1 Release published by 1 person

• v2.8.0 PyTorch 2.8.0 Release

  published Aug 6, 2025

239 Pull requests opened by 125 people

• [WIP][symm_mem] Add a wait for signal and put signal for one side API

  #159837 opened Aug 5, 2025

• Improve README.md formatting and fix documentation errors

  #159841 opened Aug 5, 2025

• [AOTInductor] ABI-Compatibility for RecordFunction.

  #159842 opened Aug 5, 2025

• Revert "[BE] Update xpu driver repo for CD used almalinux 8.10 (#1573…

  #159849 opened Aug 5, 2025

• Guard the CPU cpp wrapper tests on having a cpp wrapper

  #159850 opened Aug 5, 2025

• [dtensor] fix incorrect norm calculation for Partial DTensors

  #159856 opened Aug 5, 2025

• [Do not merge][TensorPipe] Test PR https://github.com/pytorch/tensorpipe/pull/464

  #159857 opened Aug 5, 2025

• [ROCm] Clean up CUDA state between tests

  #159858 opened Aug 5, 2025

• Added PyTorch LUT optimisation for GELU bf16 operators

  #159859 opened Aug 5, 2025

• Issue 146167 inductor type hints lowering 1

  #159861 opened Aug 5, 2025

• Fix skipIfXpu and skipIfHpu and similar skip decorators

  #159862 opened Aug 5, 2025

• Implement `list(UserDefinedObject)` via `force_unpack_var_sequence`

  #159864 opened Aug 5, 2025

• [collections.abc] Ensure that binop calls works with UserDefinedObjects

  #159865 opened Aug 5, 2025

• error message for instantiating CUDA Stream if CUDA not available

  #159868 opened Aug 5, 2025

• Add linux-aarch64 and windows python 3.14 nightly builds

  #159869 opened Aug 5, 2025

• ParamwiseScheduler for different schedulers for specific parameter groups

  #159873 opened Aug 5, 2025

• [pytorch] Moving torch.compile worker process logs to a dedicated rank based log directory

  #159874 opened Aug 5, 2025

• [ca] enable on PYTORCH_TEST_WITH_INDUCTOR

  #159875 opened Aug 5, 2025

• Automated submodule update: tensorpipe

  #159876 opened Aug 5, 2025

• Fix meta for constant_pad_nd

  #159878 opened Aug 5, 2025

• Make distributed modules importable without distributed build

  #159889 opened Aug 5, 2025

• Add tests for Gaussian Mixture Model numerical consistency

  #159893 opened Aug 5, 2025

• tools: Add option to log build output to a file

  #159895 opened Aug 5, 2025

• Allow torch.hub.load with unauthorized GITHUB_TOKEN

  #159896 opened Aug 5, 2025

• [DTensor] Migrate tests to Continuous test base

  #159898 opened Aug 5, 2025

• [pt2e] Avoid getting model device once per node

  #159901 opened Aug 5, 2025

• Wrap class definitions in `set_fullgraph(False)` in `test_collections`

  #159902 opened Aug 5, 2025

• [BE] Add linter to detect unused docker images

  #159905 opened Aug 5, 2025

• [DO NOT MERGE] Perf-testing #158137 addmm_fusion 2

  #159909 opened Aug 5, 2025

• [CUDA] Bump tolerances for `test_baddmm`

  #159915 opened Aug 5, 2025

• Fix typo

  #159919 opened Aug 6, 2025

• Close some sources of fake tensor leakages

  #159923 opened Aug 6, 2025

• [AOTI] Check if triton is installed when BUILD_AOT_INDUCTOR_TEST=1

  #159934 opened Aug 6, 2025

• added class or module info for functions blocked by weight-only load

  #159935 opened Aug 6, 2025

• Fix AdaptiveMaxPoll index error

  #159936 opened Aug 6, 2025

• [WIP][device_mesh] Move global state into class method

  #159937 opened Aug 6, 2025

• [Intel GPU PT2E] Infer runtime out dtype based on dequant node in pattern

  #159941 opened Aug 6, 2025

• [CD] Add ptl aot target in xpu windows build

  #159943 opened Aug 6, 2025

• [WIP] enable some tests in test_ops.TestCommon on Intel GPU

  #159944 opened Aug 6, 2025

• avoid bit cast for bfloat16_t

  #159946 opened Aug 6, 2025

• Update SECURITY.md - added branded name as PyTorch instead of Pytorch and others.

  #159953 opened Aug 6, 2025

• [inductor] fix triton bucketize mask propagation

  #159961 opened Aug 6, 2025

• fix fake_tensor for aten._to_copy

  #159964 opened Aug 6, 2025

• Fix redundant move warnings in dim.cpp

  #159966 opened Aug 6, 2025

• Support device ordering in Shard

  #159967 opened Aug 6, 2025

• Add mast job name from env variable

  #159971 opened Aug 6, 2025

• [bucketing] Bucket only adjacent collectives to prevent reordering

  #159983 opened Aug 6, 2025

• [ROCm][inductor][dashboard] Add GPT2ForSequenceClassification to use_larger_multiplier_for_smaller_tensor list

  #160001 opened Aug 6, 2025

• Allow setting quantized engine to none

  #160003 opened Aug 6, 2025

• [export] Fix custom ops in subgraphs

  #160004 opened Aug 6, 2025

• Set PYTHONHOME for inductor subprocesses using torch

  #160008 opened Aug 6, 2025

• [torch.mtia] fix a bug in storage.resize_()

  #160017 opened Aug 6, 2025

• [1/3][ghstack] [vllm ci build setup ]setup lumen_cli

  #160043 opened Aug 7, 2025

• [ghstack] setup torch_cli build

  #160044 opened Aug 7, 2025

• Update (base update)

  #160045 opened Aug 7, 2025

• ci: Reduce the amount of log spam for manywheel

  #160055 opened Aug 7, 2025

• ci: Remove app from GH_CHECKSUITES_FRAGMENT

  #160056 opened Aug 7, 2025

• Fix profiler stack trace names

  #160058 opened Aug 7, 2025

• Error when there is side effect in strict mode

  #160060 opened Aug 7, 2025

• Clarify EMA equation in get_ema_multi_avg_fn docstring

  #160061 opened Aug 7, 2025

• Update torch-xpu-ops commit pin

  #160062 opened Aug 7, 2025

• Fix hpu backend mapping issue

  #160063 opened Aug 7, 2025

• Move hardware_destructive_interference_size to c10/core/alignment.h

  #160067 opened Aug 7, 2025

• [DO NOT MERGE] Stress Test MI325 Capacity.

  #160073 opened Aug 7, 2025

• Bump to ONNX 1.19.0

  #160076 opened Aug 7, 2025

• Enable prioritized linker optimization for AArch64 in setup.py and clean up CI script

  #160078 opened Aug 7, 2025

• Make manylinux build.sh work for AArch64 and AArch64+CUDA builds

  #160079 opened Aug 7, 2025

• POC for VLA SVE Vectorized class

  #160080 opened Aug 7, 2025

• [ROCm] [inductor] Added test skips for small ROCm gpus

  #160081 opened Aug 7, 2025

• Generalize Block/Pool in device caching allocator

  #160082 opened Aug 7, 2025

• [cpp][inductor] Fix crash on bmm when input is used twice.

  #160087 opened Aug 7, 2025

• set up vllm build logics in torch_cli

  #160088 opened Aug 7, 2025

• [2/3 step][ vllm ci build setup] Add vlllm buld logic and dockerfile

  #160089 opened Aug 7, 2025

• Update (base update)

  #160090 opened Aug 7, 2025

• [DO NOT MERGE] Testing with TOT Kineto

  #160091 opened Aug 7, 2025

• [fx] fix split_module with symint

  #160093 opened Aug 7, 2025

• Try fix AC tag propagation in compile when not using dynamo

  #160096 opened Aug 7, 2025

• Fix flight recorder for P2P ops

  #160097 opened Aug 7, 2025

• Add ownership token when needed on GradientEdge

  #160098 opened Aug 7, 2025

• [OpenReg] Add Event&Stream Support for OpenReg Backend

  #160099 opened Aug 7, 2025

• [OpenReg] Integrate Event&Stream from OpenReg Backend into PyTorch

  #160100 opened Aug 7, 2025

• [OpenReg] Improve the Event and Stream capabilities of DeviceGuardImplInterface

  #160101 opened Aug 7, 2025

• [TEST] Revert "[ROCm][CI] upgrade to 6.4.2 patch release (#158887)"

  #160103 opened Aug 7, 2025

• [ROCm] Integrate AITER Fav3 fwd kernels

  #160105 opened Aug 7, 2025

• Add cutedsl template support to compile

  #160108 opened Aug 7, 2025

• Add flash attention impl to flex attention

  #160109 opened Aug 7, 2025

• switch prefer_deferred_runtime_asserts_over_guards in export

  #160111 opened Aug 7, 2025

• ci: Update test_trymerge for stale data

  #160112 opened Aug 7, 2025

• [inductor] Estimate peak memory allocfree and applying to reordering collectives

  #160113 opened Aug 7, 2025

• [3/3][ghstack][vllm ci build setup]vllm build workflow

  #160116 opened Aug 7, 2025

• Update (base update)

  #160117 opened Aug 7, 2025

• ci: Update permissions to include checks + actions

  #160118 opened Aug 7, 2025

• Account for triton kernel source code hidden in custom ops properly in AOTAutogradCache

  #160120 opened Aug 7, 2025

• Gh/yangw dev/13/orig

  #160125 opened Aug 7, 2025

• Gh/yangw dev/14/orig

  #160126 opened Aug 7, 2025

• [AOTI] use CudaCachingAllocator for memory allocation

  #160127 opened Aug 7, 2025

• Add `CUDA_KERNEL_ASSERT_PRINTF`, a more flexible `CUDA_KERNEL_ASSERT_MSG`

  #160129 opened Aug 7, 2025

• [AOTInductor] Add grid information for Triton Kernels

  #160131 opened Aug 7, 2025

• [inductor] TLParse tensor metadata logging + test

  #160132 opened Aug 7, 2025

• [FSDP][Collectives] skipping allgather when world size is 1

  #160135 opened Aug 7, 2025

• [FSDP][Collectives] skipping reduce_scatter when world size is 1

  #160136 opened Aug 7, 2025

• [dynamo, nested graph breaks] clean up comments and codegen

  #160138 opened Aug 7, 2025

• fixing graph break for namedtuple._replace

  #160139 opened Aug 7, 2025

• Add stable Tensor get_device_index, use more stable DeviceIndex

  #160143 opened Aug 7, 2025

• [c10d] Error out the case when registering symmetric memory without eager init

  #160145 opened Aug 7, 2025

• [Still Work in progress] vllm test cli

  #160146 opened Aug 7, 2025

• [FSDP][Replicate] replicate tests for param registration and input device movements

  #160147 opened Aug 7, 2025

• [Still WIP]setup test ci workflow

  #160149 opened Aug 7, 2025

• dynamo: Write a structured trace for supressed exceptions.

  #160151 opened Aug 8, 2025

• [dynamo] Trace nn.Module __delattr__

  #160152 opened Aug 8, 2025

• [dynamo] Make ListIterator track mutations to original list

  #160154 opened Aug 8, 2025

• Wrap class definitions in `set_fullgraph(False)` in `test_dict`/`test_ordered_dict`

  #160156 opened Aug 8, 2025

• [WIP][1/N] Port 6 fsdp distributed test cases to Intel GPU

  #160158 opened Aug 8, 2025

• [inductor] turn on windows inductor UTs

  #160161 opened Aug 8, 2025

• Support NUMA Binding for Callable Entrypoints

  #160163 opened Aug 8, 2025

• fix(inductor): show intermediate buffers for split reductions in profile

  #160166 opened Aug 8, 2025

• [CUDAGraph] Skip CUDAGraph when only 1 kernel

  #160168 opened Aug 8, 2025

• Add build support for RISCV

  #160172 opened Aug 8, 2025

• [Inductor XPU GEMM] Step 1/N: Add cutlass-sycl repro.

  #160173 opened Aug 8, 2025

• [Inductor XPU GEMM] Step 2/N: Generalize cutlass configuration.

  #160174 opened Aug 8, 2025

• Fixes #119272

  #160178 opened Aug 8, 2025

• Do not rpath CUDA stubs folder in JIT generated code

  #160179 opened Aug 8, 2025

• [cutlass backend] re-add pip cutlass path

  #160180 opened Aug 8, 2025

• [do not merge][inductor] optimize welford reduction

  #160181 opened Aug 8, 2025

• [WIP] [1/N] Introduce a generic CachingDeviceAllocatorImpl for cross backend use

  #160182 opened Aug 8, 2025

• Update triton xpu commit to support python 3.14

  #160183 opened Aug 8, 2025

• Draft: separate reqs for manywheel build and pin

  #160184 opened Aug 8, 2025

• [doc] fix spelling of word - when

  #160185 opened Aug 8, 2025

• [DO NOT MERGE][Kineto] Testing not clearing per-thread attribute at all

  #160186 opened Aug 8, 2025

• [DCP][OSS] Remove extra collective on load

  #160189 opened Aug 8, 2025

• [Triton] [Inductor] Generalize index broadcasting to handle torch.utils._sympy.functions.Identity

  #160190 opened Aug 8, 2025

• Flex Attention heuristics: a Blackwell config

  #160192 opened Aug 8, 2025

• [FR] Don't check incomplete ranks for printing

  #160195 opened Aug 8, 2025

• kill allow_complex_guards_as_runtime_asserts

  #160198 opened Aug 8, 2025

• Add CUDA installation script for CUDA 13

  #160201 opened Aug 8, 2025

• Refactors symmetric memory creation through allocator interface

  #160202 opened Aug 8, 2025

• [WIP][doc] AOTI debugging guide

  #160204 opened Aug 8, 2025

• [DCP][HF] Add option to parallelize reads in HF Storage Reader

  #160205 opened Aug 8, 2025

• [ROCm][DO NOT MERGE] treat unset variables in manywheels build_rocm.sh as an error.

  #160207 opened Aug 8, 2025

• [wip] Support `defaultdict(None, mapping | Iterable[Tuple])`

  #160209 opened Aug 8, 2025

• Add `is_cpu` and `dtype` convenience methods for stable tensor type

  #160212 opened Aug 8, 2025

• [muon] Introduce Muon optimizer to PyTorch

  #160213 opened Aug 8, 2025

• Port amax to stable ABI

  #160214 opened Aug 8, 2025

• [ROCm] Enable MI355 CI on PRs, and run full set of UTs on PRs

  #160215 opened Aug 8, 2025

• Wrap class definitions in `set_fullgraph(False)` in `test_set`

  #160216 opened Aug 8, 2025

• Wrap class definitions in `set_fullgraph(False)` in `test_operator`

  #160217 opened Aug 8, 2025

• Cap num_stages to 1 on AMD in inductor's triton_heuristics.py

  #160218 opened Aug 8, 2025

• appending the pythonpath

  #160219 opened Aug 8, 2025

• [MPS] Sparse enable indices and values

  #160223 opened Aug 8, 2025

• Remove torch.serialization entries from the doc ignore list

  #160224 opened Aug 8, 2025

• switch order of ScalarType::Undefined so it equates to -1

  #160226 opened Aug 8, 2025

• Fix undefined behavior in mul_overflows

  #160229 opened Aug 8, 2025

• torchdim Python port

  #160236 opened Aug 9, 2025

• fix(docs): wrong conversion from rst to md in torch.compiler_troubleshooting.md

  #160238 opened Aug 9, 2025

• [kernacle] add support for addmm and bmm

  #160239 opened Aug 9, 2025

• guard_or_false cat ops

  #160250 opened Aug 9, 2025

• unify broadcast_shapes functions and avoid duplicates

  #160251 opened Aug 9, 2025

• migrate more simple gso checks

  #160253 opened Aug 9, 2025

• extract shape in _view_has_unbacked_input

  #160255 opened Aug 9, 2025

• Detect torch function in lists as well

  #160256 opened Aug 9, 2025

• [inductor] Windows inductor use intel-openmp.

  #160258 opened Aug 9, 2025

• [vllm hash update] update the pinned vllm hash

  #160259 opened Aug 10, 2025

• [DO NOT MERGE] Autograd Onboarding Lab

  #160264 opened Aug 10, 2025

• Support of dtensor redistribute with device order

  #160266 opened Aug 10, 2025

• Wrap class definitions in `set_fullgraph(False)` in `test_int`/`bool`/`float`/`complex`

  #160276 opened Aug 10, 2025

• Wrap class definitions in `set_fullgraph(False)` in `test_list`/`tuple`

  #160277 opened Aug 10, 2025

• Wrap class definitions in `set_fullgraph(False)` in `test_iter`

  #160278 opened Aug 10, 2025

• [FSDP2] cast `unsharded_param_grad` to correct reduce dtype

  #160279 opened Aug 10, 2025

• [simplefsdp] add multi parallelism autobucketing

  #160282 opened Aug 10, 2025

• [C10D] Add check_rng_sync util

  #160283 opened Aug 10, 2025

• WIP summarize ranks

  #160284 opened Aug 10, 2025

• [BE] Remove modernize suppression

  #160288 opened Aug 11, 2025

• [BE] Save attributes for CppCompileError for pickleing

  #160294 opened Aug 11, 2025

• Recursively descend into lists for TF in getitem

  #160297 opened Aug 11, 2025

• [wip][claude-code] support multi kernel reductions

  #160298 opened Aug 11, 2025

• [Do not merge]Upgrade oneDNN to v3.9

  #160299 opened Aug 11, 2025

• [Intel GPU] check data alignment before contiguousness

  #160301 opened Aug 11, 2025

• Improve README.md formatting and fix documentation errors

  #160307 opened Aug 11, 2025

• Enable XPU for test_autograd_function.py

  #160309 opened Aug 11, 2025

• [inductor] Fix descriptor broadcasting for singleton dimensions

  #160310 opened Aug 11, 2025

• Fix reinplace optimization issue for index_put when self and source alias

  #160311 opened Aug 11, 2025

• Optimize `min`, `max` gradient behavior description

  #160312 opened Aug 11, 2025

• Fix get_free_symbol_uses for several nodes

  #160314 opened Aug 11, 2025

• Add sdist handling to version finding

  #160315 opened Aug 11, 2025

• [DO NOT MERGE] ACL Version Upgrade v52.3.0

  #160316 opened Aug 11, 2025

• [Windows] Update libuv version from 1.39 to 1.51

  #160318 opened Aug 11, 2025

• Draft Python API for F.linear_cross_entropy

  #160319 opened Aug 11, 2025

• [Caffe2] Enable SVE128

  #160323 opened Aug 11, 2025

• [Caffe2] Add SVE128 vectorized<bfloat16_t> template layer

  #160324 opened Aug 11, 2025

• [Caffe2] Add SVE128 vectorized<float16_t> template layer

  #160325 opened Aug 11, 2025

• [Caffe2] Add SVE128 vectorized<int##bit##_t> template layers

  #160326 opened Aug 11, 2025

• [Caffe2] Add SVE128 vectorized<T> template layers for unsigned integers

  #160327 opened Aug 11, 2025

• [Caffe2] Add SVE128 vectorized<float> template layer

  #160328 opened Aug 11, 2025

• [Caffe2] Add SVE128 vectorized<double> template layer

  #160329 opened Aug 11, 2025

• Wrap class definitions in `set_fullgraph(False)` in `test_math`/`cmath`

  #160330 opened Aug 11, 2025

• Wrap class definitions in `set_fullgraph(False)` in `test_sort`

  #160331 opened Aug 11, 2025

• Fix typo: 'complext'

  #160335 opened Aug 11, 2025

• [ROCm] Make triton build rocm agnostic

  #160336 opened Aug 11, 2025

• Return the loaded library in torch.ops.load_library

  #160338 opened Aug 11, 2025

• [MPS] Add support for log_normal_

  #160339 opened Aug 11, 2025

• Get tensor subclasses and torch.library.triton_op to dispatch correctly

  #160341 opened Aug 11, 2025

• Add batch option for send/recv_object_list

  #160342 opened Aug 11, 2025

• [FSDP][Replicate] replicate tests for casting module after init

  #160344 opened Aug 11, 2025

• [retry-land][pytorch][dynamo_compile] Log stack_trace to dynamo_compile

  #160348 opened Aug 11, 2025

• Parameterized CUDA Graph Launch 2

  #160351 opened Aug 11, 2025

• [redone][pytorch] Moving torch.compile worker process logs to a dedicated rank based log directory

  #160352 opened Aug 11, 2025

• Easy decomposeK code refactor

  #160353 opened Aug 11, 2025

• fix cpp builder to avoid missing-source compile error

  #160354 opened Aug 11, 2025

• [cutlass backend] Allow bmm use cases when batch stride is 0

  #160356 opened Aug 11, 2025

• Factor out the strings to templates for better editor integration

  #160357 opened Aug 11, 2025

• [Let's see CI] Remove use of device_guard for TensorIndex kernel

  #160360 opened Aug 11, 2025

• setup test ci

  #160361 opened Aug 11, 2025

• Typing for common.py

  #160362 opened Aug 11, 2025

• Type cudagraphs.py

  #160363 opened Aug 11, 2025

• typing debugging.py

  #160364 opened Aug 11, 2025

• typing distributed.py

  #160365 opened Aug 11, 2025

• typing inductor and placeholder backends

  #160366 opened Aug 11, 2025

• typing registry.py

  #160367 opened Aug 11, 2025

• Type backend torchxla

  #160368 opened Aug 11, 2025

• typing tvm.py

  #160369 opened Aug 11, 2025

• [kernacle] add support for addmm and bmm

  #160370 opened Aug 11, 2025

• [kernacle] add support for addmm and bmm

  #160371 opened Aug 11, 2025

• [MTIA] Add MTIA dispatch for kernel foreach_maximum (#160358)

  #160372 opened Aug 11, 2025

• [ROCm][Windows] Include native_transformers srcs to fix link errors.

  #160373 opened Aug 11, 2025

• [inductor][while_loop][be] improve the readability of output handling

  #160374 opened Aug 11, 2025

• [while_loop] autograd support

  #160375 opened Aug 11, 2025

• [while_loop][auto_grad] support dynamic shape

  #160376 opened Aug 11, 2025

• [while_loop][autograd] support carry of int tensor

  #160377 opened Aug 11, 2025

• [BE][CI] Adjust `error_inputs` for cat and complex

  #160378 opened Aug 11, 2025

• [CI] Move CUDA tests to trunk workflow

  #160379 opened Aug 11, 2025

• [AOTInductor] Add input information for Triton Kernels in AOTI

  #160380 opened Aug 11, 2025

• [Inductor][Configs] Expose autotune_num_choices_displayed through environment variable

  #160381 opened Aug 11, 2025

• Turn on part of provenance tracking by default

  #160383 opened Aug 12, 2025

• [audio hash update] update the pinned audio hash

  #160384 opened Aug 12, 2025

• [TorchScript] thread-safe ErrorReport::CallStack

  #160386 opened Aug 12, 2025

• Fix LBFGS warning convert a tensor with requires_grad=True to a scalar

  #160389 opened Aug 12, 2025

• [FSDP][Replicate] Testing replicate parity in single and multigroup

  #160390 opened Aug 12, 2025

• wip

  #160393 opened Aug 12, 2025

• [export] Refactor PT2 Archive weight saving and loading

  #160394 opened Aug 12, 2025

115 Issues closed by 37 people

• flex_attention + dynamic=True with large batch or heads causes Triton Error [CUDA]: invalid argument

  #157018 closed Aug 12, 2025

• [FlexAttention] Zero computed input gradients with torch.compile + customized autograd func

  #159299 closed Aug 12, 2025

• DISABLED test_multiple_mutations_of_buf (__main__.TestOperatorReorderForPeakMemory)

  #159952 closed Aug 12, 2025

• 2.8 flex attention kernels result in triton warning

  #158463 closed Aug 11, 2025

• Wrong-size gradients in Expert Parallel MoE

  #160285 closed Aug 11, 2025

• Use FP32 for ConvTranspose3D when using autocast on MPS

  #160332 closed Aug 11, 2025

• DISABLED test_op_has_batch_rule_tensordot_cuda_float32 (__main__.TestVmapOperatorsOpInfoCUDA)

  #142769 closed Aug 11, 2025

• DISABLED test_cuda_kernel_loop_overflow_large (__main__.TestCuda)

  #159285 closed Aug 11, 2025

• DISABLED test_deferred_runtime_asserts (__main__.ReproTests)

  #156817 closed Aug 11, 2025

• DISABLED test_graph_memory_stats_and_use_result_after_destroy_graph (__main__.TestCuda)

  #159286 closed Aug 11, 2025

• Regression on compile with backend inductor with torch 2.8

  #160084 closed Aug 11, 2025

• MPS regression on supported dtypes/scalar/alpha combinations on add/sub/rsub

  #160208 closed Aug 11, 2025

• [RFC] A Distributed CUDA Unified Memory Backend for PyTorch

  #158122 closed Aug 11, 2025

• DISABLED test_inplace_on_view_makes_base_require_grad_cpu (__main__.TestAutogradDeviceTypeCPU)

  #156209 closed Aug 11, 2025

• DISABLED test_dataclass_init_with_default_factory_with_inputs (__main__.ReproTests)

  #156799 closed Aug 11, 2025

• DISABLED test_triton_fx_graph_with_et_cuda (__main__.TestExecutionTraceCUDA)

  #159236 closed Aug 11, 2025

• Severe performance regression on deterministic algorithm in torch 2.0

  #109856 closed Aug 11, 2025

• uint64_t can't be represented as `c10::IValue` (was TorchDispatchMode + full raises an overflow error for uint64 when it's out of int64 range)

  #159168 closed Aug 11, 2025

• Add __riscv macro detection to support the scalar backend for RISCV

  #160171 closed Aug 11, 2025

• DISABLED test_qlinear_add_int8_mixed_bf16_use_relu_False_is_qat_False_is_dynamic_True (__main__.TestPatternMatcher)

  #157911 closed Aug 11, 2025

• DISABLED test_conv_transpose_unary_fusion_ops (__main__.TestMkldnnFusion)

  #158115 closed Aug 11, 2025

• MPS does not support addmm for non-float input

  #154901 closed Aug 11, 2025

• [Feature] Implement a CUDA kernel for _weight_int8pack_mm

  #158849 closed Aug 10, 2025

• aten::grid_sampler_3d'

  #160237 closed Aug 10, 2025

• `torch.compile` on `.sum() `and `.item()` calls errors from `tensorify_python_scalars`

  #158083 closed Aug 9, 2025

• torch.jit.script fails with 'undefined value' error when module has @torch.compiler.disable decorator (regression in 2.8.0)

  #160059 closed Aug 9, 2025

• Aborted (core dumped) in `reflection_pad 2d`

  #142455 closed Aug 8, 2025

• [ONNX] Exporter crashes when fx node output includes None

  #160150 closed Aug 8, 2025

• AArch64 Inductor Perf Test build fails installing torchao

  #160188 closed Aug 8, 2025

• DISABLED test_repeat_interleave_2_dynamic_shapes_xpu (__main__.DynamicShapesCodegenGPUTests)

  #159803 closed Aug 8, 2025

• error message for `Tensor.index_put_` could be improved for MPS failure mode

  #160034 closed Aug 8, 2025

• `copy_()` fails with HSDP in FSDP2

  #147568 closed Aug 8, 2025

• [dynamo, guards] Move SHAPE_ENV guard to C++

  #143309 closed Aug 8, 2025

• BUG: numpy very slow after import torch

  #158005 closed Aug 8, 2025

• [ONNX] Improve dynamic_axes to dynamic_shapes conversion in exporter

  #150940 closed Aug 8, 2025

• TypeError: (): incompatible function arguments

  #102832 closed Aug 8, 2025

• PyTorch 2.8: PYTHONPATH no longer respected when building from source?

  #160092 closed Aug 8, 2025

• HAS_CUDA in the inductor tests is really HAS_CUDA_AND_TRITON

  #159399 closed Aug 8, 2025

• lintrunner not support riscv64

  #160170 closed Aug 8, 2025

• DISABLED test_inductor_reuse_buffer_after_inplace_collective (__main__.CompileTest)

  #147950 closed Aug 8, 2025

• DISABLED test_dataclass_in_module (__main__.ReproTests)

  #156776 closed Aug 8, 2025

• DISABLED test_inductor_reduce_scatter_tensor_single (__main__.CompileTest)

  #147911 closed Aug 8, 2025

• DISABLED test_graph_two_successive (__main__.TestCuda)

  #159113 closed Aug 8, 2025

• DISABLED test_inductor_multiple_specializations_cuda (__main__.GPUTests)

  #154705 closed Aug 8, 2025

• Add the `cutlass-sycl` submodule and set it as the default cutlass path for XPU.

  #160176 closed Aug 8, 2025

• torch.multinomial sample behavior is not consist with sample numbers

  #159927 closed Aug 8, 2025

• [RFC] CUDAPluggableAllocator receives malloc request of size zero.

  #159892 closed Aug 8, 2025

• Feature Request: deterministic adaptive_avg_pool2d_backward_cuda

  #84860 closed Aug 7, 2025

• [libtorh]Consistency problem of gpu computing

  #94976 closed Aug 7, 2025

• Add nondeterministic alert to `torch.scatter_`

  #70583 closed Aug 7, 2025

• nn.LSTM gives nondeterministic results with dropout and multiple layers, OR cuDNN version mismatch

  #35661 closed Aug 7, 2025

• Add nondeterministic alert to `.scatter_`

  #133204 closed Aug 7, 2025

• Deterministic support for adaptive_avg_pool2d_backward_cuda

  #149130 closed Aug 7, 2025

• Inconsistent output for ConvTranspose3d on GPU

  #137970 closed Aug 7, 2025

• [inductor] tune_scaled_grouped_mm fails with memory layout assertion, despite memory layout assertions prior to op call passing

  #156325 closed Aug 7, 2025

• Fix the description of `alpha` in `torch.sub`

  #159637 closed Aug 7, 2025

• DISABLED test_inductor_inplace_op_on_view (__main__.CompileTest)

  #147852 closed Aug 7, 2025

• DISABLED test_cuda_kernel_loop_overflow (__main__.TestCuda)

  #159069 closed Aug 7, 2025

• DISABLED test_record_stream (__main__.TestCuda)

  #134746 closed Aug 7, 2025

• torch.cuda.empty_cache() shall support MemPool as an argument just like the C++ interface

  #160069 closed Aug 7, 2025

• Bits types cannot be used under deterministic mode

  #109802 closed Aug 7, 2025

• /usr/local/lib/python3.11/dist-packages/torch/autograd/graph.py:825: UserWarning: grid_sampler_2d_backward_cuda does not have a deterministic implementation, but you set 'torch.use_deterministic_algorithms(True, warn_only=True)'.

  #152171 closed Aug 7, 2025

• X86InductorQuantizer does not quantize anything

  #160095 closed Aug 7, 2025

• DISABLED test_cuda_memory_leak_detection_propagates_errors (__main__.TestCuda)

  #159039 closed Aug 7, 2025

• DISABLED test_inductor_all_reduce_coalesced (__main__.CompileTest)

  #147726 closed Aug 7, 2025

• DISABLED test_inductor_broadcast (__main__.CompileTest)

  #147816 closed Aug 7, 2025

• torch.sparse.to_sparse_semi_structured generate irrelevant values with dtype half or int8

  #159872 closed Aug 7, 2025

• [rocm] HIP Graph capture raises segmentation fault on AMD GPU but CUDA Graph capture succeeds on Nvidia GPU

  #155720 closed Aug 7, 2025

• `torch.unique` behaves strange for large input arrays on Windows

  #135019 closed Aug 7, 2025

• addmv bfloat16 inconsistency on x86

  #159960 closed Aug 7, 2025

• [inductor][fuzzer] Compilation Error in complex64+toint

  #157683 closed Aug 7, 2025

• Build pytorch for rocm failed

  #148167 closed Aug 7, 2025

• DISABLED test_comment_graph_fragment (__main__.TritonCodeGenTests)

  #159925 closed Aug 7, 2025

• DISABLED test_hop_eager (__main__.TorchFunctionModeTests)

  #159950 closed Aug 7, 2025

• DISABLED test_hop (__main__.TorchFunctionModeTests)

  #159951 closed Aug 7, 2025

• DISABLED test_remote_cache_load_function_device_cuda_bfloat16_dynamic_False_bundle_triton_True_use_static_cuda_launcher_False (__main__.TestFxGraphCache)

  #150444 closed Aug 7, 2025

• DISABLED test_inductor_all_reduce_non_contig_input (__main__.CompileTest)

  #147733 closed Aug 7, 2025

• DISABLED test_inductor_all_to_all_single (__main__.CompileTest)

  #147795 closed Aug 7, 2025

• DISABLED test_wait_tensor (__main__.CompileTest)

  #148014 closed Aug 7, 2025

• Release 2.8.0 validations checklist and cherry-picks

  #158939 closed Aug 6, 2025

• torch_shm_manager: undefined reference to gloo

  #146239 closed Aug 6, 2025

• Performance bug on `mode` of `torch.autograd.grad_mode.inference_mode`

  #159633 closed Aug 6, 2025

• [Docs] 2.8 version is listed twice

  #159972 closed Aug 6, 2025

• `CosineAnnealingWarmRestarts` should use integer epoch

  #69841 closed Aug 6, 2025

• `torch.export` fails on `torch.cond` that dispatches Triton kernels (`SpecViolationError: missing val`)

  #159955 closed Aug 6, 2025

• ghstack-mergeability-check and Check labels failing with 'Resource not accessible by integration'

  #159825 closed Aug 6, 2025

• DISABLED test_addmm_dtype_mismatch (__main__.TestPatternMatcher)

  #159631 closed Aug 6, 2025

• [MPS] Remove MacOS-13 support

  #159275 closed Aug 6, 2025

• Multiple runners shutdown for an autoupdate while still running jobs

  #107402 closed Aug 6, 2025

• [v.2.8.0] Release Tracker

  #156745 closed Aug 6, 2025

• PyTorch Improper Resource Shutdown or Release vulnerability

  #159963 closed Aug 6, 2025

• Inductor Perf MX to_blocked

  #153194 closed Aug 6, 2025

• Reproducibility on different platform

  #159846 closed Aug 6, 2025

• DISABLED test_distributed_checkpoint_state_dict_type0_cuda (__main__.TestDistributedCheckpointCUDA)

  #145807 closed Aug 6, 2025

• UnicodeDecodeError in torch.compile on Windows MSVC

  #159537 closed Aug 6, 2025

• DISABLED test_inplace_on_view_then_no_grad_cpu (__main__.TestAutogradDeviceTypeCPU)

  #156306 closed Aug 5, 2025

• Nightly libtorch links on website are incorrect

  #159880 closed Aug 5, 2025

• Add CUDA kernel support for RTX 5070 Ti (Ada Lovelace, SM 8.9)

  #159851 closed Aug 5, 2025

• incorrect _unsafe_index meta

  #139312 closed Aug 5, 2025

• DISABLED test_grad_with_tracer_ScheduleClass0 (__main__.ScheduleTest)

  #159034 closed Aug 5, 2025

• DISABLED test_grad_with_tracer_ScheduleClass1 (__main__.ScheduleTest)

  #159253 closed Aug 5, 2025

• DISABLED test_zero_bubble_with_model_kwargs_ScheduleClass0 (__main__.ScheduleTest)

  #154547 closed Aug 5, 2025

• DISABLED test_schedule_with_native_zero_bubble_ScheduleClass0 (__main__.ScheduleTest)

  #156088 closed Aug 5, 2025

• DISABLED test_schedule_with_native_zero_bubble_ScheduleClass1 (__main__.ScheduleTest)

  #156328 closed Aug 5, 2025

• DISABLED test_grad_with_manual_interleaved_ScheduleClass2_use_new_runtime_False (__main__.ScheduleTest)

  #154443 closed Aug 5, 2025

• DISABLED test_grad_with_manual_interleaved_ScheduleClass2_use_new_runtime_True (__main__.ScheduleTest)

  #154481 closed Aug 5, 2025

• DISABLED test_is_isnot (__main__.TestScript)

  #120694 closed Aug 5, 2025

• DISABLED test_index (__main__.TestPythonBuiltinOP)

  #119160 closed Aug 5, 2025

• DISABLED test_ranks_and_tag (__main__.CompileTest)

  #147974 closed Aug 5, 2025

• DISABLED test_tensor_subclasses (__main__.TestScript)

  #119949 closed Aug 5, 2025

• [Feature Request] Add support for CUDA sm_120 (RTX 5070 Ti) in prebuilt PyTorch binaries

  #159847 closed Aug 5, 2025

• Inconsistent Model Results and Failures on Windows with CUDA vs. CPU PyTorch Builds

  #156547 closed Aug 5, 2025

• The recorded tid in torch kineto profiler do not match with syscall(SYS_gettid) or pthread_self() in Linux

  #159771 closed Aug 5, 2025

• DISABLED test_undefined_grads_mode_warn (__main__.TestAutogradFallback)

  #159471 closed Aug 5, 2025

• DISABLED test_pin_memory_no_cuda (__main__.TestDictDataLoader)

  #159802 closed Aug 5, 2025

115 Issues opened by 87 people

• bug in libtorch optimize_for_inference

  #160392 opened Aug 12, 2025

• `torch.compile` backward pass fails with `AssertionError` in Inductor C++ codegen when model returns tuple outputs

  #160391 opened Aug 12, 2025

• Perf differences observed between AOT and JIT on RAFT model

  #160388 opened Aug 12, 2025

• DISABLED RecordDebugHandles.Basic (__main__.test_jit)

  #160387 opened Aug 12, 2025

• TTGIR error for FlexAttention on B200

  #160385 opened Aug 12, 2025

• Use `tf32x3` through Inductor

  #160359 opened Aug 11, 2025

• Max-autotune is slower than eager for specific jagged tensor shapes

  #160355 opened Aug 11, 2025

• DISABLED test_fused_all_gather_scaled_matmul_gather_dim_1_scale_mode_tensor-wise (__main__.SymmetricMemoryTest)

  #160347 opened Aug 11, 2025

• DISABLED test_custom_functions_and_tracer (__main__.TestFXNumericSuiteNShadows)

  #160346 opened Aug 11, 2025

• Meta device initialization on HF models + FSDP leads to different numerical behaviour

  #160340 opened Aug 11, 2025

• [RFC] Simplify bookkeeping of DeviceMesh slicing, (un)flattening, ...

  #160337 opened Aug 11, 2025

• Tensor subclasses don't work with torch.library.triton_op

  #160333 opened Aug 11, 2025

• `torch.Tensor` methods return type annotation too broad for `torch.Tensor` subclasses

  #160322 opened Aug 11, 2025

• FSDP does not reduce the gradient size during backward

  #160320 opened Aug 11, 2025

• Memory leak when using mark_dirty in Python custom autograd.Function

  #160317 opened Aug 11, 2025

• [DTensor] Make Redistribute autograd function twice-differentiable

  #160313 opened Aug 11, 2025

• Reserved memory is much bigger than allocated memory in multi-stream scenario

  #160308 opened Aug 11, 2025

• Many failures in inductor/test_max_autotune on H100

  #160305 opened Aug 11, 2025

• Multi device aoti compile not working

  #160303 opened Aug 11, 2025

• Redundant installation of CMake and Ninja

  #160302 opened Aug 11, 2025

• [inductor][cpu] Performance regression in 2025-08-02 nightly release

  #160296 opened Aug 11, 2025

• torch 2.9.0dev on cuda 13 producing Cuda mismatch error during flash/sage attention install

  #160293 opened Aug 11, 2025

• [FSDP2] Bug: NaN gradient when both HSDP and CPU offload are enabled

  #160291 opened Aug 11, 2025

• DISABLED test_small_stability (__main__.TestBase)

  #160290 opened Aug 11, 2025

• Symmetric memory backend bug: OOM with default, errors with CUDA/NCCL, nvshmem works

  #160289 opened Aug 11, 2025

• Improve CUDAGraph Tree heuristics on starting new generation

  #160281 opened Aug 10, 2025

• [Documentation Clarity] torch.min/torch.max gradient behavior

  #160273 opened Aug 10, 2025

• Add Green Context Support for fsdp2

  #160272 opened Aug 10, 2025

• Error when using torch.compile with dynamic=True: tensor 'guidance' stride mismatch at index 0

  #160271 opened Aug 10, 2025

• Questions about Volta Support

  #160269 opened Aug 10, 2025

• XPU out of memory on Intel iGPUs although plenty of memory is available per error message.

  #160265 opened Aug 10, 2025

• Graph partition + Fake Dependency from view op

  #160263 opened Aug 10, 2025

• CPU: Pytorch >= 2.7.0 is broken

  #160261 opened Aug 10, 2025

• PyTorch 2.8: cuda linking issue, undefined symbol

  #160248 opened Aug 9, 2025

• wan2.1 vae take more gpu memory after compile

  #160247 opened Aug 9, 2025

• pytorch.org/get-started/locally is wrong about supported Python version (3.12 is not highest)

  #160246 opened Aug 9, 2025

• DISABLED test_comprehensive_nn_functional_interpolate_trilinear_xpu_float32 (__main__.TestInductorOpInfoXPU)

  #160245 opened Aug 9, 2025

• DISABLED test_comprehensive_nn_functional_interpolate_trilinear_xpu_float64 (__main__.TestInductorOpInfoXPU)

  #160244 opened Aug 9, 2025

• DISABLED test_copy_non_blocking_is_pinned_xpu (__main__.AOTInductorTestABICompatibleGpu)

  #160243 opened Aug 9, 2025

• `torch.compile` produces output mismatch vs eager float64 for some seeds

  #160242 opened Aug 9, 2025

• `torch.export.export` fails when model is in `eval()` mode due to `torch.typename` being skipped by Dynamo

  #160241 opened Aug 9, 2025

• Migrate test.sh into python based tests

  #160240 opened Aug 9, 2025

• xpu: huggingface accelerate test_dynamo test fails on XPU

  #160232 opened Aug 8, 2025

• Offer official Pytorch Vulkan backend on pytorch.org

  #160230 opened Aug 8, 2025

• Building cuda kernels with debug information causes some cuda kernel launches to fail due to capacity constraints

  #160225 opened Aug 8, 2025

• DISABLED test_fused_all_gather_scaled_matmul_gather_dim_1_scale_mode_row-wise-sharded (__main__.SymmetricMemoryTest)

  #160203 opened Aug 8, 2025

• Shape propagation through repeated modules during dynamo export

  #160199 opened Aug 8, 2025

• LBFGS always raises warning about converting a tensor with requires_grad=True to a scalar?

  #160197 opened Aug 8, 2025

• Assertion Error When Freeing Unbacked Symbolic Shapes in Torch.Export

  #160196 opened Aug 8, 2025

• [CI] Inherit PYTHONPATH from env instead of overriding it completely

  #160193 opened Aug 8, 2025

• DISABLED test_fused_all_gather_scaled_matmul_gather_dim_1_scale_mode_row-wise-replicated (__main__.SymmetricMemoryTest)

  #160177 opened Aug 8, 2025

• [RFC] Enable cutlass to support Intel GPU into PyTorch Inductor.

  #160175 opened Aug 8, 2025

• When using DP + TP, DP only parameters diverge across TP ranks if using operations with non-deterministic implementations

  #160169 opened Aug 8, 2025

• libtorch: CMake 4.1: Error evaluating generator expression: TARGET_PROPERTY:torch,INTERFACE_LINK_LIBRARIES

  #160167 opened Aug 8, 2025

• aarch64 GPU wheel release in pypi for GH200/GB200 etc

  #160162 opened Aug 8, 2025

• Symbolic function 'aten::scaled_dot_product_attention' already registered for opset 14. Replacing the existing function with new function. This is unexpected. Please report it on github..

  #160157 opened Aug 8, 2025

• Add a "max-performance" mode to torch.compile for aggressive optimization

  #160142 opened Aug 7, 2025

• RuntimeError: miopenStatusUnknownError on PyTorch 2.8.0+rocm6.4

  #160141 opened Aug 7, 2025

• Can't use nvshmem triton device side function

  #160137 opened Aug 7, 2025

• Torchrun breaks with virtual environments

  #160130 opened Aug 7, 2025

• FX graph segments for multiple triton kernels from reduction node are broken

  #160124 opened Aug 7, 2025

• Add tensor-aware put_signal & wait_signal Triton wrappers to NVSHMEM Triton Kernels

  #160122 opened Aug 7, 2025

• cuda 13 broken

  #160104 opened Aug 7, 2025

• torch.library.infer_schema doesn't have full support for list types

  #160094 opened Aug 7, 2025

• DISABLED test_fused_all_gather_scaled_matmul_gather_dim_0_scale_mode_tensor-wise (__main__.SymmetricMemoryTest)

  #160085 opened Aug 7, 2025

• compile regression on forward pre hook with prepend=True

  #160083 opened Aug 7, 2025

• 'torch._tensor' has no attribute 'split' while using torch.compile under torch.device context.

  #160077 opened Aug 7, 2025

• FlexAttention backward compilation failure with GQA on NVIDIA B200

  #160074 opened Aug 7, 2025

• [TorchDynamo] excessive stack use: stack is 3286 deep on A100 with AMD EPYC CPU

  #160071 opened Aug 7, 2025

• FSDP2 not compatible with transformers >= 4.54.0 GenericForTokenClassification

  #160068 opened Aug 7, 2025

• `torch.compile` crashes on model exported via `torch.export`, with lifted constant from cache tensor

  #160066 opened Aug 7, 2025

• MultiheadAttention returns all NaN when using attn_mask to fully mask some heads, and using `need_weights=True`

  #160064 opened Aug 7, 2025

• Error raised when running torch.compile under FakeTensorMode context

  #160057 opened Aug 7, 2025

• `torch.Pad(mode="circular")` doesn't work for 4D or 5D input despite error msg

  #160053 opened Aug 7, 2025

• DISABLED test_flop_counter_op_options0_cuda_float16 (__main__.TestSchedulerCUDA)

  #160052 opened Aug 7, 2025

• DISABLED test_cat_extern_kernel_dynamic_shapes_mps (__main__.DynamicShapesGPUTests)

  #160051 opened Aug 7, 2025

• DISABLED test_retracibility_nested_list_out_dynamic_shapes (__main__.DynamicShapesExportTests)

  #160050 opened Aug 7, 2025

• DISABLED test_fused_all_gather_scaled_matmul_gather_dim_0_scale_mode_row-wise-sharded (__main__.SymmetricMemoryTest)

  #160049 opened Aug 7, 2025

• Triton kernel generated for torch.compile(create_block_mask) using Flex Attention throws CUDA Illegal memory access

  #160018 opened Aug 6, 2025

• Requesting improved torch.auto_grad.detect_anomaly() for NaN detection

  #160016 opened Aug 6, 2025

• Support NUMA Binding for Callable entrypoints to `elastic_launch`

  #160006 opened Aug 6, 2025

• [BE][download.pytorch.org] Index fix user navigation

  #160005 opened Aug 6, 2025

• `torch._inductor.aoti_compile_and_package` fails with CUDA kernels inside of a `torch.cond`

  #159995 opened Aug 6, 2025

• [DTensor] Decide / Document RNG semantics

  #159991 opened Aug 6, 2025

• @require_world_size(4) macro is ambiguous or buggy

  #159987 opened Aug 6, 2025

• [RFC] Cuda support matrix for Release 2.9

  #159980 opened Aug 6, 2025

• Segmentation fault when using torch.compile with PyTorch 2.8.0 XPU on Intel ARC A770

  #159974 opened Aug 6, 2025

• SAC compatibility with compiled flex attention

  #159970 opened Aug 6, 2025

• Compile error with gcc-14 when building vec_test_all_types

  #159962 opened Aug 6, 2025

• TensorMetadata not retained during DTensor convolution double backpropagation

  #159959 opened Aug 6, 2025

• torch.compile regression on side-effects between torch 2.7.1 and 2.8 (final RC)

  #159958 opened Aug 6, 2025

• CUDA initialization: CUDA unknown error

  #159954 opened Aug 6, 2025

• Add official statically linked libtorch libraries

  #159947 opened Aug 6, 2025

• Wrong backend assigned to Intel Gaudi (HPU) devices in PT 2.8.0

  #159945 opened Aug 6, 2025

• Is is possible for pytorch pipeline parallelism to support dynamic shapes(input/output)

  #159942 opened Aug 6, 2025

• scaled_mm Triton implementation causes wrong results on (at least) H100

  #159940 opened Aug 6, 2025

• contiguous has BUG ?

  #159932 opened Aug 6, 2025

• `torch.onnx.export` fails when exporting custom operator with bfloat16 constant tensor

  #159928 opened Aug 6, 2025

• DISABLED test_fused_all_gather_scaled_matmul_gather_dim_0_scale_mode_row-wise-replicated (__main__.SymmetricMemoryTest)

  #159921 opened Aug 6, 2025

• [export] Tensor subclass decomposition breaks state dict contracts

  #159918 opened Aug 6, 2025

• `test_non_contiguous_input_mm_plus_mm` looks broken on B200

  #159914 opened Aug 5, 2025

• New at::HostAllocator interface prevents using more than one allocator implementation for a device type

  #159906 opened Aug 5, 2025

• UNSTABLE Check mergeability of ghstack PR / ghstack-mergeability-check

  #159899 opened Aug 5, 2025

• UNSTABLE Check Labels / Check labels

  #159894 opened Aug 5, 2025

• torch._dynamo documentation

  #159886 opened Aug 5, 2025

• Implement `grid_sampler_3d` for MPS

  #159882 opened Aug 5, 2025

• `torch.nn.functional.sigmoid` produces inconsistent results on different device for complex inputs

  #159870 opened Aug 5, 2025

• BE: enable auto reformat List to list and similar typing

  #159866 opened Aug 5, 2025

• DISABLED test_upper_bound_i64_cuda (__main__.AOTInductorTestABICompatibleGpu)

  #159860 opened Aug 5, 2025

• [compile] Correctness and reproducibility issue

  #159855 opened Aug 5, 2025

• `torch.cond()` behaves inconsistently when using symbolic predicate

  #159852 opened Aug 5, 2025

• `tree_flatten(some_pytree, lambda x: isinstance(x, Tensor))` yields `[None], _`

  #159848 opened Aug 5, 2025

• [DTensor] nn.Embedding Compile Failure when Creating FX Graph

  #159843 opened Aug 5, 2025

• C++23 forward declaration error :`invalid application of 'sizeof' to an incomplete type 'torch::_export::Graph'`

  #159838 opened Aug 5, 2025

• Missing documentation for device mesh on DDP

  #159836 opened Aug 5, 2025

517 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Introduce Muon optimizer to PyTorch

  #159465 commented on Aug 11, 2025 • 32 new comments

• [OpenReg] Add OSX/Windows Support for OpenReg

  #159441 commented on Aug 11, 2025 • 20 new comments

• Add beginnings of torch::stable::accelerator

  #159679 commented on Aug 12, 2025 • 19 new comments

• [inductor] dont reuse buffers if it affects peak (#145883)

  #159530 commented on Aug 12, 2025 • 15 new comments

• Enable output padding when only outermost dim is dynamic

  #159404 commented on Aug 12, 2025 • 14 new comments

• [Intel GPU] Enable backward for SDPA XPU [WIP]

  #156272 commented on Aug 12, 2025 • 13 new comments

• [DeviceMesh] Add `_unflatten_` api for device mesh to support better UX for some use cases like EP and replicate

  #159482 commented on Aug 6, 2025 • 11 new comments

• Update torch::stable::Tensor() default constructor

  #159507 commented on Aug 11, 2025 • 11 new comments

• [ATen][CPU][Sparse] Use Third-Party Eigen for sparse add and addmm

  #155357 commented on Aug 8, 2025 • 10 new comments

• [PP] Add DualPipeV schedule

  #159591 commented on Aug 10, 2025 • 10 new comments

• [dynamic shapes] unbacked-safe slicing

  #157944 commented on Aug 12, 2025 • 9 new comments

• Add utility to get computed kernel in torch.library

  #158393 commented on Aug 8, 2025 • 9 new comments

• Add support for param mutation under inference mode

  #159661 commented on Aug 12, 2025 • 9 new comments

• [Draft][WIP] Enable XPU path for FlexAttention

  #143553 commented on Aug 8, 2025 • 9 new comments

• Support XPU in --nproc-per-node option to torchrun

  #159474 commented on Aug 11, 2025 • 7 new comments

• Add placeholder for the User Guide

  #159379 commented on Aug 11, 2025 • 6 new comments

• Remove accidental host synchronization in autograd cpu offload

  #159698 commented on Aug 5, 2025 • 6 new comments

• Implement OpenReg device autoload mechanism

  #158555 commented on Aug 11, 2025 • 6 new comments

• [vllm in torch ci ][step 1/3] add build logics

  #159815 commented on Aug 9, 2025 • 6 new comments

• Add support for tracing vmap in pre-dispatch export

  #154650 commented on Aug 7, 2025 • 5 new comments

• [fr] [xpu] Add FlightRecorder support for ProcessGroupXCCL

  #158568 commented on Aug 11, 2025 • 5 new comments

• [CUDA] Add experimental green context support for SM carveout

  #159104 commented on Aug 11, 2025 • 5 new comments

• Add option to assert if kernel is not fully fused in foreach_map

  #159213 commented on Aug 6, 2025 • 5 new comments

• Add `label_smoothing` param in `nn.BCELoss` and `nn.BCEWithLogitsLoss`

  #150282 commented on Aug 12, 2025 • 5 new comments

• Allow exposing more functions during initial template expansion

  #159554 commented on Aug 11, 2025 • 4 new comments

• [Draft][CUDA] Upgrade torch._scaled_grouped_mm to SM100+

  #156806 commented on Aug 7, 2025 • 4 new comments

• Replace setup.py bdist_wheel with python -m build --wheel

  #156712 commented on Aug 11, 2025 • 4 new comments

• Add fallback support for torch.mm in foreach_map_fn

  #159757 commented on Aug 9, 2025 • 4 new comments

• [DTensor] Registers sharding rule for rms_norm

  #159692 commented on Aug 8, 2025 • 4 new comments

• Fix typo in parameter name in cpp.py

  #159716 commented on Aug 11, 2025 • 3 new comments

• grid_sampler_3d for MPS

  #159421 commented on Aug 6, 2025 • 3 new comments

• [Intel GPU] Support SDPA backend selection and priority setting on XPU

  #159464 commented on Aug 12, 2025 • 3 new comments

• Add pad and narrow to torch/csrc/stable/ops.h

  #159328 commented on Aug 12, 2025 • 3 new comments

• Graph split event tracker

  #159795 commented on Aug 11, 2025 • 3 new comments

• [caffe2][IGiOS] Fix exhaustive switches

  #156832 commented on Aug 8, 2025 • 2 new comments

• [inductor] add lowering for repeat_interleave.Tensor with output size specified (#147160)

  #158462 commented on Aug 11, 2025 • 2 new comments

• Replace `std::runtime_error` with `TORCH_CHECK`

  #159344 commented on Aug 8, 2025 • 2 new comments

• Bump transformers pin

  #159291 commented on Aug 12, 2025 • 2 new comments

• [dynamo][guards] Install dict watchers for recrusive dict tag optimization

  #159796 commented on Aug 12, 2025 • 2 new comments

• [Inductor] addmm + activation function fusion

  #158137 commented on Aug 8, 2025 • 2 new comments

• add fp8 scaled_mm for XPU

  #140972 commented on Aug 9, 2025 • 2 new comments

• port 2 distributed pipeline test files for Intel GPU

  #159140 commented on Aug 12, 2025 • 2 new comments

• [Graph Partition] Pass all OSS unit tests

  #154667 commented on Aug 12, 2025 • 2 new comments

• Add torch.compile support for torch.mm(out_dtype=...)

  #159026 commented on Aug 6, 2025 • 2 new comments

• [benchmark] Add HF LLM benchmarks

  #156967 commented on Aug 11, 2025 • 1 new comment

• Add new_empty (with dtype argument only) to torch::stable

  #159508 commented on Aug 12, 2025 • 1 new comment

• [DTensor] add op support: aten.squeeze_.dim

  #159532 commented on Aug 7, 2025 • 1 new comment

• [inductor] skip bmm when converting channel last

  #159459 commented on Aug 5, 2025 • 1 new comment

• Remove usage of fsspec in HF consolidation script

  #159392 commented on Aug 11, 2025 • 1 new comment

• Fix torch.export.export() GPU failure with RNN modules.

  #155734 commented on Aug 10, 2025 • 1 new comment

• [CI][CUDA] Add periodic b200 distributed job

  #159323 commented on Aug 9, 2025 • 1 new comment

• [inductor][cpu] Fix double-offset issue in `GEMM_TEMPLATE`

  #159233 commented on Aug 12, 2025 • 1 new comment

• [dynamo] Add -> bool to functions named (_?)(is|has)_.*

  #155923 commented on Aug 11, 2025 • 1 new comment

• Use device agnostic APIs for RNG

  #159021 commented on Aug 6, 2025 • 1 new comment

• [dict] Implement `__eq__` for dict_items

  #155154 commented on Aug 12, 2025 • 1 new comment

• [inductor] initial triton static config lookup table

  #157699 commented on Aug 11, 2025 • 1 new comment

• [inductor] propagate shapes in CSEVariable

  #152198 commented on Aug 12, 2025 • 1 new comment

• [ARM] Integrate INT4→BF16 via KleidiAI, with fallback

  #158250 commented on Aug 6, 2025 • 1 new comment

• fixes #156701

  #159715 commented on Aug 6, 2025 • 1 new comment

• [cuda][cupy] Improve cupy device placement when device is provided with explicit index

  #158529 commented on Aug 10, 2025 • 1 new comment

• [WIP] Attempt to fix `torch.backends.cudnn.rnn` import

  #159828 commented on Aug 6, 2025 • 1 new comment

• [Flex Attn][CPU] support flash decoding for cpu

  #159835 commented on Aug 6, 2025 • 0 new comments

• [C10d][Gloo] Enable complex datatype support in ProcessGroupGloo

  #156633 commented on Aug 11, 2025 • 0 new comments

• [build] remove upper version pin for `setuptools<80.0`

  #156049 commented on Aug 11, 2025 • 0 new comments

• Fuse matmul

  #157743 commented on Aug 11, 2025 • 0 new comments

• unskipped mobilenet_v3 quantization and mobilenet_v2 quantization plus tests from https://github.com/pytorch/pytorch/issues/125438

  #157786 commented on Aug 7, 2025 • 0 new comments

• [test][do not merge] Upgrade oneDNN to v3.9

  #157994 commented on Aug 5, 2025 • 0 new comments

• [claude-code] Add top-level module doc for torch/distributed/tensor/_op_schema.py

  #157804 commented on Aug 11, 2025 • 0 new comments

• For sdists, replace symlink with copy for docs requirements

  #157811 commented on Aug 11, 2025 • 0 new comments

• Allow docker builds to deal with symlinks

  #157812 commented on Aug 8, 2025 • 0 new comments

• [ZENDNN] Integrate ZenDNN library, implement Linear op, add unit-tests

  #156599 commented on Aug 11, 2025 • 0 new comments

• [dynamo, nested graph breaks] implement new resume frame stack/locals/cell layout convention

  #157971 commented on Aug 11, 2025 • 0 new comments

• [cuDNN][64-bit indexing] update conv depthwise 64bit indexing dispatch condition to match native kernel

  #156140 commented on Aug 6, 2025 • 0 new comments

• docstring_linter: Fix #151692 and other issues

  #156596 commented on Aug 11, 2025 • 0 new comments

• Make functorch notebook symlinks PEP 517 valid

  #157813 commented on Aug 11, 2025 • 0 new comments

• Improve MANIFEST.in for source distribution

  #157814 commented on Aug 11, 2025 • 0 new comments

• [Doc] remove WSL2 in support matrix for Intel GPU

  #156590 commented on Aug 5, 2025 • 0 new comments

• Add PEP 517 compliant Python source distribution to release process

  #157815 commented on Aug 11, 2025 • 0 new comments

• Add functions to setup PrivateUse1 as a python backend device.

  #157859 commented on Aug 5, 2025 • 0 new comments

• [generator] Close all open generators in compile_subgraph

  #157149 commented on Aug 11, 2025 • 0 new comments

• [generator] Raise `StopIteration(value)` with value from the return stmt

  #157152 commented on Aug 12, 2025 • 0 new comments

• [SymmMem] Install NVSHMEM wheel in CI docker

  #157411 commented on Aug 6, 2025 • 0 new comments

• [build] bootstrap git repo for build for non-git-clone archive

  #157432 commented on Aug 10, 2025 • 0 new comments

• [contextlib] Fixes for CPython contextlib tests

  #157148 commented on Aug 8, 2025 • 0 new comments

• [BE][4/5] fix typos in aten/ (aten/src/ATen/native/)

  #157553 commented on Aug 10, 2025 • 0 new comments

• [BE][5/5] fix typos in aten/ (aten/src/ATen/)

  #157554 commented on Aug 10, 2025 • 0 new comments

• [dynamo] [guard] Add caching for inside torch.compile.disable function to avoid unnecessary recompilation.

  #157566 commented on Aug 11, 2025 • 0 new comments

• Introduce a new API torch.accelerator.get_mem_info

  #156812 commented on Aug 9, 2025 • 0 new comments

• [simplefsdp auto-bucketing] ir node runtime estimation

  #157572 commented on Aug 11, 2025 • 0 new comments

• [BE][1/6] fix typos in test/

  #157635 commented on Aug 10, 2025 • 0 new comments

• [BE] add `SHFMT` linter to format shell scripts

  #157685 commented on Aug 10, 2025 • 0 new comments

• [BE][1/4] format shell scripts with `SHFMT`

  #157686 commented on Aug 10, 2025 • 0 new comments

• [CI] Run all torchinductor jobs on MPS

  #156773 commented on Aug 6, 2025 • 0 new comments

• [BE][2/4] format shell scripts with `SHFMT` in .circleci/ and .github/

  #157687 commented on Aug 10, 2025 • 0 new comments

• [BE][3/4] format shell scripts with `SHFMT` in .ci/

  #157688 commented on Aug 10, 2025 • 0 new comments

• [BE][4/4] format shell scripts with `SHFMT` in scripts/

  #157689 commented on Aug 10, 2025 • 0 new comments

• Remove upper pin on setuptools

  #156713 commented on Aug 11, 2025 • 0 new comments

• Replace setup.py install with pip install

  #156711 commented on Aug 11, 2025 • 0 new comments

• Replace setup.py develop with pip install -e

  #156710 commented on Aug 11, 2025 • 0 new comments

• Stop parsing command line arguments every time common_utils is imported.

  #156703 commented on Aug 11, 2025 • 0 new comments

• [inductor] Add return types to functions named (_?)(is|has)_.*

  #155928 commented on Aug 11, 2025 • 0 new comments

• [DTensor] Fix aten.all strategy with min instead of sum as the reduce_op

  #155420 commented on Aug 9, 2025 • 0 new comments

• [Precompile] Integrate PrecompileContext with CompilePackage

  #155384 commented on Aug 9, 2025 • 0 new comments

• [export] inline into torch.jit.traced nn module

  #155381 commented on Aug 8, 2025 • 0 new comments

• Try running test_foreach sequentially

  #155366 commented on Aug 6, 2025 • 0 new comments

• Try adding bfloat16 to test_nn_lstm

  #155338 commented on Aug 5, 2025 • 0 new comments

• turn off reorder_for_peak_memory in case of collectives

  #155271 commented on Aug 8, 2025 • 0 new comments

• [WIP][dynamic shapes] if-then-else for meta_select storage offset

  #155269 commented on Aug 5, 2025 • 0 new comments

• higher_order_ops.py unimplemented_v2 migration, part1

  #155264 commented on Aug 5, 2025 • 0 new comments

• [Dynamo] Add CPython default dict tests

  #155263 commented on Aug 9, 2025 • 0 new comments

• Add more logging

  #155219 commented on Aug 6, 2025 • 0 new comments

• [WIP][fake tensor] invalidate memos for PropagateUnbackedSymInts

  #155187 commented on Aug 9, 2025 • 0 new comments

• [OrderedDict] Implement `OrderedDict.popitem(last=...)`

  #155153 commented on Aug 12, 2025 • 0 new comments

• [OrderedDict] Implement `OrderedDict.move_to_end(key, last=False)`

  #155152 commented on Aug 12, 2025 • 0 new comments

• Convert to markdown: cpp_extension.rst, cpp_index.rst, cpu.rst, cuda_environment_variables.rst, cuda._sanitizer.rst

  #155110 commented on Aug 8, 2025 • 0 new comments

• docs: link to Nvidia Container Toolkit in README

  #155102 commented on Aug 11, 2025 • 0 new comments

• [WIP][dynamic shapes] guard_or_false for are_strides_like_channel_last

  #155076 commented on Aug 10, 2025 • 0 new comments

• [dict] Implement dict.__ior__ and fix return type in dict.__or__

  #155072 commented on Aug 12, 2025 • 0 new comments

• [test] lintrunner thing

  #155062 commented on Aug 5, 2025 • 0 new comments

• [WIP] cast to bf16 before mul op in flex bwd

  #154922 commented on Aug 7, 2025 • 0 new comments

• [ROCm] SDPA fix mem fault when dropout is enabled

  #154864 commented on Aug 11, 2025 • 0 new comments

• Fix DataLoader to Pass List to getitems When Using BatchSampler. Fixes Issue_#154810

  #154844 commented on Aug 6, 2025 • 0 new comments

• Add serialization support for register_constant

  #154834 commented on Aug 8, 2025 • 0 new comments

• Fix in pytorch do_bench_using_profiling

  #154766 commented on Aug 6, 2025 • 0 new comments

• [Wheel Variant] Experimental Support

  #154733 commented on Aug 5, 2025 • 0 new comments

• [vision hash update] update the pinned vision hash

  #154694 commented on Aug 12, 2025 • 0 new comments

• libtorch_cpu.so link libm.so issue:GLIBC_2.29 not found

  #159454 commented on Aug 5, 2025 • 0 new comments

• [BE][c10d/Store]add check in pyi

  #155855 commented on Aug 12, 2025 • 0 new comments

• [ATen MTIA backend] Use aten native CPU fallback function on MTIA

  #155845 commented on Aug 11, 2025 • 0 new comments

• [CD] Move build_magma.bat to build_magma.py

  #155804 commented on Aug 11, 2025 • 0 new comments

• [Do not merge] DCP ZOC Test Changes

  #155802 commented on Aug 11, 2025 • 0 new comments

• [dynamo][guards] Skip dispatch key guards for requires_grad=False

  #155756 commented on Aug 11, 2025 • 0 new comments

• torch.distributed TCP bind address

  #155741 commented on Aug 10, 2025 • 0 new comments

• [MPS] Add regression test for memory leak in nn.MaxPool2d

  #155730 commented on Aug 11, 2025 • 0 new comments

• HOP py_impl register to tensor subclass cannot dispatch

  #155726 commented on Aug 11, 2025 • 0 new comments

• [CUDA][MAGMA][Linalg][WIP] Remove MAGMA

  #155694 commented on Aug 12, 2025 • 0 new comments

• Remove unused nonlocal declarations from checkpoint and library helper functions

  #155686 commented on Aug 11, 2025 • 0 new comments

• DOC: update CrossEntropyLoss with note and example of incorrect target specification

  #155649 commented on Aug 12, 2025 • 0 new comments

• [WIP][PGO] exclude optimizer state from PGO whitelist

  #155643 commented on Aug 10, 2025 • 0 new comments

• [Scripts] Add refresh script to clean, pull and build repo

  #155639 commented on Aug 11, 2025 • 0 new comments

• [dict] Implement dict subclass `fromkeys` classmethod

  #155608 commented on Aug 12, 2025 • 0 new comments

• [DRAFT] Evaluate feasability of using FunctionalTensor for Example Value

  #155606 commented on Aug 6, 2025 • 0 new comments

• Remove unnecessary MPSStream initialization

  #155602 commented on Aug 12, 2025 • 0 new comments

• [do NOT land] DTensor + torch_function_mode + flex_attention dispatch test

  #155600 commented on Aug 11, 2025 • 0 new comments

• [do NOT land] torch_function_mode + flex_attention dispatch test

  #155594 commented on Aug 11, 2025 • 0 new comments

• WIP add support for dynamic shapes

  #155557 commented on Aug 8, 2025 • 0 new comments

• [OrderedDict] Add `bool(OrderedDict)`

  #155503 commented on Aug 12, 2025 • 0 new comments

• [OrderedDict] Set the correct dict class in UserDefinedDictVariable

  #155502 commented on Aug 12, 2025 • 0 new comments

• [OrderedDict] Implement `hasattr(..., IteratorVariable)`

  #155501 commented on Aug 12, 2025 • 0 new comments

• [proxy_tensor] Do not clobber tensor proxies for inplace ops

  #155456 commented on Aug 10, 2025 • 0 new comments

• Quiet Inductor #135521

  #155450 commented on Aug 9, 2025 • 0 new comments

• Clean up memory management in impl_func_norm

  #155432 commented on Aug 12, 2025 • 0 new comments

• [inductor] Improve GEMM loggings

  #155427 commented on Aug 11, 2025 • 0 new comments

• [cuDNN] cuDNN frontend for LayerNorm RMSNorm

  #159682 commented on Aug 6, 2025 • 0 new comments

• [dynamo, nested graph breaks] support nested graph breaks x context managers

  #159678 commented on Aug 9, 2025 • 0 new comments

• [ROCm] Limit number of values per thread for reductions on three dimensions

  #159652 commented on Aug 5, 2025 • 0 new comments

• [OpenReg] Refactor and Bug Fix

  #159640 commented on Aug 8, 2025 • 0 new comments

• Moved Autograd Fallback Interface to Header for Use by Out-of-tree Backends

  #159639 commented on Aug 8, 2025 • 0 new comments

• setup [Do not review]

  #159636 commented on Aug 6, 2025 • 0 new comments

• [WIP] incomplete view unabcked fix to by pass vllm issue

  #159626 commented on Aug 6, 2025 • 0 new comments

• Fallback to contiguous layout in convolution lowering on stride mismatch #159462

  #159593 commented on Aug 5, 2025 • 0 new comments

• [c10d][nvshmem] add nvshmem build rules and dependency for libtorch_cuda

  #159562 commented on Aug 12, 2025 • 0 new comments

• Add dtype checks in meta dispatch for various ordering ops

  #159556 commented on Aug 8, 2025 • 0 new comments

• Move config/util to AllocatorConfig for cross-allocator sharing

  #159553 commented on Aug 11, 2025 • 0 new comments

• Editing and updating glossary to test functionality.

  #159544 commented on Aug 6, 2025 • 0 new comments

• [1/N]Port 3 distributed/_tools test cases to Intel GPU

  #159543 commented on Aug 12, 2025 • 0 new comments

• [ci][inductor dashboard] Remove torchao install as its unused

  #159501 commented on Aug 5, 2025 • 0 new comments

• Add B200 smoke test

  #159494 commented on Aug 5, 2025 • 0 new comments

• Support `next(iterator, default)`

  #159483 commented on Aug 5, 2025 • 0 new comments

• Recursively sync fbgemm submodules before build

  #159477 commented on Aug 5, 2025 • 0 new comments

• [WIP]port sevearl test files under test/distributed to Intel GPU

  #159473 commented on Aug 11, 2025 • 0 new comments

• [ROCm] Use opportunistic fastatomics based on hueristics

  #159430 commented on Aug 8, 2025 • 0 new comments

• Fixes for `collections.Counter`

  #159368 commented on Aug 5, 2025 • 0 new comments

• Fixes for `collections.NamedTuple`

  #159367 commented on Aug 5, 2025 • 0 new comments

• Change mutation type of `MutableMappingVariable` to `AttributeMutationNew`

  #159366 commented on Aug 5, 2025 • 0 new comments

• Enable trace through the collections module

  #159365 commented on Aug 5, 2025 • 0 new comments

• [dynamo] Simplify two methods in ConstDictVariable

  #159361 commented on Aug 11, 2025 • 0 new comments

• Avoid potential deadlocks in host allocator

  #159352 commented on Aug 11, 2025 • 0 new comments

• [dynamo, nested graph breaks] support very simple nested graph breaks

  #159329 commented on Aug 9, 2025 • 0 new comments

• [dynamo, nested graph breaks] use CALL_FUNCTION_EX when calling resume function

  #159281 commented on Aug 9, 2025 • 0 new comments

• [Inductor] support native_layer_norm_backward mixed dtype for privateuse1

  #159830 commented on Aug 5, 2025 • 0 new comments

• DO NOT MERGE, testing sequential builds

  #159827 commented on Aug 12, 2025 • 0 new comments

• Add support for error-ing when there is side effect

  #159826 commented on Aug 5, 2025 • 0 new comments

• [BE][Dynamo] Type improvements in `_dynamo/utils` to generics

  #159824 commented on Aug 11, 2025 • 0 new comments

• ci: Add option for sequential wheel building

  #159821 commented on Aug 11, 2025 • 0 new comments

• [dynamo, nested graph breaks] support nested closures

  #159817 commented on Aug 9, 2025 • 0 new comments

• [Inductor] Freeze matmul args layouts

  #159813 commented on Aug 5, 2025 • 0 new comments

• Replace C array with std::array in formatSockAddr

  #159812 commented on Aug 6, 2025 • 0 new comments

• [inductor] remove no_x_dim

  #159810 commented on Aug 11, 2025 • 0 new comments

• fix type documentation for context_parallel no_restore_buffers, to prevent user from passing in the wrong type

  #159808 commented on Aug 7, 2025 • 0 new comments

• [dynamo] fixes to propagate tag safeness

  #159807 commented on Aug 12, 2025 • 0 new comments

• [146643]Fixed max triton generation

  #159797 commented on Aug 6, 2025 • 0 new comments

• [dynamo, nested graph breaks] prevent excessive recompilations

  #159786 commented on Aug 12, 2025 • 0 new comments

• [Tests]: disable logspace tests correctly

  #159785 commented on Aug 5, 2025 • 0 new comments

• [MTIA Aten Backend] Migrate any.all_out (example diff for tutorial)

  #159780 commented on Aug 7, 2025 • 0 new comments

• [Caffe2] Add float batch box cox SVE128 implementation

  #159778 commented on Aug 11, 2025 • 0 new comments

• [Inductor][Triton] Support TMA before strict 3.4 cutoff

  #159777 commented on Aug 6, 2025 • 0 new comments

• [ROCm] Fix Sliding Window Attention in AOTriton integration code

  #159773 commented on Aug 8, 2025 • 0 new comments

• Add binary size check to validate current limits for binaries released to pypi

  #159768 commented on Aug 5, 2025 • 0 new comments

• [CI] Reduce XPU Windows build time

  #159763 commented on Aug 11, 2025 • 0 new comments

• [FSDP] Add FrozenParamHandle to optimize memory for frozen parameters

  #159751 commented on Aug 9, 2025 • 0 new comments

• AOT graph capture with dynamo.

  #159749 commented on Aug 6, 2025 • 0 new comments

• Build and Install Arm Compute Library in manylinux docker image

  #159737 commented on Aug 7, 2025 • 0 new comments

• Fix GroupNorm(num_groups=1) to match LayerNorm behavior

  #159736 commented on Aug 9, 2025 • 0 new comments

• Use uv run for lintrunner Python deps

  #159735 commented on Aug 11, 2025 • 0 new comments

• [Don't Review] Test XPU CI

  #159718 commented on Aug 5, 2025 • 0 new comments

• dynamo: Remove passing or deleted dynamo_expected_failures

  #159691 commented on Aug 8, 2025 • 0 new comments

• Recheck Autotune cache on Precompile serialization to prune compilation results

  #158656 commented on Aug 5, 2025 • 0 new comments

• Update persons of interest for XLA. The previous one is out of date.

  #158652 commented on Aug 9, 2025 • 0 new comments

• [NumPy] use NumPy 2.x in CI

  #158647 commented on Aug 10, 2025 • 0 new comments

• [OpenReg] Add Develop Notes for Integrating New Backend into PyTorch(Operator Aspect)

  #158644 commented on Aug 8, 2025 • 0 new comments

• unskipped flaky conv2d, rmatmul, and matmul as these now pass

  #158640 commented on Aug 11, 2025 • 0 new comments

• Don't use LLVM libraries

  #158623 commented on Aug 12, 2025 • 0 new comments

• [simplefsdp auto-bucketing] auto bucketing with greedy algorithm

  #158609 commented on Aug 11, 2025 • 0 new comments

• [cuda][complex] Use scaling to compute the absolute value of complex number to avoid overflow

  #158557 commented on Aug 6, 2025 • 0 new comments

• port 3 distributed test to Intel GPU and unified some common functions

  #158533 commented on Aug 12, 2025 • 0 new comments

• Filter out local timer tests which are unimplemented in Python on AArch64

  #158342 commented on Aug 7, 2025 • 0 new comments

• Move XPUEvent to c10

  #158336 commented on Aug 11, 2025 • 0 new comments

• [simplefsdp auto-bucketing] manual bucketing with plan

  #158321 commented on Aug 11, 2025 • 0 new comments

• autograd: Add VJP and JVP rules for aten::aminmax

  #158241 commented on Aug 11, 2025 • 0 new comments

• reuse EventPool::Event in CUDAAllocator

  #158224 commented on Aug 11, 2025 • 0 new comments

• Move EventPool::Event to c10

  #158220 commented on Aug 11, 2025 • 0 new comments

• Move CUDAEvent to c10

  #158219 commented on Aug 11, 2025 • 0 new comments

• Multi-threaded concurrent fetching in Dataloader for high-latency storage.

  #158218 commented on Aug 7, 2025 • 0 new comments

• [scan] cloned aliased input when lowering scan to while_loop

  #158168 commented on Aug 12, 2025 • 0 new comments

• [DTensor] Assert DTensorSpec has valid placements

  #158133 commented on Aug 6, 2025 • 0 new comments

• [build] pin `setuptools>=77` to enable PEP 639

  #158104 commented on Aug 11, 2025 • 0 new comments

• [simplefsdp auto-bucketing] add ir node reorder helper function

  #158098 commented on Aug 11, 2025 • 0 new comments

• [simplefsdp auto-bucketing] add ir node bucket helper function

  #158097 commented on Aug 11, 2025 • 0 new comments

• [inductor] add template hashing for template lookup table

  #158091 commented on Aug 11, 2025 • 0 new comments

• adding types to nn module init

  #158065 commented on Aug 11, 2025 • 0 new comments

• [dict] Support `dict.update()` with no args

  #158061 commented on Aug 12, 2025 • 0 new comments

• Update upstream opinfo to generate appropriately scaled sample inputs

  #158018 commented on Aug 8, 2025 • 0 new comments

• remove unnecessary sync point in AveragedModel update

  #158017 commented on Aug 11, 2025 • 0 new comments

• [Caffe2] Build perfkernels targeting SVE128

  #159274 commented on Aug 11, 2025 • 0 new comments

• add try catch around provenance tracking

  #159266 commented on Aug 6, 2025 • 0 new comments

• [WIP][2/N] Port 5 _composable distributed test to Intel GPU

  #159241 commented on Aug 8, 2025 • 0 new comments

• Remove guard_size_oblivious from default contiguity python check, and add aten.sym_is_contiguous.

  #159197 commented on Aug 9, 2025 • 0 new comments

• [TESTING] Triton pin (Aug 8) 05b2c186c1b6c9a08375389d5efe9cb4c401c075

  #159158 commented on Aug 9, 2025 • 0 new comments

• [WIP][1/N] Port 5 _composable/fsdp distributed test cases to Intel GPU

  #159118 commented on Aug 8, 2025 • 0 new comments

• outer heuristic

  #159093 commented on Aug 12, 2025 • 0 new comments

• port distributed pipeline test files for Intel GPU

  #159033 commented on Aug 12, 2025 • 0 new comments

• [while_loop] support input mutation with auto_functionalize

  #159010 commented on Aug 12, 2025 • 0 new comments

• [cond] support input mutation with auto_functionalize

  #159009 commented on Aug 12, 2025 • 0 new comments

• Docs on export joint with descriptors

  #159006 commented on Aug 12, 2025 • 0 new comments

• [inductor] add lookup table recorder

  #158987 commented on Aug 11, 2025 • 0 new comments

• [BE] create an empty shape_env for check_input_alias_and_mutation_return_outputs

  #158965 commented on Aug 12, 2025 • 0 new comments

• Ensure outer aliasing on DTensor matches inner aliasing

  #158954 commented on Aug 12, 2025 • 0 new comments

• [Caffe2] Import SVE128 PR

  #158932 commented on Aug 11, 2025 • 0 new comments

• [triton_heuristics] Optimize the triton launcher in pt2

  #158897 commented on Aug 7, 2025 • 0 new comments

• [map] support gen_schema for map

  #158884 commented on Aug 12, 2025 • 0 new comments

• [associative_scan] support gen_schema for associative_scan

  #158883 commented on Aug 12, 2025 • 0 new comments

• [CI] Switch ROCm MI300 GitHub Actions workflows from 2-GPU to 1-GPU runners

  #158882 commented on Aug 11, 2025 • 0 new comments

• [scan] support gen_schema for scan

  #158864 commented on Aug 12, 2025 • 0 new comments

• [while_loop] support gen_schema for while_loop

  #158863 commented on Aug 12, 2025 • 0 new comments

• Update RandomSampler docstring. data_source must be Sized not Dataset

  #158857 commented on Aug 7, 2025 • 0 new comments

• [ROCm] Avoid test_conv_backend_cudnn* UTs

  #158817 commented on Aug 8, 2025 • 0 new comments

• Update nullcontext to return input args

  #158776 commented on Aug 8, 2025 • 0 new comments

• Guard rocm_smi.h include with a header check

  #158771 commented on Aug 5, 2025 • 0 new comments

• [ROCm] [CK] Composable Kernel integration for ROCm

  #158747 commented on Aug 8, 2025 • 0 new comments

• [BE] Upgrade XPU support package to 2025.2

  #158733 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_empty_cpu_tensor (__main__.CudaGraphTreeTests)

  #156735 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_graph_break_unsupported_fake (__main__.ReproTests)

  #156629 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_op_has_batch_rule_nn_functional_conv2d_strided_padding_dilation_with_bias_cuda_float32 (__main__.TestVmapOperatorsOpInfoCUDA)

  #157335 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_while_loop_schema_gen (__main__.TestHopSchema)

  #141202 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_allocate_in_thread_to_pool (__main__.TestBlockStateAbsorption)

  #158764 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_rng (__main__.TestCompilerBisector)

  #139590 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_invalid_status_for_legacy_api (__main__.TestCuda)

  #157110 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_dynamic_warmup (__main__.CudaGraphTreeTests)

  #156693 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_get_parameter_dtype (__main__.ReproTests)

  #156598 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_add_sub_alpha_out (__main__.ReproTests)

  #156597 commented on Aug 7, 2025 • 0 new comments

• Inductor codegen for float8 dynamic quantization ops for scaled_grouped_mm backward pass is slow

  #159769 commented on Aug 7, 2025 • 0 new comments

• [BUG]Nan in gradients of scaled_dot_product_attention operation with mem_efficient backend

  #125674 commented on Aug 7, 2025 • 0 new comments

• non-deterministic issue of torch.einsum function on different GPU.

  #137389 commented on Aug 7, 2025 • 0 new comments

• Error : torch/utils/_sympy/interp.py:176] [0/2] failed while executing pow_by_natural([VR1, int_oo], VR[-1, -1]])

  #148003 commented on Aug 8, 2025 • 0 new comments

• [MPS] test_linalg_cholesky fails on M4

  #157364 commented on Aug 8, 2025 • 0 new comments

• Weird dataloader performance degradation caused by torch and numpy import order

  #101188 commented on Aug 8, 2025 • 0 new comments

• KeyError when using fx.split_module

  #155220 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_add_loggers_functions (__main__.TestFXNumericSuiteNShadows)

  #140380 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_op_has_batch_rule_nn_functional_conv2d_strided_padding_dilation_no_bias_cuda_float32 (__main__.TestVmapOperatorsOpInfoCUDA)

  #157315 commented on Aug 7, 2025 • 0 new comments

• `RuntimeError: UR error` with XPU

  #149953 commented on Aug 7, 2025 • 0 new comments

• Mutating a tensor while serializing with safetensors crashes free-threaded PyTorch

  #158071 commented on Aug 7, 2025 • 0 new comments

• extern kernel's get_free_symbols seems incomplete

  #159685 commented on Aug 7, 2025 • 0 new comments

• Support dict as input/output for pipeline parallelism

  #159711 commented on Aug 7, 2025 • 0 new comments

• Poor error message when trying to jit a function instead of a module (RuntimeError: Cannot insert a Tensor that requires grad as a constant.)

  #55282 commented on Aug 7, 2025 • 0 new comments

• foreach CUDA tests flaky on CUDA 12.6+ due to flaky profiler results

  #148681 commented on Aug 7, 2025 • 0 new comments

• SystemError: PY_SSIZE_T_CLEAN macro must be defined for '#' formats - PyTorch compile fails with Python 3.12

  #153737 commented on Aug 7, 2025 • 0 new comments

• Support installing Python bindings in CMake

  #159232 commented on Aug 7, 2025 • 0 new comments

• General MPS op coverage tracking issue

  #77764 commented on Aug 7, 2025 • 0 new comments

• Cannot pass through None for example_inputs in prepare_fx

  #159505 commented on Aug 7, 2025 • 0 new comments

• Feature request: `DataLoader` using multithreading instead of multiprocessing

  #158714 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_triton_barrier (__main__.NVSHMEMTritonTest)

  #158761 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_graph_concurrent_replay (__main__.TestCuda)

  #104055 commented on Aug 7, 2025 • 0 new comments

• Add official support for CUDA sm_120 (RTX 5090 / Blackwell architecture)

  #159207 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_end_recording_early (__main__.CudaGraphTreeTests)

  #156778 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_add_loggers_linear_mod_fp32_quant (__main__.TestFXNumericSuiteNShadows)

  #142860 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_dont_dce_rand (__main__.ReproTests)

  #156580 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_assigning_back_deleter_fns_to_tensor (__main__.TestBlockStateAbsorption)

  #134810 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_aot_autograd_runtime_wrapper_prologue_profiled (__main__.ReproTests)

  #156678 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_error_on_dealloc_use (__main__.CudaGraphTreeTests)

  #156801 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_op_has_batch_rule_nn_functional_conv_transpose1d_cuda_float32 (__main__.TestVmapOperatorsOpInfoCUDA)

  #142566 commented on Aug 8, 2025 • 0 new comments

• Support for _Float16/C++23 std::float16_t

  #157776 commented on Aug 8, 2025 • 0 new comments

• Add nn.GradBank for gradient scaling to prevent vanishing/exploding gradients

  #159765 commented on Aug 8, 2025 • 0 new comments

• RuntimeError: Tried to instantiate dummy base class Stream

  #159744 commented on Aug 8, 2025 • 0 new comments

• DataLoader num_workers > 0 causes CPU memory from parent process to be replicated in all worker processes

  #13246 commented on Aug 8, 2025 • 0 new comments

• [discussion, idea] Batched, vectorized base64 decoding / encoding + maybe RLE decoding / encoding

  #90560 commented on Aug 8, 2025 • 0 new comments

• canUse32BitIndexMath set to False with efficient net

  #155225 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_sort_large_cuda_float16 (__main__.TestSortAndSelectCUDA)

  #159426 commented on Aug 8, 2025 • 0 new comments

• Deterministic implementation for grid_sampler_2d_backward_cuda

  #68959 commented on Aug 8, 2025 • 0 new comments

• Obscure error: Expected a value of type 'List[int]' for argument 'sizes' but instead found type 'immutable_list'

  #122129 commented on Aug 8, 2025 • 0 new comments

• Incorrect hint calculated for expression involving unbacked SymInt

  #130456 commented on Aug 8, 2025 • 0 new comments

• Remove redundant type aliases of _device for torch.Device

  #152952 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_addr_alpha_beta_out (__main__.ReproTests)

  #156641 commented on Aug 8, 2025 • 0 new comments

• Dead link in `torch.compile` docs

  #119272 commented on Aug 8, 2025 • 0 new comments

• [feature request] Inplace downcast dtype conversion

  #158710 commented on Aug 8, 2025 • 0 new comments

• `tensordot` not working for dtype int32 and lower when there is only 1 element in the given axis

  #84530 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_triton_broadcast (__main__.NVSHMEMTritonTest)

  #158908 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_add_loggers_linear_mod_fp32_fp32 (__main__.TestFXNumericSuiteNShadows)

  #159036 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_empty_storage (__main__.CudaGraphTreeTests)

  #156755 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_relative_import (__main__.ReproTests)

  #156679 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_compile_kernel_advanced (__main__.TestCompileKernel)

  #157172 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_relative_import_no_modulename (__main__.ReproTests)

  #156691 commented on Aug 8, 2025 • 0 new comments

• UNSTABLE rocm-mi300 / linux-noble-rocm-py3.12-mi300 / test (default)

  #158182 commented on Aug 8, 2025 • 0 new comments

• UNSTABLE inductor-rocm-mi300 / rocm-py3.10-inductor-mi300 / test (inductor)

  #154884 commented on Aug 8, 2025 • 0 new comments

• Fix for special.zeta nan handling - follow-up PR #138653

  #146618 commented on Aug 8, 2025 • 0 new comments

• DISABLED test_function_compiled_multiple_times (__main__.CudaGraphTreeTests)

  #157143 commented on Aug 5, 2025 • 0 new comments

• DISABLED test_name_match (__main__.TestGuardSerialization)

  #156246 commented on Aug 5, 2025 • 0 new comments

• [BE] linter to detect unused docker images

  #158783 commented on Aug 5, 2025 • 0 new comments

• ModuleDict subscription no longer works after compile().

  #159831 commented on Aug 5, 2025 • 0 new comments

• `torch.load` can't deserialize `datetime` objects, even with the appropriate `safe_globals`

  #152985 commented on Aug 5, 2025 • 0 new comments

• DISABLED test_add_loggers_conv_bn_relu_fusion_fp32 (__main__.TestFXNumericSuiteNShadows)

  #158762 commented on Aug 6, 2025 • 0 new comments

• DISABLED test_zero_bubble_with_model_kwargs_ScheduleClass1 (__main__.ScheduleTest)

  #154579 commented on Aug 6, 2025 • 0 new comments

• DISABLED test_graph_partition (__main__.CudaGraphTreeTests)

  #157173 commented on Aug 6, 2025 • 0 new comments

• DISABLED test_graph_partition_cpu_op_and_dynamic_shapes (__main__.CudaGraphTreeTests)

  #157258 commented on Aug 6, 2025 • 0 new comments

• DISABLED test_op_has_batch_rule_nn_functional_conv2d_cuda_float32 (__main__.TestVmapOperatorsOpInfoCUDA)

  #157278 commented on Aug 6, 2025 • 0 new comments

• Inconsistent Error Message for Cross-Device Input in torch.compile

  #159133 commented on Aug 6, 2025 • 0 new comments

• CMake Error: When installing PyTorch from source, CUDA not being detected.

  #134331 commented on Aug 6, 2025 • 0 new comments

• native_layer_norm_backward supports mixed precision for PrivateUse1

  #159829 commented on Aug 6, 2025 • 0 new comments

• User Triton Kernels Are not Serialized in Fx Graph Runnable

  #153475 commented on Aug 6, 2025 • 0 new comments

• [RFC]: PyTorch Low-Precision GEMMs Public API

  #157950 commented on Aug 6, 2025 • 0 new comments

• Add mean and var operation for Nested Tensors

  #138831 commented on Aug 6, 2025 • 0 new comments

• foreach_map enhancements

  #158968 commented on Aug 5, 2025 • 0 new comments

• torch compile produces nans with GQA

  #159469 commented on Aug 5, 2025 • 0 new comments

• [Feat] Tracking for OpenReg Improvements

  #158917 commented on Aug 5, 2025 • 0 new comments

• ONNX export via Dynamo sets `dft_length = 1` in `DFT`, breaking shape-inference for `torch.fft.rfft`

  #155997 commented on Aug 5, 2025 • 0 new comments

• mmap fails on 64k page aarch64 systems for AOTI model loading

  #145610 commented on Aug 5, 2025 • 0 new comments

• DISABLED test_conv2d_api (__main__.TestQuantizedFunctionalOps)

  #157346 commented on Aug 5, 2025 • 0 new comments

• DISABLED test_op_has_batch_rule_mv_cuda_float32 (__main__.TestVmapOperatorsOpInfoCUDA)

  #142697 commented on Aug 5, 2025 • 0 new comments

• DISABLED test_mempool_limited_memory_with_allocator (__main__.TestMemPool)

  #157256 commented on Aug 5, 2025 • 0 new comments

• DDP+TP composition does not work as expected

  #157445 commented on Aug 5, 2025 • 0 new comments

• Tracking issue: Incorrect Meta Strides / Turn On PyDispatcher in FakeTensor Mode

  #145094 commented on Aug 5, 2025 • 0 new comments

• Experimental Wheel Variant Support - Technical Discussion

  #155141 commented on Aug 5, 2025 • 0 new comments

• Compilation of the post-training quantized model using Nvidia ModelOpt is failing with the error: Unsupported — 'inline in skipfiles: QuantLinearConvBase.quantize_weight

  #151450 commented on Aug 5, 2025 • 0 new comments

• torch.empty should consider np.bool while parsing args

  #159739 commented on Aug 5, 2025 • 0 new comments

• Inductor doesn't fuse outer dimension softmax into a single kernel.

  #93718 commented on Aug 5, 2025 • 0 new comments

• [NJT] can only chunk if the 2nd dimension is ragged

  #153238 commented on Aug 5, 2025 • 0 new comments

• DISABLED test_dont_aggressively_write_assert (__main__.ReproTests)

  #156570 commented on Aug 5, 2025 • 0 new comments

• DISABLED test_graph_partition_cpu_scalar1 (__main__.CudaGraphTreeTests)

  #157280 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_vmap_exhaustive_addmv_cuda_float32 (__main__.TestVmapOperatorsOpInfoCUDA)

  #157617 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_call_count_tunableop_cuda_float32 (__main__.TestLinalgCUDA)

  #155953 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_reentrant_parent_error_on_cpu_cuda (__main__.TestAutogradDeviceTypeCUDA)

  #86735 commented on Aug 7, 2025 • 0 new comments

• NotImplementedError: Could not run 'aten::q_scale' with arguments from the 'CPU'

  #159743 commented on Aug 7, 2025 • 0 new comments

• Bug: `torch.compile` triggers C++ compile error due to conflicting declaration in generated `.cpp` code

  #159245 commented on Aug 7, 2025 • 0 new comments

• [DTensor][FSDP2][DDP] benchmark dtensor cpu overhead for adam optimizer

  #159169 commented on Aug 7, 2025 • 0 new comments

• [Flex Attention] Accuracy issue with kv length not multiple of kv block size

  #159247 commented on Aug 7, 2025 • 0 new comments

• torch.export with nn.Transformer creates a non-contiguous memory tensor for aten.view

  #159126 commented on Aug 7, 2025 • 0 new comments

• Torch profiler corrupted names with Python 3.11

  #121219 commented on Aug 7, 2025 • 0 new comments

• Segmentation fault with ITIMER_REAL

  #57185 commented on Aug 7, 2025 • 0 new comments

• Foreach Where support

  #117884 commented on Aug 7, 2025 • 0 new comments

• get_ema_multi_avg_fn() equation is a little confused

  #155551 commented on Aug 7, 2025 • 0 new comments

• Autogenerate code example / tutorial outputs in documentation

  #6662 commented on Aug 7, 2025 • 0 new comments

• Is DTensor support dynamic shapes while using torch.compile ?

  #159635 commented on Aug 7, 2025 • 0 new comments

• Einsum of 2 dtensors fails in inference mode

  #157631 commented on Aug 7, 2025 • 0 new comments

• Enable CUDA 12.9 binaries

  #155196 commented on Aug 6, 2025 • 0 new comments

• extract statistics from attention weights in FlexAttention

  #159770 commented on Aug 6, 2025 • 0 new comments

• Enable CI and Build Support for PyTorch on PPC64LE Architecture

  #141235 commented on Aug 6, 2025 • 0 new comments

• Transition Adam and maybe other optimizers to fused code path by default to avoid `foreach=True`-specific VRAM peak due to temp TensorList for bias-corrected moments

  #158371 commented on Aug 6, 2025 • 0 new comments

• Different behavior between sparse and dense tensors with broadcasting multiplication.

  #158861 commented on Aug 6, 2025 • 0 new comments

• Performance bug on `inplace` of `nn.ELU`

  #159622 commented on Aug 6, 2025 • 0 new comments

• cuda memory error thrown by torch.

  #150048 commented on Aug 6, 2025 • 0 new comments

• which Pythorch suports "TORCH_USE_CUDA_DSA=1" from the shell environment?

  #121969 commented on Aug 6, 2025 • 0 new comments

• [dynamo] torch.randint_like on DTensor does not work with compile

  #156649 commented on Aug 6, 2025 • 0 new comments

• [AOTI] Severe Performance Regression with FP16 Autocast in AOTInductor for Small Batch Sizes

  #159346 commented on Aug 6, 2025 • 0 new comments

• Label tracking meta-issue (edit me to get automatically CC'ed on issues! cc bot)

  #24422 commented on Aug 6, 2025 • 0 new comments

• DISABLED test_module_and_optimizer_ids (__main__.TestTorchTidyProfiler)

  #87581 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_add_loggers_conv_bn_relu_fusion_quant (__main__.TestFXNumericSuiteNShadows)

  #127814 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_comprehensive_pca_lowrank_cuda_float32 (__main__.TestInductorOpInfoCUDA)

  #139828 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_dtensor_seq_par_shard_dim_1 (__main__.MicroPipelineTPTest)

  #153223 commented on Aug 7, 2025 • 0 new comments

• DISABLED test_triton_alltoall (__main__.NVSHMEMTritonTest)

  #158840 commented on Aug 7, 2025 • 0 new comments

• [dynamo, nested graph breaks] add nested graph break tests

  #144516 commented on Aug 9, 2025 • 0 new comments

• Replacing explicit backend search with api call

  #144944 commented on Aug 9, 2025 • 0 new comments

• Support contextlib.ExitStack

  #146506 commented on Aug 12, 2025 • 0 new comments

• Enable explicitly vectorized `_weight_int8pack_mm` op for FP16 dtype on x86_64 CPU

  #146777 commented on Aug 5, 2025 • 0 new comments

• [DO NOT MERGE][Inductor] Migrate from oneDNN Inner Product to oneDNN MatMul for mkldnn._linear_pointwise and mkldnn._linear_pointwise.binary

  #147360 commented on Aug 12, 2025 • 0 new comments

• [ONNX] Migrate onnx ops decomp functions

  #147469 commented on Aug 6, 2025 • 0 new comments

• [test] check labels

  #147470 commented on Aug 5, 2025 • 0 new comments

• Support `contextlib.suppress`

  #147990 commented on Aug 11, 2025 • 0 new comments

• [pytree] add another simplified pytree module `torch.pytree`

  #148180 commented on Aug 8, 2025 • 0 new comments

• [pytree] simplify public API exposition with `__module__`

  #148328 commented on Aug 8, 2025 • 0 new comments

• [BE][pytree] rename `NodeDef` member to match the type annotations: `*_fn -> *_func`

  #148474 commented on Aug 8, 2025 • 0 new comments

• [BE][pytree] rename argument name in register function to match the type annotations: `*_fn -> *_func`

  #148484 commented on Aug 8, 2025 • 0 new comments

• [triton hash update] update the pinned triton hash

  #148492 commented on Aug 12, 2025 • 0 new comments

• Remove shebang line from easy_install generated python scripts on Windows only

  #148673 commented on Aug 5, 2025 • 0 new comments

• Support int step for nonfused optimizer

  #148956 commented on Aug 8, 2025 • 0 new comments

• Update the heuristic for AArch64 bmm/baddbmm

  #149122 commented on Aug 8, 2025 • 0 new comments

• TorchInductor CPU Performance Dashboard

  #93531 commented on Aug 12, 2025 • 0 new comments

• [inductor] [cpu] `torch.nn.Embedding-torch.index_copy` outputs inconsistent results on cpu inductor

  #156786 commented on Aug 12, 2025 • 0 new comments

• Automated submodule update: kineto

  #106149 commented on Aug 11, 2025 • 0 new comments

• Automated submodule update: FBGEMM

  #115316 commented on Aug 12, 2025 • 0 new comments

• [AOTAutograd] tweak min-cut partitioner to avoid saving softmax output

  #126348 commented on Aug 6, 2025 • 0 new comments

• Remove deprecated torch/csrc/jit/codegen/cuda

  #131296 commented on Aug 11, 2025 • 0 new comments

• Add decompositions for median and nonmedian

  #134881 commented on Aug 7, 2025 • 0 new comments

• [pytree] Add public pytree module `torch.utils.pytree`

  #137400 commented on Aug 8, 2025 • 0 new comments

• Add TORCH_CHECK_INDEX in convert_indices_from_coo_to_csr_cpu

  #138068 commented on Aug 6, 2025 • 0 new comments

• Extending SVE VEC Backend Support in PyTorch to SVE128 and SVE512.

  #138388 commented on Aug 11, 2025 • 0 new comments

• Fix `USE_STATIC_MKL` lost functionality

  #138996 commented on Aug 11, 2025 • 0 new comments

• [Don't Review] Test CI

  #139971 commented on Aug 9, 2025 • 0 new comments

• Using acc_t for log_softmax

  #143896 commented on Aug 5, 2025 • 0 new comments

• Defaults to C++20 for torch targets

  #143959 commented on Aug 11, 2025 • 0 new comments

• [ci] Add riscv opt-int build

  #143979 commented on Aug 8, 2025 • 0 new comments

• Full static typing for `torch.distributions`

  #144219 commented on Aug 11, 2025 • 0 new comments

• [pytree] make `tree_*` functions accept both Python and C++ `PyTreeSpec`

  #152624 commented on Aug 8, 2025 • 0 new comments

• [CUTLASS][WIP] Gate rowwise matmul CUTLASS kernels by compute capability

  #152642 commented on Aug 9, 2025 • 0 new comments

• [invoke_subgraph] Force the output stride to be same as eager

  #152806 commented on Aug 5, 2025 • 0 new comments

• [MSVC] Enable updated lambda processor by setting compiler flag /Zc:lambda globally

  #152828 commented on Aug 5, 2025 • 0 new comments

• [AOTI Debugging] Add Environment Variable to control output path

  #153391 commented on Aug 11, 2025 • 0 new comments

• [AUTOCAST] FEAT: Allow passing a `torch.device` object to autocast

  #153539 commented on Aug 9, 2025 • 0 new comments

• Updates contextlib with ParamSpec

  #153623 commented on Aug 6, 2025 • 0 new comments

• Add TORCH_CHECK for group < channels for native_channel_shuffle

  #153781 commented on Aug 5, 2025 • 0 new comments

• Fix `LLONG_MIN` errors in `torch.jit.script`

  #153793 commented on Aug 5, 2025 • 0 new comments

• [Dynamo] Fixes for exceptions

  #153966 commented on Aug 11, 2025 • 0 new comments

• [cond] support gen_schema for cond

  #154193 commented on Aug 12, 2025 • 0 new comments

• Revert D74898941 (#154188)

  #154203 commented on Aug 8, 2025 • 0 new comments

• Updated padding validation in max_pool functions to account for dilation

  #154395 commented on Aug 10, 2025 • 0 new comments

• [Inductor] Fix remove_noop_ops pass where the types for the same_meta would differ

  #154460 commented on Aug 6, 2025 • 0 new comments

• [WIP][export][cond] support exporting cond with unbacked symint shaped tensor

  #154570 commented on Aug 5, 2025 • 0 new comments

• Use official CUDAToolkit module in CMake

  #154595 commented on Aug 11, 2025 • 0 new comments

• Fix `SequentialLR` deprecate warning about invoke `step(epoch)`

  #149392 commented on Aug 8, 2025 • 0 new comments

• Introduce test skip markers for Sandcastle

  #150934 commented on Aug 5, 2025 • 0 new comments

• [hop] Make base_hop share utils with control flow ops in backward

  #151146 commented on Aug 5, 2025 • 0 new comments

• [WIP] Generalize device caching allocator

  #151298 commented on Aug 8, 2025 • 0 new comments

• Add inductor backend to device interface; make minifier_tests more device agnostic

  #151314 commented on Aug 6, 2025 • 0 new comments

• Fix skipIfXpu and skipIfHpu disables tests when used on class

  #151315 commented on Aug 5, 2025 • 0 new comments

• AMD/ROCm OCP Micro-scaling Format (mx-fp8/mx-fp4) Support

  #151360 commented on Aug 11, 2025 • 0 new comments

• Add is_pinned to host allocator

  #151439 commented on Aug 10, 2025 • 0 new comments

• Allow to byteswap data when reading saved torch jit data

  #151447 commented on Aug 7, 2025 • 0 new comments

• [WIP] Deprecate getPinnedMemoryAllocator use getHostAllocator instead

  #151531 commented on Aug 10, 2025 • 0 new comments

• GEMM-template Horizontal

  #151780 commented on Aug 7, 2025 • 0 new comments

• [WIP] Deprecate AcceleratorHooksInterface isPinnedPtr, use at::getHostAllocator()->is_pinned instead

  #151916 commented on Aug 10, 2025 • 0 new comments

• Expand cache logging

  #152026 commented on Aug 10, 2025 • 0 new comments

• Work around MPSGraph issue in backward pass of nn.ReplicationPad1d/2d

  #152094 commented on Aug 11, 2025 • 0 new comments

• [jit] DeadCodeEliminator Mark(block) improvement

  #152348 commented on Aug 10, 2025 • 0 new comments

• Build libgomp (gcc-13) from src on AArch64

  #152361 commented on Aug 7, 2025 • 0 new comments

• Bitwise-perfect method for (de)serializing tensors in base64

  #93859 commented on Aug 10, 2025 • 0 new comments

• DISABLED test_op_has_batch_rule_nn_functional_conv_transpose2d_cuda_float32 (__main__.TestVmapOperatorsOpInfoCUDA)

  #157368 commented on Aug 11, 2025 • 0 new comments

• [RFC] Graph generalization

  #158827 commented on Aug 11, 2025 • 0 new comments

• [export] Unable to trace ops like min/pow

  #148389 commented on Aug 11, 2025 • 0 new comments

• DISABLED test_add_loggers_linear_mod_quant_fp32 (__main__.TestFXNumericSuiteNShadows)

  #159152 commented on Aug 11, 2025 • 0 new comments

• DISABLED test_graph_partition_cpu_scalar3 (__main__.CudaGraphTreeTests)

  #157339 commented on Aug 11, 2025 • 0 new comments

• DISABLED test_graph_partition_cpu_scalar2 (__main__.CudaGraphTreeTests)

  #157312 commented on Aug 11, 2025 • 0 new comments

• [ARM] multiple test failures in TestQuantizedConv on Aarch64

  #144770 commented on Aug 11, 2025 • 0 new comments

• Unable to Specify CUDA Stream for Collective Operations Using with torch.cuda.stream() context

  #136187 commented on Aug 11, 2025 • 0 new comments

• Allow slicing of Nested Tensors along constant dimensions

  #108567 commented on Aug 11, 2025 • 0 new comments

• torch.nn.InstanceNorm2d throws "mixed dtype" error with track_running_stats set to True

  #139140 commented on Aug 11, 2025 • 0 new comments

• Data corruption when reading data as CUDA tensor from a different process

  #134273 commented on Aug 11, 2025 • 0 new comments

• Update epsilon logic to improve numerical stability

  #151110 commented on Aug 11, 2025 • 0 new comments

• [Doc] [Win] libuv installation doc is not correct.

  #148315 commented on Aug 11, 2025 • 0 new comments

• [ONNX] broadcast_in_dim: model (ReDimNet)

  #138313 commented on Aug 11, 2025 • 0 new comments

• Most requested ops for the MPS backend

  #154052 commented on Aug 11, 2025 • 0 new comments

• Support for non-scalar valued Loss Tensor in `_export_forward_backward()`

  #159316 commented on Aug 9, 2025 • 0 new comments

• Gross mismatch in PDF between CUDA and CPU for multivariate Gaussian mixture models

  #156959 commented on Aug 9, 2025 • 0 new comments

• There is a performance drop because we have not yet implemented the batching rule for aten::_scaled_dot_product_efficient_attention_backward

  #117016 commented on Aug 9, 2025 • 0 new comments

• Performance issue on Windows with a "benchmark" comparing to Linux and WLS

  #87692 commented on Aug 9, 2025 • 0 new comments

• Get https://github.com/pytorch/benchmark working

  #87697 commented on Aug 9, 2025 • 0 new comments

• Compiling attention (SDPA) with nested tensors fails when using DDP

  #152068 commented on Aug 9, 2025 • 0 new comments

• Cannot configure dist_timeout when using device_mesh

  #119574 commented on Aug 9, 2025 • 0 new comments

• Performance regression: torch.jit.trace() significantly slower on RTX 5090 than RTX 4060 (cu128 nightly)

  #159238 commented on Aug 9, 2025 • 0 new comments

• MPS Sparse Support

  #129842 commented on Aug 9, 2025 • 0 new comments

• [feature request] Native checkpointing to/from `s3://` (DCP / `torch.load` / `torch.save`)

  #155992 commented on Aug 9, 2025 • 0 new comments

• PyTorch 2.8.0 exposes statically linked `libstdc++` CXX11 ABI symbols.

  #133437 commented on Aug 10, 2025 • 0 new comments

• strange error when distributed training

  #150081 commented on Aug 10, 2025 • 0 new comments

• torch_function objects passed as non-Tensor args should trigger overrides

  #119194 commented on Aug 10, 2025 • 0 new comments

• State of Torch Named Tensors

  #60832 commented on Aug 10, 2025 • 0 new comments

• Activation Checkpointing dtype Mismatch in Recomputed Activations Due to Skipped Type Casting in FSDPv2 _pre_forward

  #159359 commented on Aug 10, 2025 • 0 new comments

• Torch Compile. CudaGraphs Memory leak

  #159669 commented on Aug 10, 2025 • 0 new comments

• Triton 3.5 / PyTorch 2.9 Pin Update Tracker

  #159704 commented on Aug 12, 2025 • 0 new comments

• DISABLED test_remove_noop_slice1_cpu (__main__.CpuTests)

  #151379 commented on Aug 12, 2025 • 0 new comments

• DISABLED test_remove_noop_slice_scatter_cpu (__main__.CpuTests)

  #151382 commented on Aug 12, 2025 • 0 new comments

• DISABLED test_cdist_large_batch_cpu (__main__.TestTorchDeviceTypeCPU)

  #158909 commented on Aug 12, 2025 • 0 new comments

• DISABLED test_gru (__main__.TestXNNPACKQuantizer)

  #158116 commented on Aug 12, 2025 • 0 new comments

• DISABLED test_execution_into_recording (__main__.CudaGraphTreeTests)

  #156838 commented on Aug 12, 2025 • 0 new comments

• DISABLED test_remove_noop_slice_scatter_cuda (__main__.GPUTests)

  #151378 commented on Aug 12, 2025 • 0 new comments

• DISABLED test_remove_noop_view_default_cuda (__main__.GPUTests)

  #151511 commented on Aug 12, 2025 • 0 new comments

• DISABLED test_remove_noop_slice_cuda (__main__.GPUTests)

  #151383 commented on Aug 12, 2025 • 0 new comments

• DISABLED test_remove_noop_slice1_cuda (__main__.GPUTests)

  #151381 commented on Aug 12, 2025 • 0 new comments

• DISABLED test_error_on_dealloc_use2 (__main__.CudaGraphTreeTests)

  #156808 commented on Aug 12, 2025 • 0 new comments

• DISABLED test_remove_noop_slice_cpu (__main__.CpuTests)

  #151384 commented on Aug 12, 2025 • 0 new comments

• DISABLED test_against_reference_multi_input_jacfwd_cuda (__main__.TestJacCUDA)

  #156998 commented on Aug 12, 2025 • 0 new comments

• DISABLED test_op_has_batch_rule___rmatmul___cuda_float32 (__main__.TestVmapOperatorsOpInfoCUDA)

  #157003 commented on Aug 12, 2025 • 0 new comments

• DISABLED test_op_has_batch_rule_nn_functional_conv_transpose3d_cuda_float32 (__main__.TestVmapOperatorsOpInfoCUDA)

  #82340 commented on Aug 12, 2025 • 0 new comments

• Enable CUDA 13.0 binaries

  #159779 commented on Aug 12, 2025 • 0 new comments

• Immutable assignment (akin to `array.at[...]` in JAX)

  #159784 commented on Aug 11, 2025 • 0 new comments

• Re-enable Low Memory Dropout

  #102319 commented on Aug 11, 2025 • 0 new comments

• dataloader hits CUDA error: invalid argument in data.pin_memory(device) on RTX 3090

  #159447 commented on Aug 11, 2025 • 0 new comments

• RandomSampler docs wrongly states type for data_source is Dataset, when it is Sized

  #158631 commented on Aug 11, 2025 • 0 new comments

• Execution of an `ExportProgram` of a model with `torch.autograd.grad(torch.sqrt(x), x, torch.ones_like(torch.sqrt(x)))` returns a `FakeTensor` instead of a real tensor

  #155044 commented on Aug 11, 2025 • 0 new comments

• DISABLED test_resnet (__main__.TestBlockStateAbsorption)

  #157725 commented on Aug 11, 2025 • 0 new comments

• DISABLED test_bitwise_print_precedence (__main__.ReproTests)

  #156736 commented on Aug 11, 2025 • 0 new comments

• DTensor Compile w/ Dynamic Shapes Autograd - Unhashable SymInt in sharding propagation when inputs have requires_grad=True

  #159590 commented on Aug 11, 2025 • 0 new comments

• Tensor roll with different shifts

  #125422 commented on Aug 11, 2025 • 0 new comments

• Fused Linear and Cross-Entropy Loss `torch.nn.functional.linear_cross_entropy`

  #124480 commented on Aug 11, 2025 • 0 new comments

• Parallel Associative Scan

  #95408 commented on Aug 11, 2025 • 0 new comments

• Wrong meta function for constant_pad_nd

  #144187 commented on Aug 11, 2025 • 0 new comments

• C10d Elastic Training master_addr ERROR

  #74824 commented on Aug 11, 2025 • 0 new comments

• torch.export seems to emit invalid code for Tensor.split when used with meta device

  #154721 commented on Aug 11, 2025 • 0 new comments

• MPS operator coverage tracking issue (2.6+ version)

  #141287 commented on Aug 11, 2025 • 0 new comments

• [vulkan] Vulkan backend fails creating tensor on x86_64 Linux

  #72775 commented on Aug 11, 2025 • 0 new comments