #include "ggml-backend.h"
 #include "ggml-impl.h"
-#include "ggml-cpu-impl.h"
-#include "ggml-quants.h"
+#include "ggml-threading.h"
 #include "ggml.h"
-#include "ggml-aarch64.h"
+
+// FIXME: required here for quantization functions
+#include "ggml-quants.h"
+
+#ifdef GGML_USE_CPU_HBM
+#include <hbwmalloc.h>
+#endif
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <malloc.h> // using malloc.h with MSC/MINGW
@@ -47,6 +52,17 @@
 
 #define UNUSED GGML_UNUSED
 
+#if defined(_MSC_VER)
+#define m512bh(p) p
+#define m512i(p) p
+#else
+#define m512bh(p) (__m512bh)(p)
+#define m512i(p) (__m512i)(p)
+#endif
+
+// precomputed f32 table for f16 (256 KB) (ggml-impl.h)
+float ggml_table_f32_f16[1 << 16];
+
 #if (defined(__linux__) || defined(__APPLE__) || defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__)) && \
     (!defined(TARGET_OS_TV) && !defined(TARGET_OS_WATCH))
 #include <unistd.h>
@@ -363,7 +379,7 @@ void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int64_t n) {
 void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int64_t n) {
     int64_t i = 0;
 #if defined(__F16C__)
-    if (ggml_cpu_has_f16c()) {
+    //if (ggml_cpu_has_f16c()) {
         for (; i + 7 < n; i += 8) {
             __m256 x_vec = _mm256_loadu_ps(x + i);
             __m128i y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
@@ -374,7 +390,7 @@ void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int64_t n) {
             __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
             _mm_storel_epi64((__m128i *)(y + i), y_vec);
         }
-    }
+    //}
 #endif
     for (; i < n; i++) {
         y[i] = GGML_FP32_TO_FP16(x[i]);
@@ -384,7 +400,7 @@ void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int64_t n) {
 void ggml_bf16_to_fp32_row(const ggml_bf16_t * x, float * y, int64_t n) {
     int64_t i = 0;
 #if defined(__AVX512F__)
-    if (ggml_cpu_has_avx512()) {
+    //if (ggml_cpu_has_avx512()) {
         for (; i + 16 <= n; i += 16) {
             _mm512_storeu_ps(y + i,
                             _mm512_castsi512_ps(
@@ -394,10 +410,10 @@ void ggml_bf16_to_fp32_row(const ggml_bf16_t * x, float * y, int64_t n) {
                                             (const __m256i *)(x + i))),
                                     16)));
         }
-    }
+    //}
 #endif
 #if defined(__AVX2__)
-    if (ggml_cpu_has_avx2()) {
+    //if (ggml_cpu_has_avx2()) {
         for (; i + 8 <= n; i += 8) {
             _mm256_storeu_ps(y + i,
                             _mm256_castsi256_ps(
@@ -407,7 +423,7 @@ void ggml_bf16_to_fp32_row(const ggml_bf16_t * x, float * y, int64_t n) {
                                             (const __m128i *)(x + i))),
                                     16)));
         }
-    }
+    //}
 #endif
     for (; i < n; i++) {
         y[i] = GGML_BF16_TO_FP32(x[i]);
@@ -588,7 +604,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(ggml_fp16_t),
         .is_quantized             = false,
         .to_float                 = (ggml_to_float_t) ggml_fp16_to_fp32_row,
-        .from_float               = (ggml_from_float_t) ggml_fp32_to_fp16_row,
         .from_float_ref           = (ggml_from_float_t) ggml_fp32_to_fp16_row,
     },
     [GGML_TYPE_Q4_0] = {
@@ -597,7 +612,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_q4_0),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q4_0,
-        .from_float               = quantize_row_q4_0,
         .from_float_ref           = (ggml_from_float_t) quantize_row_q4_0_ref,
     },
     [GGML_TYPE_Q4_1] = {
@@ -606,7 +620,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_q4_1),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q4_1,
-        .from_float               = quantize_row_q4_1,
         .from_float_ref           = (ggml_from_float_t) quantize_row_q4_1_ref,
     },
     [4] = { // GGML_TYPE_Q4_2
@@ -614,18 +627,12 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .blck_size                = 0,
         .type_size                = 0,
         .is_quantized             = false,
-        .to_float                 = NULL,
-        .from_float               = NULL,
-        .from_float_ref           = NULL,
     },
     [5] = { // GGML_TYPE_Q4_3
         .type_name                = "DEPRECATED",
         .blck_size                = 0,
         .type_size                = 0,
         .is_quantized             = false,
-        .to_float                 = NULL,
-        .from_float               = NULL,
-        .from_float_ref           = NULL,
     },
     [GGML_TYPE_Q5_0] = {
         .type_name                = "q5_0",
@@ -633,7 +640,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_q5_0),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q5_0,
-        .from_float               = quantize_row_q5_0,
         .from_float_ref           = (ggml_from_float_t) quantize_row_q5_0_ref,
     },
     [GGML_TYPE_Q5_1] = {
@@ -642,7 +648,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_q5_1),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q5_1,
-        .from_float               = quantize_row_q5_1,
         .from_float_ref           = (ggml_from_float_t) quantize_row_q5_1_ref,
     },
     [GGML_TYPE_Q8_0] = {
@@ -651,7 +656,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_q8_0),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q8_0,
-        .from_float               = quantize_row_q8_0,
         .from_float_ref           = (ggml_from_float_t) quantize_row_q8_0_ref,
     },
     [GGML_TYPE_Q8_1] = {
@@ -659,7 +663,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .blck_size                = QK8_1,
         .type_size                = sizeof(block_q8_1),
         .is_quantized             = true,
-        .from_float               = quantize_row_q8_1,
         .from_float_ref           = (ggml_from_float_t) quantize_row_q8_1_ref,
     },
     [GGML_TYPE_Q2_K] = {
@@ -668,7 +671,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_q2_K),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q2_K,
-        .from_float               = quantize_row_q2_K,
         .from_float_ref           = (ggml_from_float_t) quantize_row_q2_K_ref,
     },
     [GGML_TYPE_Q3_K] = {
@@ -677,7 +679,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_q3_K),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q3_K,
-        .from_float               = quantize_row_q3_K,
         .from_float_ref           = (ggml_from_float_t) quantize_row_q3_K_ref,
     },
     [GGML_TYPE_Q4_K] = {
@@ -686,7 +687,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_q4_K),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q4_K,
-        .from_float               = quantize_row_q4_K,
         .from_float_ref           = (ggml_from_float_t) quantize_row_q4_K_ref,
     },
     [GGML_TYPE_Q5_K] = {
@@ -695,7 +695,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_q5_K),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q5_K,
-        .from_float               = quantize_row_q5_K,
         .from_float_ref           = (ggml_from_float_t) quantize_row_q5_K_ref,
     },
     [GGML_TYPE_Q6_K] = {
@@ -704,7 +703,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_q6_K),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q6_K,
-        .from_float               = quantize_row_q6_K,
         .from_float_ref           = (ggml_from_float_t) quantize_row_q6_K_ref,
     },
     [GGML_TYPE_IQ2_XXS] = {
@@ -713,7 +711,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_iq2_xxs),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_iq2_xxs,
-        .from_float               = NULL,
         .from_float_ref           = NULL,
     },
     [GGML_TYPE_IQ2_XS] = {
@@ -722,7 +719,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_iq2_xs),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_iq2_xs,
-        .from_float               = NULL,
         .from_float_ref           = NULL,
     },
     [GGML_TYPE_IQ3_XXS] = {
@@ -731,7 +727,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_iq3_xxs),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_iq3_xxs,
-        .from_float               = quantize_row_iq3_xxs,
         .from_float_ref           = (ggml_from_float_t)quantize_row_iq3_xxs_ref,
     },
     [GGML_TYPE_IQ3_S] = {
@@ -740,7 +735,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_iq3_s),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_iq3_s,
-        .from_float               = quantize_row_iq3_s,
         .from_float_ref           = (ggml_from_float_t)quantize_row_iq3_s_ref,
     },
     [GGML_TYPE_IQ2_S] = {
@@ -749,7 +743,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_iq2_s),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_iq2_s,
-        .from_float               = quantize_row_iq2_s,
         .from_float_ref           = (ggml_from_float_t)quantize_row_iq2_s_ref,
     },
     [GGML_TYPE_IQ1_S] = {
@@ -758,7 +751,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_iq1_s),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_iq1_s,
-        .from_float               = NULL,
         .from_float_ref           = NULL,
     },
     [GGML_TYPE_IQ1_M] = {
@@ -767,7 +759,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_iq1_m),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_iq1_m,
-        .from_float               = NULL,
         .from_float_ref           = NULL,
     },
     [GGML_TYPE_IQ4_NL] = {
@@ -776,7 +767,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_iq4_nl),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_iq4_nl,
-        .from_float               = quantize_row_iq4_nl,
         .from_float_ref           = (ggml_from_float_t)quantize_row_iq4_nl_ref,
     },
     [GGML_TYPE_IQ4_XS] = {
@@ -785,7 +775,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_iq4_xs),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_iq4_xs,
-        .from_float               = quantize_row_iq4_xs,
         .from_float_ref           = (ggml_from_float_t)quantize_row_iq4_xs_ref,
     },
     [GGML_TYPE_Q8_K] = {
@@ -793,7 +782,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .blck_size                = QK_K,
         .type_size                = sizeof(block_q8_K),
         .is_quantized             = true,
-        .from_float               = quantize_row_q8_K,
     },
     [GGML_TYPE_BF16] = {
         .type_name                = "bf16",
@@ -801,38 +789,25 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(ggml_bf16_t),
         .is_quantized             = false,
         .to_float                 = (ggml_to_float_t) ggml_bf16_to_fp32_row,
-        .from_float               = (ggml_from_float_t) ggml_fp32_to_bf16_row,
         .from_float_ref           = (ggml_from_float_t) ggml_fp32_to_bf16_row_ref,
     },
-    [GGML_TYPE_Q4_0_4_4] = {
-        .type_name                = "q4_0_4x4",
-        .blck_size                = QK4_0,
-        .blck_size_interleave     = 4,
-        .type_size                = sizeof(block_q4_0),
-        .is_quantized             = true,
-        .to_float                 = NULL,
-        .from_float               = NULL,
-        .from_float_ref           = NULL,
+    [31] = { // GGML_TYPE_Q4_0_4_4
+        .type_name                = "TYPE_Q4_0_4_4 REMOVED, use Q4_0 with runtime repacking",
+        .blck_size                = 0,
+        .type_size                = 0,
+        .is_quantized             = false,
     },
-    [GGML_TYPE_Q4_0_4_8] = {
-        .type_name                = "q4_0_4x8",
-        .blck_size                = QK4_0,
-        .blck_size_interleave     = 8,
-        .type_size                = sizeof(block_q4_0),
-        .is_quantized             = true,
-        .to_float                 = NULL,
-        .from_float               = NULL,
-        .from_float_ref           = NULL,
+    [32] = { // GGML_TYPE_Q4_0_4_8
+        .type_name                = "TYPE_Q4_0_4_8 REMOVED, use Q4_0 with runtime repacking",
+        .blck_size                = 0,
+        .type_size                = 0,
+        .is_quantized             = false,
     },
-    [GGML_TYPE_Q4_0_8_8] = {
-        .type_name                = "q4_0_8x8",
-        .blck_size                = QK4_0,
-        .blck_size_interleave     = 8,
-        .type_size                = sizeof(block_q4_0),
-        .is_quantized             = true,
-        .to_float                 = NULL,
-        .from_float               = NULL,
-        .from_float_ref           = NULL,
+    [33] = { // GGML_TYPE_Q4_0_8_8
+        .type_name                = "TYPE_Q4_0_8_8 REMOVED, use Q4_0 with runtime repacking",
+        .blck_size                = 0,
+        .type_size                = 0,
+        .is_quantized             = false,
     },
     [GGML_TYPE_TQ1_0] = {
         .type_name                = "tq1_0",
@@ -840,7 +815,6 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_tq1_0),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_tq1_0,
-        .from_float               = quantize_row_tq1_0,
         .from_float_ref           = (ggml_from_float_t) quantize_row_tq1_0_ref,
     },
     [GGML_TYPE_TQ2_0] = {
@@ -849,9 +823,26 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_tq2_0),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_tq2_0,
-        .from_float               = quantize_row_tq2_0,
         .from_float_ref           = (ggml_from_float_t) quantize_row_tq2_0_ref,
     },
+    [36] = { // GGML_TYPE_IQ4_NL_4_4
+        .type_name                = "TYPE_IQ4_NL_4_4 REMOVED, use IQ4_NL with runtime repacking",
+        .blck_size                = 0,
+        .type_size                = 0,
+        .is_quantized             = false,
+    },
+    [37] = { // GGML_TYPE_IQ4_NL_4_8
+        .type_name                = "TYPE_IQ4_NL_4_8 REMOVED, use IQ4_NL with runtime repacking",
+        .blck_size                = 0,
+        .type_size                = 0,
+        .is_quantized             = false,
+    },
+    [38] = { // GGML_TYPE_IQ4_NL_8_8
+        .type_name                = "TYPE_IQ4_NL_8_8 REMOVED, use IQ4_NL with runtime repacking",
+        .blck_size                = 0,
+        .type_size                = 0,
+        .is_quantized             = false,
+    },
 };
 
 const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type) {
@@ -962,6 +953,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "POOL_2D_BACK",
     "UPSCALE",
     "PAD",
+    "PAD_REFLECT_1D",
     "ARANGE",
     "TIMESTEP_EMBEDDING",
     "ARGSORT",
@@ -995,7 +987,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "OPT_STEP_ADAMW",
 };
 
-static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
+static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1057,6 +1049,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "pool_2d_back(x)",
     "upscale(x)",
     "pad(x)",
+    "pad_reflect_1d(x)",
     "arange(start, stop, step)",
     "timestep_embedding(timesteps, dim, max_period)",
     "argsort(x)",
@@ -1090,7 +1083,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "adamw(x)",
 };
 
-static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
+static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -1280,9 +1273,6 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_IQ4_XS:        wtype = GGML_TYPE_IQ4_XS;   break;
         case GGML_FTYPE_MOSTLY_IQ3_S:         wtype = GGML_TYPE_IQ3_S;    break;
         case GGML_FTYPE_MOSTLY_IQ2_S:         wtype = GGML_TYPE_IQ2_S;    break;
-        case GGML_FTYPE_MOSTLY_Q4_0_4_4:      wtype = GGML_TYPE_Q4_0_4_4; break;
-        case GGML_FTYPE_MOSTLY_Q4_0_4_8:      wtype = GGML_TYPE_Q4_0_4_8; break;
-        case GGML_FTYPE_MOSTLY_Q4_0_8_8:      wtype = GGML_TYPE_Q4_0_8_8; break;
         case GGML_FTYPE_UNKNOWN:              wtype = GGML_TYPE_COUNT; break;
         case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
     }
@@ -1613,14 +1603,13 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         /*.op           =*/ GGML_OP_NONE,
         /*.op_params    =*/ { 0 },
         /*.flags        =*/ 0,
-        /*.grad         =*/ NULL,
         /*.src          =*/ { NULL },
         /*.view_src     =*/ view_src,
         /*.view_offs    =*/ view_offs,
         /*.data         =*/ obj_alloc_size > 0 ? (void *)(result + 1) : data,
         /*.name         =*/ { 0 },
         /*.extra        =*/ NULL,
-        ///*.padding      =*/ { 0 },
+        /*.padding      =*/ { 0 },
     };
 
 #ifdef __clang__
@@ -2277,6 +2266,7 @@ struct ggml_tensor * ggml_argmax(
         struct ggml_context * ctx,
         struct ggml_tensor  * a) {
     GGML_ASSERT(ggml_is_matrix(a));
+    GGML_ASSERT(a->ne[0] <= INT32_MAX);
 
     struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, a->ne[1]);
 
@@ -3527,15 +3517,18 @@ static struct ggml_tensor * ggml_rope_impl(
         GGML_ASSERT(c->ne[0] >= n_dims / 2);
     }
 
+    int sections[4] = {0, 0, 0, 0};
+
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    int32_t params[11] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
+    int32_t params[15] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
     memcpy(params +  5, &freq_base,    sizeof(float));
     memcpy(params +  6, &freq_scale,   sizeof(float));
     memcpy(params +  7, &ext_factor,   sizeof(float));
     memcpy(params +  8, &attn_factor,  sizeof(float));
     memcpy(params +  9, &beta_fast,    sizeof(float));
     memcpy(params + 10, &beta_slow,    sizeof(float));
+    memcpy(params + 11, &sections,     sizeof(int)*4);
     ggml_set_op_params(result, params, sizeof(params));
 
     result->op     = GGML_OP_ROPE;
@@ -3557,6 +3550,53 @@ struct ggml_tensor * ggml_rope(
     );
 }
 
+struct ggml_tensor * ggml_rope_multi(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        struct ggml_tensor  * c,
+        int                   n_dims,
+        int                   sections[4],
+        int                   mode,
+        int                   n_ctx_orig,
+        float                 freq_base,
+        float                 freq_scale,
+        float                 ext_factor,
+        float                 attn_factor,
+        float                 beta_fast,
+        float                 beta_slow) {
+    // Multimodal Rotary Position Embedding
+    GGML_ASSERT((mode & 1) == 0 && "mode & 1 == 1 is no longer supported");
+
+    GGML_ASSERT(ggml_is_vector(b));
+    GGML_ASSERT(b->type == GGML_TYPE_I32);
+    GGML_ASSERT(a->ne[2] * 4 == b->ne[0]); // mrope expecting 4 position ids per token
+
+    if (c) {
+        GGML_ASSERT(c->type == GGML_TYPE_F32);
+        GGML_ASSERT(c->ne[0] >= n_dims / 2);
+    }
+
+    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
+
+    int32_t params[11 + 4] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
+    memcpy(params +  5, &freq_base,    sizeof(float));
+    memcpy(params +  6, &freq_scale,   sizeof(float));
+    memcpy(params +  7, &ext_factor,   sizeof(float));
+    memcpy(params +  8, &attn_factor,  sizeof(float));
+    memcpy(params +  9, &beta_fast,    sizeof(float));
+    memcpy(params + 10, &beta_slow,    sizeof(float));
+    memcpy(&params[11], sections,      sizeof(int)*4);
+    ggml_set_op_params(result, params, sizeof(params));
+
+    result->op   = GGML_OP_ROPE;
+    result->src[0] = a;
+    result->src[1] = b;
+    result->src[2] = c;
+
+    return result;
+}
+
 struct ggml_tensor * ggml_rope_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -3646,6 +3686,22 @@ struct ggml_tensor * ggml_rope_custom_inplace(
     );
 }
 
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_dim(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+static float ggml_rope_yarn_corr_dim(int n_dims, int n_ctx_orig, float n_rot, float base) {
+    return n_dims * logf(n_ctx_orig / (n_rot * 2 * (float)M_PI)) / (2 * logf(base));
+}
+
+void ggml_rope_yarn_corr_dims(
+    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]
+) {
+    // start and end correction dims
+    float start = floorf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_fast, freq_base));
+    float end   =  ceilf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_slow, freq_base));
+    dims[0] = MAX(0, start);
+    dims[1] = MIN(n_dims - 1, end);
+}
+
 // ggml_rope_back
 
 struct ggml_tensor * ggml_rope_back(
@@ -4093,6 +4149,37 @@ struct ggml_tensor * ggml_pad(
     return result;
 }
 
+// ggml_pad_reflect_1d
+
+struct ggml_tensor * ggml_pad_reflect_1d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   p0,
+        int                   p1) {
+    GGML_ASSERT(p0 >= 0);
+    GGML_ASSERT(p1 >= 0);
+
+    GGML_ASSERT(p0 < a->ne[0]); // padding length on each size must be less than the
+    GGML_ASSERT(p1 < a->ne[0]); // existing length of the dimension being padded
+
+    GGML_ASSERT(ggml_is_contiguous(a));
+    GGML_ASSERT(a->type == GGML_TYPE_F32);
+
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type,
+            a->ne[0] + p0 + p1,
+            a->ne[1],
+            a->ne[2],
+            a->ne[3]);
+
+    int32_t params[] = { p0, p1 };
+    ggml_set_op_params(result, params, sizeof(params));
+
+    result->op     = GGML_OP_PAD_REFLECT_1D;
+    result->src[0] = a;
+
+    return result;
+}
+
 // ggml_arange
 
 struct ggml_tensor * ggml_arange(
@@ -4144,6 +4231,7 @@ struct ggml_tensor * ggml_argsort(
         struct ggml_context  * ctx,
         struct ggml_tensor   * a,
         enum ggml_sort_order   order) {
+    GGML_ASSERT(a->ne[0] <= INT32_MAX);
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne);
 
     ggml_set_op_params_i32(result, 0, (int32_t) order);
@@ -4199,8 +4287,6 @@ struct ggml_tensor * ggml_flash_attn_ext(
         GGML_ASSERT(mask);
     }
 
-    bool is_node = false;
-
     // permute(0, 2, 1, 3)
     int64_t ne[4] = { q->ne[0], q->ne[2], q->ne[1], q->ne[3] };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
@@ -4208,8 +4294,7 @@ struct ggml_tensor * ggml_flash_attn_ext(
     float params[] = { scale, max_bias, logit_softcap };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op   = GGML_OP_FLASH_ATTN_EXT;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_FLASH_ATTN_EXT;
     result->src[0] = q;
     result->src[1] = k;
     result->src[2] = v;
@@ -4277,14 +4362,6 @@ struct ggml_tensor * ggml_flash_attn_back(
 
     GGML_ASSERT(ne2 % kvne2 == 0);
 
-    bool is_node = false;
-
-    if (q->grad || k->grad || v->grad) {
-        // when using this operation (in backwards pass) these grads are set.
-        // we don't want to create (big) grad of our result, so is_node is false.
-        is_node = false;
-    }
-
     // store gradients of q, k and v as continuous tensors concatenated in result.
     // note: v and gradv are actually transposed, i.e. v->ne[0] != D.
     const int64_t elem_q = ggml_nelements(q);
@@ -4307,8 +4384,7 @@ struct ggml_tensor * ggml_flash_attn_back(
     int32_t masked_i = masked ? 1 : 0;
     ggml_set_op_params(result, &masked_i, sizeof(masked_i));
 
-    result->op   = GGML_OP_FLASH_ATTN_BACK;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_FLASH_ATTN_BACK;
     result->src[0] = q;
     result->src[1] = k;
     result->src[2] = v;
@@ -4950,34 +5026,24 @@ struct ggml_tensor * ggml_opt_step_adamw(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * grad,
-        float                 alpha,
-        float                 beta1,
-        float                 beta2,
-        float                 eps,
-        float                 wd) {
+        struct ggml_tensor  * m,
+        struct ggml_tensor  * v,
+        struct ggml_tensor  * adamw_params) {
     GGML_ASSERT(a->flags & GGML_TENSOR_FLAG_PARAM);
     GGML_ASSERT(ggml_are_same_shape(a, grad));
-    GGML_ASSERT(alpha >  0.0f);
-    GGML_ASSERT(beta1 >= 0.0f && beta1 <= 1.0f);
-    GGML_ASSERT(beta2 >= 0.0f && beta2 <= 1.0f);
-    GGML_ASSERT(eps   >= 0.0f);
-    GGML_ASSERT(wd    >= 0.0f && wd    <= 1.0f);
+    GGML_ASSERT(ggml_are_same_shape(a, m));
+    GGML_ASSERT(ggml_are_same_shape(a, v));
+    GGML_ASSERT(adamw_params->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_nelements(adamw_params) == 7);
 
     struct ggml_tensor * result = ggml_view_tensor(ctx, a);
 
-    const int64_t iter = 1;
-    memcpy(&result->op_params[0], &iter, sizeof(int64_t));
-    ggml_set_op_params_f32(result, 2, alpha);
-    ggml_set_op_params_f32(result, 3, beta1);
-    ggml_set_op_params_f32(result, 4, beta2);
-    ggml_set_op_params_f32(result, 5, eps);
-    ggml_set_op_params_f32(result, 6, wd);
-
     result->op     = GGML_OP_OPT_STEP_ADAMW;
     result->src[0] = a;
     result->src[1] = grad;
-    result->src[2] = ggml_dup_tensor(ctx, grad);
-    result->src[3] = ggml_dup_tensor(ctx, grad);
+    result->src[2] = m;
+    result->src[3] = v;
+    result->src[4] = adamw_params;
 
     return result;
 }
@@ -5046,1112 +5112,526 @@ static void ggml_hash_map_free(struct hash_map * map) {
     GGML_FREE(map);
 }
 
-// gradient checkpointing
+// utility functions to change gradients
+// isrc is the index of tensor in cgraph->visited_has_set.keys
+// the corresponding gradient (accumulators) are also at position isrc
+// if tensor has a gradient accumulator, modify that accumulator in-place
+// else if there is no gradient for tensor, set the corresponding value
+// else, just add/subtract/etc. the gradients
 
-static struct ggml_tensor * ggml_recompute_graph_node(
+static void ggml_add_or_set(
         struct ggml_context * ctx,
-        struct ggml_cgraph  * graph,
-        struct hash_map     * replacements,
-        struct ggml_tensor  * node) {
-
-    if (node == NULL) {
-        return NULL;
-    }
-
-    if (node->flags & GGML_TENSOR_FLAG_PARAM) {
-        return node;
-    }
-
-    if (!ggml_hash_contains(&graph->visited_hash_set, node)) {
-        return node;
-    }
-
-    int count_children = 0;
-    for (int k = 0; k < GGML_MAX_SRC; ++k) {
-        if (node->src[k]) {
-            ++count_children;
-        }
-    }
-
-    if (count_children == 0) {
-        return node;
-    }
-
-    size_t i = ggml_hash_find(&replacements->set, node);
-    GGML_ASSERT(i != GGML_HASHSET_FULL); // assert that not full
-    if (replacements->set.keys[i] == node) {
-        return replacements->vals[i];
-    }
-
-    struct ggml_tensor * clone = ggml_new_tensor(ctx, node->type, GGML_MAX_DIMS, node->ne);
-
-    // insert clone into replacements
-    GGML_ASSERT(replacements->set.keys[i] == NULL); // assert that we don't overwrite
-    replacements->set.keys[i] = node;
-    replacements->vals[i] = clone;
-
-    clone->op       = node->op;
-    clone->grad     = node->grad;
-    clone->flags    = node->flags;
-    clone->extra    = node->extra;
-    for (int k = 0; k < GGML_MAX_DIMS; ++k) {
-        clone->nb[k] = node->nb[k];
-    }
-    for (int k = 0; k < GGML_MAX_SRC; ++k) {
-        clone->src[k] = ggml_recompute_graph_node(ctx, graph, replacements, node->src[k]);
-    }
-    if (node->view_src != NULL) {
-        clone->data = (node->view_src->data == NULL)
-                        ? NULL // view_src not yet allocated
-                        : (char *) node->view_src->data // view_src already allocated
-                                 + node->view_offs;
-        clone->view_src  = node->view_src;
-        clone->view_offs = node->view_offs;
+        struct ggml_cgraph  * cgraph,
+        size_t                isrc,
+        struct ggml_tensor  * tensor) {
+    struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
+    GGML_ASSERT(src);
+    if (cgraph->grads[isrc]) {
+        cgraph->grads[isrc] = ggml_add_impl(ctx, cgraph->grads[isrc], tensor, /*inplace =*/ cgraph->grad_accs[isrc]);
+    } else {
+        cgraph->grads[isrc] = tensor;
     }
-
-    GGML_ASSERT(sizeof(node->op_params) == sizeof(int32_t) * (GGML_MAX_OP_PARAMS / sizeof(int32_t)));
-    GGML_ASSERT(sizeof(node->name)      == GGML_MAX_NAME);
-    memcpy(clone->op_params, node->op_params, sizeof(node->op_params));
-    ggml_format_name(clone, "%s (clone)", ggml_get_name(node));
-
-    return clone;
+    ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name);
+    ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
 }
 
-void ggml_build_backward_gradient_checkpointing(
-        struct ggml_context   * ctx,
-        struct ggml_cgraph    * gf,
-        struct ggml_cgraph    * gb,
-        struct ggml_cgraph    * gb_tmp,
-        struct ggml_tensor  * * checkpoints,
-        int                     n_checkpoints) {
-    ggml_graph_cpy(gf, gb_tmp);
-    ggml_build_backward_expand(ctx, gf, gb_tmp, false);
-
-    if (n_checkpoints <= 0) {
-        ggml_graph_cpy(gb_tmp, gb);
-        return;
-    }
-
-    struct hash_map * replacements = ggml_new_hash_map(gf->n_nodes + gf->n_leafs + n_checkpoints);
-
-    // insert checkpoints in replacements
-    for (int i = 0; i < n_checkpoints; ++i) {
-        size_t k = ggml_hash_find(&replacements->set, checkpoints[i]);
-        GGML_ASSERT(k != GGML_HASHSET_FULL); // assert that not full
-        GGML_ASSERT(replacements->set.keys[k] == NULL); // assert that we don't overwrite
-        replacements->set.keys[k] = checkpoints[i];
-        replacements->vals[k]     = checkpoints[i];
-    }
-
-    ggml_graph_cpy(gf, gb);
-    // rewrite gb_tmp->nodes[gf->n_nodes:gb_tmp->n_nodes],
-    // replacing references to gb_tmp->nodes[0:gf->n_nodes] ( == gf->nodes[0:gf->n_nodes]),
-    // by recomputing them from checkpoints
-    for (int i = gf->n_nodes; i<gb_tmp->n_nodes; ++i) {
-        struct ggml_tensor * node = gb_tmp->nodes[i];
-        for (int k = 0; k < GGML_MAX_SRC; ++k) {
-            // insert new tensors recomputing src, reusing already made replacements,
-            // remember replacements: remember new tensors with mapping from corresponding gf nodes
-            // recurse for input tensors,
-            // unless (i.e. terminating when) input tensors are replacements (like checkpoints)
-            node->src[k] = ggml_recompute_graph_node(ctx, gf, replacements, node->src[k]);
-        }
-        // insert rewritten backward node with replacements made into resulting backward graph gb
-        ggml_build_forward_expand(gb, node);
+static void ggml_acc_or_set(
+        struct ggml_context * ctx,
+        struct ggml_cgraph  * cgraph,
+        size_t                isrc,
+        struct ggml_tensor  * tensor,
+        const  size_t         nb1,
+        const  size_t         nb2,
+        const  size_t         nb3,
+        const  size_t         offset) {
+    struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
+    GGML_ASSERT(src);
+    if (cgraph->grads[isrc]) {
+        cgraph->grads[isrc] = ggml_acc_impl(ctx, cgraph->grads[isrc], tensor, nb1, nb2, nb3, offset, cgraph->grad_accs[isrc]);
+    } else {
+        struct ggml_tensor * a_zero = ggml_scale(ctx, src, 0.0f); // FIXME this is going to produce NaN if a contains inf/NaN
+        cgraph->grads[isrc] = ggml_acc_impl(ctx, a_zero, tensor, nb1, nb2, nb3, offset, false);
     }
-
-    ggml_hash_map_free(replacements);
+    ggml_format_name(cgraph->grads[isrc], "grad for %s", cgraph->visited_hash_set.keys[isrc]->name);
+    ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
 }
 
-// utility functions to change gradients
-// if a is in acc_table, modify gradients in-place and mark result as gradient accumulator
-// else if a is in zero_table, replace a
-// else, just add/subtract/etc. the gradients
-
-static struct ggml_tensor * ggml_add_or_set(
-        struct ggml_context  * ctx,
-        struct ggml_tensor   * a,
-        struct ggml_tensor   * b,
-        struct ggml_hash_set * zero_table,
-        struct ggml_hash_set * acc_table) {
-    if (ggml_hash_contains(acc_table, a)) {
-        struct ggml_tensor * ret = ggml_add_impl(ctx, a, b, true);
-        const size_t insert_result = ggml_hash_insert(acc_table, ret);
-        GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
-        GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
-        return ret;
-    }
-    if (ggml_hash_contains(zero_table, a)) {
-        return b;
+static void ggml_add1_or_set(
+        struct ggml_context * ctx,
+        struct ggml_cgraph  * cgraph,
+        size_t                isrc,
+        struct ggml_tensor  * tensor) {
+    struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
+    GGML_ASSERT(src);
+    if (cgraph->grads[isrc]) {
+        cgraph->grads[isrc] = ggml_add1_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]);
+    } else {
+        cgraph->grads[isrc] = ggml_repeat(ctx, tensor, src);
     }
-    return ggml_add_impl(ctx, a, b, false);
+    ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name);
+    ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
 }
 
-static struct ggml_tensor * ggml_acc_or_set(
-        struct ggml_context  * ctx,
-        struct ggml_tensor   * a,
-        struct ggml_tensor   * b,
-        const  size_t          nb1,
-        const  size_t          nb2,
-        const  size_t          nb3,
-        const  size_t          offset,
-        struct ggml_hash_set * zero_table,
-        struct ggml_hash_set * acc_table) {
-    if (ggml_hash_contains(acc_table, a)) {
-        struct ggml_tensor * ret = ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, true);
-        const size_t insert_result = ggml_hash_insert(acc_table, ret);
-        GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
-        GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
-        return ret;
-    }
-    if (ggml_hash_contains(zero_table, a)) {
-        struct ggml_tensor * a_zero = ggml_scale(ctx, a, 0.0f); // FIXME this is going to produce NaN if a contains inf/NaN
-        return ggml_acc_impl(ctx, a_zero, b, nb1, nb2, nb3, offset, false);
+static void ggml_sub_or_set(
+        struct ggml_context * ctx,
+        struct ggml_cgraph  * cgraph,
+        size_t                isrc,
+        struct ggml_tensor  * tensor) {
+    struct ggml_tensor * src = cgraph->visited_hash_set.keys[isrc];
+    GGML_ASSERT(src);
+    if (cgraph->grads[isrc]) {
+        cgraph->grads[isrc] = ggml_sub_impl(ctx, cgraph->grads[isrc], tensor, cgraph->grad_accs[isrc]);
+    } else {
+        cgraph->grads[isrc] = ggml_neg(ctx, tensor);
     }
-    return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, false);
+    ggml_format_name(cgraph->grads[isrc], "grad for %s", src->name);
+    ggml_build_forward_expand(cgraph, cgraph->grads[isrc]);
 }
 
-static struct ggml_tensor * ggml_add1_or_set(
-        struct ggml_context  * ctx,
-        struct ggml_tensor   * a,
-        struct ggml_tensor   * b,
-        struct ggml_hash_set * zero_table,
-        struct ggml_hash_set * acc_table) {
-    if (ggml_hash_contains(acc_table, a)) {
-        struct ggml_tensor * ret = ggml_add1_impl(ctx, a, b, true);
-        const size_t insert_result = ggml_hash_insert(acc_table, ret);
-        GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
-        GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
-        return ret;
-    }
-    if (ggml_hash_contains(zero_table, a)) {
-        return ggml_repeat(ctx, b, a);
-    }
-    return ggml_add1_impl(ctx, a, b, false);
-}
+static void ggml_compute_backward(
+        struct ggml_context * ctx, struct ggml_cgraph * cgraph, int i, bool * grads_needed) {
+    struct ggml_tensor * tensor = cgraph->nodes[i];
+    struct ggml_tensor * grad   = ggml_graph_get_grad(cgraph, tensor);
 
-static struct ggml_tensor * ggml_sub_or_set(
-        struct ggml_context  * ctx,
-        struct ggml_tensor   * a,
-        struct ggml_tensor   * b,
-        struct ggml_hash_set * zero_table,
-        struct ggml_hash_set * acc_table) {
-    if (ggml_hash_contains(acc_table, a)) {
-        struct ggml_tensor * ret = ggml_sub_impl(ctx, a, b, true);
-        const size_t insert_result = ggml_hash_insert(acc_table, ret);
-        GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
-        GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
-        return ret;
-    }
-    if (ggml_hash_contains(zero_table, a)) {
-        return ggml_neg(ctx, b);
+    if (!grad) {
+        return;
     }
-    return ggml_sub_impl(ctx, a, b, false);
-}
 
-static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, struct ggml_hash_set * zero_table, struct ggml_hash_set * acc_table) {
     struct ggml_tensor * src0 = tensor->src[0];
     struct ggml_tensor * src1 = tensor->src[1];
     struct ggml_tensor * src2 = tensor->src[2];
+    struct ggml_hash_set * hash_set = &cgraph->visited_hash_set;
+    const size_t isrc0 = src0 ? ggml_hash_find(hash_set, src0) : (size_t) -1;
+    const size_t isrc1 = src1 ? ggml_hash_find(hash_set, src1) : (size_t) -1;
+    const size_t isrc2 = src2 ? ggml_hash_find(hash_set, src2) : (size_t) -1;
+    const bool src0_needs_grads = src0 && isrc0 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc0) && grads_needed[isrc0];
+    const bool src1_needs_grads = src1 && isrc1 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc1) && grads_needed[isrc1];
+    const bool src2_needs_grads = src2 && isrc2 != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, isrc2) && grads_needed[isrc2];
 
     switch (tensor->op) {
-        case GGML_OP_DUP:
-            {
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_ADD:
-            {
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    if (ggml_are_same_shape(src0, src1)) {
-                        src1->grad = ggml_add_or_set(ctx, src1->grad,                       tensor->grad,        zero_table, acc_table);
-                    } else {
-                        src1->grad = ggml_add_or_set(ctx, src1->grad, ggml_repeat_back(ctx, tensor->grad, src1), zero_table, acc_table);
-                    }
-                }
-            } break;
-        case GGML_OP_ADD1:
-            {
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    src1->grad = ggml_add_or_set(ctx,
-                        src1->grad,
-                        ggml_mean(ctx, tensor->grad), // TODO: should probably be sum instead of mean
-                        zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_ACC:
-            {
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    const size_t nb1     = ((int32_t *) tensor->op_params)[0];
-                    const size_t nb2     = ((int32_t *) tensor->op_params)[1];
-                    const size_t nb3     = ((int32_t *) tensor->op_params)[2];
-                    const size_t offset  = ((int32_t *) tensor->op_params)[3];
-
-                    struct ggml_tensor * tensor_grad_view = ggml_view_4d(ctx,
-                        tensor->grad,
-                        src1->grad->ne[0],
-                        src1->grad->ne[1],
-                        src1->grad->ne[2],
-                        src1->grad->ne[3],
-                        nb1, nb2, nb3, offset);
-
-                    src1->grad =
-                        ggml_add_or_set(ctx,
-                            src1->grad,
-                            ggml_reshape(ctx,
-                                ggml_cont(ctx, tensor_grad_view),
-                                src1->grad),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SUB:
-            {
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    src1->grad = ggml_sub_or_set(ctx, src1->grad, tensor->grad, zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_MUL:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_mul(ctx, src1, tensor->grad),
-                                zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    src1->grad =
-                        ggml_add_or_set(ctx,
-                                src1->grad,
-                                ggml_mul(ctx, src0, tensor->grad),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_DIV:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_div(ctx, tensor->grad, src1),
-                                zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    src1->grad =
-                        ggml_sub_or_set(ctx,
-                                src1->grad,
-                                ggml_mul(ctx,
-                                    tensor->grad,
-                                    ggml_div(ctx, tensor, src1)),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SQR:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_scale(ctx,
-                                    ggml_mul(ctx, src0, tensor->grad),
-                                    2.0f),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SQRT:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_scale(ctx,
-                                    ggml_div(ctx,
-                                        tensor->grad,
-                                        tensor),
-                                    0.5f),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_LOG:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_div(ctx,
-                                    tensor->grad,
-                                    src0),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SIN:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_mul(ctx,
-                                    tensor->grad,
-                                    ggml_cos(ctx, src0)),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_COS:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_sub_or_set(ctx,
-                                src0->grad,
-                                ggml_mul(ctx,
-                                    tensor->grad,
-                                    ggml_sin(ctx, src0)),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SUM:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add1_or_set(ctx,
-                                src0->grad,
-                                tensor->grad,
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SUM_ROWS:
-            {
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_repeat(ctx,
-                                    tensor->grad,
-                                    src0->grad),
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_MEAN:
-        case GGML_OP_ARGMAX:
-        case GGML_OP_COUNT_EQUAL:
-            {
-                GGML_ABORT("fatal error"); // TODO: implement
-            }
-        case GGML_OP_REPEAT:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx,
-                            src0->grad,
-                            ggml_repeat_back(ctx, tensor->grad, src0->grad),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_REPEAT_BACK:
-            {
-                if (src0->grad) {
-                    // TODO: test this
-                    src0->grad = ggml_add_or_set(ctx,
-                            src0->grad,
-                            ggml_repeat(ctx, tensor->grad, src0->grad),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_CONCAT:
-            {
-                GGML_ABORT("fatal error"); // TODO: implement
-            }
-        case GGML_OP_SILU_BACK:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        case GGML_OP_DUP: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, grad);
             }
-        case GGML_OP_NORM:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_ADD: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, grad);
             }
-        case GGML_OP_RMS_NORM:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    float eps;
-                    memcpy(&eps, tensor->op_params, sizeof(float));
-
-                    src0->grad = ggml_add_or_set(ctx,
-                            src0->grad,
-                            ggml_rms_norm_back(ctx, src0, tensor->grad, eps),
-                            zero_table, acc_table);
+            if (src1_needs_grads) {
+                struct ggml_tensor * tmp = grad;
+                if (!ggml_are_same_shape(src0, src1)) {
+                    tmp = ggml_repeat_back(ctx, tmp, src1);
                 }
-            } break;
-        case GGML_OP_RMS_NORM_BACK:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+                ggml_add_or_set(ctx, cgraph, isrc1, tmp);
             }
-        case GGML_OP_GROUP_NORM:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_ADD1: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, grad);
             }
-        case GGML_OP_MUL_MAT:
-            {
-                // https://cs231n.github.io/optimization-2/#staged
-                // # forward pass
-                // s0 = np.random.randn(5, 10)
-                // s1 = np.random.randn(10, 3)
-                // t = s0.dot(s1)
-
-                // # now suppose we had the gradient on t from above in the circuit
-                // dt = np.random.randn(*t.shape) # same shape as t
-                // ds0 = dt.dot(s1.T) #.T gives the transpose of the matrix
-                // ds1 = t.T.dot(dt)
-
-                // tensor.shape [m,p,qq,rr]
-                // src0.shape   [n,m,q1,r1]
-                // src1.shape   [n,p,qq,rr]
-
-                // necessary for llama
-                if (src0->grad) {
-                    struct ggml_tensor * s1_tg =
-                        ggml_out_prod(ctx, // [n,m,qq,rr]
-                            src1,          // [n,p,qq,rr]
-                            tensor->grad); // [m,p,qq,rr]
-                    const int64_t qq = s1_tg->ne[2];
-                    const int64_t rr = s1_tg->ne[3];
-                    const int64_t q1 = src0->ne[2];
-                    const int64_t r1 = src0->ne[3];
-                    const bool ne2_broadcasted = qq > q1;
-                    const bool ne3_broadcasted = rr > r1;
-                    if (ne2_broadcasted || ne3_broadcasted) {
-                        // sum broadcast repetitions of s1_tg into shape of src0
-                        s1_tg = ggml_repeat_back(ctx, s1_tg, src0);
-                    }
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                                src0->grad, // [n,m,q1,r1]
-                                s1_tg,      // [n,m,q1,r1]
-                                zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    src1->grad =
-                        ggml_add_or_set(ctx,
-                                src1->grad,                            // [n,p,qq,rr]
-                                // ggml_mul_mat(ctx,                   // [n,p,qq,rr]
-                                //     ggml_cont(ctx,                  // [m,n,q1,r1]
-                                //         ggml_transpose(ctx, src0)), // [m,n,q1,r1]
-                                //     tensor->grad),                  // [m,p,qq,rr]
-
-                                // // when src0 is bigger than tensor->grad (this is mostly the case in llama),
-                                // // avoid transpose of src0, rather transpose smaller tensor->grad
-                                // // and then use ggml_out_prod
-                                ggml_out_prod(ctx,                  // [n,p,qq,rr]
-                                    src0,                           // [n,m,q1,r1]
-                                    ggml_transpose(ctx,             // [p,m,qq,rr]
-                                        tensor->grad)),             // [m,p,qq,rr]
-                                zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_MUL_MAT_ID:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+            if (src1_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc1, ggml_mean(ctx, grad)); // TODO: should probably be sum instead of mean
             }
-        case GGML_OP_OUT_PROD:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_ACC: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, grad);
             }
-        case GGML_OP_SCALE:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    float s;
-                    memcpy(&s, tensor->op_params, sizeof(float));
-
-                    src0->grad =
-                        ggml_add_or_set(ctx,
-                            src0->grad,
-                            ggml_scale_impl(ctx, tensor->grad, s, false),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SET:
-            {
-                const size_t nb1     = ((int32_t *) tensor->op_params)[0];
-                const size_t nb2     = ((int32_t *) tensor->op_params)[1];
-                const size_t nb3     = ((int32_t *) tensor->op_params)[2];
-                const size_t offset  = ((int32_t *) tensor->op_params)[3];
-
-                struct ggml_tensor * tensor_grad_view = NULL;
+            if (src1_needs_grads) {
+                const size_t nb1    = ((int32_t *) tensor->op_params)[0];
+                const size_t nb2    = ((int32_t *) tensor->op_params)[1];
+                const size_t nb3    = ((int32_t *) tensor->op_params)[2];
+                const size_t offset = ((int32_t *) tensor->op_params)[3];
 
-                if (src0->grad || src1->grad) {
-                    GGML_ASSERT(src0->type == tensor->type);
-                    GGML_ASSERT(tensor->grad->type == tensor->type);
-                    GGML_ASSERT(!src1->grad || src1->grad->type == tensor->grad->type);
+                struct ggml_tensor * tensor_grad_view = ggml_view_4d(ctx,
+                    grad, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
+                    nb1, nb2, nb3, offset);
 
-                    tensor_grad_view = ggml_view_4d(ctx,
-                        tensor->grad, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
-                        nb1, nb2, nb3, offset);
+                ggml_add_or_set(ctx, cgraph, isrc1, ggml_reshape(ctx, ggml_cont(ctx, tensor_grad_view), src1));
+            }
+        } break;
+        case GGML_OP_SUB: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, grad);
+            }
+            if (src1_needs_grads) {
+                ggml_sub_or_set(ctx, cgraph, isrc1, grad);
+            }
+        } break;
+        case GGML_OP_MUL: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, src1, grad));
+            }
+            if (src1_needs_grads) {
+                struct ggml_tensor * tmp = ggml_mul(ctx, src0, grad);
+                if (!ggml_are_same_shape(src0, src1)) {
+                    tmp = ggml_repeat_back(ctx, tmp, src1);
                 }
-
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx,
-                        src0->grad,
-                        ggml_acc_impl(ctx,
-                            tensor->grad,
-                            ggml_neg(ctx, tensor_grad_view),
-                            nb1, nb2, nb3, offset, false),
-                        zero_table, acc_table);
-                }
-
-                if (src1->grad) {
-                    src1->grad =
-                        ggml_add_or_set(ctx,
-                            src1->grad,
-                            ggml_reshape(ctx,
-                                ggml_cont(ctx, tensor_grad_view),
-                                src1->grad),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_CPY:
-            {
-                // necessary for llama
-                // cpy overwrites value of src1 by src0 and returns view(src1)
-                // the overwriting is mathematically equivalent to:
-                // tensor = src0 * 1 + src1 * 0
-                if (src0->grad) {
-                    // dsrc0 = dtensor * 1
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    // dsrc1 = dtensor * 0 -> noop
-                }
-            } break;
-        case GGML_OP_CONT:
-            {
-                // same as cpy
-                if (src0->grad) {
-                    GGML_ASSERT(ggml_is_contiguous(src0->grad));
-                    GGML_ASSERT(ggml_is_contiguous(tensor->grad));
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_RESHAPE:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx, src0->grad,
-                            ggml_reshape(ctx,
-                                ggml_is_contiguous(tensor->grad)
-                                    ? tensor->grad
-                                    : ggml_cont(ctx, tensor->grad),
-                                src0->grad),
-                        zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_VIEW:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    size_t offset;
-
-                    memcpy(&offset, tensor->op_params, sizeof(offset));
-
-                    size_t nb1 = tensor->nb[1];
-                    size_t nb2 = tensor->nb[2];
-                    size_t nb3 = tensor->nb[3];
-
-                    if (src0->type != src0->grad->type) {
-                        // gradient is typically F32, but src0 could be other type
-                        size_t ng = ggml_element_size(src0->grad);
-                        size_t n0 = ggml_element_size(src0);
-                        GGML_ASSERT(offset % n0 == 0);
-                        GGML_ASSERT(nb1 % n0 == 0);
-                        GGML_ASSERT(nb2 % n0 == 0);
-                        GGML_ASSERT(nb3 % n0 == 0);
-                        offset = (offset / n0) * ng;
-                        nb1 = (nb1 / n0) * ng;
-                        nb2 = (nb2 / n0) * ng;
-                        nb3 = (nb3 / n0) * ng;
-                    }
-
-                    src0->grad = ggml_acc_or_set(ctx, src0->grad, tensor->grad, nb1, nb2, nb3, offset, zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_PERMUTE:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    int32_t * axes = (int32_t *) tensor->op_params;
-                    int axis0 = axes[0] & 0x3;
-                    int axis1 = axes[1] & 0x3;
-                    int axis2 = axes[2] & 0x3;
-                    int axis3 = axes[3] & 0x3;
-                    int axes_backward[4] = {0,0,0,0};
-                    axes_backward[axis0] = 0;
-                    axes_backward[axis1] = 1;
-                    axes_backward[axis2] = 2;
-                    axes_backward[axis3] = 3;
-                    src0->grad =
-                        ggml_add_or_set(ctx, src0->grad,
-                            ggml_permute(ctx,
-                                tensor->grad,
-                                axes_backward[0],
-                                axes_backward[1],
-                                axes_backward[2],
-                                axes_backward[3]),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_TRANSPOSE:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx, src0->grad,
-                            ggml_transpose(ctx, tensor->grad),
-                        zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_GET_ROWS:
-            {
-                // necessary for llama (only for tokenizer)
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx, src0->grad,
-                            // last ggml_get_rows_back argument src0->grad is only
-                            // necessary to setup correct output shape
-                            ggml_get_rows_back(ctx, tensor->grad, src1, src0->grad),
-                        zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    // noop
-                }
-            } break;
-        case GGML_OP_GET_ROWS_BACK:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+                ggml_add_or_set(ctx, cgraph, isrc1, tmp);
             }
-        case GGML_OP_DIAG:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_DIV: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_div(ctx, grad, src1));
             }
-        case GGML_OP_DIAG_MASK_INF:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    const int n_past = ((int32_t *) tensor->op_params)[0];
-                    src0->grad =
-                        ggml_add_or_set(ctx, src0->grad,
-                            /* ggml_diag_mask_inf_impl() shouldn't be here */
-                            /* ref:  https://github.com/ggerganov/llama.cpp/pull/4203#discussion_r1412377992 */
-                            ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false),
-                        zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_DIAG_MASK_ZERO:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    const int n_past = ((int32_t *) tensor->op_params)[0];
-                    src0->grad =
-                        ggml_add_or_set(ctx, src0->grad,
-                            ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false),
-                        zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_SOFT_MAX:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    src0->grad =
-                        ggml_add_or_set(ctx, src0->grad,
-                            ggml_soft_max_back(ctx, tensor->grad, tensor),
-                        zero_table, acc_table);
-                }
-                GGML_ASSERT((!src1 || !src1->grad) && "backward pass for softmax mask not implemented");
-            } break;
-        case GGML_OP_SOFT_MAX_BACK:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+            if (src1_needs_grads) {
+                ggml_sub_or_set(ctx, cgraph, isrc1, ggml_mul(ctx, grad, ggml_div(ctx, tensor, src1)));
             }
-        case GGML_OP_ROPE:
-            {
-                // necessary for llama
-                if (src0->grad) {
-                    //const int n_past = ((int32_t *) tensor->op_params)[0];
-                    const int n_dims     = ((int32_t *) tensor->op_params)[1];
-                    const int mode       = ((int32_t *) tensor->op_params)[2];
-                    //const int n_ctx      = ((int32_t *) tensor->op_params)[3];
-                    const int n_ctx_orig = ((int32_t *) tensor->op_params)[4];
-                    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-
-                    memcpy(&freq_base,   (int32_t *) tensor->op_params +  5, sizeof(float));
-                    memcpy(&freq_scale,  (int32_t *) tensor->op_params +  6, sizeof(float));
-                    memcpy(&ext_factor,  (int32_t *) tensor->op_params +  7, sizeof(float));
-                    memcpy(&attn_factor, (int32_t *) tensor->op_params +  8, sizeof(float));
-                    memcpy(&beta_fast,   (int32_t *) tensor->op_params +  9, sizeof(float));
-                    memcpy(&beta_slow,   (int32_t *) tensor->op_params + 10, sizeof(float));
-
-                    src0->grad = ggml_add_or_set(ctx,
-                            src0->grad,
-                            ggml_rope_back(ctx,
-                                tensor->grad,
-                                src1,
-                                src2,
-                                n_dims,
-                                mode,
-                                n_ctx_orig,
-                                freq_base,
-                                freq_scale,
-                                ext_factor,
-                                attn_factor,
-                                beta_fast,
-                                beta_slow),
-                            zero_table, acc_table);
-                }
-                GGML_ASSERT((!src2 || !src2->grad) && "gradients for freq factors not implemented");
-            } break;
-        case GGML_OP_ROPE_BACK:
-            {
-                if (src0->grad) {
-                    //const int n_past = ((int32_t *) tensor->op_params)[0];
-                    const int n_dims     = ((int32_t *) tensor->op_params)[1];
-                    const int mode       = ((int32_t *) tensor->op_params)[2];
-                    //const int n_ctx      = ((int32_t *) tensor->op_params)[3];
-                    const int n_ctx_orig = ((int32_t *) tensor->op_params)[4];
-                    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-
-                    memcpy(&freq_base,   (int32_t *) tensor->op_params +  5, sizeof(float));
-                    memcpy(&freq_scale,  (int32_t *) tensor->op_params +  6, sizeof(float));
-                    memcpy(&ext_factor,  (int32_t *) tensor->op_params +  7, sizeof(float));
-                    memcpy(&attn_factor, (int32_t *) tensor->op_params +  8, sizeof(float));
-                    memcpy(&beta_fast,   (int32_t *) tensor->op_params +  9, sizeof(float));
-                    memcpy(&beta_slow,   (int32_t *) tensor->op_params + 10, sizeof(float));
-
-                    src0->grad = ggml_add_or_set(ctx,
-                            src0->grad,
-                            ggml_rope_impl(ctx,
-                                tensor->grad,
-                                src1,
-                                src2,
-                                n_dims,
-                                mode,
-                                n_ctx_orig,
-                                freq_base,
-                                freq_scale,
-                                ext_factor,
-                                attn_factor,
-                                beta_fast,
-                                beta_slow,
-                                false),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_CLAMP:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_SQR: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_scale(ctx, ggml_mul(ctx, src0, grad), 2.0f));
             }
-        case GGML_OP_CONV_TRANSPOSE_1D:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_SQRT: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_scale(ctx, ggml_div(ctx, grad, tensor), 0.5f));
             }
-        case GGML_OP_IM2COL:
-            {
-                if (src1->grad) {
-                    const int32_t s0    = ggml_get_op_params_i32(tensor, 0);
-                    const int32_t s1    = ggml_get_op_params_i32(tensor, 1);
-                    const int32_t p0    = ggml_get_op_params_i32(tensor, 2);
-                    const int32_t p1    = ggml_get_op_params_i32(tensor, 3);
-                    const int32_t d0    = ggml_get_op_params_i32(tensor, 4);
-                    const int32_t d1    = ggml_get_op_params_i32(tensor, 5);
-                    const bool    is_2D = ggml_get_op_params_i32(tensor, 6) == 1;
-
-                    src1->grad = ggml_add_or_set(ctx,
-                            src1->grad,
-                            ggml_im2col_back(ctx, src0, tensor->grad, src1->ne, s0, s1, p0, p1, d0, d1, is_2D),
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_IM2COL_BACK:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_LOG: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_div(ctx, grad, src0));
             }
-        case GGML_OP_CONV_TRANSPOSE_2D:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_SIN: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, grad, ggml_cos(ctx, src0)));
             }
-        case GGML_OP_POOL_1D:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_COS: {
+            if (src0_needs_grads) {
+                ggml_sub_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, grad, ggml_sin(ctx, src0)));
             }
-        case GGML_OP_POOL_2D:
-            {
-                if (src0->grad) {
-                    const enum ggml_op_pool op = ggml_get_op_params_i32(tensor, 0);
-                    const      int32_t      k0 = ggml_get_op_params_i32(tensor, 1);
-                    const      int32_t      k1 = ggml_get_op_params_i32(tensor, 2);
-                    const      int32_t      s0 = ggml_get_op_params_i32(tensor, 3);
-                    const      int32_t      s1 = ggml_get_op_params_i32(tensor, 4);
-                    const      int32_t      p0 = ggml_get_op_params_i32(tensor, 5);
-                    const      int32_t      p1 = ggml_get_op_params_i32(tensor, 6);
-
-                    src0->grad = ggml_add_or_set(ctx,
-                            src0->grad,
-                            ggml_pool_2d_back(ctx, tensor->grad, src0, op, k0, k1, s0, s1, p0, p1),
-                            zero_table, acc_table);
+        } break;
+        case GGML_OP_SUM: {
+            if (src0_needs_grads) {
+                ggml_add1_or_set(ctx, cgraph, isrc0, grad);
+            }
+        } break;
+        case GGML_OP_SUM_ROWS: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_repeat(ctx, grad, src0));
+            }
+        } break;
+        case GGML_OP_MEAN: {
+            if (src0_needs_grads) {
+                ggml_add1_or_set(ctx, cgraph, isrc0, ggml_scale_impl(ctx, grad, 1.0f/src0->ne[0], false));
+            }
+        } break;
+        case GGML_OP_REPEAT: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_repeat_back(ctx, grad, src0));
+            }
+        } break;
+        case GGML_OP_REPEAT_BACK: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_repeat(ctx, grad, src0));
+            }
+        } break;
+        case GGML_OP_RMS_NORM: {
+            if (src0_needs_grads) {
+                float eps;
+                memcpy(&eps, tensor->op_params, sizeof(float));
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_rms_norm_back(ctx, src0, grad, eps));
+            }
+        } break;
+        case GGML_OP_MUL_MAT: {
+            // https://cs231n.github.io/optimization-2/#staged
+            // # forward pass
+            // s0 = np.random.randn(5, 10)
+            // s1 = np.random.randn(10, 3)
+            // t = s0.dot(s1)
+
+            // # now suppose we had the gradient on t from above in the circuit
+            // dt = np.random.randn(*t.shape) # same shape as t
+            // ds0 = dt.dot(s1.T) #.T gives the transpose of the matrix
+            // ds1 = t.T.dot(dt)
+
+            // tensor.shape [m,p,qq,rr]
+            // src0.shape   [n,m,q1,r1]
+            // src1.shape   [n,p,qq,rr]
+
+            if (src0_needs_grads) {
+                struct ggml_tensor * s1_tg =
+                    ggml_out_prod(ctx, // [n,m,qq,rr]
+                        src1,          // [n,p,qq,rr]
+                        grad);         // [m,p,qq,rr]
+                const int64_t qq = s1_tg->ne[2];
+                const int64_t rr = s1_tg->ne[3];
+                const int64_t q1 = src0->ne[2];
+                const int64_t r1 = src0->ne[3];
+                const bool ne2_broadcasted = qq > q1;
+                const bool ne3_broadcasted = rr > r1;
+                if (ne2_broadcasted || ne3_broadcasted) {
+                    // sum broadcast repetitions of s1_tg into shape of src0
+                    s1_tg = ggml_repeat_back(ctx, s1_tg, src0);
                 }
-            } break;
-        case GGML_OP_POOL_2D_BACK:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+                ggml_add_or_set(ctx, cgraph, isrc0, s1_tg /*= [n,m,q1,r1]*/);
             }
-        case GGML_OP_UPSCALE:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+            if (src1_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc1,
+                        // ggml_mul_mat(ctx,                   // [n,p,qq,rr]
+                        //     ggml_cont(ctx,                  // [m,n,q1,r1]
+                        //         ggml_transpose(ctx, src0)), // [m,n,q1,r1]
+                        //     grad),                          // [m,p,qq,rr]
+
+                        // when src0 is bigger than tensor->grad (this is mostly the case in llama),
+                        // avoid transpose of src0, rather transpose smaller tensor->grad
+                        // and then use ggml_out_prod
+                        ggml_out_prod(ctx,      // [n,p,qq,rr]
+                            src0,               // [n,m,q1,r1]
+                            ggml_transpose(ctx, // [p,m,qq,rr]
+                                grad)));        // [m,p,qq,rr]
             }
-        case GGML_OP_PAD:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_SCALE: {
+            if (src0_needs_grads) {
+                float s;
+                memcpy(&s, tensor->op_params, sizeof(float));
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_scale_impl(ctx, grad, s, false));
             }
-        case GGML_OP_ARANGE:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_SET: {
+            const size_t nb1    = ((const int32_t *) tensor->op_params)[0];
+            const size_t nb2    = ((const int32_t *) tensor->op_params)[1];
+            const size_t nb3    = ((const int32_t *) tensor->op_params)[2];
+            const size_t offset = ((const int32_t *) tensor->op_params)[3];
+
+            struct ggml_tensor * tensor_grad_view = NULL;
+
+            if (src0_needs_grads || src1_needs_grads) {
+                GGML_ASSERT(src0->type == tensor->type);
+                GGML_ASSERT(!cgraph->grads[isrc0] ||                      cgraph->grads[isrc0]->type == grad->type);
+                GGML_ASSERT(!cgraph->grads[isrc1] || !src1_needs_grads || cgraph->grads[isrc1]->type == grad->type);
+
+                tensor_grad_view = ggml_view_4d(ctx,
+                    grad, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
+                    nb1, nb2, nb3, offset);
             }
-        case GGML_OP_TIMESTEP_EMBEDDING:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+
+            if (src0_needs_grads) {
+                struct ggml_tensor * tmp = ggml_neg(ctx, tensor_grad_view);
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_acc_impl(ctx, grad, tmp, nb1, nb2, nb3, offset, false));
             }
-        case GGML_OP_ARGSORT:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+
+            if (src1_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc1, ggml_reshape(ctx, ggml_cont(ctx, tensor_grad_view), src1));
             }
-        case GGML_OP_LEAKY_RELU:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_CPY: {
+            // cpy overwrites value of src1 by src0 and returns view(src1)
+            // the overwriting is mathematically equivalent to:
+            // tensor = src0 * 1 + src1 * 0
+            if (src0_needs_grads) {
+                // dsrc0 = dtensor * 1
+                ggml_add_or_set(ctx, cgraph, isrc0, grad);
             }
-        case GGML_OP_FLASH_ATTN_EXT:
-            {
-                GGML_ABORT("FA backward pass not adapted after rework");
-                struct ggml_tensor * flash_grad = NULL;
-                if (src0->grad || src1->grad || tensor->src[2]->grad) {
-                    int32_t t = ggml_get_op_params_i32(tensor, 0);
-                    GGML_ASSERT(t == 0 || t == 1);
-                    bool masked = t != 0;
-                    flash_grad =
-                        ggml_flash_attn_back(ctx,
-                            src0,
-                            src1,
-                            tensor->src[2],
-                            tensor->grad,
-                            masked);
+            if (src1_needs_grads) {
+                // dsrc1 = dtensor * 0 -> noop
+            }
+        } break;
+        case GGML_OP_CONT: {
+            // same as cpy
+            if (src0_needs_grads) {
+                GGML_ASSERT(!cgraph->grads[isrc0] || ggml_is_contiguous(cgraph->grads[isrc0]));
+                GGML_ASSERT(ggml_is_contiguous(grad));
+                ggml_add_or_set(ctx, cgraph, isrc0, grad);
+            }
+        } break;
+        case GGML_OP_RESHAPE: {
+            if (src0_needs_grads) {
+                struct ggml_tensor * grad_cont = ggml_is_contiguous(grad) ? grad : ggml_cont(ctx, grad);
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_reshape(ctx, grad_cont, src0));
+            }
+        } break;
+        case GGML_OP_VIEW: {
+            if (src0_needs_grads) {
+                size_t offset;
+
+                memcpy(&offset, tensor->op_params, sizeof(offset));
+
+                size_t nb1 = tensor->nb[1];
+                size_t nb2 = tensor->nb[2];
+                size_t nb3 = tensor->nb[3];
+
+                if (cgraph->grads[isrc0] && src0->type != cgraph->grads[isrc0]->type) {
+                    // gradient is typically F32, but src0 could be other type
+                    size_t ng = ggml_element_size(cgraph->grads[isrc0]);
+                    size_t n0 = ggml_element_size(src0);
+                    GGML_ASSERT(offset % n0 == 0);
+                    GGML_ASSERT(nb1 % n0 == 0);
+                    GGML_ASSERT(nb2 % n0 == 0);
+                    GGML_ASSERT(nb3 % n0 == 0);
+                    offset = (offset / n0) * ng;
+                    nb1 = (nb1 / n0) * ng;
+                    nb2 = (nb2 / n0) * ng;
+                    nb3 = (nb3 / n0) * ng;
                 }
 
-                const int64_t elem_q = ggml_nelements(src0);
-                const int64_t elem_k = ggml_nelements(src1);
-                const int64_t elem_v = ggml_nelements(src2);
-
-                enum ggml_type result_type = flash_grad->type;
-                GGML_ASSERT(ggml_blck_size(result_type) == 1);
-                const size_t tsize = ggml_type_size(result_type);
-
-                const size_t offs_q = 0;
-                const size_t offs_k = offs_q + GGML_PAD(elem_q * tsize, GGML_MEM_ALIGN);
-                const size_t offs_v = offs_k + GGML_PAD(elem_k * tsize, GGML_MEM_ALIGN);
-
-                if (src0->grad) {
-                    struct ggml_tensor * view_q = ggml_view_1d(ctx, flash_grad, elem_q, offs_q);
-                    struct ggml_tensor * grad_q = ggml_reshape(ctx, view_q, src0);
-                    src0->grad = ggml_add_or_set(ctx,
-                            src0->grad,
-                            grad_q,
-                            zero_table, acc_table);
-                }
-                if (src1->grad) {
-                    struct ggml_tensor * view_k = ggml_view_1d(ctx, flash_grad, elem_k, offs_k);
-                    struct ggml_tensor * grad_k = ggml_reshape(ctx, view_k, src1);
-                    src1->grad = ggml_add_or_set(ctx,
-                            src1->grad,
-                            grad_k,
-                            zero_table, acc_table);
-                }
-                if (src2->grad) {
-                    struct ggml_tensor * view_v = ggml_view_1d(ctx, flash_grad, elem_v, offs_v);
-                    struct ggml_tensor * grad_v = ggml_reshape(ctx, view_v, src2);
-                    src2->grad = ggml_add_or_set(ctx,
-                            src2->grad,
-                            grad_v,
-                            zero_table, acc_table);
-                }
-            } break;
-        case GGML_OP_FLASH_ATTN_BACK:
-            {
-                GGML_ABORT("fatal error"); // not supported
+                ggml_acc_or_set(ctx, cgraph, isrc0, grad, nb1, nb2, nb3, offset);
             }
-        case GGML_OP_SSM_CONV:
-        case GGML_OP_SSM_SCAN:
-            {
-                GGML_ABORT("fatal error"); // TODO: not implemented
+        } break;
+        case GGML_OP_PERMUTE: {
+            if (src0_needs_grads) {
+                const int32_t * axes = (const int32_t *) tensor->op_params;
+                const int axis0 = axes[0] & 0x3;
+                const int axis1 = axes[1] & 0x3;
+                const int axis2 = axes[2] & 0x3;
+                const int axis3 = axes[3] & 0x3;
+                int axb[4] = {0,0,0,0}; // axes backward
+                axb[axis0] = 0;
+                axb[axis1] = 1;
+                axb[axis2] = 2;
+                axb[axis3] = 3;
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_permute(ctx, grad, axb[0], axb[1], axb[2], axb[3]));
+            }
+        } break;
+        case GGML_OP_TRANSPOSE: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_transpose(ctx, grad));
             }
+        } break;
+        case GGML_OP_GET_ROWS: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_get_rows_back(ctx, grad, src1, src0));
+            }
+            if (src1_needs_grads) {
+                // noop
+            }
+        } break;
+        case GGML_OP_DIAG_MASK_INF: {
+            if (src0_needs_grads) {
+                /* ggml_diag_mask_inf_impl() shouldn't be here */
+                /* ref:  https://github.com/ggerganov/llama.cpp/pull/4203#discussion_r1412377992 */
+                const int n_past = ((const int32_t *) tensor->op_params)[0];
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_diag_mask_zero_impl(ctx, grad, n_past, false));
+            }
+        } break;
+        case GGML_OP_DIAG_MASK_ZERO: {
+            if (src0_needs_grads) {
+                const int n_past = ((const int32_t *) tensor->op_params)[0];
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_diag_mask_zero_impl(ctx, grad, n_past, false));
+            }
+        } break;
+        case GGML_OP_SOFT_MAX: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_soft_max_back(ctx, grad, tensor));
+            }
+            GGML_ASSERT((!src1 || !src1_needs_grads) && "backward pass for softmax mask not implemented");
+        } break;
+        case GGML_OP_ROPE: {
+            if (src0_needs_grads) {
+                //const int n_past = ((int32_t *) tensor->op_params)[0];
+                const int n_dims     = ((const int32_t *) tensor->op_params)[1];
+                const int mode       = ((const int32_t *) tensor->op_params)[2];
+                //const int n_ctx      = ((int32_t *) tensor->op_params)[3];
+                const int n_ctx_orig = ((const int32_t *) tensor->op_params)[4];
+                float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+
+                memcpy(&freq_base,   (const float *) tensor->op_params +  5, sizeof(float));
+                memcpy(&freq_scale,  (const float *) tensor->op_params +  6, sizeof(float));
+                memcpy(&ext_factor,  (const float *) tensor->op_params +  7, sizeof(float));
+                memcpy(&attn_factor, (const float *) tensor->op_params +  8, sizeof(float));
+                memcpy(&beta_fast,   (const float *) tensor->op_params +  9, sizeof(float));
+                memcpy(&beta_slow,   (const float *) tensor->op_params + 10, sizeof(float));
+
+                ggml_add_or_set(ctx, cgraph, isrc0,
+                    ggml_rope_back(ctx, grad, src1, src2, n_dims, mode, n_ctx_orig, freq_base,
+                        freq_scale, ext_factor, attn_factor, beta_fast, beta_slow));
+            }
+            GGML_ASSERT((!src2 || !src2_needs_grads) && "gradients for freq factors not implemented");
+        } break;
+        case GGML_OP_IM2COL: {
+            if (src1_needs_grads) {
+                const int32_t s0    = ggml_get_op_params_i32(tensor, 0);
+                const int32_t s1    = ggml_get_op_params_i32(tensor, 1);
+                const int32_t p0    = ggml_get_op_params_i32(tensor, 2);
+                const int32_t p1    = ggml_get_op_params_i32(tensor, 3);
+                const int32_t d0    = ggml_get_op_params_i32(tensor, 4);
+                const int32_t d1    = ggml_get_op_params_i32(tensor, 5);
+                const bool    is_2D = ggml_get_op_params_i32(tensor, 6) == 1;
+
+                ggml_add_or_set(ctx, cgraph, isrc1, ggml_im2col_back(ctx, src0, grad, src1->ne, s0, s1, p0, p1, d0, d1, is_2D));
+            }
+        } break;
+        case GGML_OP_POOL_2D: {
+            if (src0_needs_grads) {
+                const enum ggml_op_pool op = ggml_get_op_params_i32(tensor, 0);
+                const      int32_t      k0 = ggml_get_op_params_i32(tensor, 1);
+                const      int32_t      k1 = ggml_get_op_params_i32(tensor, 2);
+                const      int32_t      s0 = ggml_get_op_params_i32(tensor, 3);
+                const      int32_t      s1 = ggml_get_op_params_i32(tensor, 4);
+                const      int32_t      p0 = ggml_get_op_params_i32(tensor, 5);
+                const      int32_t      p1 = ggml_get_op_params_i32(tensor, 6);
+
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_pool_2d_back(ctx, grad, src0, op, k0, k1, s0, s1, p0, p1));
+            }
+        } break;
         case GGML_OP_WIN_PART:
         case GGML_OP_WIN_UNPART:
-        case GGML_OP_UNARY:
-            {
-                switch (ggml_get_unary_op(tensor)) {
-                    case GGML_UNARY_OP_ABS:
-                        {
-                            if (src0->grad) {
-                                src0->grad =
-                                    ggml_add_or_set(ctx,
-                                            src0->grad,
-                                            ggml_mul(ctx,
-                                                ggml_sgn(ctx, src0),
-                                                tensor->grad),
-                                            zero_table, acc_table);
-                            }
-                        } break;
-                    case GGML_UNARY_OP_SGN:
-                        {
-                            if (src0->grad) {
-                                // noop
-                            }
-                        } break;
-                    case GGML_UNARY_OP_NEG:
-                        {
-                            if (src0->grad) {
-                                src0->grad = ggml_sub_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
-                            }
-                        } break;
-                    case GGML_UNARY_OP_STEP:
-                        {
-                            if (src0->grad) {
-                                // noop
-                            }
-                        } break;
-                    case GGML_UNARY_OP_TANH:
-                        {
-                            GGML_ABORT("fatal error"); // TODO: not implemented
-                        }
-                    case GGML_UNARY_OP_ELU:
-                        {
-                            GGML_ABORT("fatal error"); // TODO: not implemented
-                        }
-                    case GGML_UNARY_OP_RELU:
-                        {
-                            if (src0->grad) {
-                                src0->grad = ggml_add_or_set(ctx,
-                                        src0->grad,
-                                        ggml_mul(ctx,
-                                            ggml_step(ctx, src0),
-                                            tensor->grad),
-                                        zero_table, acc_table);
-                            }
-                        } break;
-                    case GGML_UNARY_OP_SIGMOID:
-                        {
-                            GGML_ABORT("fatal error"); // TODO: not implemented
-                        }
-                    case GGML_UNARY_OP_GELU:
-                        {
-                            GGML_ABORT("fatal error"); // TODO: not implemented
-                        }
-                    case GGML_UNARY_OP_GELU_QUICK:
-                        {
-                            GGML_ABORT("fatal error"); // TODO: not implemented
-                        }
-                    case GGML_UNARY_OP_SILU:
-                        {
-                            // necessary for llama
-                            if (src0->grad) {
-                                src0->grad = ggml_add_or_set(ctx,
-                                        src0->grad,
-                                        ggml_silu_back(ctx, src0, tensor->grad),
-                                        zero_table, acc_table);
-                            }
-                        } break;
-                    case GGML_UNARY_OP_EXP:
-                        {
-                            if (src0->grad) {
-                                src0->grad = ggml_add_or_set(ctx,
-                                        src0->grad,
-                                        ggml_mul(ctx, tensor, tensor->grad),
-                                        zero_table, acc_table);
-                            }
-                        } break;
-                    default:
-                        GGML_ABORT("fatal error");
-                }
-            } break;
-        case GGML_OP_GET_REL_POS:
-        case GGML_OP_ADD_REL_POS:
-        case GGML_OP_RWKV_WKV6:
-        case GGML_OP_MAP_UNARY:
-        case GGML_OP_MAP_BINARY:
-        case GGML_OP_MAP_CUSTOM1_F32:
-        case GGML_OP_MAP_CUSTOM2_F32:
-        case GGML_OP_MAP_CUSTOM3_F32:
-        case GGML_OP_MAP_CUSTOM1:
-        case GGML_OP_MAP_CUSTOM2:
-        case GGML_OP_MAP_CUSTOM3:
-            {
-                GGML_ABORT("fatal error"); // not supported
-            }
-        case GGML_OP_CROSS_ENTROPY_LOSS:
-            {
-                if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx,
-                                src0->grad,
-                                ggml_cross_entropy_loss_back(ctx,
-                                    src0,
-                                    src1,
-                                    tensor->grad),
-                                zero_table, acc_table);
-                }
-                GGML_ASSERT(!src1->grad && "backward pass for labels not implemented");
-            } break;
-        case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
-            {
-                GGML_ABORT("fatal error"); // not supported
+        case GGML_OP_UNARY: {
+            switch (ggml_get_unary_op(tensor)) {
+                case GGML_UNARY_OP_ABS: {
+                    if (src0_needs_grads) {
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, ggml_sgn(ctx, src0), grad));
+                    }
+                } break;
+                case GGML_UNARY_OP_SGN: {
+                    // noop
+                } break;
+                case GGML_UNARY_OP_NEG: {
+                    if (src0_needs_grads) {
+                        ggml_sub_or_set(ctx, cgraph, isrc0, grad);
+                    }
+                } break;
+                case GGML_UNARY_OP_STEP: {
+                    // noop
+                } break;
+                case GGML_UNARY_OP_RELU: {
+                    if (src0_needs_grads) {
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, ggml_step(ctx, src0), grad));
+                    }
+                } break;
+                case GGML_UNARY_OP_SILU: {
+                    if (src0_needs_grads) {
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_silu_back(ctx, src0, grad));
+                    }
+                } break;
+                case GGML_UNARY_OP_EXP: {
+                    if (src0_needs_grads) {
+                        ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, tensor, grad));
+                    }
+                } break;
+                default: {
+                    fprintf(stderr, "%s: unsupported unary op for backward pass: %s\n",
+                        __func__, ggml_unary_op_name(ggml_get_unary_op(tensor)));
+                    GGML_ABORT("fatal error");
+                } //break;
             }
-        case GGML_OP_OPT_STEP_ADAMW:
-            {
-                GGML_ABORT("fatal error"); // not supported
+        } break;
+        case GGML_OP_CROSS_ENTROPY_LOSS: {
+            if (src0_needs_grads) {
+                ggml_add_or_set(ctx, cgraph, isrc0, ggml_cross_entropy_loss_back(ctx, src0, src1, grad));
             }
-        case GGML_OP_NONE:
-            {
-                // nop
-            } break;
+            GGML_ASSERT(!src1_needs_grads && "backward pass for labels not implemented");
+        } break;
+        case GGML_OP_NONE: {
+            // noop
+        } break;
         case GGML_OP_COUNT:
-            {
-                GGML_ABORT("fatal error");
-            }
+        default: {
+            fprintf(stderr, "%s: unsupported ggml op for backward pass: %s\n", __func__, ggml_op_name(tensor->op));
+            GGML_ABORT("fatal error");
+        } //break;
     }
 
-    for (int i = 0; i < GGML_MAX_SRC; ++i) {
-        if (tensor->src[i] && tensor->src[i]->grad) {
-            GGML_ASSERT(ggml_are_same_shape(tensor->src[i], tensor->src[i]->grad));
-        }
-    }
+    GGML_ASSERT(!src0_needs_grads || ggml_are_same_shape(src0, cgraph->grads[isrc0]));
+    GGML_ASSERT(!src1_needs_grads || ggml_are_same_shape(src1, cgraph->grads[isrc1]));
+    GGML_ASSERT(!src2_needs_grads || ggml_are_same_shape(src2, cgraph->grads[isrc2]));
 }
 
 static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) {
-    if (node->grad == NULL) {
-        // this usually happens when we generate intermediate nodes from constants in the backward pass
-        // it can also happen during forward pass, if the user performs computations with constants
-        if (node->op != GGML_OP_NONE) {
-            //GGML_PRINT_DEBUG("%s: warning: node %p has no grad, but op %d\n", __func__, (void *) node, node->op);
-        }
-    }
-
     // check if already visited
     if (ggml_hash_insert(&cgraph->visited_hash_set, node) == GGML_HASHSET_ALREADY_EXISTS) {
         return;
@@ -6212,18 +5692,41 @@ void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor *
     ggml_build_forward_impl(cgraph, tensor, true);
 }
 
-void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate) {
-    GGML_ASSERT(gf->n_nodes > 0);
-    GGML_ASSERT(gf->grads);
+void ggml_build_backward_expand(
+        struct ggml_context * ctx_static,
+        struct ggml_context * ctx_compute,
+        struct ggml_cgraph  * cgraph,
+        bool                  accumulate) {
+    GGML_ASSERT(cgraph->n_nodes > 0);
+    GGML_ASSERT(cgraph->grads);
+    GGML_ASSERT(cgraph->grad_accs);
+
+    const int n_nodes_f = cgraph->n_nodes;
 
-    for (int i = 0; i < gf->n_nodes; ++i) {
-        struct ggml_tensor * node = gf->nodes[i];
+    memset(cgraph->grads,     0, cgraph->visited_hash_set.size*sizeof(struct ggml_tensor *));
+    memset(cgraph->grad_accs, 0, cgraph->visited_hash_set.size*sizeof(struct ggml_tensor *));
+    bool * grads_needed = calloc(cgraph->visited_hash_set.size, sizeof(bool));
+
+    {
+        bool any_params = false;
+        bool any_loss   = false;
+        for (int i = 0; i < n_nodes_f; ++i) {
+            struct ggml_tensor * node = cgraph->nodes[i];
+            any_params = any_params || (node->flags & GGML_TENSOR_FLAG_PARAM);
+            any_loss   = any_loss   || (node->flags & GGML_TENSOR_FLAG_LOSS);
+        }
+        GGML_ASSERT(any_params && "no trainable parameters found, did you forget to call ggml_set_param?");
+        GGML_ASSERT(any_loss && "no training loss found, did you forget to call ggml_set_loss?");
+    }
+
+    for (int i = 0; i < n_nodes_f; ++i) {
+        struct ggml_tensor * node = cgraph->nodes[i];
 
         if (node->type == GGML_TYPE_I32) {
             continue;
         }
 
-        bool needs_grad = node->flags & GGML_TENSOR_FLAG_PARAM;
+        bool node_needs_grad = (node->flags & GGML_TENSOR_FLAG_PARAM) || (node->flags & GGML_TENSOR_FLAG_LOSS);
         bool ignore_src[GGML_MAX_SRC] = {false};
         switch (node->op) {
             // gradients in node->src[0] for one reason or another have no effect on output gradients
@@ -6240,7 +5743,7 @@ void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph *
             } break;
 
             // gradients in node->src[1] for one reason or another have no effect on output gradients
-            case GGML_OP_CPY:           // gradients in CPY target  are irrelevant
+            case GGML_OP_CPY:           // gradients in CPY target are irrelevant
             case GGML_OP_GET_ROWS:      // row indices not differentiable
             case GGML_OP_GET_ROWS_BACK: // same as for GET_ROWS
             case GGML_OP_ROPE:          // positions not differentiable
@@ -6251,14 +5754,14 @@ void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph *
                 break;
         }
         for (int j = 0; j < GGML_MAX_SRC; ++j) {
-            if (!node->src[j] || !node->src[j]->grad || ignore_src[j]) {
+            if (!node->src[j] || ignore_src[j] || !grads_needed[ggml_hash_find(&cgraph->visited_hash_set, node->src[j])]) {
                 continue;
             }
             GGML_ASSERT(node->src[j]->type == GGML_TYPE_F32 || node->src[j]->type == GGML_TYPE_F16);
-            needs_grad = true;
+            node_needs_grad = true;
             break;
         }
-        if (!needs_grad) {
+        if (!node_needs_grad) {
             continue;
         }
 
@@ -6266,73 +5769,24 @@ void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph *
         GGML_ASSERT(!node->view_src || node->op == GGML_OP_CPY || node->op == GGML_OP_VIEW ||
             node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE);
 
-        // create a new tensor with the same type and shape as the node and set it as grad
-        node->grad = ggml_dup_tensor(ctx, node);
-    }
-
-    // keep tables of original gradients for replacement/accumulation logic
-    struct ggml_hash_set zero_table = ggml_hash_set_new(gf->size);
-    struct ggml_hash_set acc_table  = ggml_hash_set_new(gf->size);
-    for (int i = 0; i < gf->n_nodes; i++) {
-        struct ggml_tensor * node = gf->nodes[i];
-
-        if (node->grad) {
-            {
-                const size_t insert_result = ggml_hash_insert(&zero_table, node->grad);
-                GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
-                GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
-            }
-
-            // only gradients of trainable parameters should be accumulated
-            if (accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) {
-                const size_t insert_result = ggml_hash_insert(&acc_table, node->grad);
-                GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
-                GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
-            }
+        const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node);
+        GGML_ASSERT(igrad != GGML_HASHSET_FULL);
+        GGML_ASSERT(ggml_bitset_get(cgraph->visited_hash_set.used, igrad));
+        if ((accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) || (node->flags & GGML_TENSOR_FLAG_LOSS)) {
+            cgraph->grad_accs[igrad] = ggml_dup_tensor(ctx_static, node);
+            cgraph->grads[igrad]     = cgraph->grad_accs[igrad];
+            ggml_format_name(cgraph->grad_accs[igrad], "grad acc for %s", node->name);
         }
+        grads_needed[igrad] = true;
     }
 
-    for (int i = gf->n_nodes - 1; i >= 0; i--) {
-        struct ggml_tensor * node = gf->nodes[i];
-
+    for (int i = n_nodes_f - 1; i >= 0; --i) {
         // inplace operations to add gradients are not created by ggml_compute_backward except for gradient accumulation
         // use allocator to automatically make inplace operations
-        if (node->grad) {
-            ggml_compute_backward(ctx, node, &zero_table, &acc_table);
-        }
+        ggml_compute_backward(ctx_compute, cgraph, i, grads_needed);
     }
 
-    for (int i = 0; i < gf->n_nodes; i++) {
-        struct ggml_tensor * node = gf->nodes[i];
-
-        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
-            GGML_PRINT_DEBUG("%s: found root node %p\n", __func__, (void *) node);
-            ggml_build_forward_expand(gb, node->grad);
-        }
-    }
-
-    ggml_hash_set_free(&zero_table);
-    ggml_hash_set_free(&acc_table);
-}
-
-void ggml_build_opt_adamw(
-        struct ggml_context * ctx,
-        struct ggml_cgraph  * gf,
-        struct ggml_cgraph  * gb,
-        float                 alpha,
-        float                 beta1,
-        float                 beta2,
-        float                 eps,
-        float                 wd) {
-    for (int i = 0; i < gf->n_nodes; i++) {
-        struct ggml_tensor * node = gf->nodes[i];
-
-        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
-            GGML_PRINT_DEBUG("%s: found root node %p\n", __func__, (void *) node);
-            struct ggml_tensor * opt_step = ggml_opt_step_adamw(ctx, node, node->grad, alpha, beta1, beta2, eps, wd);
-            ggml_build_forward_expand(gb, opt_step);
-        }
-    }
+    free(grads_needed);
 }
 
 static void * incr_ptr_aligned(void ** p, size_t size, size_t align) {
@@ -6350,7 +5804,8 @@ static size_t ggml_graph_nbytes(size_t size, bool grads) {
     incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // leafs
     incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // hash keys
     if (grads) {
-        incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // grads
+        incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // grads
+        incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // grad_accs
     }
     incr_ptr_aligned(&p, ggml_bitset_size(hash_size) * sizeof(ggml_bitset_t), sizeof(ggml_bitset_t));
 
@@ -6376,10 +5831,12 @@ struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t siz
 
     void * p = cgraph + 1;
 
-    struct ggml_tensor ** nodes_ptr = incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
-    struct ggml_tensor ** leafs_ptr = incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
-    struct ggml_tensor ** hash_keys_ptr = incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
-    struct ggml_tensor ** grads_ptr = grads ? incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
+    struct ggml_tensor ** nodes_ptr     =         incr_ptr_aligned(&p, size      * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** leafs_ptr     =         incr_ptr_aligned(&p, size      * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** hash_keys_ptr =         incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** grads_ptr     = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
+    struct ggml_tensor ** grad_accs_ptr = grads ? incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
+
     ggml_bitset_t * hash_used = incr_ptr_aligned(&p, ggml_bitset_size(hash_size) * sizeof(ggml_bitset_t), sizeof(ggml_bitset_t));
 
     // check that we allocated the correct amount of memory
@@ -6391,12 +5848,17 @@ struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t siz
         /*.n_leafs      =*/ 0,
         /*.nodes        =*/ nodes_ptr,
         /*.grads        =*/ grads_ptr,
+        /*.grad_accs    =*/ grad_accs_ptr,
         /*.leafs        =*/ leafs_ptr,
         /*.hash_table   =*/ { hash_size, hash_used, hash_keys_ptr },
         /*.order        =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT,
     };
 
     ggml_hash_set_reset(&cgraph->visited_hash_set);
+    if (grads) {
+        memset(cgraph->grads,     0, hash_size*sizeof(struct ggml_tensor *));
+        memset(cgraph->grad_accs, 0, hash_size*sizeof(struct ggml_tensor *));
+    }
 
     return cgraph;
 }
@@ -6407,14 +5869,15 @@ struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) {
 
 struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1) {
     struct ggml_cgraph cgraph = {
-        /*.size         =*/ 0,
-        /*.n_nodes      =*/ i1 - i0,
-        /*.n_leafs      =*/ 0,
-        /*.nodes        =*/ cgraph0->nodes + i0,
-        /*.grads        =*/ cgraph0->grads ? cgraph0->grads + i0 : NULL,
-        /*.leafs        =*/ NULL,
-        /*.hash_table   =*/ { 0, NULL, NULL },
-        /*.order        =*/ cgraph0->order,
+        /*.size             =*/ 0,
+        /*.n_nodes          =*/ i1 - i0,
+        /*.n_leafs          =*/ 0,
+        /*.nodes            =*/ cgraph0->nodes + i0,
+        /*.grads            =*/ NULL, // gradients would need visited_hash_set
+        /*.grad_accs        =*/ NULL,
+        /*.leafs            =*/ NULL,
+        /*.visited_hash_set =*/ { 0, NULL, NULL },
+        /*.order            =*/ cgraph0->order,
     };
 
     return cgraph;
@@ -6437,19 +5900,33 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
         dst->nodes[i] = src->nodes[i];
     }
 
-    if (src->grads) {
-        GGML_ASSERT(dst->grads != NULL);
-        for (int i = 0; i < src->n_nodes; ++i) {
-            dst->grads[i] = src->grads[i];
-        }
-    }
-
     for (size_t i = 0; i < src->visited_hash_set.size; ++i) {
         // copy all hashset keys (tensors) that are in use
         if (ggml_bitset_get(src->visited_hash_set.used, i)) {
             ggml_hash_insert(&dst->visited_hash_set, src->visited_hash_set.keys[i]);
         }
     }
+
+    if (dst->grads) {
+        memset(dst->grads,     0, dst->visited_hash_set.size*sizeof(struct ggml_tensor *));
+        memset(dst->grad_accs, 0, dst->visited_hash_set.size*sizeof(struct ggml_tensor *));
+    }
+    if (src->grads) {
+        GGML_ASSERT(dst->grads     != NULL);
+        GGML_ASSERT(dst->grad_accs != NULL);
+        for (int i = 0; i < src->n_nodes; ++i) {
+            const size_t igrad_src = ggml_hash_find(&src->visited_hash_set, src->nodes[i]);
+            const size_t igrad_dst = ggml_hash_find(&dst->visited_hash_set, dst->nodes[i]);
+
+            GGML_ASSERT(igrad_src != GGML_HASHSET_FULL);
+            GGML_ASSERT(ggml_bitset_get(src->visited_hash_set.used, igrad_src));
+            GGML_ASSERT(igrad_dst != GGML_HASHSET_FULL);
+            GGML_ASSERT(ggml_bitset_get(dst->visited_hash_set.used, igrad_dst));
+
+            dst->grads[igrad_dst]     = src->grads[igrad_src];
+            dst->grad_accs[igrad_dst] = src->grad_accs[igrad_src];
+        }
+    }
 }
 
 struct ggml_cgraph * ggml_graph_dup(struct ggml_context * ctx, struct ggml_cgraph * cgraph) {
@@ -6475,29 +5952,32 @@ void ggml_graph_reset(struct ggml_cgraph * cgraph) {
     GGML_ASSERT(cgraph->grads != NULL);
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        struct ggml_tensor * node = cgraph->nodes[i];
+        struct ggml_tensor * node     = cgraph->nodes[i];
+        struct ggml_tensor * grad_acc = ggml_graph_get_grad_acc(cgraph, node);
+
+        if (node->op == GGML_OP_OPT_STEP_ADAMW) {
+            // clear momenta
+            ggml_set_zero(node->src[2]);
+            ggml_set_zero(node->src[3]);
+        }
 
         // initial gradients of loss should be 1, 0 otherwise
-        if (node->grad) {
+        if (grad_acc) {
             if (node->flags & GGML_TENSOR_FLAG_LOSS) {
-                GGML_ASSERT(node->grad->buffer);
-                GGML_ASSERT(node->type == GGML_TYPE_F32);
-                GGML_ASSERT(ggml_is_scalar(node));
+                GGML_ASSERT(grad_acc->type == GGML_TYPE_F32);
+                GGML_ASSERT(ggml_is_scalar(grad_acc));
 
                 const float onef = 1.0f;
-                ggml_backend_tensor_set(node->grad, &onef, 0, ggml_nbytes(node->grad));
+                if (grad_acc->buffer) {
+                    ggml_backend_tensor_set(grad_acc, &onef, 0, sizeof(float));
+                } else {
+                    GGML_ASSERT(grad_acc->data);
+                    *((float *) grad_acc->data) = onef;
+                }
             } else {
-                ggml_set_zero(node->grad);
+                ggml_set_zero(grad_acc);
             }
         }
-
-        GGML_ASSERT(node);
-        if (node->op == GGML_OP_OPT_STEP_ADAMW) {
-            // set iteration to 1 and clear momenta
-            ggml_set_op_params_i32(node, 0, 1);
-            ggml_set_zero(node->src[2]);
-            ggml_set_zero(node->src[3]);
-        }
     }
 }
 
@@ -6535,7 +6015,7 @@ void ggml_graph_add_node(struct ggml_cgraph * cgraph, struct ggml_tensor * tenso
     cgraph->n_nodes++;
 }
 
-struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name) {
+struct ggml_tensor * ggml_graph_get_tensor(const struct ggml_cgraph * cgraph, const char * name) {
     for (int i = 0; i < cgraph->n_leafs; i++) {
         struct ggml_tensor * leaf = cgraph->leafs[i];
 
@@ -6555,6 +6035,16 @@ struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const ch
     return NULL;
 }
 
+struct ggml_tensor * ggml_graph_get_grad(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
+    const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node);
+    return igrad != GGML_HASHSET_FULL && ggml_bitset_get(cgraph->visited_hash_set.used, igrad) && cgraph->grads ? cgraph->grads[igrad] : NULL;
+}
+
+struct ggml_tensor * ggml_graph_get_grad_acc(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
+    const size_t igrad = ggml_hash_find(&cgraph->visited_hash_set, node);
+    return igrad != GGML_HASHSET_FULL && ggml_bitset_get(cgraph->visited_hash_set.used, igrad) && cgraph->grad_accs ? cgraph->grad_accs[igrad] : NULL;
+}
+
 void ggml_graph_print(const struct ggml_cgraph * cgraph) {
     GGML_LOG_INFO("=== GRAPH ===\n");
 
@@ -6565,7 +6055,8 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
         GGML_LOG_INFO(" - %3d: [ %5" PRId64 ", %5" PRId64 ", %5" PRId64 "] %16s %s\n",
                 i,
                 node->ne[0], node->ne[1], node->ne[2],
-                ggml_op_name(node->op), (node->flags & GGML_TENSOR_FLAG_PARAM) ? "x" : node->grad ? "g" : " ");
+                ggml_op_name(node->op), (node->flags & GGML_TENSOR_FLAG_PARAM) ? "x" :
+                      ggml_graph_get_grad(cgraph, node) ? "g" : " ");
     }
 
     GGML_LOG_INFO("n_leafs = %d\n", cgraph->n_leafs);
@@ -6600,8 +6091,9 @@ static bool ggml_graph_find(const struct ggml_cgraph * cgraph, const struct ggml
 static struct ggml_tensor * ggml_graph_get_parent(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node) {
     for (int i = 0; i < cgraph->n_nodes; i++) {
         struct ggml_tensor * parent = cgraph->nodes[i];
+        struct ggml_tensor * grad = ggml_graph_get_grad(cgraph, parent);
 
-        if (parent->grad == node) {
+        if (grad == node) {
             return parent;
         }
     }
@@ -6641,6 +6133,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
 
     for (int i = 0; i < gb->n_nodes; i++) {
         struct ggml_tensor * node = gb->nodes[i];
+        struct ggml_tensor * grad = ggml_graph_get_grad(gb, node);
 
         if (ggml_graph_get_parent(gb, node) != NULL) {
             continue;
@@ -6648,7 +6141,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
 
         if (node->flags & GGML_TENSOR_FLAG_PARAM) {
             snprintf(color, sizeof(color), "yellow");
-        } else if (node->grad) {
+        } else if (grad) {
             if (ggml_graph_find(gf, node)) {
                 snprintf(color, sizeof(color), "green");
             } else {
@@ -6675,8 +6168,8 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
             fprintf(fp, "%d [%" PRId64 ", %" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], node->ne[2], ggml_op_symbol(node->op));
         }
 
-        if (node->grad) {
-            fprintf(fp, " | <g>%s\"; ]\n", ggml_op_symbol(node->grad->op));
+        if (grad) {
+            fprintf(fp, " | <g>%s\"; ]\n", ggml_op_symbol(grad->op));
         } else {
             fprintf(fp, "\"; ]\n");
         }
@@ -6861,9 +6354,6 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_IQ1_M:   result = quantize_iq1_m  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_NL:  result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_IQ4_XS:  result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_Q4_0_4_4: result = quantize_q4_0_4x4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_Q4_0_4_8: result = quantize_q4_0_4x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
-        case GGML_TYPE_Q4_0_8_8: result = quantize_q4_0_8x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_F16:
             {
                 size_t elemsize = sizeof(ggml_fp16_t);
@@ -7395,7 +6885,16 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                 (int64_t) info->ne[2] *
                 (int64_t) info->ne[3];
 
-            if (ggml_blck_size(info->type) == 0 || ne % ggml_blck_size(info->type) != 0) {
+            if (ggml_blck_size(info->type) == 0 ) {
+                // this tensor type support have been removed:
+                fprintf(stderr, "%s: tensor '%s' of type %d: %s\n",
+                        __func__, info->name.data, (int) info->type, ggml_type_name(info->type));
+                fclose(file);
+                gguf_free(ctx);
+                return NULL;
+            }
+
+            if (ne % ggml_blck_size(info->type) != 0) {
                 fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%" PRId64 ")\n",
                         __func__, info->name.data, (int) info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type));
                 fclose(file);
@@ -8166,222 +7665,30 @@ void gguf_get_meta_data(const struct gguf_context * ctx, void * data) {
     gguf_buf_free(buf);
 }
 
-////////////////////////////////////////////////////////////////////////////////
-
-int ggml_cpu_has_avx(void) {
-#if defined(__AVX__)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_avx_vnni(void) {
-#if defined(__AVXVNNI__)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_avx2(void) {
-#if defined(__AVX2__)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_avx512(void) {
-#if defined(__AVX512F__)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_avx512_vbmi(void) {
-#if defined(__AVX512VBMI__)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_avx512_vnni(void) {
-#if defined(__AVX512VNNI__)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_avx512_bf16(void) {
-#if defined(__AVX512BF16__)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_amx_int8(void) {
-#if defined(__AMX_INT8__)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_fma(void) {
-#if defined(__FMA__)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_arm_fma(void) {
-#if defined(__ARM_FEATURE_FMA)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_riscv_v(void) {
-#if defined(__riscv_v_intrinsic)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_metal(void) {
-#if defined(GGML_USE_METAL)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_f16c(void) {
-#if defined(__F16C__)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_fp16_va(void) {
-#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_wasm_simd(void) {
-#if defined(__wasm_simd128__)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_blas(void) {
-#if defined(GGML_USE_BLAS) || defined(GGML_USE_CUDA) || defined(GGML_USE_VULKAN) || defined(GGML_USE_SYCL)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_cuda(void) {
-#if defined(GGML_USE_CUDA)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_vulkan(void) {
-#if defined(GGML_USE_VULKAN)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_kompute(void) {
-#if defined(GGML_USE_KOMPUTE)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_sycl(void) {
-#if defined(GGML_USE_SYCL)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_rpc(void) {
-#if defined(GGML_USE_RPC)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_cann(void) {
-#if defined(GGML_USE_CANN)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-int ggml_cpu_has_llamafile(void) {
-#if defined(GGML_USE_LLAMAFILE)
-    return 1;
-#else
-    return 0;
-#endif
+void ggml_log_set(ggml_log_callback log_callback, void * user_data) {
+    g_logger_state.log_callback = log_callback ? log_callback : ggml_log_callback_default;
+    g_logger_state.log_callback_user_data = user_data;
 }
 
-int ggml_cpu_has_gpublas(void) {
-    return ggml_cpu_has_cuda() || ggml_cpu_has_vulkan() || ggml_cpu_has_kompute() || ggml_cpu_has_sycl();
+void ggml_threadpool_params_init(struct ggml_threadpool_params * p, int n_threads) {
+    p->n_threads  = n_threads;
+    p->prio       = 0;     // default priority (usually means normal or inherited)
+    p->poll       = 50;    // hybrid-polling enabled
+    p->strict_cpu = false; // no strict placement (all threads share same cpumask)
+    p->paused     = false; // threads are ready to go
+    memset(p->cpumask, 0, GGML_MAX_N_THREADS); // all-zero means use the default affinity (usually inherited)
 }
 
-int ggml_cpu_has_sse3(void) {
-#if defined(__SSE3__)
-    return 1;
-#else
-    return 0;
-#endif
+struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads) {
+    struct ggml_threadpool_params p;
+    ggml_threadpool_params_init(&p, n_threads);
+    return p;
 }
 
-int ggml_cpu_has_ssse3(void) {
-#if defined(__SSSE3__)
-    return 1;
-#else
-    return 0;
-#endif
+bool ggml_threadpool_params_match(const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1) {
+    if (p0->n_threads      != p1->n_threads  )    return false;
+    if (p0->prio           != p1->prio       )    return false;
+    if (p0->poll           != p1->poll       )    return false;
+    if (p0->strict_cpu     != p1->strict_cpu )    return false;
+    return memcmp(p0->cpumask, p1->cpumask, GGML_MAX_N_THREADS) == 0;
 }
-
-int ggml_cpu_has_vsx(void) {
-#if defined(__POWER9_VECTOR__)
-    return 1;
-#else
-    return 0;
-#endif
-}
-
-void ggml_log_set(ggml_log_callback log_callback, void * user_data) {
-    g_logger_state.log_callback = log_callback ? log_callback : ggml_log_callback_default;
-    g_logger_state.log_callback_user_data = user_data;
-}
-////////////////////////////////////////////////////////////////////////////////
diff --git a/ggml/src/kompute-shaders/op_rope_f16.comp b/ggml/src/kompute-shaders/op_rope_f16.comp
deleted file mode 100644
index 1a4058b3f1f..00000000000
--- a/ggml/src/kompute-shaders/op_rope_f16.comp
+++ /dev/null
@@ -1,73 +0,0 @@
-#version 450
-
-#include "rope_common.comp"
-
-layout(binding = 0) buffer restrict readonly  tensorInA { float16_t inA[]; };
-layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float16_t out_[]; };
-
-void main() {
-    const uint i3 = gl_WorkGroupID.z;
-    const uint i2 = gl_WorkGroupID.y;
-    const uint i1 = gl_WorkGroupID.x;
-
-    const bool is_neox = (pcs.mode & 2) != 0;
-
-    float corr_dims[2];
-    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
-
-    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
-
-    const int p = inB[pcs.inBOff + i2];
-
-    float theta = float(p);
-
-    if (!is_neox) {
-        for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) {
-            float cos_theta, sin_theta;
-            rope_yarn(theta, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            theta *= theta_scale;
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
-
-            const float x0 = float(inA[src]);
-            const float x1 = float(inA[src+1]);
-
-            out_[dst_data]   = float16_t(x0*cos_theta - x1*sin_theta);
-            out_[dst_data+1] = float16_t(x0*sin_theta + x1*cos_theta);
-        }
-    } else {
-        const float inv_ndims = -1.f/pcs.n_dims;
-        for (uint ic = 0; ic < pcs.n_dims; ic += 2) {
-            const uint cur_rot = ic;
-
-            float cos_theta, sin_theta;
-            rope_yarn(theta, pcs.freq_scale, corr_dims, cur_rot, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            theta *= theta_scale;
-
-            const uint i0 = ic/2;
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
-
-            const float x0 = float(inA[src]);
-            const float x1 = float(inA[src+pcs.n_dims/2]);
-
-            out_[dst_data]              = float16_t(x0*cos_theta - x1*sin_theta);
-            out_[dst_data+pcs.n_dims/2] = float16_t(x0*sin_theta + x1*cos_theta);
-        }
-
-        for (uint ic = pcs.n_dims; ic < pcs.ne0; ic += 2) {
-            const uint i0 = ic;
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
-
-            out_[dst_data + 0] = inA[src + 0];
-            out_[dst_data + 1] = inA[src + 1];
-        }
-    }
-}
diff --git a/ggml/src/kompute-shaders/op_rope_f32.comp b/ggml/src/kompute-shaders/op_rope_f32.comp
deleted file mode 100644
index 65e03827a26..00000000000
--- a/ggml/src/kompute-shaders/op_rope_f32.comp
+++ /dev/null
@@ -1,73 +0,0 @@
-#version 450
-
-#include "rope_common.comp"
-
-layout(binding = 0) buffer restrict readonly  tensorInA { float inA[]; };
-layout(binding = 1) buffer restrict readonly  tensorInB { int   inB[]; };
-layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
-
-void main() {
-    const uint i3 = gl_WorkGroupID.z;
-    const uint i2 = gl_WorkGroupID.y;
-    const uint i1 = gl_WorkGroupID.x;
-
-    const bool is_neox = (pcs.mode & 2) != 0;
-
-    float corr_dims[2];
-    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
-
-    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
-
-    const int p = inB[pcs.inBOff + i2];
-
-    float theta = float(p);
-
-    if (!is_neox) {
-        for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) {
-            float cos_theta, sin_theta;
-            rope_yarn(theta, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            theta *= theta_scale;
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
-
-            const float x0 = inA[src];
-            const float x1 = inA[src+1];
-
-            out_[dst_data]   = x0*cos_theta - x1*sin_theta;
-            out_[dst_data+1] = x0*sin_theta + x1*cos_theta;
-        }
-    } else {
-        const float inv_ndims = -1.f/pcs.n_dims;
-        for (uint ic = 0; ic < pcs.n_dims; ic += 2) {
-            const uint cur_rot = ic;
-
-            float cos_theta, sin_theta;
-            rope_yarn(theta, pcs.freq_scale, corr_dims, cur_rot, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
-
-            theta *= theta_scale;
-
-            const uint i0 = ic/2;
-
-            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
-
-            const float x0 = inA[src];
-            const float x1 = inA[src+pcs.n_dims/2];
-
-            out_[dst_data] = x0*cos_theta - x1*sin_theta;
-            out_[dst_data+pcs.n_dims/2] = x0*sin_theta + x1*cos_theta;
-        }
-
-        for (uint ic = pcs.n_dims; ic < pcs.ne0; ic += 2) {
-            const uint i0 = ic;
-
-            const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
-            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_
-
-            out_[dst_data + 0] = inA[src + 0];
-            out_[dst_data + 1] = inA[src + 1];
-        }
-    }
-}
diff --git a/ggml/src/sgemm.cpp b/ggml/src/sgemm.cpp
deleted file mode 100644
index 6626ceb2621..00000000000
--- a/ggml/src/sgemm.cpp
+++ /dev/null
@@ -1,1027 +0,0 @@
-// Copyright 2024 Mozilla Foundation
-//
-// Permission is hereby granted, free of charge, to any person obtaining
-// a copy of this software and associated documentation files (the
-// "Software"), to deal in the Software without restriction, including
-// without limitation the rights to use, copy, modify, merge, publish,
-// distribute, sublicense, and/or sell copies of the Software, and to
-// permit persons to whom the Software is furnished to do so, subject to
-// the following conditions:
-//
-// The above copyright notice and this permission notice shall be
-// included in all copies or substantial portions of the Software.
-//
-// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
-// EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
-// MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
-// NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
-// BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
-// ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
-// CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-// SOFTWARE.
-
-//
-//                   _   _          ___ _      _   ___
-//                  | |_(_)_ _ _  _| _ ) |    /_\ / __|
-//                  |  _| | ' \ || | _ \ |__ / _ \\__ \.
-//                   \__|_|_||_\_, |___/____/_/ \_\___/
-//                             |__/
-//
-//                    BASIC LINEAR ALGEBRA SUBPROGRAMS
-//
-//
-// This file implements multithreaded CPU matrix multiplication for the
-// common contiguous use case C = Aᵀ * B. These kernels are designed to
-// have excellent performance[1] for matrices that fit in the CPU cache
-// without imposing any overhead such as cache filling or malloc calls.
-//
-// This implementation does not guarantee any upper bound with rounding
-// errors, which grow along with k. Our goal's to maximally exploit the
-// hardware for performance, and then use whatever resources remain for
-// improving numerical accuracy.
-//
-// [1] J. Tunney, ‘LLaMA Now Goes Faster on CPUs’, Mar. 2024. [Online].
-//     Available: https://justine.lol/matmul/. [Accessed: 29-Mar-2024].
-
-#if defined(__GNUC__)
-#pragma GCC diagnostic ignored "-Wpedantic"
-#pragma GCC diagnostic ignored "-Wignored-attributes"
-#endif
-
-#include "sgemm.h"
-#include "ggml-impl.h"
-#include "ggml-quants.h"
-
-#ifdef _MSC_VER
-#define NOINLINE __declspec(noinline)
-#else
-#define NOINLINE __attribute__((__noinline__))
-#endif
-
-#if defined(__ARM_NEON) || defined(__AVX512F__)
-#define VECTOR_REGISTERS 32
-#else
-#define VECTOR_REGISTERS 16
-#endif
-
-#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
-
-namespace {
-
-inline float unhalf(ggml_fp16_t d) {
-    return GGML_FP16_TO_FP32(d);
-}
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// VECTORIZED ARITHMETIC OPERATIONS
-
-#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
-inline __m128 add(__m128 x, __m128 y) { return _mm_add_ps(x, y); }
-inline __m128 sub(__m128 x, __m128 y) { return _mm_sub_ps(x, y); }
-inline __m128 mul(__m128 x, __m128 y) { return _mm_mul_ps(x, y); }
-#endif  // __SSE__
-
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
-inline __m256 add(__m256 x, __m256 y) { return _mm256_add_ps(x, y); }
-inline __m256 sub(__m256 x, __m256 y) { return _mm256_sub_ps(x, y); }
-inline __m256 mul(__m256 x, __m256 y) { return _mm256_mul_ps(x, y); }
-#endif // __AVX__
-
-#if defined(__AVX512F__)
-inline __m512 add(__m512 x, __m512 y) { return _mm512_add_ps(x, y); }
-inline __m512 sub(__m512 x, __m512 y) { return _mm512_sub_ps(x, y); }
-inline __m512 mul(__m512 x, __m512 y) { return _mm512_mul_ps(x, y); }
-#endif // __AVX512F__
-
-#if defined(__ARM_NEON)
-inline float32x4_t add(float32x4_t x, float32x4_t y) { return vaddq_f32(x, y); }
-inline float32x4_t sub(float32x4_t x, float32x4_t y) { return vsubq_f32(x, y); }
-inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vmulq_f32(x, y); }
-#endif // __ARM_NEON
-
-#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
-inline float16x8_t add(float16x8_t x, float16x8_t y) { return vaddq_f16(x, y); }
-inline float16x8_t sub(float16x8_t x, float16x8_t y) { return vsubq_f16(x, y); }
-inline float16x8_t mul(float16x8_t x, float16x8_t y) { return vmulq_f16(x, y); }
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// VECTORIZED FUSED MULTIPLY ADD
-
-/**
- * Computes a * b + c.
- */
-template <typename T, typename U>
-inline U madd(T a, T b, U c) {
-    return add(mul(a, b), c);
-}
-
-#if defined(__FMA__)
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
-template <>
-inline __m256 madd(__m256 a, __m256 b, __m256 c) {
-    return _mm256_fmadd_ps(a, b, c);
-}
-#endif
-#if defined(__AVX512F__)
-template <>
-inline __m512 madd(__m512 a, __m512 b, __m512 c) {
-    return _mm512_fmadd_ps(a, b, c);
-}
-#endif
-#endif
-
-#if defined(__ARM_FEATURE_FMA)
-template <>
-inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
-    return vfmaq_f32(c, b, a);
-}
-#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
-template <>
-inline float16x8_t madd(float16x8_t a, float16x8_t b, float16x8_t c) {
-    return vfmaq_f16(c, b, a);
-}
-#endif
-#endif
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// VECTORIZED HORIZONTAL SUM
-
-#if defined(__ARM_NEON)
-inline float hsum(float32x4_t x) {
-    return vaddvq_f32(x);
-}
-#endif // __ARM_NEON
-
-#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
-inline float hsum(float16x8_t x) {
-    return vaddvq_f32(vaddq_f32(vcvt_f32_f16(vget_low_f16(x)),
-                                vcvt_f32_f16(vget_high_f16(x))));
-}
-#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-
-#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
-inline float hsum(__m128 x) {
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
-    x = _mm_add_ps(x, _mm_movehl_ps(x, x));
-    x = _mm_add_ss(x, _mm_movehdup_ps(x));
-#else
-    __m128 t;
-    t = _mm_shuffle_ps(x, x, _MM_SHUFFLE(2, 3, 0, 1));
-    x = _mm_add_ps(x, t);
-    t = _mm_movehl_ps(t, x);
-    x = _mm_add_ss(x, t);
-#endif
-    return _mm_cvtss_f32(x);
-}
-#endif
-
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
-inline float hsum(__m256 x) {
-    return hsum(_mm_add_ps(_mm256_extractf128_ps(x, 1),
-                           _mm256_castps256_ps128(x)));
-}
-#endif // __AVX__
-
-#if defined(__AVX512F__)
-inline float hsum(__m512 x) {
-    return _mm512_reduce_add_ps(x);
-}
-#endif // __AVX512F__
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// VECTORIZED MEMORY LOADING
-
-template <typename T, typename U> T load(const U *);
-
-#if defined(__ARM_NEON)
-template <> inline float32x4_t load(const float *p) {
-    return vld1q_f32(p);
-}
-#if !defined(_MSC_VER)
-template <> inline float16x8_t load(const ggml_fp16_t *p) {
-    return vld1q_f16((const float16_t *)p);
-}
-template <> inline float32x4_t load(const ggml_fp16_t *p) {
-    return vcvt_f32_f16(vld1_f16((const float16_t *)p));
-}
-#endif // _MSC_VER
-#endif // __ARM_NEON
-
-#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
-template <> inline __m128 load(const float *p) {
-    return _mm_loadu_ps(p);
-}
-#endif  // __SSE__
-
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
-template <> inline __m256 load(const float *p) {
-    return _mm256_loadu_ps(p);
-}
-#endif // __AVX__
-
-#if defined(__F16C__)
-template <> inline __m256 load(const ggml_fp16_t *p) {
-    return _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)p));
-}
-#endif // __F16C__
-
-#if defined(__AVX512F__)
-template <> inline __m512 load(const float *p) {
-    return _mm512_loadu_ps(p);
-}
-template <> inline __m512 load(const ggml_fp16_t *p) {
-    return _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)p));
-}
-#endif // __AVX512F__
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// FLOATING POINT MATRIX MULTIPLICATION
-
-template <int KN, typename D, typename V, typename TA, typename TB, typename TC>
-class tinyBLAS {
-  public:
-    tinyBLAS(int64_t k,
-             const TA *A, int64_t lda,
-             const TB *B, int64_t ldb,
-             TC *C, int64_t ldc,
-             int ith, int nth)
-        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
-    }
-
-    void matmul(int64_t m, int64_t n) {
-        mnpack(0, m, 0, n);
-    }
-
-  private:
-    NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
-        int64_t mc, nc, mp, np;
-        switch ((MIN(m - m0, 5) << 4) | MIN(n - n0, 5)) {
-#if VECTOR_REGISTERS == 32
-        case 0x55:
-            mc = 5;
-            nc = 5;
-            gemm<5, 5>(m0, m, n0, n);
-            break;
-        case 0x45:
-            mc = 4;
-            nc = 5;
-            gemm<4, 5>(m0, m, n0, n);
-            break;
-        case 0x54:
-            mc = 5;
-            nc = 4;
-            gemm<5, 4>(m0, m, n0, n);
-            break;
-        case 0x44:
-            mc = 4;
-            nc = 4;
-            gemm<4, 4>(m0, m, n0, n);
-            break;
-        case 0x53:
-            mc = 5;
-            nc = 3;
-            gemm<5, 3>(m0, m, n0, n);
-            break;
-        case 0x35:
-            mc = 3;
-            nc = 5;
-            gemm<3, 5>(m0, m, n0, n);
-            break;
-        case 0x43:
-            mc = 4;
-            nc = 3;
-            gemm<4, 3>(m0, m, n0, n);
-            break;
-#else
-        case 0x55:
-        case 0x54:
-        case 0x53:
-        case 0x45:
-        case 0x44:
-        case 0x43:
-            mc = 4;
-            nc = 3;
-            gemm<4, 3>(m0, m, n0, n);
-            break;
-        case 0x35:
-#endif
-        case 0x34:
-            mc = 3;
-            nc = 4;
-            gemm<3, 4>(m0, m, n0, n);
-            break;
-        case 0x52:
-            mc = 5;
-            nc = 2;
-            gemm<5, 2>(m0, m, n0, n);
-            break;
-        case 0x33:
-            mc = 3;
-            nc = 3;
-            gemm<3, 3>(m0, m, n0, n);
-            break;
-        case 0x25:
-            mc = 2;
-            nc = 5;
-            gemm<2, 5>(m0, m, n0, n);
-            break;
-        case 0x42:
-            mc = 4;
-            nc = 2;
-            gemm<4, 2>(m0, m, n0, n);
-            break;
-        case 0x24:
-            mc = 2;
-            nc = 4;
-            gemm<2, 4>(m0, m, n0, n);
-            break;
-        case 0x32:
-            mc = 3;
-            nc = 2;
-            gemm<3, 2>(m0, m, n0, n);
-            break;
-        case 0x23:
-            mc = 2;
-            nc = 3;
-            gemm<2, 3>(m0, m, n0, n);
-            break;
-        case 0x51:
-            mc = 5;
-            nc = 1;
-            gemm<5, 1>(m0, m, n0, n);
-            break;
-        case 0x41:
-            mc = 4;
-            nc = 1;
-            gemm<4, 1>(m0, m, n0, n);
-            break;
-        case 0x22:
-            mc = 2;
-            nc = 2;
-            gemm<2, 2>(m0, m, n0, n);
-            break;
-        case 0x15:
-            mc = 1;
-            nc = 5;
-            gemm<1, 5>(m0, m, n0, n);
-            break;
-        case 0x14:
-            mc = 1;
-            nc = 4;
-            gemm<1, 4>(m0, m, n0, n);
-            break;
-        case 0x31:
-            mc = 3;
-            nc = 1;
-            gemm<3, 1>(m0, m, n0, n);
-            break;
-        case 0x13:
-            mc = 1;
-            nc = 3;
-            gemm<1, 3>(m0, m, n0, n);
-            break;
-        case 0x21:
-            mc = 2;
-            nc = 1;
-            gemm<2, 1>(m0, m, n0, n);
-            break;
-        case 0x12:
-            mc = 1;
-            nc = 2;
-            gemm<1, 2>(m0, m, n0, n);
-            break;
-        case 0x11:
-            mc = 1;
-            nc = 1;
-            gemm<1, 1>(m0, m, n0, n);
-            break;
-        default:
-            return;
-        }
-        mp = m0 + (m - m0) / mc * mc;
-        np = n0 + (n - n0) / nc * nc;
-        mnpack(mp, m, n0, np);
-        mnpack(m0, m, np, n);
-    }
-
-    template <int RM, int RN>
-    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
-        int64_t ytiles = (m - m0) / RM;
-        int64_t xtiles = (n - n0) / RN;
-        int64_t tiles = xtiles * ytiles;
-        int64_t duty = (tiles + nth - 1) / nth;
-        int64_t start = duty * ith;
-        int64_t end = start + duty;
-        if (end > tiles)
-            end = tiles;
-        for (int64_t job = start; job < end; ++job) {
-            int64_t ii = m0 + job / xtiles * RM;
-            int64_t jj = n0 + job % xtiles * RN;
-            D Cv[RN][RM] = {};
-            for (int64_t l = 0; l < k; l += KN)
-                for (int64_t j = 0; j < RN; ++j)
-                    for (int64_t i = 0; i < RM; ++i)
-                        Cv[j][i] = madd(load<V>(A + lda * (ii + i) + l),
-                                        load<V>(B + ldb * (jj + j) + l),
-                                        Cv[j][i]);
-            for (int64_t j = 0; j < RN; ++j)
-                for (int64_t i = 0; i < RM; ++i)
-                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
-        }
-    }
-
-    const TA *const A;
-    const TB *const B;
-    TC *const C;
-    const int64_t k;
-    const int64_t lda;
-    const int64_t ldb;
-    const int64_t ldc;
-    const int ith;
-    const int nth;
-};
-
-//////////////////////////////////////////////////////////////////////////////////////////
-// QUANT ZERO MATRIX MULTIPLICATION
-
-#if defined(__ARM_FEATURE_DOTPROD)
-template <typename TA>
-class tinyBLAS_Q0_ARM {
-  public:
-    tinyBLAS_Q0_ARM(int64_t k,
-                    const TA *A, int64_t lda,
-                    const block_q8_0 *B, int64_t ldb,
-                    float *C, int64_t ldc,
-                    int ith, int nth)
-        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
-    }
-
-    void matmul(int64_t m, int64_t n) {
-        mnpack(0, m, 0, n);
-    }
-
-  private:
-    NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
-        int64_t mc, nc, mp, np;
-        switch ((MIN(m - m0, 3) << 4) | MIN(n - n0, 3ll)) {
-        case 0x33:
-            mc = 3;
-            nc = 3;
-            gemm<3, 3>(m0, m, n0, n);
-            break;
-        case 0x32:
-            mc = 3;
-            nc = 2;
-            gemm<3, 2>(m0, m, n0, n);
-            break;
-        case 0x23:
-            mc = 2;
-            nc = 3;
-            gemm<2, 3>(m0, m, n0, n);
-            break;
-        case 0x22:
-            mc = 2;
-            nc = 2;
-            gemm<2, 2>(m0, m, n0, n);
-            break;
-        case 0x31:
-            mc = 3;
-            nc = 1;
-            gemm<3, 1>(m0, m, n0, n);
-            break;
-        case 0x13:
-            mc = 1;
-            nc = 3;
-            gemm<1, 3>(m0, m, n0, n);
-            break;
-        case 0x21:
-            mc = 2;
-            nc = 1;
-            gemm<2, 1>(m0, m, n0, n);
-            break;
-        case 0x12:
-            mc = 1;
-            nc = 2;
-            gemm<1, 2>(m0, m, n0, n);
-            break;
-        case 0x11:
-            mc = 1;
-            nc = 1;
-            gemm<1, 1>(m0, m, n0, n);
-            break;
-        default:
-            return;
-        }
-        mp = m0 + (m - m0) / mc * mc;
-        np = n0 + (n - n0) / nc * nc;
-        mnpack(mp, m, n0, np);
-        mnpack(m0, m, np, n);
-    }
-
-    template <int RM, int RN>
-    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
-        int64_t ytiles = (m - m0) / RM;
-        int64_t xtiles = (n - n0) / RN;
-        int64_t tiles = xtiles * ytiles;
-        int64_t duty = (tiles + nth - 1) / nth;
-        int64_t start = duty * ith;
-        int64_t end = start + duty;
-        if (end > tiles)
-            end = tiles;
-        for (int64_t job = start; job < end; ++job) {
-            int64_t ii = m0 + job / xtiles * RM;
-            int64_t jj = n0 + job % xtiles * RN;
-            float32x4_t Cv[RN][RM] = {};
-            for (int64_t l = 0; l < k; ++l)
-                for (int64_t j = 0; j < RN; ++j)
-                    for (int64_t i = 0; i < RM; ++i)
-                        Cv[j][i] = vmlaq_n_f32(Cv[j][i],
-                                               vcvtq_f32_s32(vdotq_s32(
-                                                   vdotq_s32(vdupq_n_s32(0),
-                                                             load_lo(A + lda * (ii + i) + l),
-                                                             load_lo(B + ldb * (jj + j) + l)),
-                                                   load_hi(A + lda * (ii + i) + l),
-                                                   load_hi(B + ldb * (jj + j) + l))),
-                                               unhalf(A[lda * (ii + i) + l].d) *
-                                               unhalf(B[ldb * (jj + j) + l].d));
-            for (int64_t j = 0; j < RN; ++j)
-                for (int64_t i = 0; i < RM; ++i)
-                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
-        }
-    }
-
-    inline int8x16_t load_lo(const block_q8_0 *b) {
-        return vld1q_s8(b->qs);
-    }
-
-    inline int8x16_t load_hi(const block_q8_0 *b) {
-        return vld1q_s8(b->qs + 16);
-    }
-
-    inline int8x16_t load_lo(const block_q4_0 *b) {
-        return vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vld1q_u8(b->qs),
-                                                     vdupq_n_u8(0x0f))),
-                        vdupq_n_s8(0x8));
-    }
-
-    inline int8x16_t load_hi(const block_q4_0 *b) {
-        return vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(vld1q_u8(b->qs), 4)),
-                        vdupq_n_s8(0x8));
-    }
-
-    const TA *const A;
-    const block_q8_0 *const B;
-    float *const C;
-    const int64_t k;
-    const int64_t lda;
-    const int64_t ldb;
-    const int64_t ldc;
-    const int ith;
-    const int nth;
-};
-#endif // __ARM_FEATURE_DOTPROD
-
-#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
-template <typename TA, typename TB, typename TC>
-class tinyBLAS_Q0_AVX {
-  public:
-    tinyBLAS_Q0_AVX(int64_t k,
-                    const TA *A, int64_t lda,
-                    const TB *B, int64_t ldb,
-                    TC *C, int64_t ldc,
-                    int ith, int nth)
-        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
-    }
-
-    void matmul(int64_t m, int64_t n) {
-        mnpack(0, m, 0, n);
-    }
-
-  private:
-    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
-        int64_t mc, nc, mp, np;
-        switch ((MIN(m - m0, 4) << 4) | MIN(n - n0, 4)) {
-#if VECTOR_REGISTERS == 32
-        case 0x44:
-            mc = 4;
-            nc = 4;
-            gemm<4, 4>(m0, m, n0, n);
-            break;
-        case 0x43:
-            mc = 4;
-            nc = 3;
-            gemm<4, 3>(m0, m, n0, n);
-            break;
-        case 0x34:
-            mc = 3;
-            nc = 4;
-            gemm<3, 4>(m0, m, n0, n);
-            break;
-        case 0x33:
-            mc = 3;
-            nc = 3;
-            gemm<3, 3>(m0, m, n0, n);
-            break;
-        case 0x42:
-            mc = 4;
-            nc = 2;
-            gemm<4, 2>(m0, m, n0, n);
-            break;
-        case 0x24:
-            mc = 2;
-            nc = 4;
-            gemm<2, 4>(m0, m, n0, n);
-            break;
-#else
-        case 0x44:
-        case 0x43:
-        case 0x42:
-            mc = 4;
-            nc = 2;
-            gemm<4, 2>(m0, m, n0, n);
-            break;
-        case 0x34:
-        case 0x24:
-            mc = 2;
-            nc = 4;
-            gemm<2, 4>(m0, m, n0, n);
-            break;
-        case 0x33:
-#endif
-        case 0x32:
-            mc = 3;
-            nc = 2;
-            gemm<3, 2>(m0, m, n0, n);
-            break;
-        case 0x23:
-            mc = 2;
-            nc = 3;
-            gemm<2, 3>(m0, m, n0, n);
-            break;
-        case 0x41:
-            mc = 4;
-            nc = 1;
-            gemm<4, 1>(m0, m, n0, n);
-            break;
-        case 0x22:
-            mc = 2;
-            nc = 2;
-            gemm<2, 2>(m0, m, n0, n);
-            break;
-        case 0x14:
-            mc = 1;
-            nc = 4;
-            gemm<1, 4>(m0, m, n0, n);
-            break;
-        case 0x31:
-            mc = 3;
-            nc = 1;
-            gemm<3, 1>(m0, m, n0, n);
-            break;
-        case 0x13:
-            mc = 1;
-            nc = 3;
-            gemm<1, 3>(m0, m, n0, n);
-            break;
-        case 0x21:
-            mc = 2;
-            nc = 1;
-            gemm<2, 1>(m0, m, n0, n);
-            break;
-        case 0x12:
-            mc = 1;
-            nc = 2;
-            gemm<1, 2>(m0, m, n0, n);
-            break;
-        case 0x11:
-            mc = 1;
-            nc = 1;
-            gemm<1, 1>(m0, m, n0, n);
-            break;
-        default:
-            return;
-        }
-        mp = m0 + (m - m0) / mc * mc;
-        np = n0 + (n - n0) / nc * nc;
-        mnpack(mp, m, n0, np);
-        mnpack(m0, m, np, n);
-    }
-
-    template <int RM, int RN>
-    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
-        int64_t ytiles = (m - m0) / RM;
-        int64_t xtiles = (n - n0) / RN;
-        int64_t tiles = xtiles * ytiles;
-        int64_t duty = (tiles + nth - 1) / nth;
-        int64_t start = duty * ith;
-        int64_t end = start + duty;
-        if (end > tiles)
-            end = tiles;
-        for (int64_t job = start; job < end; ++job) {
-            int64_t ii = m0 + job / xtiles * RM;
-            int64_t jj = n0 + job % xtiles * RN;
-            __m256 Cv[RN][RM] = {};
-            for (int64_t l = 0; l < k; ++l)
-                for (int64_t j = 0; j < RN; ++j)
-                    for (int64_t i = 0; i < RM; ++i) {
-#if defined(__AVX2__)
-                        __m256 udTmp = updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
-                                                              load(A + lda * (ii + i) + l)),
-                                             _mm256_sign_epi8(load(B + ldb * (jj + j) + l),
-                                                              load(A + lda * (ii + i) + l)));
-#else
-                        __m128i ali0 = load0(A + lda * (ii + i) + l);
-                        __m128i ali1 = load1(A + lda * (ii + i) + l);
-                        __m128i blj0 = load0(B + ldb * (jj + j) + l);
-                        __m128i blj1 = load1(B + ldb * (jj + j) + l);
-
-                        __m128i sepAA0 = _mm_sign_epi8(ali0, ali0);
-                        __m128i sepAA1 = _mm_sign_epi8(ali1, ali1);
-                        __m128i sepBA0 = _mm_sign_epi8(blj0, ali0);
-                        __m128i sepBA1 = _mm_sign_epi8(blj1, ali1);
-
-                        // updot
-                        const __m128i oneFill = _mm_set1_epi16(1);
-                        __m128i mad0 = _mm_maddubs_epi16(sepAA0, sepBA0);
-                        __m128i mad1 = _mm_maddubs_epi16(sepAA1, sepBA1);
-                        __m256 udTmp = _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_madd_epi16(oneFill, mad1), _mm_madd_epi16(oneFill, mad0)));
-#endif
-                        Cv[j][i] = madd(_mm256_set1_ps(unhalf(A[lda * (ii + i) + l].d) *
-                                                       unhalf(B[ldb * (jj + j) + l].d)),
-                                                       udTmp,
-                                                       Cv[j][i]);
-                    }
-            for (int64_t j = 0; j < RN; ++j)
-                for (int64_t i = 0; i < RM; ++i)
-                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
-        }
-    }
-
-    inline __m256i load(const block_q8_0 *b) {
-        return _mm256_loadu_si256((const __m256i *)b->qs);
-    }
-
-    inline __m128i load0(const block_q8_0 *b) {
-        return _mm_loadu_si128((const __m128i *)b->qs);
-    }
-
-    inline __m128i load1(const block_q8_0 *b) {
-        return _mm_loadu_si128(((const __m128i *)b->qs) + 1);
-    }
-
-    inline __m256i load(const block_q4_0 *b) {
-        return _mm256_sub_epi8(denibble(b->qs), _mm256_set1_epi8(8));
-    }
-
-    inline __m128i load0(const block_q4_0 *b) {
-        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
-        return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), x), _mm_set1_epi8(8));
-    }
-
-    inline __m128i load1(const block_q4_0 *b) {
-        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
-        return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)), _mm_set1_epi8(8));
-    }
-
-    inline __m256 updot(__m256i u, __m256i s) {
-        __m256i res;
-#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
-        res = _mm256_dpbusd_epi32(_mm256_setzero_si256(), u, s);
-#else
-        res = _mm256_madd_epi16(_mm256_set1_epi16(1), _mm256_maddubs_epi16(u, s));
-#endif
-        return _mm256_cvtepi32_ps(res);
-    }
-
-    static inline __m256i denibble(const uint8_t *p) {
-        __m128i x = _mm_loadu_si128((const __m128i *)p);
-        return _mm256_and_si256(_mm256_set1_epi8(15),
-                                _mm256_insertf128_si256(_mm256_castsi128_si256(x),
-                                                        _mm_srli_epi16(x, 4), 1));
-    }
-
-    const TA *const A;
-    const TB *const B;
-    TC *const C;
-    const int64_t k;
-    const int64_t lda;
-    const int64_t ldb;
-    const int64_t ldc;
-    const int ith;
-    const int nth;
-};
-#endif // __AVX__
-
-} // namespace
-
-/**
- * Performs optimized matrix multiplication on CPU.
- *
- * This subroutine may compute C = Aᵀ * B with column major ordering.
- * Despite its name, this isn't a generalized implementation. Work is
- * only performed when a handwritten kernel is written and available.
- * Otherwise the caller should fall back to a general matmul routine.
- *
- * For example, for single-threaded single-precision GEMM you can say
- *
- *     llamafile_sgemm(m, n, k, A, lda, B, ldb, C, ldc,
- *                     0, 1,
- *                     GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32);
- *
- * @param m is rows in `A` and `C`
- * @param n is cols in `B` and `C`
- * @param k is cols in `A` and rows in `B`
- * @param A is first input matrix (always transposed)
- * @param lda is row stride of `A`
- * @param B is second input matrix (never transposed)
- * @param ldb is row stride of `B`
- * @param C is input/output array of output matrices
- * @param ldc is row stride of `C`
- * @param ith is thread id (must be less than `nth`)
- * @param nth is number of threads (must be greater than zero)
- * @param Atype is GGML data type of `A`
- * @param Btype is GGML data type of `B`
- * @param Ctype is GGML data type of `C`
- * @return true if this function was able to service the matmul request
- */
-bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
-                     int64_t ldc, int ith, int nth, int Atype, int Btype, int Ctype) {
-
-    assert(m >= 0);
-    assert(n >= 0);
-    assert(k >= 0);
-    assert(lda >= k);
-    assert(ldb >= k);
-    assert(ldc >= m);
-    assert(nth > 0);
-    assert(ith < nth);
-
-    if (Ctype != GGML_TYPE_F32)
-        return false;
-
-    switch (Atype) {
-
-    case GGML_TYPE_F32: {
-        if (Btype != GGML_TYPE_F32)
-            return false;
-#if defined(__AVX512F__)
-        if (k % 16)
-            return false;
-        tinyBLAS<16, __m512, __m512, float, float, float> tb{
-            k, (const float *)A, lda,
-            (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
-#elif defined(__AVX__) || defined(__AVX2__)
-        if (k % 8)
-            return false;
-        tinyBLAS<8, __m256, __m256, float, float, float> tb{
-            k, (const float *)A, lda,
-            (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
-#elif defined(__ARM_NEON)
-        if (n < 4)
-            return false;
-        if (k % 4)
-            return false;
-        tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{
-            k, (const float *)A, lda,
-            (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
-#else
-        return false;
-#endif
-    }
-
-    case GGML_TYPE_F16: {
-#if defined(__AVX512F__)
-        if (k % 16)
-            return false;
-        if (Btype != GGML_TYPE_F32)
-            return false;
-        tinyBLAS<16, __m512, __m512, ggml_fp16_t, float, float> tb{
-            k, (const ggml_fp16_t *)A, lda,
-            (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
-#elif (defined(__AVX__) || defined(__AVX2__)) && defined(__F16C__)
-        if (k % 8)
-            return false;
-        if (Btype != GGML_TYPE_F32)
-            return false;
-        tinyBLAS<8, __m256, __m256, ggml_fp16_t, float, float> tb{
-            k, (const ggml_fp16_t *)A, lda,
-            (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
-#elif defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
-        if (n < 8)
-            return false;
-        if (k % 8)
-            return false;
-        if (Btype != GGML_TYPE_F16)
-            return false;
-        tinyBLAS<8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, float> tb{
-            k, (const ggml_fp16_t *)A, lda,
-            (const ggml_fp16_t *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
-#elif defined(__ARM_NEON) && !defined(_MSC_VER)
-        if (k % 4)
-            return false;
-        if (Btype != GGML_TYPE_F32)
-            return false;
-        tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, float, float> tb{
-            k, (const ggml_fp16_t *)A, lda,
-            (const float *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
-#else
-        return false;
-#endif
-    }
-
-    case GGML_TYPE_Q8_0: {
-        if (Btype != GGML_TYPE_Q8_0)
-           return false;
-#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
-        tinyBLAS_Q0_AVX<block_q8_0, block_q8_0, float> tb{
-            k, (const block_q8_0 *)A, lda,
-            (const block_q8_0 *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
-#elif defined(__ARM_FEATURE_DOTPROD)
-        tinyBLAS_Q0_ARM<block_q8_0> tb{
-            k, (const block_q8_0 *)A, lda,
-            (const block_q8_0 *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
-#else
-        return false;
-#endif
-    }
-
-    case GGML_TYPE_Q4_0: {
-        if (Btype != GGML_TYPE_Q8_0)
-            return false;
-#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
-        tinyBLAS_Q0_AVX<block_q4_0, block_q8_0, float> tb{
-            k, (const block_q4_0 *)A, lda,
-            (const block_q8_0 *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
-#elif defined(__ARM_FEATURE_DOTPROD)
-        tinyBLAS_Q0_ARM<block_q4_0> tb{
-            k, (const block_q4_0 *)A, lda,
-            (const block_q8_0 *)B, ldb,
-            (float *)C, ldc,
-            ith, nth};
-        tb.matmul(m, n);
-        return true;
-#else
-        return false;
-#endif
-    }
-
-    default:
-        return false;
-    }
-
-    (void)m;
-    (void)n;
-    (void)k;
-    (void)A;
-    (void)lda;
-    (void)B;
-    (void)ldb;
-    (void)C;
-    (void)ldc;
-    (void)ith;
-    (void)nth;
-    (void)Atype;
-    (void)Btype;
-    (void)Ctype;
-}
diff --git a/ggml/src/vulkan-shaders/add.comp b/ggml/src/vulkan-shaders/add.comp
deleted file mode 100644
index 3974845d637..00000000000
--- a/ggml/src/vulkan-shaders/add.comp
+++ /dev/null
@@ -1,14 +0,0 @@
-#version 450
-
-#include "types.comp"
-#include "generic_binary_head.comp"
-
-void main() {
-    const uint idx = get_idx();
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) + FLOAT_TYPE(data_b[src1_idx(idx)]));
-}
diff --git a/ggml/src/vulkan-shaders/dequant_funcs.comp b/ggml/src/vulkan-shaders/dequant_funcs.comp
deleted file mode 100644
index d5b989735bc..00000000000
--- a/ggml/src/vulkan-shaders/dequant_funcs.comp
+++ /dev/null
@@ -1,68 +0,0 @@
-#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
-#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
-#endif
-
-#if defined(DATA_A_F32)
-vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
-}
-#endif
-
-#if defined(DATA_A_F16)
-vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
-}
-#endif
-
-#if defined(DATA_A_Q4_0)
-vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
-    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return (vec2(vui & 0xF, vui >> 4) - 8.0f) * d;
-}
-#endif
-
-#if defined(DATA_A_Q4_1)
-vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
-    const float m = float(data_a[a_offset + ib].m);
-    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return vec2(vui & 0xF, vui >> 4) * d + m;
-}
-#endif
-
-#if defined(DATA_A_Q5_0)
-vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
-    const uint uint_qh = uint(data_a[a_offset + ib].qh[1]) << 16 | data_a[a_offset + ib].qh[0];
-    const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
-    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d;
-}
-#endif
-
-#if defined(DATA_A_Q5_1)
-vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
-    const float m = float(data_a[a_offset + ib].m);
-    const uint uint_qh = data_a[a_offset + ib].qh;
-    const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
-    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m;
-}
-#endif
-
-#if defined(DATA_A_Q8_0)
-vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
-    return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])) * d;
-}
-#endif
-
-#if defined(DATA_A_IQ4_NL)
-vec2 dequantize(uint ib, uint iqs, uint a_offset) {
-    const float d = float(data_a[a_offset + ib].d);
-    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
-    return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d;
-}
-#endif
diff --git a/ggml/src/vulkan-shaders/dequant_q4_k.comp b/ggml/src/vulkan-shaders/dequant_q4_k.comp
deleted file mode 100644
index 92acb75406d..00000000000
--- a/ggml/src/vulkan-shaders/dequant_q4_k.comp
+++ /dev/null
@@ -1,56 +0,0 @@
-#version 450
-
-#include "dequant_head.comp"
-
-layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
-layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
-
-void main() {
-    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
-        const uint i = gl_WorkGroupID.x * 256 + wgy;
-        if (i >= p.M * p.K / QUANT_K) {
-            return;
-        }
-
-        const uint tid = gl_LocalInvocationID.x;
-        const uint il = tid / 8;
-        const uint ir = tid % 8;
-        const uint is = 2 * il;
-        const uint n = 4;
-
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
-
-        const uint y_idx = i * QUANT_K + 64 * il + n * ir;
-        const uint qs_idx = 32*il + n * ir;
-
-        uint8_t sc;
-        uint8_t m;
-        if (is < 4) {
-            sc = uint8_t(data_a[i].scales[is] & 63);
-            m  = uint8_t(data_a[i].scales[is + 4] & 63);
-        } else {
-            sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
-            m  = uint8_t((data_a[i].scales[is + 4] >>  4) | ((data_a[i].scales[is    ] >> 6) << 4));
-        }
-        const FLOAT_TYPE d1 = dall * sc;
-        const FLOAT_TYPE m1 = dmin * m;
-
-        if (is < 4) {
-            sc = uint8_t(data_a[i].scales[is + 1] & 63);
-            m  = uint8_t(data_a[i].scales[is + 5] & 63);
-        } else {
-            sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4));
-            m  = uint8_t((data_a[i].scales[is + 5] >>  4) | ((data_a[i].scales[is + 1] >> 6) << 4));
-        }
-        const FLOAT_TYPE d2 = dall * sc;
-        const FLOAT_TYPE m2 = dmin * m;
-
-        [[unroll]] for (uint l = 0; l < n; ++l) {
-            data_b[y_idx + l     ] = D_TYPE(d1 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] & 0xF) - m1);
-            data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] >>  4) - m2);
-        }
-    }
-}
diff --git a/ggml/src/vulkan-shaders/dequant_q5_k.comp b/ggml/src/vulkan-shaders/dequant_q5_k.comp
deleted file mode 100644
index f314a76d105..00000000000
--- a/ggml/src/vulkan-shaders/dequant_q5_k.comp
+++ /dev/null
@@ -1,58 +0,0 @@
-#version 450
-
-#include "dequant_head.comp"
-
-layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
-layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
-
-void main() {
-    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
-        const uint i = gl_WorkGroupID.x * 256 + wgy;
-        if (i >= p.M * p.K / QUANT_K) {
-            return;
-        }
-
-        const uint tid = gl_LocalInvocationID.x;
-        const uint il = tid / 16;
-        const uint ir = tid % 16;
-        const uint is = 2 * il;
-
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
-
-        const uint y_idx = i * QUANT_K + 64 * il + 2 * ir;
-        const uint qs_idx = 32*il + 2 * ir;
-        const uint qh_idx = 2 * ir;
-
-        uint8_t sc;
-        uint8_t m;
-        if (is < 4) {
-            sc = uint8_t(data_a[i].scales[is] & 63);
-            m  = uint8_t(data_a[i].scales[is + 4] & 63);
-        } else {
-            sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
-            m  = uint8_t((data_a[i].scales[is + 4] >>  4) | ((data_a[i].scales[is    ] >> 6) << 4));
-        }
-        const FLOAT_TYPE d1 = dall * sc;
-        const FLOAT_TYPE m1 = dmin * m;
-
-        if (is < 4) {
-            sc = uint8_t(data_a[i].scales[is + 1] & 63);
-            m  = uint8_t(data_a[i].scales[is + 5] & 63);
-        } else {
-            sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4));
-            m  = uint8_t((data_a[i].scales[is + 5] >>  4) | ((data_a[i].scales[is + 1] >> 6) << 4));
-        }
-        const FLOAT_TYPE d2 = dall * sc;
-        const FLOAT_TYPE m2 = dmin * m;
-
-        const uint8_t hm1 = uint8_t(1 << (2 * il    ));
-        const uint8_t hm2 = uint8_t(1 << (2 * il + 1));
-        data_b[y_idx     ] = D_TYPE(d1 * FLOAT_TYPE((data_a[i].qs[qs_idx    ] & 0xF) + (((data_a[i].qh[qh_idx    ] & hm1) != 0) ? 16 : 0)) - m1);
-        data_b[y_idx +  1] = D_TYPE(d1 * FLOAT_TYPE((data_a[i].qs[qs_idx + 1] & 0xF) + (((data_a[i].qh[qh_idx + 1] & hm1) != 0) ? 16 : 0)) - m1);
-        data_b[y_idx + 32] = D_TYPE(d2 * FLOAT_TYPE((data_a[i].qs[qs_idx    ]  >> 4) + (((data_a[i].qh[qh_idx    ] & hm2) != 0) ? 16 : 0)) - m2);
-        data_b[y_idx + 33] = D_TYPE(d2 * FLOAT_TYPE((data_a[i].qs[qs_idx + 1]  >> 4) + (((data_a[i].qh[qh_idx + 1] & hm2) != 0) ? 16 : 0)) - m2);
-    }
-}
diff --git a/ggml/src/vulkan-shaders/div.comp b/ggml/src/vulkan-shaders/div.comp
deleted file mode 100644
index 8cfce58b150..00000000000
--- a/ggml/src/vulkan-shaders/div.comp
+++ /dev/null
@@ -1,14 +0,0 @@
-#version 450
-
-#include "types.comp"
-#include "generic_binary_head.comp"
-
-void main() {
-    const uint idx = get_idx();
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) / FLOAT_TYPE(data_b[src1_idx(idx)]));
-}
diff --git a/ggml/src/vulkan-shaders/generic_binary_head.comp b/ggml/src/vulkan-shaders/generic_binary_head.comp
deleted file mode 100644
index b6beaff1cf6..00000000000
--- a/ggml/src/vulkan-shaders/generic_binary_head.comp
+++ /dev/null
@@ -1,52 +0,0 @@
-#extension GL_EXT_shader_16bit_storage : require
-
-layout (push_constant) uniform parameter
-{
-    uint ne;
-    uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
-    uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
-    uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
-    uint d_offset;
-    float param1; float param2; int param3;
-} p;
-
-layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
-layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
-layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
-
-uint get_idx() {
-    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
-}
-
-uint src0_idx(uint idx) {
-    const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
-    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
-    const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
-    const uint i02_offset = i02*p.ne01*p.ne00;
-    const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
-    const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
-    return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00;
-}
-
-uint src1_idx(uint idx) {
-    const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
-    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
-    const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
-    const uint i02_offset = i02*p.ne01*p.ne00;
-    const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
-    const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
-
-    return (i03 % p.ne13)*p.nb13 + (i02 % p.ne12)*p.nb12 + (i01 % p.ne11)*p.nb11 + (i00 % p.ne10)*p.nb10;
-}
-
-uint dst_idx(uint idx) {
-    const uint i23 = idx / (p.ne22*p.ne21*p.ne20);
-    const uint i23_offset = i23 * p.ne22*p.ne21*p.ne20;
-    const uint i22 = (idx - i23_offset) / (p.ne21*p.ne20);
-    const uint i22_offset = i22*p.ne21*p.ne20;
-    const uint i21 = (idx - i23_offset - i22_offset) / p.ne20;
-    const uint i20 = idx - i23_offset - i22_offset - i21*p.ne20;
-    return i23*p.nb23 + i22*p.nb22 + i21*p.nb21 + i20*p.nb20;
-}
diff --git a/ggml/src/vulkan-shaders/mul.comp b/ggml/src/vulkan-shaders/mul.comp
deleted file mode 100644
index bfb61c92d68..00000000000
--- a/ggml/src/vulkan-shaders/mul.comp
+++ /dev/null
@@ -1,14 +0,0 @@
-#version 450
-
-#include "types.comp"
-#include "generic_binary_head.comp"
-
-void main() {
-    const uint idx = get_idx();
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) * FLOAT_TYPE(data_b[src1_idx(idx)]));
-}
diff --git a/ggml/src/vulkan-shaders/mul_mat_split_k_reduce.comp b/ggml/src/vulkan-shaders/mul_mat_split_k_reduce.comp
deleted file mode 100644
index 825b91031f5..00000000000
--- a/ggml/src/vulkan-shaders/mul_mat_split_k_reduce.comp
+++ /dev/null
@@ -1,29 +0,0 @@
-#version 450
-
-#extension GL_EXT_control_flow_attributes : enable
-
-layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer A {float data_a[];};
-layout (binding = 1) writeonly buffer D {float data_d[];};
-
-layout (push_constant) uniform parameter {
-    uint ne;
-    uint k_num;
-} p;
-
-void main() {
-    const uint idx = gl_GlobalInvocationID.x;
-
-    if (idx >= p.ne) {
-        return;
-    }
-
-    float result = 0.0f;
-
-    [[unroll]] for (uint i = 0; i < p.k_num; i++) {
-        result += data_a[i * p.ne + idx];
-    }
-
-    data_d[idx] = result;
-}
diff --git a/ggml/src/vulkan-shaders/mul_mat_vec.comp b/ggml/src/vulkan-shaders/mul_mat_vec.comp
deleted file mode 100644
index d3ccba7fcb3..00000000000
--- a/ggml/src/vulkan-shaders/mul_mat_vec.comp
+++ /dev/null
@@ -1,56 +0,0 @@
-#version 450
-
-#ifdef FLOAT16
-#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
-#endif
-
-#include "mul_mat_vec_base.comp"
-
-layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
-
-layout (constant_id = 0) const uint BLOCK_SIZE = 32;
-
-shared FLOAT_TYPE tmp[BLOCK_SIZE];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-    const uint tid = gl_LocalInvocationID.x;
-
-    // There are not enough cols to use all threads
-    if (tid >= p.ncols) {
-        return;
-    }
-
-    const uint block_size = min(p.ncols, BLOCK_SIZE);
-
-    uint a_offset, b_offset, d_offset;
-    get_offsets(a_offset, b_offset, d_offset);
-
-    const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
-
-    tmp[tid] = FLOAT_TYPE(0.0f);
-
-    [[unroll]] for (uint i = 0; i < p.ncols/block_size; i += 2) {
-        const uint col = i*block_size + 2*tid;
-        const uint ib = (row*p.ncols + col)/QUANT_K; // block index
-        const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
-        const uint iybs = col - col%QUANT_K; // y block start index
-
-        vec2 v = dequantize(ib, iqs, a_offset / QUANT_K);
-
-        // matrix multiplication
-        tmp[tid] = fma(FLOAT_TYPE(v.x), FLOAT_TYPE(data_b[b_offset + iybs + iqs]), fma(FLOAT_TYPE(v.y), FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]), tmp[tid]));
-    }
-
-    // sum up partial sums and write back result
-    barrier();
-    [[unroll]] for (uint s = block_size/2; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-        }
-        barrier();
-    }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
-    }
-}
diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_q2_k.comp b/ggml/src/vulkan-shaders/mul_mat_vec_q2_k.comp
deleted file mode 100644
index ec8eadcd582..00000000000
--- a/ggml/src/vulkan-shaders/mul_mat_vec_q2_k.comp
+++ /dev/null
@@ -1,74 +0,0 @@
-#version 450
-
-#include "mul_mat_vec_base.comp"
-
-layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
-
-shared FLOAT_TYPE tmp[32];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    uint a_offset, b_offset, d_offset;
-    get_offsets(a_offset, b_offset, d_offset);
-
-    const uint num_blocks_per_row = p.ncols / QUANT_K;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
-
-    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
-
-    const uint step = 16/K_QUANTS_PER_ITERATION;            // 16 or 8
-
-    const uint v_im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const uint v_in = tid - step*v_im;                      // 0...15 or 0...7
-
-    const uint l0 = K_QUANTS_PER_ITERATION*v_in;            // 0...15
-    const uint q_offset = 32*v_im + l0;
-    const uint s_offset = 8*v_im;
-    const uint y_offset = 128*v_im + l0;
-
-    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
-
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-        const uint y_idx = i * QUANT_K + y_offset;
-
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
-
-        FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
-        FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum1 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 0) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 0) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 2) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 2) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 4) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 4) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 6) & 3),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 6) & 3), sum1))))))));
-            sum2 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 0] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 1] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 2] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 3] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 4] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 5] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 6] >> 4) & 0xF),
-                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 7] >> 4) & 0xF), sum2))))))));
-        }
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(dall, sum1, fma(-dmin, sum2, tmp[tmp_idx]));
-    }
-
-    // sum up partial sums and write back result
-    barrier();
-    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-        }
-        barrier();
-    }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
-    }
-}
diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_q3_k.comp b/ggml/src/vulkan-shaders/mul_mat_vec_q3_k.comp
deleted file mode 100644
index 3ca4ad85a5c..00000000000
--- a/ggml/src/vulkan-shaders/mul_mat_vec_q3_k.comp
+++ /dev/null
@@ -1,67 +0,0 @@
-#version 450
-
-#include "mul_mat_vec_base.comp"
-
-layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
-
-shared FLOAT_TYPE tmp[32];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    uint a_offset, b_offset, d_offset;
-    get_offsets(a_offset, b_offset, d_offset);
-
-    const uint num_blocks_per_row = p.ncols / QUANT_K;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
-
-    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
-
-    const uint step = 16/K_QUANTS_PER_ITERATION;            // 16 or 8
-
-    const uint v_im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const uint v_in = tid - step*v_im;                      // 0...15 or 0...7
-
-    const uint8_t m = uint8_t(1 << (4 * v_im));
-
-    const uint l0 = K_QUANTS_PER_ITERATION*v_in;            // 0...15
-    const uint q_offset = 32*v_im + l0;
-    const uint y_offset = 128*v_im + l0;
-
-    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
-
-    const uint s_shift = 4 * v_im;
-
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-        const uint y_idx = i * QUANT_K + y_offset;
-
-        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
-
-        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[0] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[2] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[4] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[6] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[1] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[3] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[5] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[7] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
-        }
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(d, sum, tmp[tmp_idx]);
-    }
-
-    // sum up partial sums and write back result
-    barrier();
-    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-        }
-        barrier();
-    }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
-    }
-}
diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_q4_k.comp b/ggml/src/vulkan-shaders/mul_mat_vec_q4_k.comp
deleted file mode 100644
index d91e00e1006..00000000000
--- a/ggml/src/vulkan-shaders/mul_mat_vec_q4_k.comp
+++ /dev/null
@@ -1,118 +0,0 @@
-#version 450
-
-#include "mul_mat_vec_base.comp"
-
-layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
-
-shared FLOAT_TYPE tmp[32];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    uint a_offset, b_offset, d_offset;
-    get_offsets(a_offset, b_offset, d_offset);
-
-    const uint num_blocks_per_row = p.ncols / QUANT_K;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
-
-    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
-
-    const uint step = 8/K_QUANTS_PER_ITERATION;             // 8 or 4
-
-    const uint il = tid/step;                               // 0...3
-    const uint ir = tid - step*il;                          // 0...7 or 0...3
-    const uint n =  2 * K_QUANTS_PER_ITERATION;             // 2 or 4
-
-    const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const uint v_in = il % 2;
-
-    const uint l0 = n * (2 * ir + v_in);            // 0...15
-    const uint q_offset = 32*v_im + l0;
-    const uint y_offset = 64*v_im + l0;
-
-    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
-
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-        const uint y1_idx = i * QUANT_K + y_offset;
-        const uint y2_idx = y1_idx + 128;
-
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
-
-        const uint8_t sc0 = uint8_t(  data_a[ib0 + i].scales[v_im * 2    ]       & 0x3f);
-        const uint8_t sc1 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 1]       & 0x3f);
-        const uint8_t sc2 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 4]       & 0x3f);
-        const uint8_t sc3 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 5]       & 0x3f);
-        const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2    ] & 0xc0) >> 2));
-        const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2));
-        const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2));
-        const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2));
-
-#if K_QUANTS_PER_ITERATION == 2
-        const uint8_t q4_0  = uint8_t(data_a[ib0 + i].qs[q_offset     ] & 0xf);
-        const uint8_t q4_1  = uint8_t(data_a[ib0 + i].qs[q_offset +  1] & 0xf);
-        const uint8_t q4_2  = uint8_t(data_a[ib0 + i].qs[q_offset +  2] & 0xf);
-        const uint8_t q4_3  = uint8_t(data_a[ib0 + i].qs[q_offset +  3] & 0xf);
-        const uint8_t q4_4  = uint8_t(data_a[ib0 + i].qs[q_offset     ]  >> 4);
-        const uint8_t q4_5  = uint8_t(data_a[ib0 + i].qs[q_offset +  1]  >> 4);
-        const uint8_t q4_6  = uint8_t(data_a[ib0 + i].qs[q_offset +  2]  >> 4);
-        const uint8_t q4_7  = uint8_t(data_a[ib0 + i].qs[q_offset +  3]  >> 4);
-        const uint8_t q4_8  = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
-        const uint8_t q4_9  = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
-        const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 66] & 0xf);
-        const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 67] & 0xf);
-        const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64]  >> 4);
-        const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65]  >> 4);
-        const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 66]  >> 4);
-        const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 67]  >> 4);
-
-        const FLOAT_TYPE sx = fma(FLOAT_TYPE(data_b[b_offset + y1_idx]),      q4_0,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]),  q4_1,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 2]),  q4_2,  FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) *  q4_3)));
-        const FLOAT_TYPE sy = fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), q4_4,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), q4_5,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 34]), q4_6,  FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * q4_7)));
-        const FLOAT_TYPE sz = fma(FLOAT_TYPE(data_b[b_offset + y2_idx]),      q4_8,  fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]),  q4_9,  fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 2]),  q4_10, FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) *  q4_11)));
-        const FLOAT_TYPE sw = fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), q4_12, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), q4_13, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 34]), q4_14, FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * q4_15)));
-        const FLOAT_TYPE smin =
-            fma(FLOAT_TYPE(data_b[b_offset + y1_idx    ]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx    ]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), sc7,
-            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), sc7,
-            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 2]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 34]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 2]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 34]), sc7,
-            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 3]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 35]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 3]), sc6,     FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * sc7)))))))))))))));
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, tmp[tmp_idx]));
-#else
-        const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset     ] & 0xf);
-        const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset +  1] & 0xf);
-        const uint8_t q4_2 = uint8_t(data_a[ib0 + i].qs[q_offset     ]  >> 4);
-        const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset +  1]  >> 4);
-        const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
-        const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
-        const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 64]  >> 4);
-        const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 65]  >> 4);
-
-        const FLOAT_TYPE sx = fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]), q4_0, FLOAT_TYPE(data_b[b_offset + y1_idx +  1]) * q4_1);
-        const FLOAT_TYPE sy = fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), q4_2, FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_3);
-        const FLOAT_TYPE sz = fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]), q4_4, FLOAT_TYPE(data_b[b_offset + y2_idx +  1]) * q4_5);
-        const FLOAT_TYPE sw = fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), q4_6, FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_7);
-        const FLOAT_TYPE smin =
-            fma(FLOAT_TYPE(data_b[b_offset + y1_idx    ]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx    ]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), sc7,
-          + fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), sc6, FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7)))))));
-
-        tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f) + sy * FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f) +
-                        sz * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)) + sw * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))) - dmin * smin);
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(dall, (fma(sx, FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f), fma(sy, FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f),
-                       fma(sz, FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)), fma(sw, FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))))))), fma(-dmin, smin, tmp[tmp_idx]));
-#endif
-    }
-
-    // sum up partial sums and write back result
-    barrier();
-    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-        }
-        barrier();
-    }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
-    }
-}
diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_q5_k.comp b/ggml/src/vulkan-shaders/mul_mat_vec_q5_k.comp
deleted file mode 100644
index 2306785af42..00000000000
--- a/ggml/src/vulkan-shaders/mul_mat_vec_q5_k.comp
+++ /dev/null
@@ -1,109 +0,0 @@
-#version 450
-
-#include "mul_mat_vec_base.comp"
-
-layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
-
-shared FLOAT_TYPE tmp[32];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    uint a_offset, b_offset, d_offset;
-    get_offsets(a_offset, b_offset, d_offset);
-
-    const uint num_blocks_per_row = p.ncols / QUANT_K;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
-
-    const uint tid = gl_LocalInvocationID.x/2;  // 0...31 or 0...16
-    const uint ix  = gl_LocalInvocationID.x%2;  // 0 or 0, 1
-
-    const uint il = tid/4;                           // 0...3
-    const uint ir = tid - 4*il;                      // 0...7 or 0...3
-
-    const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const uint v_in = il % 2;
-
-    const uint l0 = 4*ir + 2*v_in;                   // 0...15
-    const uint q_offset = 32*v_im + l0;
-    const uint y_offset = 64*v_im + l0;
-
-    const uint8_t hm1 = uint8_t(1 << (2*v_im));
-    const uint8_t hm2 = uint8_t(hm1 << 4);
-
-    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
-
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += 2) {
-        const uint y1_idx = i * QUANT_K + y_offset;
-        const uint y2_idx = y1_idx + 128;
-
-        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
-        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
-
-        const uint8_t sc0 = uint8_t(  data_a[ib0 + i].scales[v_im * 2    ]       & 0x3f);
-        const uint8_t sc1 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 1]       & 0x3f);
-        const uint8_t sc2 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 4]       & 0x3f);
-        const uint8_t sc3 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 5]       & 0x3f);
-        const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2    ] & 0xc0) >> 2));
-        const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2));
-        const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2));
-        const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2));
-
-        const uint8_t q4_0  = uint8_t(data_a[ib0 + i].qs[q_offset     ] & 0xf);
-        const uint8_t q4_1  = uint8_t(data_a[ib0 + i].qs[q_offset +  1] & 0xf);
-        const uint8_t q4_2  = uint8_t(data_a[ib0 + i].qs[q_offset + 16] & 0xf);
-        const uint8_t q4_3  = uint8_t(data_a[ib0 + i].qs[q_offset + 17] & 0xf);
-        const uint8_t q4_4  = uint8_t(data_a[ib0 + i].qs[q_offset     ]  >> 4);
-        const uint8_t q4_5  = uint8_t(data_a[ib0 + i].qs[q_offset +  1]  >> 4);
-        const uint8_t q4_6  = uint8_t(data_a[ib0 + i].qs[q_offset + 16]  >> 4);
-        const uint8_t q4_7  = uint8_t(data_a[ib0 + i].qs[q_offset + 17]  >> 4);
-        const uint8_t q4_8  = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
-        const uint8_t q4_9  = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
-        const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] & 0xf);
-        const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] & 0xf);
-        const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64]  >> 4);
-        const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65]  >> 4);
-        const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 80]  >> 4);
-        const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 81]  >> 4);
-
-        const FLOAT_TYPE sx =
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]), (q4_0 + (((data_a[ib0 + i].qh[l0     ] & hm1) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx +  1]), (q4_1 + (((data_a[ib0 + i].qh[l0 +  1] & hm1) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 16]), (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0)),
-             FLOAT_TYPE(data_b[b_offset + y1_idx + 17]) * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0)))));
-        const FLOAT_TYPE sy =
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), (q4_4 + (((data_a[ib0 + i].qh[l0     ] & (hm1 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), (q4_5 + (((data_a[ib0 + i].qh[l0 +  1] & (hm1 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 48]), (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0)),
-             FLOAT_TYPE(data_b[b_offset + y1_idx + 49]) * (q4_7 + (((data_a[ib0 + i].qh[l0 + 17] & (hm1 << 1)) != 0) ? 16 : 0)))));
-        const FLOAT_TYPE sz =
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]), (q4_8  + (((data_a[ib0 + i].qh[l0     ] & hm2) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx +  1]), (q4_9  + (((data_a[ib0 + i].qh[l0 +  1] & hm2) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 16]), (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0)),
-             FLOAT_TYPE(data_b[b_offset + y2_idx + 17]) * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0)))));
-        const FLOAT_TYPE sw =
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), (q4_12 + (((data_a[ib0 + i].qh[l0     ] & (hm2 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), (q4_13 + (((data_a[ib0 + i].qh[l0 +  1] & (hm2 << 1)) != 0) ? 16 : 0)),
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 48]), (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0)),
-             FLOAT_TYPE(data_b[b_offset + y2_idx + 49]) * (q4_15 + (((data_a[ib0 + i].qh[l0 + 17] & (hm2 << 1)) != 0) ? 16 : 0)))));
-        const FLOAT_TYPE smin =
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 17]), sc2,
-          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 49]), sc3,
-          fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 17]), sc6,
-              (FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 49])) * sc7)));
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, tmp[tmp_idx]));
-    }
-
-    // sum up partial sums and write back result
-    barrier();
-    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-        }
-        barrier();
-    }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
-    }
-}
diff --git a/ggml/src/vulkan-shaders/mul_mat_vec_q6_k.comp b/ggml/src/vulkan-shaders/mul_mat_vec_q6_k.comp
deleted file mode 100644
index 95c286eeb17..00000000000
--- a/ggml/src/vulkan-shaders/mul_mat_vec_q6_k.comp
+++ /dev/null
@@ -1,79 +0,0 @@
-#version 450
-
-#include "mul_mat_vec_base.comp"
-
-layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
-
-shared FLOAT_TYPE tmp[32];
-
-void main() {
-    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
-
-    uint a_offset, b_offset, d_offset;
-    get_offsets(a_offset, b_offset, d_offset);
-
-    const uint num_blocks_per_row = p.ncols / QUANT_K;
-    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
-
-    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
-
-    const uint step = 16/K_QUANTS_PER_ITERATION;            // 16 or 8
-
-    const uint v_im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const uint v_in = tid - step*v_im;                      // 0...15 or 0...7
-
-#if K_QUANTS_PER_ITERATION == 1
-    const uint l0 = v_in;                                   // 0...15
-    const uint is = 0;
-#else
-    const uint l0 = 4 * v_in;                               // 0, 4, 8, ..., 28
-    const uint is = v_in / 4;
-#endif
-
-    const uint ql_offset = 64*v_im + l0;
-    const uint qh_offset = 32*v_im + l0;
-    const uint s_offset  =  8*v_im + is;
-    const uint y_offset = 128*v_im + l0;
-
-    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
-
-    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-        const uint y_idx   = i * QUANT_K + y_offset;
-
-        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
-
-#if K_QUANTS_PER_ITERATION == 1
-        const uint tmp_idx = 16 * ix + tid;
-        tmp[tmp_idx] = fma(FLOAT_TYPE(data_b[b_offset + y_idx +  0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x03) << 4)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x0c) << 2)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x30) >> 0)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0xc0) >> 2)) - 32),
-                       fma(FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32), tmp[tmp_idx]))))))));
-#else
-        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
-        [[unroll]] for (int l = 0; l < 4; ++l) {
-            sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32),
-                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32), sum))));
-        }
-        tmp[16 * ix + tid] += sum;
-#endif
-    }
-
-    // sum up partial sums and write back result
-    barrier();
-    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-       }
-        barrier();
-    }
-    if (tid == 0) {
-        data_d[d_offset + row] = D_TYPE(tmp[0]);
-    }
-}
diff --git a/ggml/src/vulkan-shaders/soft_max.comp b/ggml/src/vulkan-shaders/soft_max.comp
deleted file mode 100644
index 0bd51ecab58..00000000000
--- a/ggml/src/vulkan-shaders/soft_max.comp
+++ /dev/null
@@ -1,106 +0,0 @@
-#version 450
-
-#extension GL_EXT_shader_16bit_storage : require
-
-layout (push_constant) uniform parameter
-{
-    uint KX;
-    uint KY;
-    float scale;
-    float max_bias;
-    float m0;
-    float m1;
-    uint n_head_log2;
-} p;
-
-#include "types.comp"
-
-#extension GL_EXT_control_flow_attributes : enable
-#define BLOCK_SIZE 512
-
-layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
-
-layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
-layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
-layout (binding = 2) buffer D {D_TYPE data_d[];};
-
-shared FLOAT_TYPE vals[BLOCK_SIZE];
-
-void main() {
-    const uint tid = gl_LocalInvocationID.x;
-    const uint rowx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
-    const uint rowy = rowx % p.KY;
-
-    float slope = 1.0f;
-
-    // ALiBi
-    if (p.max_bias > 0.0f) {
-        const uint h = rowx/p.KY; // head index
-
-        const float base = h < p.n_head_log2 ? p.m0 : p.m1;
-        const uint   exp  = h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1;
-
-        slope = pow(base, exp);
-    }
-
-    // Find max
-    FLOAT_TYPE max_val = uintBitsToFloat(0xFF800000);
-
-    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
-        const uint col = col0 + tid;
-
-        if (col >= p.KX) {
-            break;
-        }
-
-        max_val = max(max_val, FLOAT_TYPE(data_a[rowx * p.KX + col]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)));
-    }
-    vals[tid] = max_val;
-
-    barrier();
-    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
-        if (tid < s) {
-            vals[tid] = max(vals[tid], vals[tid + s]);
-        }
-        barrier();
-    }
-
-    max_val = vals[0];
-    barrier();
-
-    // Sum up values
-    vals[tid] = FLOAT_TYPE(0.0f);
-
-    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
-        const uint col = col0 + tid;
-
-        if (col >= p.KX) {
-            break;
-        }
-
-        const uint i = rowx * p.KX + col;
-        const FLOAT_TYPE val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
-        vals[tid] += val;
-        data_d[i] = D_TYPE(val);
-    }
-
-    barrier();
-    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
-        if (tid < s) {
-            vals[tid] += vals[tid + s];
-        }
-        barrier();
-    }
-
-    const D_TYPE divisor = D_TYPE(vals[0]);
-
-    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
-        const uint col = col0 + tid;
-
-        if (col >= p.KX) {
-            break;
-        }
-
-        data_d[rowx*p.KX + col] /= divisor;
-    }
-}
diff --git a/ggml/src/vulkan-shaders/types.comp b/ggml/src/vulkan-shaders/types.comp
deleted file mode 100644
index 21dce72fc7d..00000000000
--- a/ggml/src/vulkan-shaders/types.comp
+++ /dev/null
@@ -1,200 +0,0 @@
-#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
-#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
-#endif
-
-#if defined(DATA_A_F32)
-#define QUANT_K 1
-#define QUANT_R 1
-
-#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
-#define A_TYPE float
-#elif LOAD_VEC_A == 4
-#define A_TYPE vec4
-#elif LOAD_VEC_A == 8
-#define A_TYPE mat2x4
-#endif
-#endif
-
-#if defined(DATA_A_F16)
-#define QUANT_K 1
-#define QUANT_R 1
-
-#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
-#define A_TYPE float16_t
-#elif LOAD_VEC_A == 4
-#define A_TYPE f16vec4
-#elif LOAD_VEC_A == 8
-#define A_TYPE f16mat2x4
-#endif
-#endif
-
-#if defined(DATA_A_Q4_0)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 32
-#define QUANT_R 2
-
-struct block_q4_0
-{
-    float16_t d;
-    uint8_t qs[16];
-};
-
-#define A_TYPE block_q4_0
-#endif
-
-#if defined(DATA_A_Q4_1)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 32
-#define QUANT_R 2
-
-struct block_q4_1
-{
-    float16_t d;
-    float16_t m;
-    uint8_t qs[16];
-};
-
-#define A_TYPE block_q4_1
-#endif
-
-#if defined(DATA_A_Q5_0)
-#extension GL_EXT_shader_16bit_storage : require
-#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
-#define QUANT_K 32
-#define QUANT_R 2
-
-struct block_q5_0
-{
-    float16_t d;
-    uint16_t qh[2];
-    uint8_t qs[16];
-};
-
-#define A_TYPE block_q5_0
-#endif
-
-#if defined(DATA_A_Q5_1)
-#extension GL_EXT_shader_16bit_storage : require
-#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
-#define QUANT_K 32
-#define QUANT_R 2
-
-struct block_q5_1
-{
-    float16_t d;
-    float16_t m;
-    uint qh;
-    uint8_t qs[16];
-};
-
-#define A_TYPE block_q5_1
-#endif
-
-#if defined(DATA_A_Q8_0)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 32
-#define QUANT_R 1
-
-struct block_q8_0
-{
-    float16_t d;
-    int8_t qs[32];
-};
-
-#define A_TYPE block_q8_0
-#endif
-
-// K-quants
-#if defined(DATA_A_Q2_K)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 256
-
-struct block_q2_K
-{
-    uint8_t scales[QUANT_K/16];
-    uint8_t qs[QUANT_K/4];
-    f16vec2 d;
-};
-
-#define A_TYPE block_q2_K
-#endif
-
-#if defined(DATA_A_Q3_K)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 256
-
-struct block_q3_K
-{
-    uint8_t hmask[QUANT_K/8];
-    uint8_t qs[QUANT_K/4];
-    uint8_t scales[12];
-    float16_t d;
-};
-
-#define A_TYPE block_q3_K
-#endif
-
-#if defined(DATA_A_Q4_K)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 256
-
-struct block_q4_K
-{
-    f16vec2 d;
-    uint8_t scales[3*QUANT_K/64];
-    uint8_t qs[QUANT_K/2];
-};
-
-#define A_TYPE block_q4_K
-#endif
-
-#if defined(DATA_A_Q5_K)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 256
-
-struct block_q5_K
-{
-    f16vec2 d;
-    uint8_t scales[12];
-    uint8_t qh[QUANT_K/8];
-    uint8_t qs[QUANT_K/2];
-};
-
-#define A_TYPE block_q5_K
-#endif
-
-#if defined(DATA_A_Q6_K)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 256
-
-struct block_q6_K
-{
-    uint8_t ql[QUANT_K/2];
-    uint8_t qh[QUANT_K/4];
-    int8_t scales[QUANT_K/16];
-    float16_t d;
-};
-
-#define A_TYPE block_q6_K
-#endif
-
-// IQuants
-
-#if defined(DATA_A_IQ4_NL)
-#extension GL_EXT_shader_16bit_storage : require
-#define QUANT_K 32
-#define QUANT_R 2
-
-struct block_iq4_nl
-{
-    float16_t d;
-    uint8_t qs[QUANT_K/2];
-};
-
-#define A_TYPE block_iq4_nl
-
-const int8_t kvalues_iq4nl[16] = {
-    int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10),
-    int8_t(1), int8_t(13), int8_t(25), int8_t(38), int8_t(53), int8_t(69), int8_t(89), int8_t(113)
-};
-#endif
diff --git a/include/whisper.h b/include/whisper.h
index 9188d686a31..71949bdd397 100644
--- a/include/whisper.h
+++ b/include/whisper.h
@@ -534,7 +534,7 @@ extern "C" {
         float temperature_inc;
         float entropy_thold;    // similar to OpenAI's "compression_ratio_threshold"
         float logprob_thold;
-        float no_speech_thold;  // TODO: not implemented
+        float no_speech_thold;
 
         struct {
             int best_of;    // ref: https://github.com/openai/whisper/blob/f82bc59f5ea234d4b97fb2860842ed38519f7e65/whisper/transcribe.py#L264
diff --git a/models/download-ggml-model.sh b/models/download-ggml-model.sh
index e555aee8bdb..09690ecd497 100755
--- a/models/download-ggml-model.sh
+++ b/models/download-ggml-model.sh
@@ -29,26 +29,32 @@ models="tiny
 tiny.en
 tiny-q5_1
 tiny.en-q5_1
+tiny-q8_0
 base
 base.en
 base-q5_1
 base.en-q5_1
+base-q8_0
 small
 small.en
 small.en-tdrz
 small-q5_1
 small.en-q5_1
+small-q8_0
 medium
 medium.en
 medium-q5_0
 medium.en-q5_0
+medium-q8_0
 large-v1
 large-v2
 large-v2-q5_0
+large-v2-q8_0
 large-v3
 large-v3-q5_0
 large-v3-turbo
-large-v3-turbo-q5_0"
+large-v3-turbo-q5_0
+large-v3-turbo-q8_0"
 
 # list available models
 list_models() {
diff --git a/scripts/bench-all.sh b/scripts/bench-all.sh
index 469ec04d5bf..c7dc52b1e00 100755
--- a/scripts/bench-all.sh
+++ b/scripts/bench-all.sh
@@ -2,7 +2,7 @@
 
 # Helper script to run the bench tool on all models and print the results in share-able format
 
-printf "Usage: ./bench.sh [n_threads] [encoder-only] [flash-attn]\n"
+printf "Usage: ./scripts/bench-all.sh [n_threads] [encoder-only] [flash-attn]\n"
 
 if [ -z "$1" ]; then
     n_threads=4
@@ -38,13 +38,13 @@ if [ "$encoder_only" -eq 0 ]; then
     printf "Running memcpy benchmark\n"
     printf "\n"
 
-    ./bench -w 1 -t $n_threads 2>&1
+    ./build/bin/bench -w 1 -t $n_threads 2>&1
 
     printf "\n"
     printf "Running ggml_mul_mat benchmark with $n_threads threads\n"
     printf "\n"
 
-    ./bench -w 2 -t $n_threads 2>&1
+    ./build/bin/bench -w 2 -t $n_threads 2>&1
 
     printf "\n"
     printf "Running benchmark for all models\n"
@@ -64,7 +64,7 @@ printf "| %6s | %6s | %16s | %13s | %3s | %3s | %7s | %7s | %7s | %7s | %7s |\n"
 for model in "${models[@]}"; do
     # actual run
     # store stderr output in a variable in order to parse it later
-    output=$(./bench -m ./models/ggml-$model.bin -t $n_threads $fattn 2>&1)
+    output=$(./build/bin/bench -m ./models/ggml-$model.bin -t $n_threads $fattn 2>&1)
     ret=$?
 
     # parse the output:
diff --git a/scripts/sync-ggml-am.sh b/scripts/sync-ggml-am.sh
index e5e360386ef..0af5477a108 100755
--- a/scripts/sync-ggml-am.sh
+++ b/scripts/sync-ggml-am.sh
@@ -59,14 +59,17 @@ while read c; do
         src/ggml*.h \
         src/ggml*.c \
         src/ggml*.cpp \
-        src/ggml*.m \
-        src/ggml*.metal \
-        src/ggml*.cu \
-        src/ggml-amx/* \
+        src/ggml-blas/* \
         src/ggml-cann/* \
+        src/ggml-cpu/* \
         src/ggml-cuda/* \
+        src/ggml-hip/* \
+        src/ggml-kompute/* \
+        src/ggml-metal/* \
+        src/ggml-musa/* \
+        src/ggml-rpc/* \
         src/ggml-sycl/* \
-        src/vulkan-shaders/* \
+        src/ggml-vulkan/* \
         include/ggml*.h \
         examples/common.h \
         examples/common.cpp \
@@ -103,16 +106,20 @@ if [ -f $SRC_WHISPER/ggml-src.patch ]; then
     # src/CMakeLists.txt   -> ggml/src/CMakeLists.txt
     # cmake/FindSIMD.cmake -> ggml/cmake/FindSIMD.cmake
     #
-    # src/ggml*.c             -> ggml/src/ggml*.c
-    # src/ggml*.cpp           -> ggml/src/ggml*.cpp
-    # src/ggml*.h             -> ggml/src/ggml*.h
-    # src/ggml*.cu            -> ggml/src/ggml*.cu
-    # src/ggml*.m             -> ggml/src/ggml*.m
-    # src/ggml-amx/*          -> ggml/src/ggml-amx/
-    # src/ggml-cann/*         -> ggml/src/ggml-cann/
-    # src/ggml-cuda/*         -> ggml/src/ggml-cuda/
-    # src/ggml-sycl/*         -> ggml/src/ggml-sycl/*
-    # src/vulkan-shaders/*    -> ggml/src/vulkan-shaders/*
+    # src/ggml*.c          -> ggml/src/ggml*.c
+    # src/ggml*.cpp        -> ggml/src/ggml*.cpp
+    # src/ggml*.h          -> ggml/src/ggml*.h
+    # src/ggml-blas/*      -> ggml/src/ggml-blas/*
+    # src/ggml-cann/*      -> ggml/src/ggml-cann/*
+    # src/ggml-cpu/*       -> ggml/src/ggml-cpu/*
+    # src/ggml-cuda/*      -> ggml/src/ggml-cuda/*
+    # src/ggml-hip/*       -> ggml/src/ggml-hip/*
+    # src/ggml-kompute/*   -> ggml/src/ggml-kompute/*
+    # src/ggml-metal/*     -> ggml/src/ggml-metal/*
+    # src/ggml-musa/*      -> ggml/src/ggml-musa/*
+    # src/ggml-rpc/*       -> ggml/src/ggml-rpc/*
+    # src/ggml-sycl/*      -> ggml/src/ggml-sycl/*
+    # src/ggml-vulkan/*    -> ggml/src/ggml-vulkan/*
     #
     # include/ggml*.h -> ggml/include/ggml*.h
     #
@@ -128,18 +135,21 @@ if [ -f $SRC_WHISPER/ggml-src.patch ]; then
         -e 's/(^[[:space:]]|[ab]\/)CMakeLists.txt/\1ggml\/CMakeLists.txt/g' \
         -e 's/(^[[:space:]]|[ab]\/)src\/CMakeLists.txt/\1ggml\/src\/CMakeLists.txt/g' \
         -e 's/(^[[:space:]]|[ab]\/)cmake\/FindSIMD.cmake/\1ggml\/cmake\/FindSIMD.cmake/g' \
-        -e 's/(^[[:space:]]|[ab]\/)src\/ggml(.*)\.c/\1ggml\/src\/ggml\2.c/g' \
-        -e 's/(^[[:space:]]|[ab]\/)src\/ggml(.*)\.cpp/\1ggml\/src\/ggml\2.cpp/g' \
-        -e 's/(^[[:space:]]|[ab]\/)src\/ggml(.*)\.h/\1ggml\/src\/ggml\2.h/g' \
-        -e 's/(^[[:space:]]|[ab]\/)src\/ggml(.*)\.cu/\1ggml\/src\/ggml\2.cu/g' \
-        -e 's/(^[[:space:]]|[ab]\/)src\/ggml(.*)\.m/\1ggml\/src\/ggml\2.m/g' \
-        -e 's/(^[[:space:]]|[ab]\/)src\/ggml-amx\//\1ggml\/src\/ggml-amx\//g' \
-        -e 's/(^[[:space:]]|[ab]\/)src\/ggml-cann\//\1ggml\/src\/ggml-cann\//g' \
-        -e 's/(^[[:space:]]|[ab]\/)src\/ggml-cuda\//\1ggml\/src\/ggml-cuda\//g' \
-        -e 's/(^[[:space:]]|[ab]\/)src\/ggml-kompute\.cpp/\1ggml\/src\/ggml-kompute.cpp/g' \
-        -e 's/(^[[:space:]]|[ab]\/)src\/ggml-sycl\//\1ggml\/src\/ggml-sycl\//g' \
-        -e 's/(^[[:space:]]|[ab]\/)src\/vulkan-shaders\//\1ggml\/src\/vulkan-shaders\//g' \
-        -e 's/(^[[:space:]]|[ab]\/)include\/ggml(.*)\.h/\1ggml\/include\/ggml\2.h/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.c/\1ggml\/src\/ggml\2.c/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.cpp/\1ggml\/src\/ggml\2.cpp/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml(.*)\.h/\1ggml\/src\/ggml\2.h/g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-blas\//\1ggml\/src\/ggml-blas\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-cann\//\1ggml\/src\/ggml-cann\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-cpu\//\1ggml\/src\/ggml-cpu\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-cuda\//\1ggml\/src\/ggml-cuda\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-hip\//\1ggml\/src\/ggml-hip\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-kompute\//\1ggml\/src\/ggml-kompute\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-metal\//\1ggml\/src\/ggml-metal\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-musa\//\1ggml\/src\/ggml-musa\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-rpc\//\1ggml\/src\/ggml-rpc\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-sycl\//\1ggml\/src\/ggml-sycl\//g' \
+        -e 's/([[:space:]]|[ab]\/)src\/ggml-vulkan\//\1ggml\/src\/ggml-vulkan\//g' \
+        -e 's/([[:space:]]|[ab]\/)include\/ggml(.*)\.h/\1ggml\/include\/ggml\2.h/g' \
         -e 's/(^[[:space:]]|[ab]\/)examples\/common\.h/\1examples\/common.h/g' \
         -e 's/(^[[:space:]]|[ab]\/)examples\/common\.cpp/\1examples\/common.cpp/g' \
         -e 's/(^[[:space:]]|[ab]\/)examples\/common-ggml\.h/\1examples\/common-ggml.h/g' \
diff --git a/scripts/sync-ggml.last b/scripts/sync-ggml.last
index 2135424bf20..fc6e1fd46e6 100644
--- a/scripts/sync-ggml.last
+++ b/scripts/sync-ggml.last
@@ -1 +1 @@
-e178a213638f7cebd96a6ddac0151351ca5da066
+cc1ab854224b1723961863c7f73bc7905cc22bf2
diff --git a/scripts/sync-ggml.sh b/scripts/sync-ggml.sh
index 5c23e3407fa..72adec853a0 100755
--- a/scripts/sync-ggml.sh
+++ b/scripts/sync-ggml.sh
@@ -4,16 +4,20 @@ cp -rpv ../ggml/CMakeLists.txt       ./ggml/CMakeLists.txt
 cp -rpv ../ggml/src/CMakeLists.txt   ./ggml/src/CMakeLists.txt
 cp -rpv ../ggml/cmake/FindSIMD.cmake ./ggml/cmake/FindSIMD.cmake
 
-cp -rpv ../ggml/src/ggml*.c          ./ggml/src/
-cp -rpv ../ggml/src/ggml*.cpp        ./ggml/src/
-cp -rpv ../ggml/src/ggml*.h          ./ggml/src/
-cp -rpv ../ggml/src/ggml*.cu         ./ggml/src/
-cp -rpv ../ggml/src/ggml*.m          ./ggml/src/
-cp -rpv ../ggml/src/ggml-amx/*       ./ggml/src/ggml-amx/
-cp -rpv ../ggml/src/ggml-cann/*      ./ggml/src/ggml-cann/
-cp -rpv ../ggml/src/ggml-cuda/*      ./ggml/src/ggml-cuda/
-cp -rpv ../ggml/src/ggml-sycl/*      ./ggml/src/ggml-sycl/
-cp -rpv ../ggml/src/vulkan-shaders/* ./ggml/src/vulkan-shaders/
+cp -rpv ../ggml/src/ggml*.c        ./ggml/src/
+cp -rpv ../ggml/src/ggml*.cpp      ./ggml/src/
+cp -rpv ../ggml/src/ggml*.h        ./ggml/src/
+cp -rpv ../ggml/src/ggml-blas/*    ./ggml/src/ggml-blas/
+cp -rpv ../ggml/src/ggml-cann/*    ./ggml/src/ggml-cann/
+cp -rpv ../ggml/src/ggml-cpu/*     ./ggml/src/ggml-cpu/
+cp -rpv ../ggml/src/ggml-cuda/*    ./ggml/src/ggml-cuda/
+cp -rpv ../ggml/src/ggml-hip/*     ./ggml/src/ggml-hip/
+cp -rpv ../ggml/src/ggml-kompute/* ./ggml/src/ggml-kompute/
+cp -rpv ../ggml/src/ggml-metal/*   ./ggml/src/ggml-metal/
+cp -rpv ../ggml/src/ggml-musa/*    ./ggml/src/ggml-musa/
+cp -rpv ../ggml/src/ggml-rpc/*     ./ggml/src/ggml-rpc/
+cp -rpv ../ggml/src/ggml-sycl/*    ./ggml/src/ggml-sycl/
+cp -rpv ../ggml/src/ggml-vulkan/*  ./ggml/src/ggml-vulkan/
 
 cp -rpv ../ggml/include/ggml*.h ./ggml/include/
 
diff --git a/src/ggml-cpu-impl.h b/src/ggml-cpu-impl.h
deleted file mode 100644
index 5b45155b028..00000000000
--- a/src/ggml-cpu-impl.h
+++ /dev/null
@@ -1,614 +0,0 @@
-#pragma once
-
-// GGML CPU internal header
-
-#include "ggml.h"
-#include "ggml-impl.h"
-#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
-//#include <stddef.h>
-#include <stdbool.h>
-#include <string.h> // memcpy
-#include <math.h>   // fabsf
-
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-#if defined(_MSC_VER)
-
-#define m512bh(p) p
-#define m512i(p) p
-
-#else
-
-#define m512bh(p) (__m512bh)(p)
-#define m512i(p) (__m512i)(p)
-
-#endif
-
-/**
- * Converts brain16 to float32.
- *
- * The bfloat16 floating point format has the following structure:
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───┐
- *     0b0000000000000000 brain16
- *
- * Since bf16 has the same number of exponent bits as a 32bit float,
- * encoding and decoding numbers becomes relatively straightforward.
- *
- *       ┌sign
- *       │
- *       │   ┌exponent
- *       │   │
- *       │   │      ┌mantissa
- *       │   │      │
- *       │┌──┴───┐┌─┴───────────────────┐
- *     0b00000000000000000000000000000000 IEEE binary32
- *
- * For comparison, the standard fp16 format has fewer exponent bits.
- *
- *       ┌sign
- *       │
- *       │  ┌exponent
- *       │  │
- *       │  │    ┌mantissa
- *       │  │    │
- *       │┌─┴─┐┌─┴──────┐
- *     0b0000000000000000 IEEE binary16
- *
- * @see IEEE 754-2008
- */
-static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.i = (uint32_t)h.bits << 16;
-    return u.f;
-}
-
-/**
- * Converts float32 to brain16.
- *
- * This is binary identical with Google Brain float conversion.
- * Floats shall round to nearest even, and NANs shall be quiet.
- * Subnormals aren't flushed to zero, except perhaps when used.
- * This code should vectorize nicely if using modern compilers.
- */
-static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
-    ggml_bf16_t h;
-    union {
-        float f;
-        uint32_t i;
-    } u;
-    u.f = s;
-    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
-        h.bits = (u.i >> 16) | 64; /* force to quiet */
-        return h;
-    }
-    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
-    return h;
-}
-
-#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
-#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
-
-// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
-#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __FMA__
-#define __FMA__
-#endif
-#ifndef __F16C__
-#define __F16C__
-#endif
-#endif
-
-// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
-#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __SSE3__
-#define __SSE3__
-#endif
-#ifndef __SSSE3__
-#define __SSSE3__
-#endif
-#endif
-
-#if defined(__ARM_FEATURE_SVE)
-#include <arm_sve.h>
-#include <sys/prctl.h>
-#endif
-
-// 16-bit float
-// on Arm, we use __fp16
-// on x86, we use uint16_t
-#if defined(__ARM_NEON)
-
-// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
-//
-//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
-//
-#include <arm_neon.h>
-
-#ifdef _MSC_VER
-
-typedef uint16_t ggml_fp16_internal_t;
-
-#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
-
-#else
-
-typedef __fp16 ggml_fp16_internal_t;
-
-#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
-
-#endif // _MSC_VER
-
-#if !defined(__aarch64__)
-
-// 32-bit ARM compatibility
-
-// vaddlvq_s16
-// vpaddq_s16
-// vpaddq_s32
-// vaddvq_s32
-// vaddvq_f32
-// vmaxvq_f32
-// vcvtnq_s32_f32
-// vzip1_u8
-// vzip2_u8
-
-inline static int32_t vaddlvq_s16(int16x8_t v) {
-    int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
-    return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
-}
-
-inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
-    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
-    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
-    return vcombine_s16(a0, b0);
-}
-
-inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
-    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
-    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
-    return vcombine_s32(a0, b0);
-}
-
-inline static int32_t vaddvq_s32(int32x4_t v) {
-    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
-}
-
-inline static float vaddvq_f32(float32x4_t v) {
-    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
-}
-
-inline static float vmaxvq_f32(float32x4_t v) {
-    return
-        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
-            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
-}
-
-inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
-    int32x4_t res;
-
-    res[0] = roundf(vgetq_lane_f32(v, 0));
-    res[1] = roundf(vgetq_lane_f32(v, 1));
-    res[2] = roundf(vgetq_lane_f32(v, 2));
-    res[3] = roundf(vgetq_lane_f32(v, 3));
-
-    return res;
-}
-
-inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
-    uint8x8_t res;
-
-    res[0] = a[0]; res[1] = b[0];
-    res[2] = a[1]; res[3] = b[1];
-    res[4] = a[2]; res[5] = b[2];
-    res[6] = a[3]; res[7] = b[3];
-
-    return res;
-}
-
-inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
-    uint8x8_t res;
-
-    res[0] = a[4]; res[1] = b[4];
-    res[2] = a[5]; res[3] = b[5];
-    res[4] = a[6]; res[5] = b[6];
-    res[6] = a[7]; res[7] = b[7];
-
-    return res;
-}
-
-// vld1q_s16_x2
-// vld1q_u8_x2
-// vld1q_u8_x4
-// vld1q_s8_x2
-// vld1q_s8_x4
-// TODO: double-check these work correctly
-
-typedef struct ggml_int16x8x2_t {
-    int16x8_t val[2];
-} ggml_int16x8x2_t;
-
-inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
-    ggml_int16x8x2_t res;
-
-    res.val[0] = vld1q_s16(ptr + 0);
-    res.val[1] = vld1q_s16(ptr + 8);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x2_t {
-    uint8x16_t val[2];
-} ggml_uint8x16x2_t;
-
-inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
-    ggml_uint8x16x2_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x4_t {
-    uint8x16_t val[4];
-} ggml_uint8x16x4_t;
-
-inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
-    ggml_uint8x16x4_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-    res.val[2] = vld1q_u8(ptr + 32);
-    res.val[3] = vld1q_u8(ptr + 48);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x2_t {
-    int8x16_t val[2];
-} ggml_int8x16x2_t;
-
-inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
-    ggml_int8x16x2_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x4_t {
-    int8x16_t val[4];
-} ggml_int8x16x4_t;
-
-inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
-    ggml_int8x16x4_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-    res.val[2] = vld1q_s8(ptr + 32);
-    res.val[3] = vld1q_s8(ptr + 48);
-
-    return res;
-}
-
-// NOTE: not tested
-inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
-    int8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-// NOTE: not tested
-inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
-    uint8x16_t res;
-
-    res[ 0] = a[b[ 0]];
-    res[ 1] = a[b[ 1]];
-    res[ 2] = a[b[ 2]];
-    res[ 3] = a[b[ 3]];
-    res[ 4] = a[b[ 4]];
-    res[ 5] = a[b[ 5]];
-    res[ 6] = a[b[ 6]];
-    res[ 7] = a[b[ 7]];
-    res[ 8] = a[b[ 8]];
-    res[ 9] = a[b[ 9]];
-    res[10] = a[b[10]];
-    res[11] = a[b[11]];
-    res[12] = a[b[12]];
-    res[13] = a[b[13]];
-    res[14] = a[b[14]];
-    res[15] = a[b[15]];
-
-    return res;
-}
-
-#else
-
-#define ggml_int16x8x2_t  int16x8x2_t
-#define ggml_uint8x16x2_t uint8x16x2_t
-#define ggml_uint8x16x4_t uint8x16x4_t
-#define ggml_int8x16x2_t  int8x16x2_t
-#define ggml_int8x16x4_t  int8x16x4_t
-
-#define ggml_vld1q_s16_x2 vld1q_s16_x2
-#define ggml_vld1q_u8_x2  vld1q_u8_x2
-#define ggml_vld1q_u8_x4  vld1q_u8_x4
-#define ggml_vld1q_s8_x2  vld1q_s8_x2
-#define ggml_vld1q_s8_x4  vld1q_s8_x4
-#define ggml_vqtbl1q_s8   vqtbl1q_s8
-#define ggml_vqtbl1q_u8   vqtbl1q_u8
-
-#endif // !defined(__aarch64__)
-
-#if !defined(__ARM_FEATURE_DOTPROD)
-
-inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
-    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
-    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
-
-    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
-}
-
-#else
-
-#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
-
-#endif // !defined(__ARM_FEATURE_DOTPROD)
-
-#endif // defined(__ARM_NEON)
-
-#if defined(__ARM_NEON) && !defined(_MSC_VER)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    ggml_fp16_internal_t tmp;
-    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-    return (float)tmp;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    ggml_fp16_t res;
-    ggml_fp16_internal_t tmp = f;
-    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-    return res;
-}
-
-#else
-
-#ifdef __wasm_simd128__
-#include <wasm_simd128.h>
-#else
-#ifdef __POWER9_VECTOR__
-#include <altivec.h>
-#undef bool
-#define bool _Bool
-#else
-#if defined(_MSC_VER) || defined(__MINGW32__)
-#include <intrin.h>
-#else
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
-#if !defined(__riscv)
-#include <immintrin.h>
-#endif
-#endif
-#endif
-#endif
-#endif
-
-#ifdef __riscv_v_intrinsic
-#include <riscv_vector.h>
-#endif
-
-#if defined(__loongarch64)
-#if defined(__loongarch_asx)
-#include <lasxintrin.h>
-#endif
-#if defined(__loongarch_sx)
-#include <lsxintrin.h>
-#endif
-#endif
-
-#if defined(__loongarch_asx)
-
-typedef union {
-    int32_t i;
-    float f;
-} ft_union;
-
-/* float type data load instructions */
-static __m128 __lsx_vreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
-}
-
-static __m256 __lasx_xvreplfr2vr_s(float val) {
-    ft_union fi_tmpval = {.f = val};
-    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
-}
-#endif
-
-#ifdef __F16C__
-
-#ifdef _MSC_VER
-#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-#else
-#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-#endif
-
-#elif defined(__POWER9_VECTOR__)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-/* the inline asm below is about 12% faster than the lookup method */
-#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    register float f;
-    register double d;
-    __asm__(
-        "mtfprd %0,%2\n"
-        "xscvhpdp %0,%0\n"
-        "frsp %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=f"(f):
-        /* in */   "r"(h));
-    return f;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    register double d;
-    register ggml_fp16_t r;
-    __asm__( /* xscvdphp can work on double or single precision */
-        "xscvdphp %0,%2\n"
-        "mffprd %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=r"(r):
-        /* in */   "f"(f));
-    return r;
-}
-
-#else
-
-// FP16 <-> FP32
-// ref: https://github.com/Maratyszcza/FP16
-
-static inline float fp32_from_bits(uint32_t w) {
-    union {
-        uint32_t as_bits;
-        float as_value;
-    } fp32;
-    fp32.as_bits = w;
-    return fp32.as_value;
-}
-
-static inline uint32_t fp32_to_bits(float f) {
-    union {
-        float as_value;
-        uint32_t as_bits;
-    } fp32;
-    fp32.as_value = f;
-    return fp32.as_bits;
-}
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    const uint32_t w = (uint32_t) h << 16;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    const uint32_t two_w = w + w;
-
-    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float exp_scale = 0x1.0p-112f;
-#else
-    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-#endif
-    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-    const uint32_t magic_mask = UINT32_C(126) << 23;
-    const float magic_bias = 0.5f;
-    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-    const uint32_t result = sign |
-        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-    return fp32_from_bits(result);
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float scale_to_inf = 0x1.0p+112f;
-    const float scale_to_zero = 0x1.0p-110f;
-#else
-    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-#endif
-    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-    const uint32_t w = fp32_to_bits(f);
-    const uint32_t shl1_w = w + w;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-    if (bias < UINT32_C(0x71000000)) {
-        bias = UINT32_C(0x71000000);
-    }
-
-    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-    const uint32_t bits = fp32_to_bits(base);
-    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-    const uint32_t nonsign = exp_bits + mantissa_bits;
-    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-}
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#endif // __F16C__
-
-#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
-
-#ifdef __ARM_FEATURE_SVE
-#include <arm_sve.h>
-#endif // __ARM_FEATURE_SVE
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in ggml_init()
-extern float ggml_table_f32_f16[1 << 16];
-
-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
-// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
-inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
-    uint16_t s;
-    memcpy(&s, &f, sizeof(uint16_t));
-    return ggml_table_f32_f16[s];
-}
-
-#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(GGML_FP32_TO_FP16)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-#endif
-
-#ifdef __cplusplus
-}
-#endif
diff --git a/src/whisper.cpp b/src/whisper.cpp
index 1c4965b6d4b..bcc530ae891 100644
--- a/src/whisper.cpp
+++ b/src/whisper.cpp
@@ -1,43 +1,19 @@
 #include "whisper.h"
 
-#ifdef WHISPER_USE_COREML
-#include "coreml/whisper-encoder.h"
-#endif
-
 #include "ggml-cpu.h"
 
-#ifdef GGML_USE_METAL
-#include "ggml-metal.h"
-#endif
-
-#ifdef GGML_USE_CUDA
-#include "ggml-cuda.h"
-#endif
-
-#ifdef GGML_USE_SYCL
-#include "ggml-sycl.h"
-#endif
-
-#ifdef GGML_USE_VULKAN
-#include "ggml-vulkan.h"
-#endif
+#include "ggml.h"
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
 
-#ifdef GGML_USE_BLAS
-#include "ggml-blas.h"
+#ifdef WHISPER_USE_COREML
+#include "coreml/whisper-encoder.h"
 #endif
 
 #ifdef WHISPER_USE_OPENVINO
 #include "openvino/whisper-openvino-encoder.h"
 #endif
 
-#ifdef GGML_USE_CANN
-#include "ggml-cann.h"
-#endif
-
-#include "ggml.h"
-#include "ggml-alloc.h"
-#include "ggml-backend.h"
-
 #include <atomic>
 #include <algorithm>
 #include <cassert>
@@ -195,14 +171,13 @@ static bool ggml_graph_compute_helper(
 
     for (int i = 0; i < ggml_backend_sched_get_n_backends(sched); ++i) {
         ggml_backend_t backend = ggml_backend_sched_get_backend(sched, i);
-        if (ggml_backend_is_cpu(backend)) {
-            ggml_backend_cpu_set_n_threads(backend, n_threads);
-        }
-#ifdef GGML_USE_BLAS
-        if (ggml_backend_is_blas(backend)) {
-            ggml_backend_blas_set_n_threads(backend, n_threads);
+        ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+        ggml_backend_reg_t reg = dev ? ggml_backend_dev_backend_reg(dev) : nullptr;
+
+        auto * fn_set_n_threads = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
+        if (fn_set_n_threads) {
+            fn_set_n_threads(backend, n_threads);
         }
-#endif
     }
 
     bool t = ggml_backend_sched_graph_compute(sched, graph) == GGML_STATUS_SUCCESS;
@@ -892,6 +867,7 @@ struct whisper_state {
     whisper_token tid_last;
 
     std::vector<float> energy; // PCM signal energy
+    float no_speech_prob = 0.0f;
 
     // [EXPERIMENTAL] Token-level timestamps with DTW
     whisper_aheads_masks aheads_masks;
@@ -1256,67 +1232,23 @@ static size_t aheads_masks_nbytes(struct whisper_aheads_masks & aheads_masks) {
 }
 
 static ggml_backend_t whisper_backend_init_gpu(const whisper_context_params & params) {
-    ggml_backend_t result = NULL;
-
     ggml_log_set(g_state.log_callback, g_state.log_callback_user_data);
 
-#ifdef GGML_USE_CUDA
     if (params.use_gpu) {
-        WHISPER_LOG_INFO("%s: using CUDA backend\n", __func__);
-        result = ggml_backend_cuda_init(params.gpu_device);
-        if (!result) {
-            WHISPER_LOG_ERROR("%s: ggml_backend_cuda_init() failed\n", __func__);
-        }
-    }
-#endif
-
-#ifdef GGML_USE_METAL
-    if (params.use_gpu) {
-        WHISPER_LOG_INFO("%s: using Metal backend\n", __func__);
-        result = ggml_backend_metal_init();
-        if (!result) {
-            WHISPER_LOG_ERROR("%s: ggml_backend_metal_init() failed\n", __func__);
-        } else if (!ggml_backend_metal_supports_family(result, 7)) {
-            WHISPER_LOG_ERROR("%s: Metal GPU does not support family 7 - falling back to CPU\n", __func__);
-            ggml_backend_free(result);
-            result = NULL;
-        }
-    }
-#endif
-
-#ifdef GGML_USE_SYCL
-    if (params.use_gpu) {
-        WHISPER_LOG_INFO("%s: using SYCL backend\n", __func__);
-        result = ggml_backend_sycl_init(params.gpu_device);
-        if (!result) {
-            WHISPER_LOG_ERROR("%s: ggml_backend_sycl_init() failed\n", __func__);
-        }
-    }
-#endif
-
-#ifdef GGML_USE_VULKAN
-    if (params.use_gpu) {
-        WHISPER_LOG_INFO("%s: using Vulkan backend\n", __func__);
-        result = ggml_backend_vk_init(params.gpu_device);
-        if (!result) {
-            WHISPER_LOG_ERROR("%s: ggml_backend_vk_init() failed\n", __func__);
-        }
-    }
-#endif
-
-#ifdef GGML_USE_CANN
-    if (params.use_gpu) {
-        WHISPER_LOG_INFO("%s: using CANN backend\n", __func__);
-        result = ggml_backend_cann_init(params.gpu_device);
-        if (!result) {
-            WHISPER_LOG_ERROR("%s: ggml_backend_cann_init() failed\n", __func__);
+        for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+            ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+            if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
+                WHISPER_LOG_INFO("%s: using %s backend\n", __func__, ggml_backend_dev_name(dev));
+                ggml_backend_t result = ggml_backend_dev_init(dev, nullptr);
+                if (!result) {
+                    WHISPER_LOG_ERROR("%s: failed to initialize %s backend\n", __func__, ggml_backend_dev_name(dev));
+                }
+                return result;
+            }
         }
     }
-#endif
 
-    GGML_UNUSED(params);
-
-    return result;
+    return nullptr;
 }
 
 static std::vector<ggml_backend_t> whisper_backend_init(const whisper_context_params & params) {
@@ -1328,17 +1260,19 @@ static std::vector<ggml_backend_t> whisper_backend_init(const whisper_context_pa
         result.push_back(backend_gpu);
     }
 
-#ifdef GGML_USE_BLAS
-    {
-        WHISPER_LOG_INFO("%s: using BLAS backend\n", __func__);
-        ggml_backend_t backend_blas = ggml_backend_blas_init();
-        if (!backend_blas) {
-            WHISPER_LOG_ERROR("%s: ggml_backend_blas_init() failed\n", __func__);
-        } else {
-            result.push_back(backend_blas);
+    // ACCEL backends
+    for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_ACCEL) {
+            WHISPER_LOG_INFO("%s: using %s backend\n", __func__, ggml_backend_dev_name(dev));
+            ggml_backend_t backend = ggml_backend_dev_init(dev, nullptr);
+            if (!backend) {
+                WHISPER_LOG_ERROR("%s: failed to initialize %s backend\n", __func__, ggml_backend_dev_name(dev));
+                continue;
+            }
+            result.push_back(backend);
         }
     }
-#endif
 
     GGML_UNUSED(params);
 
@@ -1348,33 +1282,20 @@ static std::vector<ggml_backend_t> whisper_backend_init(const whisper_context_pa
 }
 
 static ggml_backend_buffer_type_t whisper_default_buffer_type(const whisper_context_params & params) {
-    ggml_backend_buffer_type_t result = nullptr;
-
-    params.use_gpu || (result = ggml_backend_cpu_buffer_type());
-
-#ifdef GGML_USE_CUDA
-    result || (result = ggml_backend_cuda_buffer_type(params.gpu_device));
-#endif
-
-#ifdef GGML_USE_METAL
-    result || (result = ggml_backend_metal_buffer_type());
-#endif
-
-#ifdef GGML_USE_SYCL
-    result || (result = ggml_backend_sycl_buffer_type(params.gpu_device));
-#endif
-
-#ifdef GGML_USE_VULKAN
-    result || (result = ggml_backend_vk_buffer_type(params.gpu_device));
-#endif
-
-#ifdef GGML_USE_CANN
-    result || (result == ggml_backend_cann_buffer_type(params.gpu_device));
-#endif
+    if (!params.use_gpu) {
+        return ggml_backend_cpu_buffer_type();
+    }
 
-    result || (result = ggml_backend_cpu_buffer_type());
+    // if we have a GPU device - use it
+    for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
+            WHISPER_LOG_INFO("%s: using device %s (%s)\n", __func__, ggml_backend_dev_name(dev), ggml_backend_dev_description(dev));
+            return ggml_backend_dev_buffer_type(dev);
+        }
+    }
 
-    return result;
+    return ggml_backend_cpu_buffer_type();
 }
 
 // load the model from a ggml file
@@ -3668,8 +3589,7 @@ struct whisper_context * whisper_init_with_params_no_state(struct whisper_model_
     WHISPER_LOG_INFO("%s: flash attn = %d\n", __func__, params.flash_attn);
     WHISPER_LOG_INFO("%s: gpu_device = %d\n", __func__, params.gpu_device);
     WHISPER_LOG_INFO("%s: dtw        = %d\n", __func__, params.dtw_token_timestamps);
-
-    // TODO: temporary call to force backend registry initialization
+    WHISPER_LOG_INFO("%s: devices    = %zu\n", __func__, ggml_backend_dev_count());
     WHISPER_LOG_INFO("%s: backends   = %zu\n", __func__, ggml_backend_reg_count());
 
     whisper_context * ctx = new whisper_context;
@@ -4268,18 +4188,15 @@ const char * whisper_print_system_info(void) {
     s += "FMA = "       + std::to_string(ggml_cpu_has_fma())       + " | ";
     s += "NEON = "      + std::to_string(ggml_cpu_has_neon())      + " | ";
     s += "ARM_FMA = "   + std::to_string(ggml_cpu_has_arm_fma())   + " | ";
-    s += "METAL = "     + std::to_string(ggml_cpu_has_metal())     + " | ";
     s += "F16C = "      + std::to_string(ggml_cpu_has_f16c())      + " | ";
     s += "FP16_VA = "   + std::to_string(ggml_cpu_has_fp16_va())   + " | ";
     s += "WASM_SIMD = " + std::to_string(ggml_cpu_has_wasm_simd()) + " | ";
-    s += "BLAS = "      + std::to_string(ggml_cpu_has_blas())      + " | ";
     s += "SSE3 = "      + std::to_string(ggml_cpu_has_sse3())      + " | ";
     s += "SSSE3 = "     + std::to_string(ggml_cpu_has_ssse3())     + " | ";
     s += "VSX = "       + std::to_string(ggml_cpu_has_vsx())       + " | ";
-    s += "CUDA = "      + std::to_string(ggml_cpu_has_cuda())      + " | ";
     s += "COREML = "    + std::to_string(whisper_has_coreml())     + " | ";
     s += "OPENVINO = "  + std::to_string(whisper_has_openvino())   + " | ";
-    s += "CANN = "      + std::to_string(ggml_cpu_has_cann())             ;
+
     return s.c_str();
 }
 
@@ -4909,6 +4826,42 @@ static const std::vector<std::string> non_speech_tokens = {
     "♪♪♪","♩", "♪", "♫", "♬", "♭", "♮", "♯"
 };
 
+static void whisper_compute_logprobs(
+                const std::vector<float> & logits,
+                              const int    n_logits,
+                      std::vector<float> & logprobs) {
+    const float logit_max = *std::max_element(logits.begin(), logits.end());
+    float logsumexp = 0.0f;
+    for (int i = 0; i < n_logits; ++i) {
+        if (logits[i] > -INFINITY) {
+            logsumexp += expf(logits[i] - logit_max);
+        }
+    }
+    logsumexp = logf(logsumexp) + logit_max;
+
+    for (int i = 0; i < n_logits; ++i) {
+        if (logits[i] > -INFINITY) {
+            logprobs[i] = logits[i] - logsumexp;
+        } else {
+            logprobs[i] = -INFINITY;
+        }
+    }
+}
+
+static void whisper_compute_probs(
+    const std::vector<float> & logits,
+                  const int    n_logits,
+    const std::vector<float> & logprobs,
+          std::vector<float> & probs)     {
+    for (int i = 0; i < n_logits; ++i) {
+        if (logits[i] == -INFINITY) {
+            probs[i] = 0.0f;
+        } else {
+            probs[i] = expf(logprobs[i]);
+        }
+    }
+}
+
 // process the logits for the selected decoder
 // - applies logit filters
 // - computes logprobs and probs
@@ -4970,7 +4923,7 @@ static void whisper_process_logits(
 
         // suppress sot and nosp tokens
         logits[vocab.token_sot]  = -INFINITY;
-        logits[vocab.token_nosp] = -INFINITY; // TODO: ignore this token for now
+        logits[vocab.token_nosp] = -INFINITY;
 
         // [TDRZ] when tinydiarize is disabled, suppress solm token
         if (params.tdrz_enable == false) {
@@ -5069,24 +5022,7 @@ static void whisper_process_logits(
         }
 
         // populate the logprobs array (log_softmax)
-        {
-            const float logit_max = *std::max_element(logits.begin(), logits.end());
-            float logsumexp = 0.0f;
-            for (int i = 0; i < n_logits; ++i) {
-                if (logits[i] > -INFINITY) {
-                    logsumexp += expf(logits[i] - logit_max);
-                }
-            }
-            logsumexp = logf(logsumexp) + logit_max;
-
-            for (int i = 0; i < n_logits; ++i) {
-                if (logits[i] > -INFINITY) {
-                    logprobs[i] = logits[i] - logsumexp;
-                } else {
-                    logprobs[i] = -INFINITY;
-                }
-            }
-        }
+        whisper_compute_logprobs(logits, n_logits, logprobs);
 
         // if sum of probability over timestamps is above any other token, sample timestamp
         // ref: https://github.com/openai/whisper/blob/0b1ba3d46ebf7fe6f953acfd8cad62a4f851b49f/whisper/decoding.py#L431-L437
@@ -5144,15 +5080,7 @@ static void whisper_process_logits(
     }
 
     // compute probs
-    {
-        for (int i = 0; i < n_logits; ++i) {
-            if (logits[i] == -INFINITY) {
-                probs[i] = 0.0f;
-            } else {
-                probs[i] = expf(logprobs[i]);
-            }
-        }
-    }
+    whisper_compute_probs(logits, n_logits, logprobs, probs);
 
 #if 0
     // print first 100 logits - token string : logit
@@ -5731,6 +5659,18 @@ int whisper_full_with_state(
                     return -8;
                 }
 
+                // Calculate no_speech probability after first decode.
+                // This has to be done before any logit filtering. Hence we cannot use the probs from the whisper_process_logits.
+                {
+                    const int n_logits = ctx->vocab.id_to_token.size();
+                    std::vector<float> logprobs(n_logits);
+                    std::vector<float> probs(n_logits);
+
+                    whisper_compute_logprobs(state->logits, n_logits, logprobs);
+                    whisper_compute_probs(state->logits, n_logits, logprobs, probs);
+                    state->no_speech_prob = probs[whisper_token_nosp(ctx)];
+                }
+
                 {
                     const int64_t t_start_sample_us = ggml_time_us();
 
@@ -6122,8 +6062,9 @@ int whisper_full_with_state(
             if (it != (int) temperatures.size() - 1) {
                 const auto & decoder = state->decoders[best_decoder_id];
 
-                if (decoder.failed || decoder.sequence.avg_logprobs < params.logprob_thold) {
-                    WHISPER_LOG_DEBUG("%s: failed due to avg_logprobs %8.5f < %8.5f\n", __func__, decoder.sequence.avg_logprobs, params.logprob_thold);
+                if (decoder.failed ||
+                    (decoder.sequence.avg_logprobs < params.logprob_thold && state->no_speech_prob < params.no_speech_thold)) {
+                    WHISPER_LOG_DEBUG("%s: failed due to avg_logprobs %8.5f < %8.5f and no_speech_prob %8.5f < %8.5f\n", __func__, decoder.sequence.avg_logprobs, params.logprob_thold, state->no_speech_prob, params.no_speech_thold);
                     success = false;
                     state->n_fail_p++;
                 }
@@ -6144,7 +6085,7 @@ int whisper_full_with_state(
         {
             const auto & best_decoder = state->decoders[best_decoder_id];
 
-            const auto seek_delta = best_decoder.seek_delta;
+            auto seek_delta = best_decoder.seek_delta;
             const auto result_len = best_decoder.sequence.result_len;
 
             const auto & tokens_cur = best_decoder.sequence.tokens;
@@ -6152,6 +6093,9 @@ int whisper_full_with_state(
             // [EXPERIMENTAL] Token-level timestamps with DTW
             const auto n_segments_before = state->result_all.size();
 
+            const bool is_no_speech = (state->no_speech_prob > params.no_speech_thold &&
+                best_decoder.sequence.avg_logprobs < params.logprob_thold);
+
             //WHISPER_LOG_DEBUG("prompt_init.size() = %d, prompt.size() = %d, result_len = %d, seek_delta = %d\n", prompt_init.size(), prompt.size(), result_len, seek_delta);
 
             // update prompt_past
@@ -6160,11 +6104,11 @@ int whisper_full_with_state(
                 prompt_past.insert(prompt_past.end(), prompt.begin() + 1, prompt.end() - prompt_init.size());
             }
 
-            for (int i = 0; i < result_len; ++i) {
+            for (int i = 0; i < result_len && !is_no_speech; ++i) {
                 prompt_past.push_back(tokens_cur[i].id);
             }
 
-            if (!tokens_cur.empty() && ctx->model.n_loaded > 0) {
+            if (!tokens_cur.empty() && ctx->model.n_loaded > 0 && !is_no_speech) {
                 int  i0 = 0;
                 auto t0 = seek + 2*(tokens_cur.front().tid - whisper_token_beg(ctx));
 
@@ -6285,6 +6229,15 @@ int whisper_full_with_state(
                 }
             }
 
+            // ref: https://github.com/ggerganov/whisper.cpp/pull/2629
+            const bool single_timestamp_ending = tokens_cur.size() > 1 &&
+                tokens_cur[tokens_cur.size() - 2].id < whisper_token_beg(ctx) &&
+                tokens_cur[tokens_cur.size() - 1].id > whisper_token_beg(ctx);
+            if (single_timestamp_ending) {
+                WHISPER_LOG_DEBUG("single timestamp ending - skip entire chunk\n");
+                seek_delta = std::min(seek_end - seek, WHISPER_CHUNK_SIZE * 100);
+            }
+
             // update audio window
             seek += seek_delta;
 
@@ -7430,6 +7383,11 @@ static void whisper_log_internal(ggml_log_level level, const char * format, ...)
 static void whisper_log_callback_default(ggml_log_level level, const char * text, void * user_data) {
     (void) level;
     (void) user_data;
+#ifndef WHISPER_DEBUG
+    if (level == GGML_LOG_LEVEL_DEBUG) {
+        return;
+    }
+#endif
     fputs(text, stderr);
     fflush(stderr);
 }
diff --git a/tests/run-tests.sh b/tests/run-tests.sh
index ae76b4b7cfb..bb8aa5e21f1 100755
--- a/tests/run-tests.sh
+++ b/tests/run-tests.sh
@@ -39,7 +39,7 @@ if [ $# -eq 0 ]; then
 fi
 
 model=$1
-main="../main"
+main="../build/bin/main"
 
 threads=""
 if [ $# -eq 2 ]; then