From 2e7ef98f18090b382611c135efc417200b23780b Mon Sep 17 00:00:00 2001 From: Aman Gupta Date: Fri, 28 Nov 2025 20:34:51 +0800 Subject: [PATCH 1/6] ggml-cuda: add stricter checking for fusion (#17568) * ggml-cuda: make conditions for fusion more explicit * ggml-cuda: remove size check as std::equal already does it --- ggml/src/ggml-cuda/ggml-cuda.cu | 26 +++++++++++++++----------- 1 file changed, 15 insertions(+), 11 deletions(-) diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 6463921a6..a844a3d99 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -3050,7 +3050,12 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list topk_moe_ops_delayed_softmax = ggml_cuda_topk_moe_ops(/*with_norm=*/false, /*delayed_softmax=*/true); - if (ops.size() == topk_moe_ops_with_norm.size() && + const auto is_equal = [](const std::initializer_list & list1, + const std::initializer_list & list2) { + return std::equal(list1.begin(), list1.end(), list2.begin(), list2.end()); + }; + + if (is_equal(topk_moe_ops_with_norm, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 9 })) { ggml_tensor * softmax = cgraph->nodes[node_idx]; ggml_tensor * weights = cgraph->nodes[node_idx + 9]; @@ -3060,8 +3065,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, } } - if (ops.size() == topk_moe_ops.size() && - ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 4 })) { + if (is_equal(topk_moe_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 3, node_idx + 4 })) { ggml_tensor * softmax = cgraph->nodes[node_idx]; ggml_tensor * weights = cgraph->nodes[node_idx + 4]; if (ggml_cuda_should_use_topk_moe(softmax, weights)) { @@ -3069,7 +3073,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, } } - if (ops.size() == topk_moe_ops_delayed_softmax.size() && + if (is_equal(topk_moe_ops_delayed_softmax, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 1, node_idx + 5 })) { ggml_tensor * softmax = cgraph->nodes[node_idx + 4]; ggml_tensor * weights = cgraph->nodes[node_idx + 5]; @@ -3085,9 +3089,8 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list mul_mat_id_glu_ops = { GGML_OP_MUL_MAT_ID, GGML_OP_MUL_MAT_ID, GGML_OP_GLU }; std::initializer_list mul_mat_glu_ops = { GGML_OP_MUL_MAT, GGML_OP_MUL_MAT, GGML_OP_GLU }; - if (ops.size() == 5 && (ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 4}) || - ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 4}))) { - + if ((is_equal(mul_mat_bias_glu_ops, ops) || is_equal(mul_mat_id_bias_glu_ops, ops)) && + ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 4 })) { const ggml_tensor * ffn_gate = cgraph->nodes[node_idx]; const ggml_tensor * ffn_gate_bias = cgraph->nodes[node_idx + 1]; const ggml_tensor * ffn_up = cgraph->nodes[node_idx + 2]; @@ -3099,9 +3102,8 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, } } - if (ops.size() == 3 && (ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 2}) || - ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 2}))) { - + if ((is_equal(mul_mat_id_glu_ops, ops) || is_equal(mul_mat_glu_ops, ops)) && + ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) { const ggml_tensor * ffn_gate = cgraph->nodes[node_idx]; const ggml_tensor * ffn_up = cgraph->nodes[node_idx + 1]; const ggml_tensor * glu = cgraph->nodes[node_idx + 2]; @@ -3111,7 +3113,9 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, } } - if (ops.size() == 3 && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) { + std::initializer_list rope_set_rows_ops = { GGML_OP_ROPE, GGML_OP_VIEW, GGML_OP_SET_ROWS }; + + if (is_equal(rope_set_rows_ops, ops) && ggml_can_fuse_subgraph(cgraph, node_idx, ops, { node_idx + 2 })) { const ggml_tensor * rope = cgraph->nodes[node_idx]; const ggml_tensor * view = cgraph->nodes[node_idx + 1]; const ggml_tensor * set_rows = cgraph->nodes[node_idx + 2]; From c6f7a423c8c87748ef563a99d81c3b1b05cecff0 Mon Sep 17 00:00:00 2001 From: R0CKSTAR Date: Fri, 28 Nov 2025 21:08:29 +0800 Subject: [PATCH 2/6] [MUSA] enable fp16/fast_fp16/bf16_mma on PH1 (#17551) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit * [MUSA] enable fp16/fast_fp16/bf16_mma on PH1 Signed-off-by: Xiaodong Ye * Update ggml/src/ggml-cuda/fattn-vec.cuh Co-authored-by: Johannes Gäßler * Update ggml/src/ggml-cuda/fattn-vec.cuh Co-authored-by: Johannes Gäßler * Update ggml/src/ggml-cuda/fattn-tile.cuh Co-authored-by: Johannes Gäßler * Address review comments Signed-off-by: Xiaodong Ye --------- Signed-off-by: Xiaodong Ye Co-authored-by: Johannes Gäßler --- ggml/src/ggml-cuda/common.cuh | 20 ++++++++++++-------- ggml/src/ggml-cuda/cpy.cu | 5 ++++- ggml/src/ggml-cuda/fattn-tile.cuh | 2 +- ggml/src/ggml-cuda/fattn-vec.cuh | 4 ++-- ggml/src/ggml-cuda/mma.cuh | 4 ++-- 5 files changed, 21 insertions(+), 14 deletions(-) diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh index 81fb31240..0b10e5f6a 100644 --- a/ggml/src/ggml-cuda/common.cuh +++ b/ggml/src/ggml-cuda/common.cuh @@ -84,12 +84,12 @@ #define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000 #define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000 -#define GGML_CUDA_CC_NG (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD +#define GGML_CUDA_CC_PH1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // MTT S5000 #define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD) #define GGML_CUDA_CC_IS_QY1(cc) (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2) -#define GGML_CUDA_CC_IS_QY2(cc) (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NG) -#define GGML_CUDA_CC_IS_NG(cc) (cc >= GGML_CUDA_CC_NG) +#define GGML_CUDA_CC_IS_QY2(cc) (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_PH1) +#define GGML_CUDA_CC_IS_PH1(cc) (cc >= GGML_CUDA_CC_PH1) #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070 # define GGML_CUDA_USE_CUB @@ -212,9 +212,9 @@ static const char * cu_get_error_str(CUresult err) { #define GGML_USE_VMM #endif // (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM)) -#if defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL +#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL #define FP16_AVAILABLE -#endif // defined(GGML_USE_HIP) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL +#endif // defined(GGML_USE_HIP) || defined(GGML_USE_MUSA) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL #if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610 #define FAST_FP16_AVAILABLE @@ -250,12 +250,14 @@ static const char * cu_get_error_str(CUresult err) { #endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220) static bool fp16_available(const int cc) { - return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL; + return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL || + (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1); } static bool fast_fp16_available(const int cc) { return GGML_CUDA_CC_IS_AMD(cc) || - (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610); + (GGML_CUDA_CC_IS_NVIDIA(cc) && fp16_available(cc) && ggml_cuda_highest_compiled_arch(cc) != 610) || + (GGML_CUDA_CC_IS_MTHREADS(cc) && fp16_available(cc)); } // To be used for feature selection of external libraries, e.g. cuBLAS. @@ -272,7 +274,9 @@ static bool fp16_mma_hardware_available(const int cc) { } static bool bf16_mma_hardware_available(const int cc) { - return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE) || GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3; + return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE) || + GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3 || + (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1); } static bool fp32_mma_hardware_available(const int cc) { diff --git a/ggml/src/ggml-cuda/cpy.cu b/ggml/src/ggml-cuda/cpy.cu index 82367ad3f..c4ceb4fc5 100644 --- a/ggml/src/ggml-cuda/cpy.cu +++ b/ggml/src/ggml-cuda/cpy.cu @@ -86,6 +86,9 @@ static __global__ void cpy_scalar_transpose(const char * cx, char * cdst, const } } } + + GGML_UNUSED_VARS(ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, + nb12, nb13); } static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) { @@ -202,7 +205,7 @@ static void ggml_cpy_scalar_cuda( ne00n = ne00; ne01n = ne01; ne02n = ne02; - } else if (nb00 > nb02) { + } else { ne00n = ne00; ne01n = ne01*ne02; ne02n = 1; diff --git a/ggml/src/ggml-cuda/fattn-tile.cuh b/ggml/src/ggml-cuda/fattn-tile.cuh index c358aa1e8..3e58d64ff 100644 --- a/ggml/src/ggml-cuda/fattn-tile.cuh +++ b/ggml/src/ggml-cuda/fattn-tile.cuh @@ -609,7 +609,7 @@ static __device__ __forceinline__ void flash_attn_tile_iter( float KQ_sum_add = 0.0f; #pragma unroll for (int i0 = 0; i0 < nbatch_fa; i0 += np*warp_size) { - const float val = !oob_check || i0 + (threadIdx.y % np)*warp_size + threadIdx.x < k_VKQ_sup ? + const float val = !oob_check || i0 + (threadIdx.y % np)*warp_size + threadIdx.x < static_cast(k_VKQ_sup) ? expf(KQ_acc[(i0/(np*warp_size))*cpw + jc] - KQ_max[jc]) : 0.0f; KQ_sum_add += val; tmp[i0/(np*warp_size)][jc1] = val; diff --git a/ggml/src/ggml-cuda/fattn-vec.cuh b/ggml/src/ggml-cuda/fattn-vec.cuh index 6e63e860a..67aa67ecb 100644 --- a/ggml/src/ggml-cuda/fattn-vec.cuh +++ b/ggml/src/ggml-cuda/fattn-vec.cuh @@ -155,7 +155,7 @@ static __global__ void flash_attn_ext_vec( for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) { const int i = i0 + threadIdx.x; - if (i0 + WARP_SIZE <= D/sizeof(int) || i < D/sizeof(int)) { + if (i0 + WARP_SIZE <= int(D/sizeof(int)) || i < int(D/sizeof(int))) { tmp_q_i32[i] = 0; } } @@ -272,7 +272,7 @@ static __global__ void flash_attn_ext_vec( KQ_max_new[j] = fmaxf(KQ_max_new[j], sum); - if ((nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ) == i_KQ_0) { + if ((nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ) == uint32_t(i_KQ_0)) { KQ_reg[j] = sum; } } diff --git a/ggml/src/ggml-cuda/mma.cuh b/ggml/src/ggml-cuda/mma.cuh index c0a9c2c08..0ed42e87d 100644 --- a/ggml/src/ggml-cuda/mma.cuh +++ b/ggml/src/ggml-cuda/mma.cuh @@ -889,8 +889,8 @@ namespace ggml_cuda_mma { : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7]) : "r"(Axi[6]), "r"(Axi[7]), "r"(Bxi[6]), "r"(Bxi[7])); #else - tile<16, 8, float> * D16 = (tile<16, 8, float> *) &D; - tile<16, 8, half2> * A16 = (tile<16, 8, half2> *) &A; + tile <16, 8, float> * D16 = reinterpret_cast *>(&D); + const tile<16, 8, half2> * A16 = reinterpret_cast *>(&A); mma(D16[0], A16[0], B); mma(D16[1], A16[1], B); #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE From e072b2052e9250395e4a28a28d37806342ac5db1 Mon Sep 17 00:00:00 2001 From: Diego Devesa Date: Fri, 28 Nov 2025 07:33:23 -0800 Subject: [PATCH 3/6] ggml : add GGML_SCHED_NO_REALLOC option to disable reallocations in ggml_backend_sched (#17276) * ggml : add GGML_SCHED_NO_REALLOC option to disable reallocations in ggml_backend_sched Enabled in ggml-ci for testing. * llama : update worst-case graph for unified cache * ci : disable op offload in some tests * fix spelling --------- Co-authored-by: Georgi Gerganov --- ci/run.sh | 16 ++++++++-------- examples/embedding/embedding.cpp | 7 ++++--- ggml/CMakeLists.txt | 1 + ggml/src/CMakeLists.txt | 4 ++++ ggml/src/ggml-alloc.c | 11 ++++++++--- ggml/src/ggml-backend.cpp | 12 +++++++++--- src/llama-context.cpp | 4 ++-- tests/CMakeLists.txt | 2 +- 8 files changed, 37 insertions(+), 20 deletions(-) diff --git a/ci/run.sh b/ci/run.sh index 3fec8e911..1dd65adea 100755 --- a/ci/run.sh +++ b/ci/run.sh @@ -45,7 +45,7 @@ sd=`dirname $0` cd $sd/../ SRC=`pwd` -CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON" +CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_SCHED_NO_REALLOC=ON" if [ ! -z ${GG_BUILD_METAL} ]; then CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON" @@ -428,10 +428,10 @@ function gg_run_qwen3_0_6b { (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -ngl 99 -c 1024 -b 512 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log - (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log - (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa off --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 1024 -fa on --no-op-offload) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa off ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log + (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 1024 -fa on ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log function check_ppl { qnt="$1" @@ -523,8 +523,8 @@ function gg_run_embd_bge_small { ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0 - (time ./bin/llama-embedding --model ${model_f16} -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log - (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log + (time ./bin/llama-embedding --model ${model_f16} -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-f16.log + (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 --no-op-offload) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log set +e } @@ -564,7 +564,7 @@ function gg_run_rerank_tiny { model_f16="${path_models}/ggml-model-f16.gguf" # for this model, the SEP token is "" - (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log + (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --no-op-offload --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log # sample output # rerank score 0: 0.029 diff --git a/examples/embedding/embedding.cpp b/examples/embedding/embedding.cpp index 9e3ab5905..fe91b308c 100644 --- a/examples/embedding/embedding.cpp +++ b/examples/embedding/embedding.cpp @@ -104,12 +104,16 @@ int main(int argc, char ** argv) { params.embedding = true; + // get max number of sequences per batch + const int n_seq_max = llama_max_parallel_sequences(); + // if the number of prompts that would be encoded is known in advance, it's more efficient to specify the // --parallel argument accordingly. for convenience, if not specified, we fallback to unified KV cache // in order to support any number of prompts if (params.n_parallel == 1) { LOG_INF("%s: n_parallel == 1 -> unified KV cache is enabled\n", __func__); params.kv_unified = true; + params.n_parallel = n_seq_max; } // utilize the full context @@ -123,9 +127,6 @@ int main(int argc, char ** argv) { params.n_ubatch = params.n_batch; } - // get max number of sequences per batch - const int n_seq_max = llama_max_parallel_sequences(); - llama_backend_init(); llama_numa_init(params.numa); diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt index 0211255a7..9b10df00d 100644 --- a/ggml/CMakeLists.txt +++ b/ggml/CMakeLists.txt @@ -183,6 +183,7 @@ endif() # ggml core set(GGML_SCHED_MAX_COPIES "4" CACHE STRING "ggml: max input copies for pipeline parallelism") option(GGML_CPU "ggml: enable CPU backend" ON) +option(GGML_SCHED_NO_REALLOC "ggml: disallow reallocations in ggml-alloc (for debugging)" OFF) # 3rd party libs / backends option(GGML_ACCELERATE "ggml: enable Accelerate framework" ON) diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index a4499509e..a36f5b664 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -221,6 +221,10 @@ if (GGML_BACKEND_DL) target_compile_definitions(ggml-base PUBLIC GGML_BACKEND_DL) endif() +if (GGML_SCHED_NO_REALLOC) + target_compile_definitions(ggml-base PUBLIC GGML_SCHED_NO_REALLOC) +endif() + add_library(ggml ggml-backend-reg.cpp) add_library(ggml::ggml ALIAS ggml) diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c index 91aff205f..218222ece 100644 --- a/ggml/src/ggml-alloc.c +++ b/ggml/src/ggml-alloc.c @@ -921,10 +921,15 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c } if (realloc) { #ifndef NDEBUG - size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0; - GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0); + { + size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0; + if (cur_size > 0) { + GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", + __func__, ggml_backend_buft_name(galloc->bufts[i]), + cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0); + } + } #endif - ggml_vbuffer_free(galloc->buffers[i]); galloc->buffers[i] = ggml_vbuffer_alloc(galloc->bufts[i], galloc->buf_tallocs[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE); if (galloc->buffers[i] == NULL) { diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp index eeaf35c16..4cf377e7f 100644 --- a/ggml/src/ggml-backend.cpp +++ b/ggml/src/ggml-backend.cpp @@ -1395,14 +1395,20 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) { // allocate graph if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) { +#ifdef GGML_SCHED_NO_REALLOC + GGML_ABORT("%s: failed to allocate graph, but graph re-allocation is disabled by GGML_SCHED_NO_REALLOC\n", __func__); +#endif + +#ifndef NDEBUG + GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed); +#endif + // the re-allocation may cause the split inputs to be moved to a different address // synchronize without ggml_backend_sched_synchronize to avoid changing cur_copy for (int i = 0; i < sched->n_backends; i++) { ggml_backend_synchronize(sched->backends[i]); } -#ifndef NDEBUG - GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed); -#endif + ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids); if (!ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) { GGML_LOG_ERROR("%s: failed to allocate graph\n", __func__); diff --git a/src/llama-context.cpp b/src/llama-context.cpp index a58914429..e04f0fc4f 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -300,7 +300,7 @@ llama_context::llama_context( cross.v_embd.clear(); - const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max; + const uint32_t n_seqs = cparams.n_seq_max; const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch); // avoid reserving graphs with zero outputs - assume one output per sequence @@ -543,7 +543,7 @@ bool llama_context::memory_update(bool optimize) { throw std::runtime_error("failed to initialize memory context"); } - const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max; + const uint32_t n_seqs = cparams.n_seq_max; const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch); auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get()); diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index d9cc5e933..9361a113a 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -196,7 +196,7 @@ if (NOT WIN32) llama_build_and_test(test-arg-parser.cpp) endif() -if (NOT LLAMA_SANITIZE_ADDRESS) +if (NOT LLAMA_SANITIZE_ADDRESS AND NOT GGML_SCHED_NO_REALLOC) # TODO: repair known memory leaks llama_build_and_test(test-opt.cpp) endif() From 3ce7a65c2f2529a8fc566b4aead53b088f7faec2 Mon Sep 17 00:00:00 2001 From: o7si <32285332+o7si@users.noreply.github.com> Date: Sat, 29 Nov 2025 02:14:00 +0800 Subject: [PATCH 4/6] server: fix: /metrics endpoint returning JSON-escaped Prometheus format (#17386) * fix: /metrics endpoint returning JSON-escaped Prometheus format * mod: remove string overload from ok() method --- tools/server/server.cpp | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/tools/server/server.cpp b/tools/server/server.cpp index 05bbe648c..96b2df27f 100644 --- a/tools/server/server.cpp +++ b/tools/server/server.cpp @@ -2713,7 +2713,8 @@ public: res->headers["Process-Start-Time-Unix"] = std::to_string(res_task->t_start); res->content_type = "text/plain; version=0.0.4"; - res->ok(prometheus.str()); + res->status = 200; + res->data = prometheus.str(); return res; }; From 03914c7ef826caf0b6371a6d1de270cda102b542 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?DAN=E2=84=A2?= Date: Fri, 28 Nov 2025 13:29:36 -0500 Subject: [PATCH 5/6] common : move all common_chat_parse_* to chat-parser.cpp. (#17481) --- common/chat-parser.cpp | 968 +++++++++++++++++++++++++++++++++++++++++ common/chat.cpp | 952 ---------------------------------------- 2 files changed, 968 insertions(+), 952 deletions(-) diff --git a/common/chat-parser.cpp b/common/chat-parser.cpp index ff8310278..301f439a6 100644 --- a/common/chat-parser.cpp +++ b/common/chat-parser.cpp @@ -13,6 +13,120 @@ using json = nlohmann::ordered_json; +static void parse_prefixed_json_tool_call_array(common_chat_msg_parser & builder, + const common_regex & prefix, + size_t rstrip_prefix = 0) { + static const std::vector> args_paths = { { "arguments" } }; + if (auto res = builder.try_find_regex(prefix)) { + builder.move_back(rstrip_prefix); + auto tool_calls = builder.consume_json_with_dumped_args(args_paths); + if (!builder.add_tool_calls(tool_calls.value) || tool_calls.is_partial) { + throw common_chat_msg_partial_exception("incomplete tool call array"); + } + } else { + builder.add_content(builder.consume_rest()); + } +} + +static std::string wrap_code_as_arguments(common_chat_msg_parser & builder, const std::string & code) { + std::string arguments; + if (builder.is_partial()) { + arguments = (json{ + { "code", code + builder.healing_marker() } + }) + .dump(); + auto idx = arguments.find(builder.healing_marker()); + if (idx != std::string::npos) { + arguments.resize(idx); + } + } else { + arguments = (json{ + { "code", code } + }) + .dump(); + } + return arguments; +} + +/** + * Takes a prefix regex that must have 1 group to capture the function name, a closing suffix, and expects json parameters in between. + * Aggregates the prefix, suffix and in-between text into the content. + */ +static void parse_json_tool_calls( + common_chat_msg_parser & builder, + const std::optional & block_open, + const std::optional & function_regex_start_only, + const std::optional & function_regex, + const common_regex & close_regex, + const std::optional & block_close, + bool allow_raw_python = false, + const std::function & get_function_name = + nullptr) { + auto parse_tool_calls = [&]() { + size_t from = std::string::npos; + auto first = true; + while (true) { + auto start_pos = builder.pos(); + auto res = function_regex_start_only && first ? builder.try_consume_regex(*function_regex_start_only) : + function_regex ? builder.try_find_regex(*function_regex, from) : + std::nullopt; + + if (res) { + std::string name; + if (get_function_name) { + name = get_function_name(*res); + } else { + GGML_ASSERT(res->groups.size() == 2); + name = builder.str(res->groups[1]); + } + first = false; + if (name.empty()) { + // get_function_name signalled us that we should skip this match and treat it as content. + from = res->groups[0].begin + 1; + continue; + } + from = std::string::npos; + + auto maybe_raw_python = name == "python" && allow_raw_python; + if (builder.input()[builder.pos()] == '{' || !maybe_raw_python) { + if (auto arguments = builder.try_consume_json_with_dumped_args({ {} })) { + if (!builder.add_tool_call(name, "", arguments->value) || arguments->is_partial) { + throw common_chat_msg_partial_exception("incomplete tool call"); + } + builder.consume_regex(close_regex); + } + continue; + } + if (maybe_raw_python) { + auto arguments = wrap_code_as_arguments(builder, builder.consume_rest()); + if (!builder.add_tool_call(name, "", arguments)) { + throw common_chat_msg_partial_exception("incomplete tool call"); + } + return; + } + throw common_chat_msg_partial_exception("incomplete tool call"); + } else { + builder.move_to(start_pos); + } + break; + } + if (block_close) { + builder.consume_regex(*block_close); + } + builder.consume_spaces(); + builder.add_content(builder.consume_rest()); + }; + if (block_open) { + if (auto res = builder.try_find_regex(*block_open)) { + parse_tool_calls(); + } else { + builder.add_content(builder.consume_rest()); + } + } else { + parse_tool_calls(); + } +} + common_chat_msg_parser::common_chat_msg_parser(const std::string & input, bool is_partial, const common_chat_syntax & syntax) : input_(input), is_partial_(is_partial), syntax_(syntax) { @@ -532,3 +646,857 @@ std::optional common_chat_msg_parse void common_chat_msg_parser::clear_tools() { result_.tool_calls.clear(); } + +/** + * All common_chat_parse_* moved from chat.cpp to chat-parser.cpp below + * to reduce incremental compile time for parser changes. + */ +static void common_chat_parse_generic(common_chat_msg_parser & builder) { + if (!builder.syntax().parse_tool_calls) { + builder.add_content(builder.consume_rest()); + return; + } + static const std::vector> content_paths = { + {"response"}, + }; + static const std::vector> args_paths = { + {"tool_call", "arguments"}, + {"tool_calls", "arguments"}, + }; + auto data = builder.consume_json_with_dumped_args(args_paths, content_paths); + if (data.value.contains("tool_calls")) { + if (!builder.add_tool_calls(data.value.at("tool_calls")) || data.is_partial) { + throw common_chat_msg_partial_exception("incomplete tool calls"); + } + } else if (data.value.contains("tool_call")) { + if (!builder.add_tool_call(data.value.at("tool_call")) || data.is_partial) { + throw common_chat_msg_partial_exception("incomplete tool call"); + } + } else if (data.value.contains("response")) { + const auto & response = data.value.at("response"); + builder.add_content(response.is_string() ? response.template get() : response.dump(2)); + if (data.is_partial) { + throw common_chat_msg_partial_exception("incomplete response"); + } + } else { + throw common_chat_msg_partial_exception("Expected 'tool_call', 'tool_calls' or 'response' in JSON"); + } +} + +static void common_chat_parse_mistral_nemo(common_chat_msg_parser & builder) { + if (!builder.syntax().parse_tool_calls) { + builder.add_content(builder.consume_rest()); + return; + } + + static const common_regex prefix(regex_escape("[TOOL_CALLS]")); + parse_prefixed_json_tool_call_array(builder, prefix); +} + +static void common_chat_parse_magistral(common_chat_msg_parser & builder) { + builder.try_parse_reasoning("[THINK]", "[/THINK]"); + + if (!builder.syntax().parse_tool_calls) { + builder.add_content(builder.consume_rest()); + return; + } + + static const common_regex prefix(regex_escape("[TOOL_CALLS]")); + parse_prefixed_json_tool_call_array(builder, prefix); +} + +static void common_chat_parse_command_r7b(common_chat_msg_parser & builder) { + builder.try_parse_reasoning("<|START_THINKING|>", "<|END_THINKING|>"); + + static const common_regex start_action_regex("<\\|START_ACTION\\|>"); + static const common_regex end_action_regex("<\\|END_ACTION\\|>"); + static const common_regex start_response_regex("<\\|START_RESPONSE\\|>"); + static const common_regex end_response_regex("<\\|END_RESPONSE\\|>"); + + if (auto res = builder.try_find_regex(start_action_regex)) { + // If we didn't extract thoughts, prelude includes them. + auto tool_calls = builder.consume_json_with_dumped_args({{"parameters"}}); + for (const auto & tool_call : tool_calls.value) { + std::string name = tool_call.contains("tool_name") ? tool_call.at("tool_name") : ""; + std::string id = tool_call.contains("tool_call_id") ? tool_call.at("tool_call_id") : ""; + std::string arguments = tool_call.contains("parameters") ? tool_call.at("parameters") : ""; + if (!builder.add_tool_call(name, id, arguments) || tool_calls.is_partial) { + throw common_chat_msg_partial_exception("incomplete tool call"); + } + } + if (tool_calls.is_partial) { + throw common_chat_msg_partial_exception("incomplete tool call"); + } + builder.consume_regex(end_action_regex); + } else if (auto res = builder.try_find_regex(start_response_regex)) { + if (!builder.try_find_regex(end_response_regex)) { + builder.add_content(builder.consume_rest()); + throw common_chat_msg_partial_exception(end_response_regex.str()); + } + } else { + builder.add_content(builder.consume_rest()); + } +} + +static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool with_builtin_tools = false) { + builder.try_parse_reasoning("", ""); + + if (!builder.syntax().parse_tool_calls) { + builder.add_content(builder.consume_rest()); + return; + } + + static const common_regex function_regex( + "\\s*\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"([^\"]+)\"\\s*,\\s*\"parameters\"\\s*: "); + static const common_regex close_regex("\\}\\s*"); + + static const common_regex function_name_regex("\\s*(\\w+)\\s*\\.\\s*call\\("); + static const common_regex arg_name_regex("\\s*(\\w+)\\s*=\\s*"); + + if (with_builtin_tools) { + static const common_regex builtin_call_regex("<\\|python_tag\\|>"); + if (auto res = builder.try_find_regex(builtin_call_regex)) { + auto fun_res = builder.consume_regex(function_name_regex); + auto function_name = builder.str(fun_res.groups[1]); + + common_healing_marker healing_marker; + json args = json::object(); + while (true) { + if (auto arg_res = builder.try_consume_regex(arg_name_regex)) { + auto arg_name = builder.str(arg_res->groups[1]); + auto partial = builder.consume_json(); + args[arg_name] = partial.json; + healing_marker.marker = partial.healing_marker.marker; + healing_marker.json_dump_marker = partial.healing_marker.json_dump_marker; + builder.consume_spaces(); + if (!builder.try_consume_literal(",")) { + break; + } + } else { + break; + } + } + builder.consume_literal(")"); + builder.consume_spaces(); + + auto arguments = args.dump(); + if (!builder.add_tool_call(function_name, "", arguments)) { + throw common_chat_msg_partial_exception("Incomplete tool call"); + } + return; + } + } + parse_json_tool_calls( + builder, + /* block_open= */ std::nullopt, + /* function_regex_start_only= */ function_regex, + /* function_regex= */ std::nullopt, + close_regex, + std::nullopt); + +} + +static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) { + builder.try_parse_reasoning("", ""); + if (!builder.syntax().parse_tool_calls) { + builder.add_content(builder.consume_rest()); + return; + } + + static const common_regex tool_calls_begin("(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)"); + static const common_regex tool_calls_end("<｜tool▁calls▁end｜>"); + static const common_regex function_regex("(?:<｜tool▁call▁begin｜>)?function<｜tool▁sep｜>([^\n]+)\n```json\n"); + static const common_regex close_regex("```[\\s\\r\\n]*<｜tool▁call▁end｜>"); + + parse_json_tool_calls( + builder, + /* block_open= */ tool_calls_begin, + /* function_regex_start_only= */ std::nullopt, + function_regex, + close_regex, + tool_calls_end); +} + +static void common_chat_parse_deepseek_v3_1_content(common_chat_msg_parser & builder) { + static const common_regex function_regex("(?:<｜tool▁call▁begin｜>)?([^\\n<]+)(?:<｜tool▁sep｜>)"); + + static const common_regex close_regex("(?:[\\s]*)?<｜tool▁call▁end｜>"); + static const common_regex tool_calls_begin("(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)"); + static const common_regex tool_calls_end("<｜tool▁calls▁end｜>"); + + if (!builder.syntax().parse_tool_calls) { + LOG_DBG("%s: not parse_tool_calls\n", __func__); + builder.add_content(builder.consume_rest()); + return; + } + + LOG_DBG("%s: parse_tool_calls\n", __func__); + + parse_json_tool_calls( + builder, + /* block_open= */ tool_calls_begin, + /* function_regex_start_only= */ std::nullopt, + function_regex, + close_regex, + tool_calls_end); +} + +static void common_chat_parse_deepseek_v3_1(common_chat_msg_parser & builder) { + // DeepSeek V3.1 outputs reasoning content between "" and "" tags, followed by regular content + // First try to parse using the standard reasoning parsing method + LOG_DBG("%s: thinking_forced_open: %s\n", __func__, std::to_string(builder.syntax().thinking_forced_open).c_str()); + + auto start_pos = builder.pos(); + auto found_end_think = builder.try_find_literal(""); + builder.move_to(start_pos); + + if (builder.syntax().thinking_forced_open && !builder.is_partial() && !found_end_think) { + LOG_DBG("%s: no end_think, not partial, adding content\n", __func__); + common_chat_parse_deepseek_v3_1_content(builder); + } else if (builder.try_parse_reasoning("", "")) { + // If reasoning was parsed successfully, the remaining content is regular content + LOG_DBG("%s: parsed reasoning, adding content\n", __func__); + // <｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>NAME\n```json\nJSON\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜> + common_chat_parse_deepseek_v3_1_content(builder); + } else { + if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE) { + LOG_DBG("%s: reasoning_format none, adding content\n", __func__); + common_chat_parse_deepseek_v3_1_content(builder); + return; + } + // If no reasoning tags found, check if we should treat everything as reasoning + if (builder.syntax().thinking_forced_open) { + // If thinking is forced open but no tags found, treat everything as reasoning + LOG_DBG("%s: thinking_forced_open, adding reasoning content\n", __func__); + builder.add_reasoning_content(builder.consume_rest()); + } else { + LOG_DBG("%s: no thinking_forced_open, adding content\n", __func__); + // <｜tool▁call▁begin｜>NAME<｜tool▁sep｜>JSON<｜tool▁call▁end｜> + common_chat_parse_deepseek_v3_1_content(builder); + } + } +} + +static void common_chat_parse_minimax_m2(common_chat_msg_parser & builder) { + static const xml_tool_call_format form { + /* form.scope_start = */ "", + /* form.tool_start = */ "", + /* form.key_start = */ "", + /* form.val_end = */ "", + /* form.tool_end = */ "", + /* form.scope_end = */ "", + }; + builder.consume_reasoning_with_xml_tool_calls(form, "", ""); +} + +static void common_chat_parse_qwen3_coder_xml(common_chat_msg_parser & builder) { + static const xml_tool_call_format form = ([]() { + xml_tool_call_format form {}; + form.scope_start = ""; + form.tool_start = "", ""); +} + +static void common_chat_parse_apriel_1_5(common_chat_msg_parser & builder) { + static const xml_tool_call_format form = ([]() { + xml_tool_call_format form {}; + form.scope_start = "["; + form.tool_start = "{\"name\": \""; + form.tool_sep = "\", \"arguments\": {"; + form.key_start = "\""; + form.key_val_sep = "\": "; + form.val_end = ", "; + form.tool_end = "}, "; + form.scope_end = "]"; + form.raw_argval = false; + form.last_val_end = ""; + form.last_tool_end = "}"; + return form; + })(); + builder.consume_reasoning_with_xml_tool_calls(form, "", ""); +} + +static void common_chat_parse_xiaomi_mimo(common_chat_msg_parser & builder) { + static const xml_tool_call_format form = ([]() { + xml_tool_call_format form {}; + form.scope_start = ""; + form.tool_start = "\n{\"name\": \""; + form.tool_sep = "\", \"arguments\": {"; + form.key_start = "\""; + form.key_val_sep = "\": "; + form.val_end = ", "; + form.tool_end = "}\n"; + form.scope_end = ""; + form.raw_argval = false; + form.last_val_end = ""; + return form; + })(); + builder.consume_reasoning_with_xml_tool_calls(form); +} + +static void common_chat_parse_gpt_oss(common_chat_msg_parser & builder) { + static const std::string constraint = "(?: (<\\|constrain\\|>)?([a-zA-Z0-9_-]+))"; + static const std::string recipient("(?: to=functions\\.([^<\\s]+))"); + + static const common_regex start_regex("<\\|start\\|>assistant"); + static const common_regex analysis_regex("<\\|channel\\|>analysis"); + static const common_regex final_regex("<\\|channel\\|>final" + constraint + "?"); + static const common_regex preamble_regex("<\\|channel\\|>commentary"); + static const common_regex tool_call1_regex(recipient + "<\\|channel\\|>(analysis|commentary)" + constraint + "?"); + static const common_regex tool_call2_regex("<\\|channel\\|>(analysis|commentary)" + recipient + constraint + "?"); + + auto consume_end = [&](bool include_end = false) { + if (auto res = builder.try_find_literal("<|end|>")) { + return res->prelude + (include_end ? builder.str(res->groups[0]) : ""); + } + return builder.consume_rest(); + }; + + auto handle_tool_call = [&](const std::string & name) { + if (auto args = builder.try_consume_json_with_dumped_args({{}})) { + if (builder.syntax().parse_tool_calls) { + if (!builder.add_tool_call(name, "", args->value) || args->is_partial) { + throw common_chat_msg_partial_exception("incomplete tool call"); + } + } else if (args->is_partial) { + throw common_chat_msg_partial_exception("incomplete tool call"); + } + } + }; + + auto regex_match = [](const common_regex & regex, const std::string & input) -> std::optional { + auto match = regex.search(input, 0, true); + if (match.type == COMMON_REGEX_MATCH_TYPE_FULL) { + return match; + } + return std::nullopt; + }; + + do { + auto header_start_pos = builder.pos(); + auto content_start = builder.try_find_literal("<|message|>"); + if (!content_start) { + throw common_chat_msg_partial_exception("incomplete header"); + } + + auto header = content_start->prelude; + + if (auto match = regex_match(tool_call1_regex, header)) { + auto group = match->groups[1]; + auto name = header.substr(group.begin, group.end - group.begin); + handle_tool_call(name); + continue; + } + + if (auto match = regex_match(tool_call2_regex, header)) { + auto group = match->groups[2]; + auto name = header.substr(group.begin, group.end - group.begin); + handle_tool_call(name); + continue; + } + + if (regex_match(analysis_regex, header)) { + builder.move_to(header_start_pos); + if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE || builder.syntax().reasoning_in_content) { + builder.add_content(consume_end(true)); + } else { + builder.try_parse_reasoning("<|channel|>analysis<|message|>", "<|end|>"); + } + continue; + } + + if(regex_match(final_regex, header) || regex_match(preamble_regex, header)) { + builder.add_content(consume_end()); + continue; + } + + // Possibly a malformed message, attempt to recover by rolling + // back to pick up the next <|start|> + LOG_DBG("%s: unknown header from message: %s\n", __func__, header.c_str()); + builder.move_to(header_start_pos); + } while (builder.try_find_regex(start_regex, std::string::npos, false)); + + auto remaining = builder.consume_rest(); + if (!remaining.empty()) { + LOG_DBG("%s: content after last message: %s\n", __func__, remaining.c_str()); + } +} + +static void common_chat_parse_glm_4_5(common_chat_msg_parser & builder) { + static const xml_tool_call_format form { + /* form.scope_start = */ "", + /* form.tool_start = */ "", + /* form.tool_sep = */ "", + /* form.key_start = */ "", + /* form.key_val_sep = */ "", + /* form.val_end = */ "", + /* form.tool_end = */ "", + /* form.scope_end = */ "", + /* form.key_val_sep2 = */ "", + }; + builder.consume_reasoning_with_xml_tool_calls(form, "", ""); +} + +static void common_chat_parse_firefunction_v2(common_chat_msg_parser & builder) { + if (!builder.syntax().parse_tool_calls) { + builder.add_content(builder.consume_rest()); + return; + } + static const common_regex prefix(regex_escape(" functools[")); + parse_prefixed_json_tool_call_array(builder, prefix, /* rstrip_prefix= */ 1); +} + +static void common_chat_parse_functionary_v3_2(common_chat_msg_parser & builder) { + static const common_regex function_regex_start_only(R"((\w+\n\{|python\n|all\n))"); + static const common_regex function_regex(R"(>>>(\w+\n\{|python\n|all\n))"); + static const common_regex close_regex(R"(\s*)"); + + parse_json_tool_calls( + builder, + std::nullopt, + function_regex_start_only, + function_regex, + close_regex, + std::nullopt, + /* allow_raw_python= */ true, + /* get_function_name= */ [&](const auto & res) -> std::string { + auto at_start = res.groups[0].begin == 0; + auto name = builder.str(res.groups[1]); + if (!name.empty() && name.back() == '{') { + // Unconsume the opening brace '{' to ensure the JSON parsing goes well. + builder.move_back(1); + } + auto idx = name.find_last_not_of("\n{"); + name = name.substr(0, idx + 1); + if (at_start && name == "all") { + return ""; + } + return name; + }); +} + +static void common_chat_parse_functionary_v3_1_llama_3_1(common_chat_msg_parser & builder) { + if (!builder.syntax().parse_tool_calls) { + builder.add_content(builder.consume_rest()); + return; + } + // This version of Functionary still supports the llama 3.1 tool call format for the python tool. + static const common_regex python_tag_regex(regex_escape("<|python_tag|>")); + + static const common_regex function_regex(R"()"); + static const common_regex close_regex(R"()"); + + parse_json_tool_calls( + builder, + /* block_open= */ std::nullopt, + /* function_regex_start_only= */ std::nullopt, + function_regex, + close_regex, + std::nullopt); + + if (auto res = builder.try_find_regex(python_tag_regex)) { + auto arguments = wrap_code_as_arguments(builder, builder.consume_rest()); + builder.add_tool_call("python", "", arguments); + return; + } +} + +static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) { + builder.try_parse_reasoning("", ""); + if (!builder.syntax().parse_tool_calls) { + builder.add_content(builder.consume_rest()); + return; + } + + static const common_regex open_regex( + "(?:" + "(```(?:xml|json)?\\n\\s*)?" // match 1 (block_start) + "(" // match 2 (open_tag) + "" + "|" + "|" + "|" + "|" + "|" + "|" + "|" + ")?" + "(\\s*\\{\\s*\"name\")" // match 3 (named tool call) + ")" + "|]+)>" // match 4 (function name) + "|" // match 5 (function name again) + ); + + while (auto res = builder.try_find_regex(open_regex)) { + const auto & block_start = res->groups[1]; + std::string block_end = block_start.empty() ? "" : "```"; + + const auto & open_tag = res->groups[2]; + std::string close_tag; + + if (!res->groups[3].empty()) { + builder.move_to(res->groups[3].begin); + close_tag = open_tag.empty() ? "" : "value) || tool_call->is_partial) { + throw common_chat_msg_partial_exception("incomplete tool call"); + } + builder.consume_spaces(); + builder.consume_literal(close_tag); + builder.consume_spaces(); + if (!block_end.empty()) { + builder.consume_literal(block_end); + builder.consume_spaces(); + } + } else { + throw common_chat_msg_partial_exception("failed to parse tool call"); + } + } else { + auto function_name = builder.str(res->groups[4]); + if (function_name.empty()) { + function_name = builder.str(res->groups[5]); + } + GGML_ASSERT(!function_name.empty()); + + close_tag = ""; + + if (auto arguments = builder.try_consume_json_with_dumped_args({{}})) { + if (!builder.add_tool_call(function_name, "", arguments->value) || arguments->is_partial) { + throw common_chat_msg_partial_exception("incomplete tool call"); + } + builder.consume_spaces(); + builder.consume_literal(close_tag); + builder.consume_spaces(); + if (!block_end.empty()) { + builder.consume_literal(block_end); + builder.consume_spaces(); + } + } + } + } + + builder.add_content(builder.consume_rest()); +} + +static void common_chat_parse_granite(common_chat_msg_parser & builder) { + // Parse thinking tags + static const common_regex start_think_regex(regex_escape("")); + static const common_regex end_think_regex(regex_escape("")); + // Granite models output partial tokens such as "<" and "groups[0].begin); + builder.try_find_regex(end_think_regex, std::string::npos, false); + // Restore position for try_parse_reasoning() + builder.move_to(res->groups[0].begin); + } + builder.try_parse_reasoning("", ""); + + // Parse response tags + static const common_regex start_response_regex(regex_escape("")); + static const common_regex end_response_regex(regex_escape("")); + // Granite models output partial tokens such as "<" and "")); + if (auto res = builder.try_find_regex(tool_call_regex)) { + builder.move_to(res->groups[0].end); + + // Expect JSON array of tool calls + if (auto tool_call = builder.try_consume_json_with_dumped_args({{{"arguments"}}})) { + if (!builder.add_tool_calls(tool_call->value) || tool_call->is_partial) { + throw common_chat_msg_partial_exception("incomplete tool call"); + } + } + } else { + builder.add_content(builder.consume_rest()); + } +} + +static void common_chat_parse_nemotron_v2(common_chat_msg_parser & builder) { + // Parse thinking tags + builder.try_parse_reasoning("", ""); + if (!builder.syntax().parse_tool_calls) { + builder.add_content(builder.consume_rest()); + return; + } + + // Look for tool calls + static const common_regex tool_call_regex(regex_escape("")); + if (auto res = builder.try_find_regex(tool_call_regex)) { + builder.move_to(res->groups[0].end); + + // Expect JSON array of tool calls + auto tool_calls_data = builder.consume_json(); + if (tool_calls_data.json.is_array()) { + if (!builder.try_consume_literal("")) { + throw common_chat_msg_partial_exception("Incomplete tool call"); + } + builder.add_tool_calls(tool_calls_data.json); + } else { + throw common_chat_msg_partial_exception("Incomplete tool call"); + } + } + builder.add_content(builder.consume_rest()); +} + +static void common_chat_parse_apertus(common_chat_msg_parser & builder) { + // Parse thinking tags + builder.try_parse_reasoning("<|inner_prefix|>", "<|inner_suffix|>"); + if (!builder.syntax().parse_tool_calls) { + builder.add_content(builder.consume_rest()); + return; + } + + // Look for tool calls + static const common_regex tool_call_regex(regex_escape("<|tools_prefix|>")); + if (auto res = builder.try_find_regex(tool_call_regex)) { + builder.move_to(res->groups[0].end); + + auto tool_calls_data = builder.consume_json(); + if (tool_calls_data.json.is_array()) { + builder.consume_spaces(); + if (!builder.try_consume_literal("<|tools_suffix|>")) { + throw common_chat_msg_partial_exception("Incomplete tool call"); + } + for (const auto & value : tool_calls_data.json) { + if (value.is_object()) { + builder.add_tool_call_short_form(value); + } + } + } else { + throw common_chat_msg_partial_exception("Incomplete tool call"); + } + } + builder.add_content(builder.consume_rest()); +} + + +static void common_chat_parse_lfm2(common_chat_msg_parser & builder) { + if (!builder.syntax().parse_tool_calls) { + builder.add_content(builder.consume_rest()); + return; + } + + // LFM2 format: <|tool_call_start|>[{"name": "get_current_time", "arguments": {"location": "Paris"}}]<|tool_call_end|> + static const common_regex tool_call_start_regex(regex_escape("<|tool_call_start|>")); + static const common_regex tool_call_end_regex(regex_escape("<|tool_call_end|>")); + + // Loop through all tool calls + while (auto res = builder.try_find_regex(tool_call_start_regex, std::string::npos, /* add_prelude_to_content= */ true)) { + builder.move_to(res->groups[0].end); + + // Parse JSON array format: [{"name": "...", "arguments": {...}}] + auto tool_calls_data = builder.consume_json(); + + // Consume end marker + builder.consume_spaces(); + if (!builder.try_consume_regex(tool_call_end_regex)) { + throw common_chat_msg_partial_exception("Expected <|tool_call_end|>"); + } + + // Process each tool call in the array + if (tool_calls_data.json.is_array()) { + for (const auto & tool_call : tool_calls_data.json) { + if (!tool_call.is_object()) { + throw common_chat_msg_partial_exception("Tool call must be an object"); + } + + if (!tool_call.contains("name")) { + throw common_chat_msg_partial_exception("Tool call missing 'name' field"); + } + + std::string function_name = tool_call.at("name"); + std::string arguments = "{}"; + + if (tool_call.contains("arguments")) { + if (tool_call.at("arguments").is_object()) { + arguments = tool_call.at("arguments").dump(); + } else if (tool_call.at("arguments").is_string()) { + arguments = tool_call.at("arguments"); + } + } + + if (!builder.add_tool_call(function_name, "", arguments)) { + throw common_chat_msg_partial_exception("Incomplete tool call"); + } + } + } else { + throw common_chat_msg_partial_exception("Expected JSON array for tool calls"); + } + + // Consume any trailing whitespace after this tool call + builder.consume_spaces(); + } + + // Consume any remaining content after all tool calls + auto remaining = builder.consume_rest(); + if (!string_strip(remaining).empty()) { + builder.add_content(remaining); + } +} + +static void common_chat_parse_seed_oss(common_chat_msg_parser & builder) { + static const xml_tool_call_format form { + /* form.scope_start = */ "", + /* form.tool_start = */ "", + /* form.key_start = */ "", + /* form.val_end = */ "", + /* form.tool_end = */ "", + /* form.scope_end = */ "", + }; + builder.consume_reasoning_with_xml_tool_calls(form, "", ""); +} + +static void common_chat_parse_content_only(common_chat_msg_parser & builder) { + builder.try_parse_reasoning("", ""); + builder.add_content(builder.consume_rest()); +} + +static void common_chat_parse(common_chat_msg_parser & builder) { + LOG_DBG("Parsing input with format %s: %s\n", common_chat_format_name(builder.syntax().format), builder.input().c_str()); + + switch (builder.syntax().format) { + case COMMON_CHAT_FORMAT_CONTENT_ONLY: + common_chat_parse_content_only(builder); + break; + case COMMON_CHAT_FORMAT_GENERIC: + common_chat_parse_generic(builder); + break; + case COMMON_CHAT_FORMAT_MISTRAL_NEMO: + common_chat_parse_mistral_nemo(builder); + break; + case COMMON_CHAT_FORMAT_MAGISTRAL: + common_chat_parse_magistral(builder); + break; + case COMMON_CHAT_FORMAT_LLAMA_3_X: + common_chat_parse_llama_3_1(builder); + break; + case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS: + common_chat_parse_llama_3_1(builder, /* with_builtin_tools= */ true); + break; + case COMMON_CHAT_FORMAT_DEEPSEEK_R1: + common_chat_parse_deepseek_r1(builder); + break; + case COMMON_CHAT_FORMAT_DEEPSEEK_V3_1: + common_chat_parse_deepseek_v3_1(builder); + break; + case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2: + common_chat_parse_functionary_v3_2(builder); + break; + case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1: + common_chat_parse_functionary_v3_1_llama_3_1(builder); + break; + case COMMON_CHAT_FORMAT_HERMES_2_PRO: + common_chat_parse_hermes_2_pro(builder); + break; + case COMMON_CHAT_FORMAT_FIREFUNCTION_V2: + common_chat_parse_firefunction_v2(builder); + break; + case COMMON_CHAT_FORMAT_COMMAND_R7B: + common_chat_parse_command_r7b(builder); + break; + case COMMON_CHAT_FORMAT_GRANITE: + common_chat_parse_granite(builder); + break; + case COMMON_CHAT_FORMAT_GPT_OSS: + common_chat_parse_gpt_oss(builder); + break; + case COMMON_CHAT_FORMAT_SEED_OSS: + common_chat_parse_seed_oss(builder); + break; + case COMMON_CHAT_FORMAT_NEMOTRON_V2: + common_chat_parse_nemotron_v2(builder); + break; + case COMMON_CHAT_FORMAT_APERTUS: + common_chat_parse_apertus(builder); + break; + case COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS: + common_chat_parse_lfm2(builder); + break; + case COMMON_CHAT_FORMAT_MINIMAX_M2: + common_chat_parse_minimax_m2(builder); + break; + case COMMON_CHAT_FORMAT_GLM_4_5: + common_chat_parse_glm_4_5(builder); + break; + case COMMON_CHAT_FORMAT_KIMI_K2: + common_chat_parse_kimi_k2(builder); + break; + case COMMON_CHAT_FORMAT_QWEN3_CODER_XML: + common_chat_parse_qwen3_coder_xml(builder); + break; + case COMMON_CHAT_FORMAT_APRIEL_1_5: + common_chat_parse_apriel_1_5(builder); + break; + case COMMON_CHAT_FORMAT_XIAOMI_MIMO: + common_chat_parse_xiaomi_mimo(builder); + break; + default: + throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format)); + } + builder.finish(); +} + +common_chat_msg common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax) { + common_chat_msg_parser builder(input, is_partial, syntax); + try { + common_chat_parse(builder); + } catch (const common_chat_msg_partial_exception & ex) { + LOG_DBG("Partial parse: %s\n", ex.what()); + if (!is_partial) { + builder.clear_tools(); + builder.move_to(0); + common_chat_parse_content_only(builder); + } + } + auto msg = builder.result(); + if (!is_partial) { + LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat({msg}).at(0).dump().c_str()); + } + return msg; +} diff --git a/common/chat.cpp b/common/chat.cpp index 6fa05a604..b4a0f985e 100644 --- a/common/chat.cpp +++ b/common/chat.cpp @@ -678,114 +678,6 @@ common_reasoning_format common_reasoning_format_from_name(const std::string & fo throw std::runtime_error("Unknown reasoning format: " + format); } -static std::string wrap_code_as_arguments(common_chat_msg_parser & builder, const std::string & code) { - std::string arguments; - if (builder.is_partial()) { - arguments = (json {{"code", code + builder.healing_marker()}}).dump(); - auto idx = arguments.find(builder.healing_marker()); - if (idx != std::string::npos) { - arguments.resize(idx); - } - } else { - arguments = (json {{"code", code}}).dump(); - } - return arguments; -} - -/** - * Takes a prefix regex that must have 1 group to capture the function name, a closing suffix, and expects json parameters in between. - * Aggregates the prefix, suffix and in-between text into the content. - */ -static void parse_json_tool_calls( - common_chat_msg_parser & builder, - const std::optional & block_open, - const std::optional & function_regex_start_only, - const std::optional & function_regex, - const common_regex & close_regex, - const std::optional & block_close, - bool allow_raw_python = false, - const std::function & get_function_name = nullptr) { - - auto parse_tool_calls = [&]() { - size_t from = std::string::npos; - auto first = true; - while (true) { - auto start_pos = builder.pos(); - auto res = function_regex_start_only && first - ? builder.try_consume_regex(*function_regex_start_only) - : function_regex - ? builder.try_find_regex(*function_regex, from) - : std::nullopt; - - if (res) { - std::string name; - if (get_function_name) { - name = get_function_name(*res); - } else { - GGML_ASSERT(res->groups.size() == 2); - name = builder.str(res->groups[1]); - } - first = false; - if (name.empty()) { - // get_function_name signalled us that we should skip this match and treat it as content. - from = res->groups[0].begin + 1; - continue; - } - from = std::string::npos; - - auto maybe_raw_python = name == "python" && allow_raw_python; - if (builder.input()[builder.pos()] == '{' || !maybe_raw_python) { - if (auto arguments = builder.try_consume_json_with_dumped_args({{}})) { - if (!builder.add_tool_call(name, "", arguments->value) || arguments->is_partial) { - throw common_chat_msg_partial_exception("incomplete tool call"); - } - builder.consume_regex(close_regex); - } - continue; - } - if (maybe_raw_python) { - auto arguments = wrap_code_as_arguments(builder, builder.consume_rest()); - if (!builder.add_tool_call(name, "", arguments)) { - throw common_chat_msg_partial_exception("incomplete tool call"); - } - return; - } - throw common_chat_msg_partial_exception("incomplete tool call"); - } else { - builder.move_to(start_pos); - } - break; - } - if (block_close) { - builder.consume_regex(*block_close); - } - builder.consume_spaces(); - builder.add_content(builder.consume_rest()); - }; - if (block_open) { - if (auto res = builder.try_find_regex(*block_open)) { - parse_tool_calls(); - } else { - builder.add_content(builder.consume_rest()); - } - } else { - parse_tool_calls(); - } -} - -static void parse_prefixed_json_tool_call_array(common_chat_msg_parser & builder, const common_regex & prefix, size_t rstrip_prefix = 0) { - static const std::vector> args_paths = {{"arguments"}}; - if (auto res = builder.try_find_regex(prefix)) { - builder.move_back(rstrip_prefix); - auto tool_calls = builder.consume_json_with_dumped_args(args_paths); - if (!builder.add_tool_calls(tool_calls.value) || tool_calls.is_partial) { - throw common_chat_msg_partial_exception("incomplete tool call array"); - } - } else { - builder.add_content(builder.consume_rest()); - } -} - static void foreach_function(const json & tools, const std::function & fn) { for (const auto & tool : tools) { if (!tool.contains("type") || tool.at("type") != "function" || !tool.contains("function")) { @@ -918,37 +810,6 @@ static common_chat_params common_chat_params_init_generic(const common_chat_temp data.format = COMMON_CHAT_FORMAT_GENERIC; return data; } -static void common_chat_parse_generic(common_chat_msg_parser & builder) { - if (!builder.syntax().parse_tool_calls) { - builder.add_content(builder.consume_rest()); - return; - } - static const std::vector> content_paths = { - {"response"}, - }; - static const std::vector> args_paths = { - {"tool_call", "arguments"}, - {"tool_calls", "arguments"}, - }; - auto data = builder.consume_json_with_dumped_args(args_paths, content_paths); - if (data.value.contains("tool_calls")) { - if (!builder.add_tool_calls(data.value.at("tool_calls")) || data.is_partial) { - throw common_chat_msg_partial_exception("incomplete tool calls"); - } - } else if (data.value.contains("tool_call")) { - if (!builder.add_tool_call(data.value.at("tool_call")) || data.is_partial) { - throw common_chat_msg_partial_exception("incomplete tool call"); - } - } else if (data.value.contains("response")) { - const auto & response = data.value.at("response"); - builder.add_content(response.is_string() ? response.template get() : response.dump(2)); - if (data.is_partial) { - throw common_chat_msg_partial_exception("incomplete response"); - } - } else { - throw common_chat_msg_partial_exception("Expected 'tool_call', 'tool_calls' or 'response' in JSON"); - } -} static common_chat_params common_chat_params_init_mistral_nemo(const common_chat_template & tmpl, const struct templates_params & inputs) { common_chat_params data; @@ -1173,28 +1034,6 @@ static common_chat_params common_chat_params_init_magistral(const common_chat_te return data; } -static void common_chat_parse_mistral_nemo(common_chat_msg_parser & builder) { - if (!builder.syntax().parse_tool_calls) { - builder.add_content(builder.consume_rest()); - return; - } - - static const common_regex prefix(regex_escape("[TOOL_CALLS]")); - parse_prefixed_json_tool_call_array(builder, prefix); -} - -static void common_chat_parse_magistral(common_chat_msg_parser & builder) { - builder.try_parse_reasoning("[THINK]", "[/THINK]"); - - if (!builder.syntax().parse_tool_calls) { - builder.add_content(builder.consume_rest()); - return; - } - - static const common_regex prefix(regex_escape("[TOOL_CALLS]")); - parse_prefixed_json_tool_call_array(builder, prefix); -} - static common_chat_params common_chat_params_init_command_r7b(const common_chat_template & tmpl, const struct templates_params & inputs) { common_chat_params data; @@ -1275,39 +1114,6 @@ static common_chat_params common_chat_params_init_command_r7b(const common_chat_ return data; } -static void common_chat_parse_command_r7b(common_chat_msg_parser & builder) { - builder.try_parse_reasoning("<|START_THINKING|>", "<|END_THINKING|>"); - - static const common_regex start_action_regex("<\\|START_ACTION\\|>"); - static const common_regex end_action_regex("<\\|END_ACTION\\|>"); - static const common_regex start_response_regex("<\\|START_RESPONSE\\|>"); - static const common_regex end_response_regex("<\\|END_RESPONSE\\|>"); - - if (auto res = builder.try_find_regex(start_action_regex)) { - // If we didn't extract thoughts, prelude includes them. - auto tool_calls = builder.consume_json_with_dumped_args({{"parameters"}}); - for (const auto & tool_call : tool_calls.value) { - std::string name = tool_call.contains("tool_name") ? tool_call.at("tool_name") : ""; - std::string id = tool_call.contains("tool_call_id") ? tool_call.at("tool_call_id") : ""; - std::string arguments = tool_call.contains("parameters") ? tool_call.at("parameters") : ""; - if (!builder.add_tool_call(name, id, arguments) || tool_calls.is_partial) { - throw common_chat_msg_partial_exception("incomplete tool call"); - } - } - if (tool_calls.is_partial) { - throw common_chat_msg_partial_exception("incomplete tool call"); - } - builder.consume_regex(end_action_regex); - } else if (auto res = builder.try_find_regex(start_response_regex)) { - if (!builder.try_find_regex(end_response_regex)) { - builder.add_content(builder.consume_rest()); - throw common_chat_msg_partial_exception(end_response_regex.str()); - } - } else { - builder.add_content(builder.consume_rest()); - } -} - static void expect_tool_parameters(const std::string & name, const json & parameters, const std::vector & expected_properties) { if (!parameters.is_object() || !parameters.contains("type") || parameters.at("type") != "object" || !parameters.contains("properties") || !parameters.contains("required")) { throw std::runtime_error("Parameters of tool " + name + " must be an object w/ required properties"); @@ -1536,63 +1342,6 @@ static common_chat_params common_chat_params_init_apertus(const common_chat_temp } return data; } -static void common_chat_parse_llama_3_1(common_chat_msg_parser & builder, bool with_builtin_tools = false) { - builder.try_parse_reasoning("", ""); - - if (!builder.syntax().parse_tool_calls) { - builder.add_content(builder.consume_rest()); - return; - } - - static const common_regex function_regex( - "\\s*\\{\\s*(?:\"type\"\\s*:\\s*\"function\"\\s*,\\s*)?\"name\"\\s*:\\s*\"([^\"]+)\"\\s*,\\s*\"parameters\"\\s*: "); - static const common_regex close_regex("\\}\\s*"); - - static const common_regex function_name_regex("\\s*(\\w+)\\s*\\.\\s*call\\("); - static const common_regex arg_name_regex("\\s*(\\w+)\\s*=\\s*"); - - if (with_builtin_tools) { - static const common_regex builtin_call_regex("<\\|python_tag\\|>"); - if (auto res = builder.try_find_regex(builtin_call_regex)) { - auto fun_res = builder.consume_regex(function_name_regex); - auto function_name = builder.str(fun_res.groups[1]); - - common_healing_marker healing_marker; - json args = json::object(); - while (true) { - if (auto arg_res = builder.try_consume_regex(arg_name_regex)) { - auto arg_name = builder.str(arg_res->groups[1]); - auto partial = builder.consume_json(); - args[arg_name] = partial.json; - healing_marker.marker = partial.healing_marker.marker; - healing_marker.json_dump_marker = partial.healing_marker.json_dump_marker; - builder.consume_spaces(); - if (!builder.try_consume_literal(",")) { - break; - } - } else { - break; - } - } - builder.consume_literal(")"); - builder.consume_spaces(); - - auto arguments = args.dump(); - if (!builder.add_tool_call(function_name, "", arguments)) { - throw common_chat_msg_partial_exception("Incomplete tool call"); - } - return; - } - } - parse_json_tool_calls( - builder, - /* block_open= */ std::nullopt, - /* function_regex_start_only= */ function_regex, - /* function_regex= */ std::nullopt, - close_regex, - std::nullopt); - -} static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct templates_params & inputs) { common_chat_params data; @@ -1732,88 +1481,6 @@ static common_chat_params common_chat_params_init_deepseek_v3_1(const common_cha return data; } -static void common_chat_parse_deepseek_r1(common_chat_msg_parser & builder) { - builder.try_parse_reasoning("", ""); - if (!builder.syntax().parse_tool_calls) { - builder.add_content(builder.consume_rest()); - return; - } - - static const common_regex tool_calls_begin("(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)"); - static const common_regex tool_calls_end("<｜tool▁calls▁end｜>"); - static const common_regex function_regex("(?:<｜tool▁call▁begin｜>)?function<｜tool▁sep｜>([^\n]+)\n```json\n"); - static const common_regex close_regex("```[\\s\\r\\n]*<｜tool▁call▁end｜>"); - - parse_json_tool_calls( - builder, - /* block_open= */ tool_calls_begin, - /* function_regex_start_only= */ std::nullopt, - function_regex, - close_regex, - tool_calls_end); -} - -static void common_chat_parse_deepseek_v3_1_content(common_chat_msg_parser & builder) { - static const common_regex function_regex("(?:<｜tool▁call▁begin｜>)?([^\\n<]+)(?:<｜tool▁sep｜>)"); - - static const common_regex close_regex("(?:[\\s]*)?<｜tool▁call▁end｜>"); - static const common_regex tool_calls_begin("(?:<｜tool▁calls▁begin｜>|<｜tool_calls_begin｜>|<｜tool calls begin｜>|<｜tool\\\\_calls\\\\_begin｜>|<｜tool▁calls｜>)"); - static const common_regex tool_calls_end("<｜tool▁calls▁end｜>"); - - if (!builder.syntax().parse_tool_calls) { - LOG_DBG("%s: not parse_tool_calls\n", __func__); - builder.add_content(builder.consume_rest()); - return; - } - - LOG_DBG("%s: parse_tool_calls\n", __func__); - - parse_json_tool_calls( - builder, - /* block_open= */ tool_calls_begin, - /* function_regex_start_only= */ std::nullopt, - function_regex, - close_regex, - tool_calls_end); -} - -static void common_chat_parse_deepseek_v3_1(common_chat_msg_parser & builder) { - // DeepSeek V3.1 outputs reasoning content between "" and "" tags, followed by regular content - // First try to parse using the standard reasoning parsing method - LOG_DBG("%s: thinking_forced_open: %s\n", __func__, std::to_string(builder.syntax().thinking_forced_open).c_str()); - - auto start_pos = builder.pos(); - auto found_end_think = builder.try_find_literal(""); - builder.move_to(start_pos); - - if (builder.syntax().thinking_forced_open && !builder.is_partial() && !found_end_think) { - LOG_DBG("%s: no end_think, not partial, adding content\n", __func__); - common_chat_parse_deepseek_v3_1_content(builder); - } else if (builder.try_parse_reasoning("", "")) { - // If reasoning was parsed successfully, the remaining content is regular content - LOG_DBG("%s: parsed reasoning, adding content\n", __func__); - // <｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function<｜tool▁sep｜>NAME\n```json\nJSON\n```<｜tool▁call▁end｜><｜tool▁calls▁end｜> - common_chat_parse_deepseek_v3_1_content(builder); - } else { - if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE) { - LOG_DBG("%s: reasoning_format none, adding content\n", __func__); - common_chat_parse_deepseek_v3_1_content(builder); - return; - } - // If no reasoning tags found, check if we should treat everything as reasoning - if (builder.syntax().thinking_forced_open) { - // If thinking is forced open but no tags found, treat everything as reasoning - LOG_DBG("%s: thinking_forced_open, adding reasoning content\n", __func__); - builder.add_reasoning_content(builder.consume_rest()); - } else { - LOG_DBG("%s: no thinking_forced_open, adding content\n", __func__); - // <｜tool▁call▁begin｜>NAME<｜tool▁sep｜>JSON<｜tool▁call▁end｜> - common_chat_parse_deepseek_v3_1_content(builder); - } - } -} - - static common_chat_params common_chat_params_init_minimax_m2(const common_chat_template & tmpl, const struct templates_params & params) { common_chat_params data; data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED; @@ -1856,20 +1523,6 @@ static common_chat_params common_chat_params_init_minimax_m2(const common_chat_t return data; } -static void common_chat_parse_minimax_m2(common_chat_msg_parser & builder) { - static const xml_tool_call_format form { - /* form.scope_start = */ "", - /* form.tool_start = */ "", - /* form.key_start = */ "", - /* form.val_end = */ "", - /* form.tool_end = */ "", - /* form.scope_end = */ "", - }; - builder.consume_reasoning_with_xml_tool_calls(form, "", ""); -} - static common_chat_params common_chat_params_init_qwen3_coder_xml(const common_chat_template & tmpl, const struct templates_params & params) { common_chat_params data; data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED; @@ -1902,23 +1555,6 @@ static common_chat_params common_chat_params_init_qwen3_coder_xml(const common_c return data; } -static void common_chat_parse_qwen3_coder_xml(common_chat_msg_parser & builder) { - static const xml_tool_call_format form = ([]() { - xml_tool_call_format form {}; - form.scope_start = ""; - form.tool_start = "", ""); -} - static common_chat_params common_chat_params_init_apriel_1_5(const common_chat_template & tmpl, const struct templates_params & params) { common_chat_params data; data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED; @@ -2016,25 +1634,6 @@ static common_chat_params common_chat_params_init_apriel_1_5(const common_chat_t return data; } -static void common_chat_parse_apriel_1_5(common_chat_msg_parser & builder) { - static const xml_tool_call_format form = ([]() { - xml_tool_call_format form {}; - form.scope_start = "["; - form.tool_start = "{\"name\": \""; - form.tool_sep = "\", \"arguments\": {"; - form.key_start = "\""; - form.key_val_sep = "\": "; - form.val_end = ", "; - form.tool_end = "}, "; - form.scope_end = "]"; - form.raw_argval = false; - form.last_val_end = ""; - form.last_tool_end = "}"; - return form; - })(); - builder.consume_reasoning_with_xml_tool_calls(form, "", ""); -} - static common_chat_params common_chat_params_init_xiaomi_mimo(const common_chat_template & tmpl, const struct templates_params & params) { common_chat_params data; data.grammar_lazy = params.tools.is_array() && !params.tools.empty() && params.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED; @@ -2067,24 +1666,6 @@ static common_chat_params common_chat_params_init_xiaomi_mimo(const common_chat_ return data; } -static void common_chat_parse_xiaomi_mimo(common_chat_msg_parser & builder) { - static const xml_tool_call_format form = ([]() { - xml_tool_call_format form {}; - form.scope_start = ""; - form.tool_start = "\n{\"name\": \""; - form.tool_sep = "\", \"arguments\": {"; - form.key_start = "\""; - form.key_val_sep = "\": "; - form.val_end = ", "; - form.tool_end = "}\n"; - form.scope_end = ""; - form.raw_argval = false; - form.last_val_end = ""; - return form; - })(); - builder.consume_reasoning_with_xml_tool_calls(form); -} - static common_chat_params common_chat_params_init_gpt_oss(const common_chat_template & tmpl, const struct templates_params & inputs) { common_chat_params data; @@ -2231,93 +1812,6 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp return data; } -static void common_chat_parse_gpt_oss(common_chat_msg_parser & builder) { - static const std::string constraint = "(?: (<\\|constrain\\|>)?([a-zA-Z0-9_-]+))"; - static const std::string recipient("(?: to=functions\\.([^<\\s]+))"); - - static const common_regex start_regex("<\\|start\\|>assistant"); - static const common_regex analysis_regex("<\\|channel\\|>analysis"); - static const common_regex final_regex("<\\|channel\\|>final" + constraint + "?"); - static const common_regex preamble_regex("<\\|channel\\|>commentary"); - static const common_regex tool_call1_regex(recipient + "<\\|channel\\|>(analysis|commentary)" + constraint + "?"); - static const common_regex tool_call2_regex("<\\|channel\\|>(analysis|commentary)" + recipient + constraint + "?"); - - auto consume_end = [&](bool include_end = false) { - if (auto res = builder.try_find_literal("<|end|>")) { - return res->prelude + (include_end ? builder.str(res->groups[0]) : ""); - } - return builder.consume_rest(); - }; - - auto handle_tool_call = [&](const std::string & name) { - if (auto args = builder.try_consume_json_with_dumped_args({{}})) { - if (builder.syntax().parse_tool_calls) { - if (!builder.add_tool_call(name, "", args->value) || args->is_partial) { - throw common_chat_msg_partial_exception("incomplete tool call"); - } - } else if (args->is_partial) { - throw common_chat_msg_partial_exception("incomplete tool call"); - } - } - }; - - auto regex_match = [](const common_regex & regex, const std::string & input) -> std::optional { - auto match = regex.search(input, 0, true); - if (match.type == COMMON_REGEX_MATCH_TYPE_FULL) { - return match; - } - return std::nullopt; - }; - - do { - auto header_start_pos = builder.pos(); - auto content_start = builder.try_find_literal("<|message|>"); - if (!content_start) { - throw common_chat_msg_partial_exception("incomplete header"); - } - - auto header = content_start->prelude; - - if (auto match = regex_match(tool_call1_regex, header)) { - auto group = match->groups[1]; - auto name = header.substr(group.begin, group.end - group.begin); - handle_tool_call(name); - continue; - } - - if (auto match = regex_match(tool_call2_regex, header)) { - auto group = match->groups[2]; - auto name = header.substr(group.begin, group.end - group.begin); - handle_tool_call(name); - continue; - } - - if (regex_match(analysis_regex, header)) { - builder.move_to(header_start_pos); - if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE || builder.syntax().reasoning_in_content) { - builder.add_content(consume_end(true)); - } else { - builder.try_parse_reasoning("<|channel|>analysis<|message|>", "<|end|>"); - } - continue; - } - - if(regex_match(final_regex, header) || regex_match(preamble_regex, header)) { - builder.add_content(consume_end()); - continue; - } - - // Possibly a malformed message, attempt to recover by rolling - // back to pick up the next <|start|> - LOG_DBG("%s: unknown header from message: %s\n", __func__, header.c_str()); - builder.move_to(header_start_pos); - } while (builder.try_find_regex(start_regex, std::string::npos, false)); - - auto remaining = builder.consume_rest(); - if (!remaining.empty()) { - LOG_DBG("%s: content after last message: %s\n", __func__, remaining.c_str()); - } -} static common_chat_params common_chat_params_init_glm_4_5(const common_chat_template & tmpl, const struct templates_params & inputs) { common_chat_params data; @@ -2398,21 +1892,6 @@ static common_chat_params common_chat_params_init_glm_4_5(const common_chat_temp return data; } -static void common_chat_parse_glm_4_5(common_chat_msg_parser & builder) { - static const xml_tool_call_format form { - /* form.scope_start = */ "", - /* form.tool_start = */ "", - /* form.tool_sep = */ "", - /* form.key_start = */ "", - /* form.key_val_sep = */ "", - /* form.val_end = */ "", - /* form.tool_end = */ "", - /* form.scope_end = */ "", - /* form.key_val_sep2 = */ "", - }; - builder.consume_reasoning_with_xml_tool_calls(form, "", ""); -} - static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct templates_params & inputs) { LOG_DBG("%s\n", __func__); common_chat_params data; @@ -2460,14 +1939,6 @@ static common_chat_params common_chat_params_init_firefunction_v2(const common_c } return data; } -static void common_chat_parse_firefunction_v2(common_chat_msg_parser & builder) { - if (!builder.syntax().parse_tool_calls) { - builder.add_content(builder.consume_rest()); - return; - } - static const common_regex prefix(regex_escape(" functools[")); - parse_prefixed_json_tool_call_array(builder, prefix, /* rstrip_prefix= */ 1); -} static common_chat_params common_chat_params_init_functionary_v3_2(const common_chat_template & tmpl, const struct templates_params & inputs) { // >>>all\nlet's call functions>>>fn1\n{"arg1": 1...}\n>>>fn2\n{"arg1": 1...}... @@ -2518,34 +1989,6 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_ } return data; } -static void common_chat_parse_functionary_v3_2(common_chat_msg_parser & builder) { - static const common_regex function_regex_start_only(R"((\w+\n\{|python\n|all\n))"); - static const common_regex function_regex(R"(>>>(\w+\n\{|python\n|all\n))"); - static const common_regex close_regex(R"(\s*)"); - - parse_json_tool_calls( - builder, - std::nullopt, - function_regex_start_only, - function_regex, - close_regex, - std::nullopt, - /* allow_raw_python= */ true, - /* get_function_name= */ [&](const auto & res) -> std::string { - auto at_start = res.groups[0].begin == 0; - auto name = builder.str(res.groups[1]); - if (!name.empty() && name.back() == '{') { - // Unconsume the opening brace '{' to ensure the JSON parsing goes well. - builder.move_back(1); - } - auto idx = name.find_last_not_of("\n{"); - name = name.substr(0, idx + 1); - if (at_start && name == "all") { - return ""; - } - return name; - }); -} static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(const common_chat_template & tmpl, const struct templates_params & inputs) { // https://github.com/MeetKai/functionary/blob/main/tests/prompt_test_v3-llama3.1.txt @@ -2605,31 +2048,6 @@ static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(con // TODO: if (has_raw_python) return data; } -static void common_chat_parse_functionary_v3_1_llama_3_1(common_chat_msg_parser & builder) { - if (!builder.syntax().parse_tool_calls) { - builder.add_content(builder.consume_rest()); - return; - } - // This version of Functionary still supports the llama 3.1 tool call format for the python tool. - static const common_regex python_tag_regex(regex_escape("<|python_tag|>")); - - static const common_regex function_regex(R"()"); - static const common_regex close_regex(R"()"); - - parse_json_tool_calls( - builder, - /* block_open= */ std::nullopt, - /* function_regex_start_only= */ std::nullopt, - function_regex, - close_regex, - std::nullopt); - - if (auto res = builder.try_find_regex(python_tag_regex)) { - auto arguments = wrap_code_as_arguments(builder, builder.consume_rest()); - builder.add_tool_call("python", "", arguments); - return; - } -} static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat_template & tmpl, const struct templates_params & inputs) { common_chat_params data; @@ -2746,83 +2164,6 @@ static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat return data; } -static void common_chat_parse_hermes_2_pro(common_chat_msg_parser & builder) { - builder.try_parse_reasoning("", ""); - if (!builder.syntax().parse_tool_calls) { - builder.add_content(builder.consume_rest()); - return; - } - - static const common_regex open_regex( - "(?:" - "(```(?:xml|json)?\\n\\s*)?" // match 1 (block_start) - "(" // match 2 (open_tag) - "" - "|" - "|" - "|" - "|" - "|" - "|" - "|" - ")?" - "(\\s*\\{\\s*\"name\")" // match 3 (named tool call) - ")" - "|]+)>" // match 4 (function name) - "|" // match 5 (function name again) - ); - - while (auto res = builder.try_find_regex(open_regex)) { - const auto & block_start = res->groups[1]; - std::string block_end = block_start.empty() ? "" : "```"; - - const auto & open_tag = res->groups[2]; - std::string close_tag; - - if (!res->groups[3].empty()) { - builder.move_to(res->groups[3].begin); - close_tag = open_tag.empty() ? "" : "value) || tool_call->is_partial) { - throw common_chat_msg_partial_exception("incomplete tool call"); - } - builder.consume_spaces(); - builder.consume_literal(close_tag); - builder.consume_spaces(); - if (!block_end.empty()) { - builder.consume_literal(block_end); - builder.consume_spaces(); - } - } else { - throw common_chat_msg_partial_exception("failed to parse tool call"); - } - } else { - auto function_name = builder.str(res->groups[4]); - if (function_name.empty()) { - function_name = builder.str(res->groups[5]); - } - GGML_ASSERT(!function_name.empty()); - - close_tag = ""; - - if (auto arguments = builder.try_consume_json_with_dumped_args({{}})) { - if (!builder.add_tool_call(function_name, "", arguments->value) || arguments->is_partial) { - throw common_chat_msg_partial_exception("incomplete tool call"); - } - builder.consume_spaces(); - builder.consume_literal(close_tag); - builder.consume_spaces(); - if (!block_end.empty()) { - builder.consume_literal(block_end); - builder.consume_spaces(); - } - } - } - } - - builder.add_content(builder.consume_rest()); -} static common_chat_params common_chat_params_init_granite(const common_chat_template & tmpl, const struct templates_params & inputs) { common_chat_params data; @@ -2905,190 +2246,6 @@ static common_chat_params common_chat_params_init_granite(const common_chat_temp return data; } -static void common_chat_parse_granite(common_chat_msg_parser & builder) { - // Parse thinking tags - static const common_regex start_think_regex(regex_escape("")); - static const common_regex end_think_regex(regex_escape("")); - // Granite models output partial tokens such as "<" and "groups[0].begin); - builder.try_find_regex(end_think_regex, std::string::npos, false); - // Restore position for try_parse_reasoning() - builder.move_to(res->groups[0].begin); - } - builder.try_parse_reasoning("", ""); - - // Parse response tags - static const common_regex start_response_regex(regex_escape("")); - static const common_regex end_response_regex(regex_escape("")); - // Granite models output partial tokens such as "<" and "")); - if (auto res = builder.try_find_regex(tool_call_regex)) { - builder.move_to(res->groups[0].end); - - // Expect JSON array of tool calls - if (auto tool_call = builder.try_consume_json_with_dumped_args({{{"arguments"}}})) { - if (!builder.add_tool_calls(tool_call->value) || tool_call->is_partial) { - throw common_chat_msg_partial_exception("incomplete tool call"); - } - } - } else { - builder.add_content(builder.consume_rest()); - } -} - -static void common_chat_parse_nemotron_v2(common_chat_msg_parser & builder) { - // Parse thinking tags - builder.try_parse_reasoning("", ""); - if (!builder.syntax().parse_tool_calls) { - builder.add_content(builder.consume_rest()); - return; - } - - // Look for tool calls - static const common_regex tool_call_regex(regex_escape("")); - if (auto res = builder.try_find_regex(tool_call_regex)) { - builder.move_to(res->groups[0].end); - - // Expect JSON array of tool calls - auto tool_calls_data = builder.consume_json(); - if (tool_calls_data.json.is_array()) { - if (!builder.try_consume_literal("")) { - throw common_chat_msg_partial_exception("Incomplete tool call"); - } - builder.add_tool_calls(tool_calls_data.json); - } else { - throw common_chat_msg_partial_exception("Incomplete tool call"); - } - } - builder.add_content(builder.consume_rest()); -} - -static void common_chat_parse_apertus(common_chat_msg_parser & builder) { - // Parse thinking tags - builder.try_parse_reasoning("<|inner_prefix|>", "<|inner_suffix|>"); - if (!builder.syntax().parse_tool_calls) { - builder.add_content(builder.consume_rest()); - return; - } - - // Look for tool calls - static const common_regex tool_call_regex(regex_escape("<|tools_prefix|>")); - if (auto res = builder.try_find_regex(tool_call_regex)) { - builder.move_to(res->groups[0].end); - - auto tool_calls_data = builder.consume_json(); - if (tool_calls_data.json.is_array()) { - builder.consume_spaces(); - if (!builder.try_consume_literal("<|tools_suffix|>")) { - throw common_chat_msg_partial_exception("Incomplete tool call"); - } - for (const auto & value : tool_calls_data.json) { - if (value.is_object()) { - builder.add_tool_call_short_form(value); - } - } - } else { - throw common_chat_msg_partial_exception("Incomplete tool call"); - } - } - builder.add_content(builder.consume_rest()); -} - - -static void common_chat_parse_lfm2(common_chat_msg_parser & builder) { - if (!builder.syntax().parse_tool_calls) { - builder.add_content(builder.consume_rest()); - return; - } - - // LFM2 format: <|tool_call_start|>[{"name": "get_current_time", "arguments": {"location": "Paris"}}]<|tool_call_end|> - static const common_regex tool_call_start_regex(regex_escape("<|tool_call_start|>")); - static const common_regex tool_call_end_regex(regex_escape("<|tool_call_end|>")); - - // Loop through all tool calls - while (auto res = builder.try_find_regex(tool_call_start_regex, std::string::npos, /* add_prelude_to_content= */ true)) { - builder.move_to(res->groups[0].end); - - // Parse JSON array format: [{"name": "...", "arguments": {...}}] - auto tool_calls_data = builder.consume_json(); - - // Consume end marker - builder.consume_spaces(); - if (!builder.try_consume_regex(tool_call_end_regex)) { - throw common_chat_msg_partial_exception("Expected <|tool_call_end|>"); - } - - // Process each tool call in the array - if (tool_calls_data.json.is_array()) { - for (const auto & tool_call : tool_calls_data.json) { - if (!tool_call.is_object()) { - throw common_chat_msg_partial_exception("Tool call must be an object"); - } - - if (!tool_call.contains("name")) { - throw common_chat_msg_partial_exception("Tool call missing 'name' field"); - } - - std::string function_name = tool_call.at("name"); - std::string arguments = "{}"; - - if (tool_call.contains("arguments")) { - if (tool_call.at("arguments").is_object()) { - arguments = tool_call.at("arguments").dump(); - } else if (tool_call.at("arguments").is_string()) { - arguments = tool_call.at("arguments"); - } - } - - if (!builder.add_tool_call(function_name, "", arguments)) { - throw common_chat_msg_partial_exception("Incomplete tool call"); - } - } - } else { - throw common_chat_msg_partial_exception("Expected JSON array for tool calls"); - } - - // Consume any trailing whitespace after this tool call - builder.consume_spaces(); - } - - // Consume any remaining content after all tool calls - auto remaining = builder.consume_rest(); - if (!string_strip(remaining).empty()) { - builder.add_content(remaining); - } -} - -static void common_chat_parse_seed_oss(common_chat_msg_parser & builder) { - static const xml_tool_call_format form { - /* form.scope_start = */ "", - /* form.tool_start = */ "", - /* form.key_start = */ "", - /* form.val_end = */ "", - /* form.tool_end = */ "", - /* form.scope_end = */ "", - }; - builder.consume_reasoning_with_xml_tool_calls(form, "", ""); -} - static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) { common_chat_params data; data.prompt = apply(tmpl, inputs); @@ -3428,112 +2585,3 @@ common_chat_params common_chat_templates_apply( ? common_chat_templates_apply_jinja(tmpls, inputs) : common_chat_templates_apply_legacy(tmpls, inputs); } - -static void common_chat_parse_content_only(common_chat_msg_parser & builder) { - builder.try_parse_reasoning("", ""); - builder.add_content(builder.consume_rest()); -} - -static void common_chat_parse(common_chat_msg_parser & builder) { - LOG_DBG("Parsing input with format %s: %s\n", common_chat_format_name(builder.syntax().format), builder.input().c_str()); - - switch (builder.syntax().format) { - case COMMON_CHAT_FORMAT_CONTENT_ONLY: - common_chat_parse_content_only(builder); - break; - case COMMON_CHAT_FORMAT_GENERIC: - common_chat_parse_generic(builder); - break; - case COMMON_CHAT_FORMAT_MISTRAL_NEMO: - common_chat_parse_mistral_nemo(builder); - break; - case COMMON_CHAT_FORMAT_MAGISTRAL: - common_chat_parse_magistral(builder); - break; - case COMMON_CHAT_FORMAT_LLAMA_3_X: - common_chat_parse_llama_3_1(builder); - break; - case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS: - common_chat_parse_llama_3_1(builder, /* with_builtin_tools= */ true); - break; - case COMMON_CHAT_FORMAT_DEEPSEEK_R1: - common_chat_parse_deepseek_r1(builder); - break; - case COMMON_CHAT_FORMAT_DEEPSEEK_V3_1: - common_chat_parse_deepseek_v3_1(builder); - break; - case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2: - common_chat_parse_functionary_v3_2(builder); - break; - case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1: - common_chat_parse_functionary_v3_1_llama_3_1(builder); - break; - case COMMON_CHAT_FORMAT_HERMES_2_PRO: - common_chat_parse_hermes_2_pro(builder); - break; - case COMMON_CHAT_FORMAT_FIREFUNCTION_V2: - common_chat_parse_firefunction_v2(builder); - break; - case COMMON_CHAT_FORMAT_COMMAND_R7B: - common_chat_parse_command_r7b(builder); - break; - case COMMON_CHAT_FORMAT_GRANITE: - common_chat_parse_granite(builder); - break; - case COMMON_CHAT_FORMAT_GPT_OSS: - common_chat_parse_gpt_oss(builder); - break; - case COMMON_CHAT_FORMAT_SEED_OSS: - common_chat_parse_seed_oss(builder); - break; - case COMMON_CHAT_FORMAT_NEMOTRON_V2: - common_chat_parse_nemotron_v2(builder); - break; - case COMMON_CHAT_FORMAT_APERTUS: - common_chat_parse_apertus(builder); - break; - case COMMON_CHAT_FORMAT_LFM2_WITH_JSON_TOOLS: - common_chat_parse_lfm2(builder); - break; - case COMMON_CHAT_FORMAT_MINIMAX_M2: - common_chat_parse_minimax_m2(builder); - break; - case COMMON_CHAT_FORMAT_GLM_4_5: - common_chat_parse_glm_4_5(builder); - break; - case COMMON_CHAT_FORMAT_KIMI_K2: - common_chat_parse_kimi_k2(builder); - break; - case COMMON_CHAT_FORMAT_QWEN3_CODER_XML: - common_chat_parse_qwen3_coder_xml(builder); - break; - case COMMON_CHAT_FORMAT_APRIEL_1_5: - common_chat_parse_apriel_1_5(builder); - break; - case COMMON_CHAT_FORMAT_XIAOMI_MIMO: - common_chat_parse_xiaomi_mimo(builder); - break; - default: - throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format)); - } - builder.finish(); -} - -common_chat_msg common_chat_parse(const std::string & input, bool is_partial, const common_chat_syntax & syntax) { - common_chat_msg_parser builder(input, is_partial, syntax); - try { - common_chat_parse(builder); - } catch (const common_chat_msg_partial_exception & ex) { - LOG_DBG("Partial parse: %s\n", ex.what()); - if (!is_partial) { - builder.clear_tools(); - builder.move_to(0); - common_chat_parse_content_only(builder); - } - } - auto msg = builder.result(); - if (!is_partial) { - LOG_DBG("Parsed message: %s\n", common_chat_msgs_to_json_oaicompat({msg}).at(0).dump().c_str()); - } - return msg; -} From d82b7a7c1d73c0674698d9601b1bbb0200933f29 Mon Sep 17 00:00:00 2001 From: Aleksei Nikiforov <103434461+AlekseiNikiforovIBM@users.noreply.github.com> Date: Fri, 28 Nov 2025 20:53:01 +0100 Subject: [PATCH 6/6] gguf-py : fix passing non-native endian tensors (editor-gui and new-metadata) (#17553) gguf_new_metadata.py reads data from reader. Reader doesn't byteswap tensors to native endianness. But writer does expect tensors in native endianness to convert them into requested endianness. There are two ways to fix this: update reader and do conversion to native endianness and back, or skip converting endianness in writer in this particular USE-case. gguf_editor_gui.py doesn't allow editing or viewing tensor data. Let's go with skipping excessive byteswapping. If eventually capability to view or edit tensor data is added, tensor data should be instead byteswapped when reading it. --- gguf-py/gguf/gguf_writer.py | 18 ++++++++++++------ gguf-py/gguf/scripts/gguf_editor_gui.py | 2 +- gguf-py/gguf/scripts/gguf_new_metadata.py | 2 +- 3 files changed, 14 insertions(+), 8 deletions(-) diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 57ca2035f..8ddd895cb 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -371,10 +371,13 @@ class GGUFWriter: def add_tensor( self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None, - raw_dtype: GGMLQuantizationType | None = None, + raw_dtype: GGMLQuantizationType | None = None, tensor_endianess: GGUFEndian | None = None ) -> None: - if (self.endianess == GGUFEndian.BIG and sys.byteorder != 'big') or \ - (self.endianess == GGUFEndian.LITTLE and sys.byteorder != 'little'): + # if tensor endianness is not passed, assume it's native to system + if tensor_endianess is None: + tensor_endianess = GGUFEndian.BIG if sys.byteorder == 'big' else GGUFEndian.LITTLE + + if tensor_endianess != self.endianess: # Don't byteswap inplace since lazy copies cannot handle it tensor = tensor.byteswap(inplace=False) if self.use_temp_file and self.temp_file is None: @@ -397,13 +400,16 @@ class GGUFWriter: if pad != 0: fp.write(bytes([0] * pad)) - def write_tensor_data(self, tensor: np.ndarray[Any, Any]) -> None: + def write_tensor_data(self, tensor: np.ndarray[Any, Any], tensor_endianess: GGUFEndian | None = None) -> None: if self.state is not WriterState.TI_DATA and self.state is not WriterState.WEIGHTS: raise ValueError(f'Expected output file to contain tensor info or weights, got {self.state}') assert self.fout is not None - if (self.endianess == GGUFEndian.BIG and sys.byteorder != 'big') or \ - (self.endianess == GGUFEndian.LITTLE and sys.byteorder != 'little'): + # if tensor endianness is not passed, assume it's native to system + if tensor_endianess is None: + tensor_endianess = GGUFEndian.BIG if sys.byteorder == 'big' else GGUFEndian.LITTLE + + if tensor_endianess != self.endianess: # Don't byteswap inplace since lazy copies cannot handle it tensor = tensor.byteswap(inplace=False) diff --git a/gguf-py/gguf/scripts/gguf_editor_gui.py b/gguf-py/gguf/scripts/gguf_editor_gui.py index 05f4db0f8..293316afe 100755 --- a/gguf-py/gguf/scripts/gguf_editor_gui.py +++ b/gguf-py/gguf/scripts/gguf_editor_gui.py @@ -1552,7 +1552,7 @@ class GGUFEditorWindow(QMainWindow): # Add tensors (including data) for tensor in self.reader.tensors: - writer.add_tensor(tensor.name, tensor.data, raw_shape=tensor.data.shape, raw_dtype=tensor.tensor_type) + writer.add_tensor(tensor.name, tensor.data, raw_shape=tensor.data.shape, raw_dtype=tensor.tensor_type, tensor_endianess=self.reader.endianess) # Write header and metadata writer.open_output_file(Path(file_path)) diff --git a/gguf-py/gguf/scripts/gguf_new_metadata.py b/gguf-py/gguf/scripts/gguf_new_metadata.py index 2fa5800cf..c67436bad 100755 --- a/gguf-py/gguf/scripts/gguf_new_metadata.py +++ b/gguf-py/gguf/scripts/gguf_new_metadata.py @@ -94,7 +94,7 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new writer.write_ti_data_to_file() for tensor in reader.tensors: - writer.write_tensor_data(tensor.data) + writer.write_tensor_data(tensor.data, tensor_endianess=reader.endianess) bar.update(tensor.n_bytes) writer.close()