From 344c106a6961b50f734b95f084fbf02057d4b475 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 12 Mar 2025 04:11:42 -0400 Subject: [PATCH 1/7] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 74 ++++++++++++++++++++++++++++++++++++++++-- vendor/llama.cpp | 2 +- 2 files changed, 72 insertions(+), 4 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 205d89a0b..f3985ad2f 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -227,6 +227,7 @@ # LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26, # LLAMA_VOCAB_PRE_TYPE_MINERVA = 27, # LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28, +# LLAMA_VOCAB_PRE_TYPE_GPT4O = 29, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -257,6 +258,7 @@ LLAMA_VOCAB_PRE_TYPE_CHAMELEON = 26 LLAMA_VOCAB_PRE_TYPE_MINERVA = 27 LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28 +LLAMA_VOCAB_PRE_TYPE_GPT4O = 29 # // note: these values should be synchronized with ggml_rope @@ -1357,6 +1359,12 @@ def llama_model_n_head(model: llama_model_p, /) -> int: ... +# LLAMA_API int32_t llama_model_n_head_kv (const struct llama_model * model); +@ctypes_function("llama_model_n_head_kv", [llama_model_p_ctypes], ctypes.c_int32) +def llama_model_n_head_kv(model: llama_model_p, /) -> int: + ... + + # // Get the model's RoPE frequency scaling factor # LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model); @ctypes_function("llama_model_rope_freq_scale_train", [llama_model_p_ctypes], ctypes.c_float) @@ -3375,8 +3383,8 @@ class llama_sampler_i(ctypes.Structure): # struct llama_sampler { -# struct llama_sampler_i * iface; -# llama_sampler_context_t ctx; +# const struct llama_sampler_i * iface; +# llama_sampler_context_t ctx; # }; class llama_sampler(ctypes.Structure): _fields_ = [ @@ -3410,6 +3418,18 @@ class llama_sampler(ctypes.Structure): # // mirror of llama_sampler_i: +# LLAMA_API struct llama_sampler * llama_sampler_init (const struct llama_sampler_i * iface, llama_sampler_context_t ctx); +@ctypes_function( + "llama_sampler_init", + [ctypes.POINTER(llama_sampler_i), llama_sampler_context_t], + llama_sampler_p_ctypes, +) +def llama_sampler_init( + iface: ctypes.POINTER(llama_sampler_i), ctx: llama_sampler_context_t, / +) -> llama_sampler_p: + ... + + # LLAMA_API const char * llama_sampler_name (const struct llama_sampler * smpl); @ctypes_function( "llama_sampler_name", @@ -3627,6 +3647,17 @@ def llama_sampler_init_xtc( ... +# /// @details Top n sigma sampling as described in academic paper "Top-nσ: Not All Logits Are You Need" https://arxiv.org/pdf/2411.07641 +# LLAMA_API struct llama_sampler * llama_sampler_init_top_n_sigma(float n); +@ctypes_function( + "llama_sampler_init_top_n_sigma", + [ctypes.c_float], + llama_sampler_p_ctypes, +) +def llama_sampler_init_top_n_sigma(n: float, /) -> llama_sampler_p: + ... + + # /// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words. # /// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text. # /// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. @@ -3685,6 +3716,43 @@ def llama_sampler_init_grammar( ... +# /// @details Lazy grammar sampler, introduced in https://github.com/ggml-org/llama.cpp/pull/9639 +# /// @param trigger_patterns A list of patterns that will trigger the grammar sampler. Pattern will be matched from the start of the generation output, and grammar sampler will be fed content starting from its first match group. +# /// @param trigger_tokens A list of tokens that will trigger the grammar sampler. Grammar sampler will be fed content starting from the trigger token included. +# LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy_patterns( +# const struct llama_vocab * vocab, +# const char * grammar_str, +# const char * grammar_root, +# const char ** trigger_patterns, +# size_t num_trigger_patterns, +# const llama_token * trigger_tokens, +# size_t num_trigger_tokens); +@ctypes_function( + "llama_sampler_init_grammar_lazy_patterns", + [ + llama_vocab_p_ctypes, + ctypes.c_char_p, + ctypes.c_char_p, + ctypes.POINTER(ctypes.c_char_p), + ctypes.c_size_t, + ctypes.POINTER(llama_token), + ctypes.c_size_t, + ], + llama_sampler_p_ctypes, +) +def llama_sampler_init_grammar_lazy_patterns( + vocab: llama_vocab_p, + grammar_str: bytes, + grammar_root: bytes, + trigger_patterns: CtypesArray[bytes], + num_trigger_patterns: int, + trigger_tokens: CtypesArray[llama_token], + num_trigger_tokens: int, + /, +) -> llama_sampler_p: + ... + + # /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first. # LLAMA_API struct llama_sampler * llama_sampler_init_penalties( # int32_t penalty_last_n, // last n tokens to penalize (0 = disable penalty, -1 = context size) @@ -3737,7 +3805,7 @@ def llama_sampler_init_dry( dry_base: float, dry_allowed_length: int, dry_penalty_last_n: int, - seq_breakers: CtypesArray[bytes], + seq_breakers, num_breakers: int, /, ) -> llama_sampler_p: diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 794fe23f2..2c9f833d1 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 794fe23f29fb40104975c91fe19f23798f7c726e +Subproject commit 2c9f833d17bb5b8ea89dec663b072b5420fc5438 From e232fae1a35c0bdaabf230a5d71780c0178e5d8f Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 12 Mar 2025 04:44:18 -0400 Subject: [PATCH 2/7] feat: Update llama.cpp --- vendor/llama.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 2c9f833d1..7841fc723 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 2c9f833d17bb5b8ea89dec663b072b5420fc5438 +Subproject commit 7841fc723e059d1fd9640e5c0ef19050fcc7c698 From 37eb5f0a4c2a8706b89ead1406b1577c4602cdec Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Wed, 12 Mar 2025 05:30:21 -0400 Subject: [PATCH 3/7] chore: Bump version --- CHANGELOG.md | 4 ++++ llama_cpp/__init__.py | 2 +- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 53365e368..605370e7d 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.3.8] + +- feat: Update llama.cpp to ggerganov/llama.cpp@7841fc723e059d1fd9640e5c0ef19050fcc7c698 + ## [0.3.7] - feat: Update llama.cpp to ggerganov/llama.cpp@794fe23f29fb40104975c91fe19f23798f7c726e diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index fc1fcbcf6..b1a8b9baa 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.3.7" +__version__ = "0.3.8" From 99f2ebfde18912adeb7f714b49c1ddb624df3087 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Fri, 11 Apr 2025 03:28:20 -0400 Subject: [PATCH 4/7] feat: Update llama.cpp --- llama_cpp/llama_cpp.py | 268 ++++++++++++++++++++++++++++++++++++++--- vendor/llama.cpp | 2 +- 2 files changed, 255 insertions(+), 15 deletions(-) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index f3985ad2f..710bd83c8 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -165,6 +165,10 @@ # llama_sampler_p = NewType("llama_sampler_p", int) # llama_sampler_p_ctypes = ctypes.c_void_p +# struct llama_kv_cache; +llama_kv_cache_p = NewType("llama_kv_cache_p", int) +llama_kv_cache_p_ctypes = ctypes.c_void_p + # typedef int32_t llama_pos; llama_pos = ctypes.c_int32 # typedef int32_t llama_token; @@ -228,6 +232,9 @@ # LLAMA_VOCAB_PRE_TYPE_MINERVA = 27, # LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28, # LLAMA_VOCAB_PRE_TYPE_GPT4O = 29, +# LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30, +# LLAMA_VOCAB_PRE_TYPE_TRILLION = 31, +# LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -259,6 +266,9 @@ LLAMA_VOCAB_PRE_TYPE_MINERVA = 27 LLAMA_VOCAB_PRE_TYPE_DEEPSEEK3_LLM = 28 LLAMA_VOCAB_PRE_TYPE_GPT4O = 29 +LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30 +LLAMA_VOCAB_PRE_TYPE_TRILLION = 31 +LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32 # // note: these values should be synchronized with ggml_rope @@ -630,10 +640,19 @@ class llama_model_kv_override(ctypes.Structure): value: Union[int, float, bool, bytes] +# struct llama_model_tensor_buft_override { +# const char * pattern; +# ggml_backend_buffer_type_t buft; +# }; + + # struct llama_model_params { # // NULL-terminated list of devices to use for offloading (if NULL, all available devices are used) # ggml_backend_dev_t * devices; +# // NULL-terminated list of buffer types to use for tensors that match a pattern +# const struct llama_model_tensor_buft_override * tensor_buft_overrides; + # int32_t n_gpu_layers; // number of layers to store in VRAM # enum llama_split_mode split_mode; // how to split the model across multiple GPUs @@ -668,6 +687,8 @@ class llama_model_params(ctypes.Structure): """Parameters for llama_model Attributes: + devices (ctypes.Array[ggml_backend_dev_t]): NULL-terminated list of devices to use for offloading (if NULL, all available devices are used) + tensor_buft_overrides (ctypes.Array[llama_model_tensor_buft_override]): NULL-terminated list of buffer types to use for tensors that match a pattern n_gpu_layers (int): number of layers to store in VRAM split_mode (int): how to split the model across multiple GPUs main_gpu (int): the GPU that is used for the entire model. main_gpu interpretation depends on split_mode: LLAMA_SPLIT_NONE: the GPU that is used for the entire model LLAMA_SPLIT_ROW: the GPU that is used for small tensors and intermediate results LLAMA_SPLIT_LAYER: ignored @@ -681,6 +702,8 @@ class llama_model_params(ctypes.Structure): check_tensors (bool): validate model tensor data""" if TYPE_CHECKING: + devices: CtypesArray[ctypes.c_void_p] # NOTE: unused + tensor_buft_overrides: CtypesArray[llama_model_tensor_buft_override] # NOTE: unused n_gpu_layers: int split_mode: int main_gpu: int @@ -695,6 +718,7 @@ class llama_model_params(ctypes.Structure): _fields_ = [ ("devices", ctypes.c_void_p), # NOTE: unnused + ("tensor_buft_overrides", ctypes.c_void_p), # NOTE: unused ("n_gpu_layers", ctypes.c_int32), ("split_mode", ctypes.c_int), ("main_gpu", ctypes.c_int32), @@ -1317,7 +1341,18 @@ def llama_get_model(ctx: llama_context_p, /) -> Optional[llama_model_p]: ... -# LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); +# LLAMA_API struct llama_kv_cache * llama_get_kv_self ( struct llama_context * ctx); +@ctypes_function( + "llama_get_kv_self", + [llama_context_p_ctypes], + llama_kv_cache_p_ctypes, +) +def llama_get_kv_self(ctx: llama_context_p, /) -> Optional[llama_kv_cache_p]: + """Get the KV cache for self-attention""" + ... + + +# LLAMA_API enum llama_pooling_type llama_pooling_type(const struct llama_context * ctx); @ctypes_function("llama_pooling_type", [llama_context_p_ctypes], ctypes.c_int) def llama_pooling_type(ctx: llama_context_p, /) -> int: ... @@ -1810,7 +1845,19 @@ def llama_kv_cache_view_update(ctx: llama_context_p, view: CtypesPointerOrRef[ll # // Returns the number of tokens in the KV cache (slow, use only for debug) # // If a KV cell has multiple sequences assigned to it, it will be counted multiple times -# LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx); +# LLAMA_API int32_t llama_kv_self_n_tokens(const struct llama_context * ctx); +@ctypes_function( + "llama_kv_self_n_tokens", [llama_context_p_ctypes], ctypes.c_int32 +) +def llama_kv_self_n_tokens(ctx: llama_context_p, /) -> int: + """Returns the number of tokens in the KV cache (slow, use only for debug) + If a KV cell has multiple sequences assigned to it, it will be counted multiple times + """ + ... + + +# DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_token_count(const struct llama_context * ctx), +# "use llama_kv_self_n_tokens instead"); @ctypes_function( "llama_get_kv_cache_token_count", [llama_context_p_ctypes], ctypes.c_int32 ) @@ -1822,7 +1869,17 @@ def llama_get_kv_cache_token_count(ctx: llama_context_p, /) -> int: # // Returns the number of used KV cells (i.e. have at least one sequence assigned to them) -# LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx); +# LLAMA_API int32_t llama_kv_self_used_cells(const struct llama_context * ctx); +@ctypes_function( + "llama_kv_self_used_cells", [llama_context_p_ctypes], ctypes.c_int32 +) +def llama_kv_self_used_cells(ctx: llama_context_p, /) -> int: + """Returns the number of used KV cells (i.e. have at least one sequence assigned to them)""" + ... + + +# DEPRECATED(LLAMA_API int32_t llama_get_kv_cache_used_cells(const struct llama_context * ctx), +# "use llama_kv_self_used_cells instead"); @ctypes_function( "llama_get_kv_cache_used_cells", [llama_context_p_ctypes], ctypes.c_int32 ) @@ -1832,9 +1889,17 @@ def llama_get_kv_cache_used_cells(ctx: llama_context_p, /) -> int: # // Clear the KV cache - both cell info is erased and KV data is zeroed -# LLAMA_API void llama_kv_cache_clear( +# LLAMA_API void llama_kv_self_clear( # struct llama_context * ctx); -@ctypes_function("llama_kv_cache_clear", [llama_context_p_ctypes], None) +@ctypes_function( + "llama_kv_self_clear", [llama_context_p_ctypes], None +) +def llama_kv_self_clear(ctx: llama_context_p, /): + """Clear the KV cache - both cell info is erased and KV data is zeroed""" + ... + +# NOTE: Deprecated +@ctypes_function("llama_kv_self_clear", [llama_context_p_ctypes], None) def llama_kv_cache_clear(ctx: llama_context_p, /): """Clear the KV cache""" ... @@ -1881,14 +1946,41 @@ def llama_kv_cache_seq_rm( # // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence # // p0 < 0 : [0, p1] # // p1 < 0 : [p0, inf) -# LLAMA_API void llama_kv_cache_seq_cp( +# LLAMA_API void llama_kv_self_seq_cp( # struct llama_context * ctx, # llama_seq_id seq_id_src, # llama_seq_id seq_id_dst, # llama_pos p0, # llama_pos p1); @ctypes_function( - "llama_kv_cache_seq_cp", + "llama_kv_self_seq_cp", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_seq_id, + llama_pos, + llama_pos, + ], + None, +) +def llama_kv_self_seq_cp( + ctx: llama_context_p, + seq_id_src: Union[llama_seq_id, int], + seq_id_dst: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + /, +): + """Copy all tokens that belong to the specified sequence to another sequence + Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence + p0 < 0 : [0, p1] + p1 < 0 : [p0, inf)""" + ... + + +# NOTE: Deprecated +@ctypes_function( + "llama_kv_self_seq_cp", [ llama_context_p_ctypes, llama_seq_id, @@ -1914,17 +2006,27 @@ def llama_kv_cache_seq_cp( # // Removes all tokens that do not belong to the specified sequence -# LLAMA_API void llama_kv_cache_seq_keep( +# LLAMA_API void llama_kv_self_seq_keep( # struct llama_context * ctx, # llama_seq_id seq_id); @ctypes_function( - "llama_kv_cache_seq_keep", [llama_context_p_ctypes, llama_seq_id], None + "llama_kv_self_seq_keep", [llama_context_p_ctypes, llama_seq_id], None +) +def llama_kv_self_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, int], /): + """Removes all tokens that do not belong to the specified sequence""" + ... + + +# NOTE: Deprecated +@ctypes_function( + "llama_kv_self_seq_keep", [llama_context_p_ctypes, llama_seq_id], None ) def llama_kv_cache_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, int], /): """Removes all tokens that do not belong to the specified sequence""" ... + # // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) # // If the KV cache is RoPEd, the KV data is updated accordingly: # // - lazily on next llama_decode() @@ -1938,7 +2040,48 @@ def llama_kv_cache_seq_keep(ctx: llama_context_p, seq_id: Union[llama_seq_id, in # llama_pos p1, # llama_pos delta); @ctypes_function( - "llama_kv_cache_seq_add", + "llama_kv_self_seq_add", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + llama_pos, + ], + None, +) +def llama_kv_self_seq_add( + ctx: llama_context_p, + seq_id: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + delta: Union[llama_pos, int], + /, +): + """Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) + If the KV cache is RoPEd, the KV data is updated accordingly: + - lazily on next llama_decode() + - explicitly with llama_kv_cache_update() + p0 < 0 : [0, p1] + p1 < 0 : [p0, inf)""" + ... + + +# // NOTE: Deprecated +# // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1) +# // If the KV cache is RoPEd, the KV data is updated accordingly: +# // - lazily on next llama_decode() +# // - explicitly with llama_kv_cache_update() +# // p0 < 0 : [0, p1] +# // p1 < 0 : [p0, inf) +# LLAMA_API void llama_kv_cache_seq_add( +# struct llama_context * ctx, +# llama_seq_id seq_id, +# llama_pos p0, +# llama_pos p1, +# llama_pos delta); +@ctypes_function( + "llama_kv_self_seq_add", [ llama_context_p_ctypes, llama_seq_id, @@ -1976,7 +2119,44 @@ def llama_kv_cache_seq_add( # llama_pos p1, # int d); @ctypes_function( - "llama_kv_cache_seq_div", + "llama_kv_self_seq_div", + [ + llama_context_p_ctypes, + llama_seq_id, + llama_pos, + llama_pos, + ctypes.c_int, + ], + None, +) +def llama_kv_self_seq_div( + ctx: llama_context_p, + seq_id: Union[llama_seq_id, int], + p0: Union[llama_pos, int], + p1: Union[llama_pos, int], + d: Union[ctypes.c_int, int], + /, +): + """Integer division of the positions by factor of `d > 1` + If the KV cache is RoPEd, the KV data is updated accordingly + p0 < 0 : [0, p1] + p1 < 0 : [p0, inf)""" + ... + + +# // NOTE: Deprecated +# // Integer division of the positions by factor of `d > 1` +# // If the KV cache is RoPEd, the KV data is updated accordingly +# // p0 < 0 : [0, p1] +# // p1 < 0 : [p0, inf) +# LLAMA_API void llama_kv_cache_seq_div( +# struct llama_context * ctx, +# llama_seq_id seq_id, +# llama_pos p0, +# llama_pos p1, +# int d); +@ctypes_function( + "llama_kv_self_seq_div", [ llama_context_p_ctypes, llama_seq_id, @@ -2001,10 +2181,39 @@ def llama_kv_cache_seq_div( ... +# // Returns the largest position present in the KV cache for the specified sequence +# LLAMA_API llama_pos llama_kv_self_seq_pos_max( +# struct llama_context * ctx, +# llama_seq_id seq_id); +@ctypes_function( + "llama_kv_self_seq_pos_max", [llama_context_p_ctypes, llama_seq_id], llama_pos +) +def llama_kv_self_seq_pos_max( + ctx: llama_context_p, seq_id: Union[llama_seq_id, int], / +) -> int: + """Returns the largest position present in the KV cache for the specified sequence""" + ... + + # // Defragment the KV cache # // This will be applied: # // - lazily on next llama_decode() -# // - explicitly with llama_kv_cache_update() +# // - explicitly with llama_kv_self_update() +# LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx); +@ctypes_function("llama_kv_self_defrag", [llama_context_p_ctypes], None) +def llama_kv_self_defrag(ctx: llama_context_p, /): + """Defragment the KV cache + This will be applied: + - lazily on next llama_decode() + - explicitly with llama_kv_cache_update()""" + ... + + +# NOTE: Deprecated +# // Defragment the KV cache +# // This will be applied: +# // - lazily on next llama_decode() +# // - explicitly with llama_kv_self_update() # LLAMA_API void llama_kv_cache_defrag(struct llama_context * ctx); @ctypes_function("llama_kv_cache_defrag", [llama_context_p_ctypes], None) def llama_kv_cache_defrag(ctx: llama_context_p, /): @@ -2017,7 +2226,15 @@ def llama_kv_cache_defrag(ctx: llama_context_p, /): # // Apply the KV cache updates (such as K-shifts, defragmentation, etc.) # LLAMA_API void llama_kv_cache_update(struct llama_context * ctx); -@ctypes_function("llama_kv_cache_update", [llama_context_p_ctypes], None) +@ctypes_function("llama_kv_self_update", [llama_context_p_ctypes], None) +def llama_kv_self_update(ctx: llama_context_p, /): + """Apply the KV cache updates (such as K-shifts, defragmentation, etc.)""" + ... + +# // NOTE: Deprecated +# // Apply the KV cache updates (such as K-shifts, defragmentation, etc.) +# LLAMA_API void llama_kv_cache_update(struct llama_context * ctx); +@ctypes_function("llama_kv_self_update", [llama_context_p_ctypes], None) def llama_kv_cache_update(ctx: llama_context_p, /): """Apply the KV cache updates (such as K-shifts, defragmentation, etc.)""" ... @@ -2025,7 +2242,16 @@ def llama_kv_cache_update(ctx: llama_context_p, /): # // Check if the context supports KV cache shifting # LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx); -@ctypes_function("llama_kv_cache_can_shift", [llama_context_p_ctypes], ctypes.c_bool) +@ctypes_function("llama_kv_self_can_shift", [llama_context_p_ctypes], ctypes.c_bool) +def llama_kv_self_can_shift(ctx: llama_context_p, /) -> bool: + """Check if the context supports KV cache shifting""" + ... + + +# // NOTE: Deprecated +# // Check if the context supports KV cache shifting +# LLAMA_API bool llama_kv_cache_can_shift(struct llama_context * ctx); +@ctypes_function("llama_kv_self_can_shift", [llama_context_p_ctypes], ctypes.c_bool) def llama_kv_cache_can_shift(ctx: llama_context_p, /) -> bool: """Check if the context supports KV cache shifting""" ... @@ -2547,6 +2773,16 @@ def llama_set_causal_attn(ctx: llama_context_p, causal_attn: bool, /): ... +# // Set whether the model is in warmup mode or not +# // If true, all model tensors are activated during llama_decode() to load and cache their weights. +# LLAMA_API void llama_set_warmup(struct llama_context * ctx, bool warmup); +@ctypes_function("llama_set_warmup", [llama_context_p_ctypes, ctypes.c_bool], None) +def llama_set_warmup(ctx: llama_context_p, warmup: bool, /): + """Set whether the model is in warmup mode or not + If true, all model tensors are activated during llama_decode() to load and cache their weights.""" + ... + + # // Set abort callback # LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data); @ctypes_function( @@ -3701,6 +3937,10 @@ def llama_sampler_init_mirostat_v2( ... +# /// @details Intializes a GBNF grammar, see grammars/README.md for details. +# /// @param vocab The vocabulary that this grammar will be used with. +# /// @param grammar_str The production rules for the grammar, encoded as a string. Returns an empty grammar if empty. Returns NULL if parsing of grammar_str fails. +# /// @param grammar_root The name of the start symbol for the grammar. # LLAMA_API struct llama_sampler * llama_sampler_init_grammar( # const struct llama_vocab * vocab, # const char * grammar_str, diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 7841fc723..6bf28f011 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 7841fc723e059d1fd9640e5c0ef19050fcc7c698 +Subproject commit 6bf28f0111ff9f21b3c1b1eace20c590281e7ba6 From 4c6514d365cb4f00910424c2feb8e45e68a870a6 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 8 May 2025 06:52:52 -0400 Subject: [PATCH 5/7] feat: Update llama.cpp --- CMakeLists.txt | 2 +- llama_cpp/llama_cpp.py | 33 +++++++++++++++++++++------------ vendor/llama.cpp | 2 +- 3 files changed, 23 insertions(+), 14 deletions(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 64a0304a1..443d1ee53 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -143,7 +143,7 @@ if (LLAMA_BUILD) endif() # Building llava - add_subdirectory(vendor/llama.cpp/examples/llava) + add_subdirectory(vendor/llama.cpp/tools/mtmd) set_target_properties(llava_shared PROPERTIES OUTPUT_NAME "llava") if (WIN32) diff --git a/llama_cpp/llama_cpp.py b/llama_cpp/llama_cpp.py index 710bd83c8..63de3a93a 100644 --- a/llama_cpp/llama_cpp.py +++ b/llama_cpp/llama_cpp.py @@ -235,6 +235,8 @@ # LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30, # LLAMA_VOCAB_PRE_TYPE_TRILLION = 31, # LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32, +# LLAMA_VOCAB_PRE_TYPE_LLAMA4 = 33, +# LLAMA_VOCAB_PRE_TYPE_PIXTRAL = 34, # }; LLAMA_VOCAB_PRE_TYPE_DEFAULT = 0 LLAMA_VOCAB_PRE_TYPE_LLAMA3 = 1 @@ -252,7 +254,7 @@ LLAMA_VOCAB_PRE_TYPE_DBRX = 13 LLAMA_VOCAB_PRE_TYPE_SMAUG = 14 LLAMA_VOCAB_PRE_TYPE_PORO = 15 -LLAMA_VOCAV_PRE_TYPE_CHATGLM3 = 16 +LLAMA_VOCAB_PRE_TYPE_CHATGLM3 = 16 LLAMA_VOCAB_PRE_TYPE_CHATGLM4 = 17 LLAMA_VOCAB_PRE_TYPE_VIKING = 18 LLAMA_VOCAB_PRE_TYPE_JAIS = 19 @@ -269,6 +271,8 @@ LLAMA_VOCAB_PRE_TYPE_SUPERBPE = 30 LLAMA_VOCAB_PRE_TYPE_TRILLION = 31 LLAMA_VOCAB_PRE_TYPE_BAILINGMOE = 32 +LLAMA_VOCAB_PRE_TYPE_LLAMA4 = 33 +LLAMA_VOCAB_PRE_TYPE_PIXTRAL = 34 # // note: these values should be synchronized with ggml_rope @@ -891,17 +895,18 @@ class llama_context_params(ctypes.Structure): # // model quantization parameters # typedef struct llama_model_quantize_params { -# int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() -# enum llama_ftype ftype; // quantize to this llama_ftype -# enum ggml_type output_tensor_type; // output tensor type -# enum ggml_type token_embedding_type; // token embeddings tensor type -# bool allow_requantize; // allow quantizing non-f32/f16 tensors -# bool quantize_output_tensor; // quantize output.weight -# bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored -# bool pure; // quantize all tensors to the default type -# bool keep_split; // quantize to the same number of shards -# void * imatrix; // pointer to importance matrix data -# void * kv_overrides; // pointer to vector containing overrides +# int32_t nthread; // number of threads to use for quantizing, if <=0 will use std::thread::hardware_concurrency() +# enum llama_ftype ftype; // quantize to this llama_ftype +# enum ggml_type output_tensor_type; // output tensor type +# enum ggml_type token_embedding_type; // token embeddings tensor type +# bool allow_requantize; // allow quantizing non-f32/f16 tensors +# bool quantize_output_tensor; // quantize output.weight +# bool only_copy; // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored +# bool pure; // quantize all tensors to the default type +# bool keep_split; // quantize to the same number of shards +# void * imatrix; // pointer to importance matrix data +# void * kv_overrides; // pointer to vector containing overrides +# void * tensor_types; // pointer to vector containing tensor types # } llama_model_quantize_params; class llama_model_quantize_params(ctypes.Structure): """Parameters for llama_model_quantize @@ -918,6 +923,7 @@ class llama_model_quantize_params(ctypes.Structure): keep_split (bool): quantize to the same number of shards imatrix (ctypes.c_void_p): pointer to importance matrix data kv_overrides (ctypes.c_void_p): pointer to vector containing overrides + tensor_types (ctypes.c_void_p): pointer to vector containing tensor types """ if TYPE_CHECKING: @@ -932,6 +938,7 @@ class llama_model_quantize_params(ctypes.Structure): keep_split: bool imatrix: ctypes.c_void_p kv_overrides: ctypes.c_void_p + tensor_types: ctypes.c_void_p _fields_ = [ ("nthread", ctypes.c_int32), @@ -945,6 +952,7 @@ class llama_model_quantize_params(ctypes.Structure): ("keep_split", ctypes.c_bool), ("imatrix", ctypes.c_void_p), ("kv_overrides", ctypes.c_void_p), + ("tensor_types", ctypes.c_void_p), ] @@ -3812,6 +3820,7 @@ def llama_sampler_init_softmax() -> llama_sampler_p: # /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 +# /// Setting k <= 0 makes this a noop # LLAMA_API struct llama_sampler * llama_sampler_init_top_k (int32_t k); @ctypes_function("llama_sampler_init_top_k", [ctypes.c_int32], llama_sampler_p_ctypes) def llama_sampler_init_top_k(k: int) -> llama_sampler_p: diff --git a/vendor/llama.cpp b/vendor/llama.cpp index 6bf28f011..8733e0cf6 160000 --- a/vendor/llama.cpp +++ b/vendor/llama.cpp @@ -1 +1 @@ -Subproject commit 6bf28f0111ff9f21b3c1b1eace20c590281e7ba6 +Subproject commit 8733e0cf6eefc7c7752297cc22d0836706f4222c From cb2edb9c9fef90fe680e528fadc86806c6f385e5 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 8 May 2025 06:59:55 -0400 Subject: [PATCH 6/7] chore: Bump version --- CHANGELOG.md | 4 ++++ llama_cpp/__init__.py | 2 +- 2 files changed, 5 insertions(+), 1 deletion(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 605370e7d..affbd5db7 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ## [Unreleased] +## [0.3.9] + +- feat: Update llama.cpp to ggerganov/llama.cpp@8733e0cf6eefc7c7752297cc22d0836706f4222c + ## [0.3.8] - feat: Update llama.cpp to ggerganov/llama.cpp@7841fc723e059d1fd9640e5c0ef19050fcc7c698 diff --git a/llama_cpp/__init__.py b/llama_cpp/__init__.py index b1a8b9baa..2c9c527cd 100644 --- a/llama_cpp/__init__.py +++ b/llama_cpp/__init__.py @@ -1,4 +1,4 @@ from .llama_cpp import * from .llama import * -__version__ = "0.3.8" +__version__ = "0.3.9" From b1d23df0bbd327b774083b5cf88e67ca0dd52b92 Mon Sep 17 00:00:00 2001 From: Andrei Betlen Date: Thu, 8 May 2025 07:02:20 -0400 Subject: [PATCH 7/7] hotfix: Disable curl support --- CMakeLists.txt | 3 +++ 1 file changed, 3 insertions(+) diff --git a/CMakeLists.txt b/CMakeLists.txt index 443d1ee53..b9178e856 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -62,6 +62,9 @@ if (LLAMA_BUILD) # Enable building of the common library set(LLAMA_BUILD_COMMON ON CACHE BOOL "Build llama.cpp common library" FORCE) + # Disable building curl support + set(LLAMA_CURL OFF CACHE BOOL "llama.cpp: enable curl" FORCE) + # Architecture detection and settings for Apple platforms if (APPLE) # Get the target architecture