From b61de2b2df4ff07e6d6de96320fb311d96908b7a Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Mon, 24 Nov 2025 15:50:55 +0100 Subject: [PATCH 01/18] convert : allow quantizing lora again (#17453) --- convert_hf_to_gguf.py | 2 +- convert_lora_to_gguf.py | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 6cbaee03df..d24a4682f3 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -565,7 +565,7 @@ class ModelBase: gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF, ) ) - or not new_name.endswith(".weight") + or new_name[-7:] not in (".weight", ".lora_a", ".lora_b") ): data_qtype = gguf.GGMLQuantizationType.F32 diff --git a/convert_lora_to_gguf.py b/convert_lora_to_gguf.py index 57c6cd0df1..b0adde8a8b 100755 --- a/convert_lora_to_gguf.py +++ b/convert_lora_to_gguf.py @@ -242,7 +242,7 @@ def parse_args() -> argparse.Namespace: help="path to write to; default: based on input. {ftype} will be replaced by the outtype.", ) parser.add_argument( - "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16", + "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f32", help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type", ) parser.add_argument( From 0543f928a3ae576e6e16d3bbf02c0bf9fddba688 Mon Sep 17 00:00:00 2001 From: "Jiacheng (Jason) Chen" <76919340+jiachengjason@users.noreply.github.com> Date: Mon, 24 Nov 2025 14:00:10 -0500 Subject: [PATCH 02/18] HIP: WMMA-MMQ kernels for RDNA 4 (#17156) * first commit naive test to enable mmq for RDNA4 * adding appropriate WMMA instructions * git rebase on top of master: fixing the correctness of the mat mul operations, updating layout mappings for RDNA4 * clean up merge conflicts * add comments and code clean up * PR clean up, addressed comments * enable MMQ fallback on RDNA4 * addressed comments: add guards in load generic, separate wmma branch for use_mmq function * Revert build-xcframework.sh * Formating: remove trailing whitespace * revert CMake files * clean up after rebase: remove duplicated change, revert cmake files * clean up after rebase: revert changes from build-xcframework.sh * clean up: remove extra space line in mma.cuh * Revert "clean up: remove extra space line in mma.cuh" This reverts commit b39ed57c4529906466bd0bc7c2a86e08fc2f8bee. --- ggml/src/ggml-cuda/mma.cuh | 131 ++++++++--- ggml/src/ggml-cuda/mmq.cu | 8 +- ggml/src/ggml-cuda/mmq.cuh | 437 +++++++++++++++++++++++++------------ 3 files changed, 408 insertions(+), 168 deletions(-) diff --git a/ggml/src/ggml-cuda/mma.cuh b/ggml/src/ggml-cuda/mma.cuh index c3c4b77996..caa08b360b 100644 --- a/ggml/src/ggml-cuda/mma.cuh +++ b/ggml/src/ggml-cuda/mma.cuh @@ -73,34 +73,7 @@ namespace ggml_cuda_mma { static constexpr int I = I_; static constexpr int J = J_; -#if defined(GGML_USE_HIP) -#if defined(RDNA4) - static constexpr int ne = I * J / 32; - T x[ne] = {0}; - - static constexpr __device__ bool supported() { - if (I == 16 && J == 16) return true; - return false; - } - - static __device__ __forceinline__ int get_i(const int l) { - if constexpr (I == 16 && J == 16) { - return 8 * (threadIdx.x / 16) + l; - } else { - NO_DEVICE_CODE; - return -1; - } - } - - static __device__ __forceinline__ int get_j(const int l) { - if constexpr (I == 16 && J == 16) { - return threadIdx.x % 16; - } else { - NO_DEVICE_CODE; - return -1; - } - } -#else +#if defined(AMD_MFMA_AVAILABLE) static constexpr int ne = I * J / 64; T x[ne] = {0}; @@ -146,7 +119,6 @@ namespace ggml_cuda_mma { return -1; } } -#endif // defined(RDNA4) #elif __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA static constexpr int ne = I * J / 32; T x[ne] = {0}; @@ -177,6 +149,34 @@ namespace ggml_cuda_mma { return -1; } } +#elif defined(AMD_WMMA_AVAILABLE) +#if defined(RDNA4) + static constexpr int ne = I * J / 32; + T x[ne] = {0}; + + static constexpr __device__ bool supported() { + if (I == 16 && J == 16) return true; + return false; + } + + static __device__ __forceinline__ int get_i(const int l) { + if constexpr (I == 16 && J == 16) { + return 8 * (threadIdx.x / 16) + l; + } else { + NO_DEVICE_CODE; + return -1; + } + } + + static __device__ __forceinline__ int get_j(const int l) { + if constexpr (I == 16 && J == 16) { + return threadIdx.x % 16; + } else { + NO_DEVICE_CODE; + return -1; + } + } +#endif #else static constexpr int ne = I * J / 32; T x[ne] = {0}; @@ -437,7 +437,20 @@ namespace ggml_cuda_mma { xi[0] = xs[0]; } #elif defined(AMD_WMMA_AVAILABLE) - ggml_cuda_memcpy_1(t.x, xs0 + t.get_i(0) * stride + t.get_j(0)); + if constexpr (I == 16 && J == 4) { + int64_t * xi = (int64_t *) t.x; + const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 2 * (threadIdx.x / t.I)); + xi[0] = xs[0]; + }else if constexpr (I == 16 && J == 8) { + int64_t * xi = (int64_t *) t.x; + const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 4 * (threadIdx.x / t.I)); + xi[0] = xs[0]; + + const int64_t * xs1 = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 4 * (threadIdx.x / t.I) + 2); + xi[1] = xs1[0]; + }else{ + NO_DEVICE_CODE; + } #else #pragma unroll for (int l = 0; l < t.ne; ++l) { @@ -772,6 +785,36 @@ namespace ggml_cuda_mma { acc[0], 0, 0, 0); #endif // defined(CDNA3) + +#elif defined(AMD_WMMA_AVAILABLE) + using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int; + int32x2_t * a_vec = (int32x2_t *) A.x; + int32x2_t * b_vec = (int32x2_t *) B.x; + + using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int; + int32x8_t * acc = (int32x8_t *) D.x; + +#if defined(RDNA4) + + acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12( + true, + a_vec[0], + true, + b_vec[0], + acc[0], + true + ); + + acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12( + true, + a_vec[1], + true, + b_vec[1], + acc[0], + true + ); +#endif // defined(RDNA4) + #else GGML_UNUSED_VARS(D, A, B); NO_DEVICE_CODE; @@ -798,6 +841,7 @@ namespace ggml_cuda_mma { acc[0], 0, 0, 0); #endif // defined(CDNA3) + #else GGML_UNUSED_VARS(D, A, B); NO_DEVICE_CODE; @@ -842,4 +886,31 @@ namespace ggml_cuda_mma { mma(D16[1], A16[1], B); #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE } + +static __device__ __forceinline__ void mma( + tile<16, 16, int> & D, const tile<16, 4, int> & A, const tile<16, 4, int> & B) { +#if defined(AMD_WMMA_AVAILABLE) + using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int; + int32x2_t * a_vec = (int32x2_t *) A.x; + int32x2_t * b_vec = (int32x2_t *) B.x; + + using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int; + int32x8_t * acc = (int32x8_t *) D.x; + + acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12( + true, + a_vec[0], + true, + b_vec[0], + acc[0], + false + ); +#else + GGML_UNUSED(D); + GGML_UNUSED(A); + GGML_UNUSED(B); + NO_DEVICE_CODE; +#endif + } } + diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu index a2c8760abe..03ceba874d 100644 --- a/ggml/src/ggml-cuda/mmq.cu +++ b/ggml/src/ggml-cuda/mmq.cu @@ -306,5 +306,11 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) { return false; } - return (!GGML_CUDA_CC_IS_RDNA4(cc) && !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE; + if (amd_wmma_available(cc)) { + if (GGML_CUDA_CC_IS_RDNA4(cc)) { + return true; + } + } + + return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE; } diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh index 2e133b6bda..99760d56c7 100644 --- a/ggml/src/ggml-cuda/mmq.cuh +++ b/ggml/src/ggml-cuda/mmq.cuh @@ -92,7 +92,7 @@ struct tile_x_sizes { }; static int get_mmq_x_max_host(const int cc) { - return (amd_mfma_available(cc) || turing_mma_available(cc)) ? 128 : + return (amd_mfma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc)) ? 128 : GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ? #ifdef GGML_CUDA_FORCE_MMQ 128 : 64; @@ -102,7 +102,7 @@ static int get_mmq_x_max_host(const int cc) { } static constexpr __device__ int get_mmq_x_max_device() { -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) return 128; #else // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) @@ -121,7 +121,7 @@ static constexpr __device__ int get_mmq_x_max_device() { #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA #endif // defined(GGML_USE_HIP) -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } static int get_mmq_y_host(const int cc) { @@ -231,7 +231,7 @@ static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) { #define MMQ_TILE_Y_K (MMQ_TILE_NE_K + MMQ_TILE_NE_K/QI8_1) static int mmq_get_granularity_host(const int mmq_x, const int cc) { - if (amd_mfma_available(cc)) { + if (amd_mfma_available(cc) || amd_wmma_available(cc)) { return mmq_x >= 128 ? 32 : 16; } else if (turing_mma_available(cc) && mmq_x >= 48) { return 16; @@ -240,7 +240,7 @@ static int mmq_get_granularity_host(const int mmq_x, const int cc) { } } -#if defined(AMD_MFMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) static constexpr __device__ int mmq_get_granularity_device(const int mmq_x) { return mmq_x >= 128 ? 32 : 16; } @@ -265,7 +265,7 @@ static int mmq_get_nwarps_host(const int /*cc*/, const int warp_size) { #endif // (GGML_USE_HIP) static constexpr __device__ int mmq_get_nwarps_device() { -#if defined(AMD_MFMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) return 8; #else return 256/ggml_cuda_get_physical_warp_size(); @@ -279,14 +279,14 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K); #else constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y); int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_0); constexpr int nrows = warp_size / threads_per_row; @@ -305,7 +305,7 @@ template static __device__ __forceinline__ void loa const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbx; const int qs0 = get_int_b2(bxi->qs, kqsx); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI4_0) + kqsx + 0] = __vsubss4((qs0 >> 0) & 0x0F0F0F0F, 0x08080808); x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI4_0) + kqsx + QI4_0] = __vsubss4((qs0 >> 4) & 0x0F0F0F0F, 0x08080808); #else @@ -327,11 +327,11 @@ template static __device__ __forceinline__ void loa const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbxd; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = bxi->d; #else x_df[i*(MMQ_TILE_NE_K/QI4_0) + i/QI4_0 + kbxd] = bxi->d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -382,14 +382,14 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); #else constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y); int * x_qs = (int *) x_tile; half2 * x_dm = (half2 *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_1); constexpr int nrows = warp_size / threads_per_row; @@ -408,12 +408,12 @@ template static __device__ __forceinline__ void loa const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbx; const int qs0 = get_int_b4(bxi->qs, kqsx); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI4_1) + kqsx + 0] = (qs0 >> 0) & 0x0F0F0F0F; x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI4_1) + kqsx + QI4_1] = (qs0 >> 4) & 0x0F0F0F0F; #else x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_1; @@ -430,11 +430,11 @@ template static __device__ __forceinline__ void loa const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbxd; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + kbxd] = bxi->dm; #else x_dm[i*(MMQ_TILE_NE_K/QI4_1) + i/QI4_1 + kbxd] = bxi->dm; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -485,14 +485,14 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); #else constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_0, mmq_y); int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_0); constexpr int nrows = warp_size / threads_per_row; @@ -527,13 +527,13 @@ template static __device__ __forceinline__ void loa qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 qs1 = __vsubss4(qs1, 0x10101010); // subtract 16 -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI5_0) + kqsx + 0] = qs0; x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1; #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + 0] = qs0; x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_0; @@ -550,11 +550,11 @@ template static __device__ __forceinline__ void loa const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbxd; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = bxi->d; #else x_df[i*(MMQ_TILE_NE_K/QI5_0) + i/QI5_0 + kbxd] = bxi->d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -563,14 +563,14 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); #else constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y); int * x_qs = (int *) x_tile; half2 * x_dm = (half2 *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = MMQ_ITER_K / (4 * QR5_1); constexpr int nrows = warp_size / threads_per_row; @@ -603,13 +603,13 @@ template static __device__ __forceinline__ void loa qs1 |= (qh << 2) & 0x00100000; // 18 -> 20 qs1 |= (qh << 9) & 0x10000000; // 19 -> 28 -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI5_1) + kqsx + 0] = qs0; x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1; #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + 0] = qs0; x_qs[i*(2*MMQ_TILE_NE_K + 1) + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI5_1; @@ -626,11 +626,11 @@ template static __device__ __forceinline__ void loa const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbxd; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + kbxd] = bxi->dm; #else x_dm[i*(MMQ_TILE_NE_K/QI5_1) + i/QI5_1 + kbxd] = bxi->dm; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -639,14 +639,14 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; float * x_df = (float *) (x_tile + 2*MMQ_TILE_NE_K); #else constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y); int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) // MMQ_ITER_K / (4 * QR8_0) == 64 required. but NV has only 32 threads per warp constexpr int threads_per_row = 32; @@ -665,13 +665,13 @@ template static __device__ __forceinline__ void loa const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbx; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 0 + txi] = get_int_b2(bxi[0].qs, kqsx); x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx); #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + 0 + txi] = get_int_b2(bxi[0].qs, kqsx); x_qs[i*(2*MMQ_TILE_NE_K + 1) + MMQ_TILE_NE_K + txi] = get_int_b2(bxi[MMQ_TILE_NE_K/QI8_0].qs, kqsx); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } constexpr int blocks_per_tile_x_row = 2*MMQ_TILE_NE_K / QI8_0; @@ -688,11 +688,11 @@ template static __device__ __forceinline__ void loa const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbxd; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = bxi->d; #else x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + kbxd] = bxi->d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -701,14 +701,14 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); #else constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_MXFP4, mmq_y); int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = MMQ_ITER_K / (4 * QR_MXFP4); constexpr int nrows = warp_size / threads_per_row; @@ -730,13 +730,13 @@ template static __device__ __forceinline__ void loa const int2 v = get_int_from_table_16(aux_q4, kvalues_mxfp4); const int k0 = kbx * (2 * QI_MXFP4) + kqsx; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + k0 + 0] = v.x; x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + k0 + QI_MXFP4] = v.y; #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI_MXFP4] = v.y; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI_MXFP4; @@ -753,11 +753,11 @@ template static __device__ __forceinline__ void loa const block_mxfp4 * bxi = (const block_mxfp4 *) x + kbx0 + i*stride + kbxd; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_df[i*MMQ_MMA_TILE_X_K_Q8_1 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f; #else x_df[i*(MMQ_TILE_NE_K/QI_MXFP4) + i/QI_MXFP4 + kbxd] = ggml_cuda_e8m0_to_fp32(bxi->e)*0.5f; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -796,7 +796,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a( template static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma( const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { -#if defined(AMD_MFMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) typedef tile<16, 8, int> tile_A; typedef tile<16, 8, int> tile_B; typedef tile<16, 16, int> tile_C; @@ -927,7 +927,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma( } } } -#endif // defined(AMD_MFMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } template @@ -965,7 +965,7 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a( template static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma( const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) { -#if defined(AMD_MFMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) typedef tile<16, 8, int> tile_A; typedef tile<16, 8, int> tile_B; typedef tile<16, 16, int> tile_C; @@ -1087,7 +1087,7 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma( } } } -#endif // defined(AMD_MFMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } // Used for Q3_K, IQ2_S, and IQ2_XS @@ -1170,6 +1170,54 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma( tile_C C; mma(C, A[n], B[0]); +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(l); + sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * x_df[i*MMQ_MMA_TILE_X_K_Q3_K + k0/4] * dB; + } + } + } + } +#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles + typedef tile<16, 4, int> tile_A; + typedef tile<16, 4, int> tile_B; + typedef tile<16, 16, int> tile_C; + + constexpr int granularity = mmq_get_granularity_device(mmq_x); + constexpr int rows_per_warp = granularity; + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { + const int k0 = k00 + k01; + + tile_A A[ntx]; +#pragma unroll + for (int n = 0; n < ntx; ++n) { + load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K); + } + +#pragma unroll + for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { + tile_B B; + load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); + + const int j = j0 + tile_C::get_j(0); + const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C; + mma(C, A[n], B); + #pragma unroll for (int l = 0; l < tile_C::ne; ++l) { const int i = i0 + n*tile_C::I + tile_C::get_i(l); @@ -1257,21 +1305,21 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma( #else GGML_UNUSED_VARS(x, y, sum, k00); NO_DEVICE_CODE; -#endif // AMD_MFMA_AVAILABLE +#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE } template static __device__ __forceinline__ void load_tiles_q2_K( const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) { constexpr int nwarps = mmq_get_nwarps_device(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); #else constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y); int * x_qs = (int *) x_tile; half2 * x_dm = (half2 *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = MMQ_ITER_K / (4 * QR2_K); constexpr int nrows = ggml_cuda_get_physical_warp_size() / threads_per_row; @@ -1295,11 +1343,11 @@ template static __device__ __forceinline__ void loa const int x_qs_k = (x_ql_0 >> (2*l)) & 0x03030303; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q2_K + k] = x_qs_k; #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } const int sc_m = bxi->scales[kqsx]; @@ -1310,11 +1358,11 @@ template static __device__ __forceinline__ void loa const half2 x_dm_ik = make_half2(bxi_dmf.x*(sc_m & 0x0F), bxi_dmf.y*(sc_m >> 4)); #endif // FAST_FP16_AVAILABLE -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + kqsx] = x_dm_ik; #else x_dm[i*(MMQ_TILE_NE_K + 1) + kqsx] = x_dm_ik; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -1438,6 +1486,72 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma( tile_C Cd; mma(Cd, A[n], B[0]); +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(l); + const float2 dm = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + k0/4]); + float tmp = Cd.x[l]*dm.x; + if (k01 >= MMQ_TILE_NE_K * 3/4) { + tmp -= Cm.x[l]*dm.y; + } + sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp*dB; + sum[(j0/tile_C::J + n)*tile_C::ne + l] -= dm.y*sB; + } + } + } + } +#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles + + typedef tile<16, 4, int> tile_A; + typedef tile<16, 4, int> tile_B; + typedef tile<16, 16, int> tile_C; + + constexpr int granularity = mmq_get_granularity_device(mmq_x); + constexpr int rows_per_warp = granularity; + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const half2 * x_dm = (const half2 *) x_qs + MMQ_TILE_NE_K*2; + const int * y_qs = (const int *) y + 4; + const half2 * y_ds = (const half2 *) y; + + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { + const int k0 = k00 + k01; + + tile_A A[ntx]; +#pragma unroll + for (int n = 0; n < ntx; ++n) { + load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K); + } + +#pragma unroll + for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { + tile_B B; + load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); + + const int j = j0 + tile_C::get_j(0); + const float dB = (k01 < MMQ_TILE_NE_K/2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K]).x : __half22float2(y_ds[j*MMQ_TILE_Y_K]).y; + const float sB = (k01 >= MMQ_TILE_NE_K * 3/4) ? 0 + : (((k01/4)%2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).y + : __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).x); + + tile_C Cm; + if (k01 >= MMQ_TILE_NE_K * 3/4) { + tile_A A1; + A1.x[0] = 0x01010101; + A1.x[1] = 0x01010101; + mma(Cm, A1, B); + } + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C Cd; + mma(Cd, A[n], B); + #pragma unroll for (int l = 0; l < tile_C::ne; ++l) { const int i = i0 + n*tile_C::I + tile_C::get_i(l); @@ -1574,7 +1688,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma( #else GGML_UNUSED_VARS(x, y, sum, k00); NO_DEVICE_CODE; -#endif // AMD_MFMA_AVAILABLE +#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE } template static __device__ __forceinline__ void load_tiles_q3_K( @@ -1582,7 +1696,7 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); #else @@ -1618,11 +1732,11 @@ template static __device__ __forceinline__ void loa const int x_qs_k = __vsubss4(x_ql_k | x_qh_k, 0x04040404); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + k] = x_qs_k; #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + k] = x_qs_k; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -1649,7 +1763,7 @@ template static __device__ __forceinline__ void loa const int sc = __vsubss4(sc_low | sc_high, 0x20202020); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) const int8_t * sc8 = (const int8_t *) ≻ const float d = bxi->d; @@ -1659,10 +1773,10 @@ template static __device__ __forceinline__ void loa } #else x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = sc; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } -#if !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)) +#if !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)) #pragma unroll for (int i0 = 0; i0 < mmq_y; i0 += nwarps*warp_size) { int i = (i0 + threadIdx.y*warp_size + threadIdx.x) % mmq_y; @@ -1675,7 +1789,7 @@ template static __device__ __forceinline__ void loa x_df[i] = bxi->d; } -#endif // !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)) +#endif // !(defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)) || defined(AMD_WMMA_AVAILABLE) } template @@ -1728,7 +1842,7 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; half2 * x_dm = (half2 *) (x_qs + 2*MMQ_TILE_NE_K); #else @@ -1736,7 +1850,7 @@ template static __device__ __forceinline__ void loa int * x_qs = (int *) x_tile; half2 * x_dm = (half2 *) (x_qs + txs.qs); int * x_sc = (int *) (x_dm + txs.dm); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_K); constexpr int nrows = warp_size / threads_per_row; @@ -1753,19 +1867,19 @@ template static __device__ __forceinline__ void loa const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride; const int qs0 = get_int_b4(bxi->qs, txi); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 16*(txi/8) + txi % 8 + 0] = (qs0 >> 0) & 0x0F0F0F0F; x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 16*(txi/8) + txi % 8 + 8] = (qs0 >> 4) & 0x0F0F0F0F; #else x_qs[i*(MMQ_TILE_NE_K + 1) + txi] = qs0; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int rows_per_warp = warp_size / 2; #pragma unroll for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) { -#if defined(AMD_MFMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) // Need if on AMD instead of % because warp_size == 64 // This causes double work and throughput loss (MI300X) // H100 loses about 100 t/s with 'if' condition over '%' @@ -1774,7 +1888,7 @@ template static __device__ __forceinline__ void loa #else int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/2) % mmq_y; { -#endif // defined(AMD_MFMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) if (need_check) { i = min(i, i_max); } @@ -1829,7 +1943,7 @@ template static __device__ __forceinline__ void loa x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8; } -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } template @@ -1872,7 +1986,7 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; half2 * x_dm = (half2 *) (x_qs + MMQ_TILE_NE_K*2); #else @@ -1908,16 +2022,16 @@ template static __device__ __forceinline__ void loa const int kq0 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + 0; const int kq1 = ky - ky % (QI5_K/2) + txi % (QI5_K/4) + QI5_K/4; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kq0] = ql0 | qh0; x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kq1] = ql1 | qh1; #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = ql0 | qh0; x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = ql1 | qh1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int rows_per_warp = warp_size / 2; #pragma unroll for (int i0 = 0; i0 < mmq_y; i0 += nwarps*rows_per_warp) { @@ -1930,7 +2044,7 @@ template static __device__ __forceinline__ void loa #else int i = (i0 + threadIdx.y*rows_per_warp + threadIdx.x/2) % mmq_y; { -#endif // defined(AMD_MFMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) if (need_check) { i = min(i, i_max); } @@ -1986,7 +2100,7 @@ template static __device__ __forceinline__ void loa x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + ksc] = scales8; } -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } template @@ -2029,7 +2143,7 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); int * x_sc = (int *) (x_df + MMQ_TILE_NE_K/QI6_K); @@ -2038,7 +2152,7 @@ template static __device__ __forceinline__ void loa int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + txs.qs); int * x_sc = (int *) (x_df + txs.dm); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = MMQ_ITER_K / (4 * QR6_K); constexpr int nrows = warp_size / threads_per_row; @@ -2065,13 +2179,13 @@ template static __device__ __forceinline__ void loa const int kq0 = 2*txi - txi % (QI6_K/2) + 0; const int kq1 = 2*txi - txi % (QI6_K/2) + QI6_K/2; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq0] = __vsubss4(ql0 | qh0, 0x20202020); x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq1] = __vsubss4(ql1 | qh1, 0x20202020); #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq0] = __vsubss4(ql0 | qh0, 0x20202020); x_qs[i*(2*MMQ_TILE_NE_K + 1) + kq1] = __vsubss4(ql1 | qh1, 0x20202020); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } #pragma unroll @@ -2084,11 +2198,11 @@ template static __device__ __forceinline__ void loa const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_df[i*MMQ_MMA_TILE_X_K_Q6_K] = bxi->d; #else x_df[i*(MMQ_TILE_NE_K/QI6_K) + i/QI6_K] = bxi->d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } constexpr int rows_per_warp = warp_size / 4; @@ -2102,11 +2216,11 @@ template static __device__ __forceinline__ void loa const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + (threadIdx.x % (MMQ_TILE_NE_K/8)) / 4; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + threadIdx.x%4] = get_int_b2(bxi->scales, threadIdx.x % (MMQ_TILE_NE_K/8)); #else x_sc[i*(MMQ_TILE_NE_K/8) + i/8 + threadIdx.x%(MMQ_TILE_NE_K/8)] = get_int_b2(bxi->scales, threadIdx.x%(QI6_K/8)); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -2190,6 +2304,56 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma( tile_C C; mma(C, A[n], B[0]); +#pragma unroll + for (int l = 0; l < tile_C::ne; ++l) { + const int i = i0 + n*tile_C::I + tile_C::get_i(l); + const int8_t * sc = (const int8_t *) (x_sc + i*MMQ_MMA_TILE_X_K_Q6_K + k00/16); + sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * sc[k01/4] * x_df[i*MMQ_MMA_TILE_X_K_Q6_K] * dB; + } + } + } + } +#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles + typedef tile<16, 4, int> tile_A; + typedef tile<16, 4, int> tile_B; + typedef tile<16, 16, int> tile_C; + + constexpr int granularity = mmq_get_granularity_device(mmq_x); + constexpr int rows_per_warp = granularity; + constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. + + y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K); + + const int * x_qs = (const int *) x; + const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2; + const int * x_sc = (const int *) x_df + MMQ_TILE_NE_K/QI6_K; + const int * y_qs = (const int *) y + 4; + const float * y_df = (const float *) y; + + const int i0 = (threadIdx.y / ntx) * rows_per_warp; + + for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) { + const int k0 = k00 + k01; + + tile_A A[ntx]; +#pragma unroll + for (int n = 0; n < ntx; ++n) { + load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + k0, MMQ_MMA_TILE_X_K_Q6_K); + } + +#pragma unroll + for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { + tile_B B; + load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K); + + const int j = j0 + tile_C::get_j(0); + const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1]; + +#pragma unroll + for (int n = 0; n < ntx; ++n) { + tile_C C; + mma(C, A[n], B); + #pragma unroll for (int l = 0; l < tile_C::ne; ++l) { const int i = i0 + n*tile_C::I + tile_C::get_i(l); @@ -2303,7 +2467,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma( #else GGML_UNUSED_VARS(x, y, sum, k00); NO_DEVICE_CODE; -#endif // AMD_MFMA_AVAILABLE +#endif // AMD_MFMA_AVAILABLE || AMD_WMMA_AVAILABLE } template static __device__ __forceinline__ void load_tiles_iq4_nl( @@ -2311,14 +2475,14 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); #else constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_NL, mmq_y); int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_NL); constexpr int nrows = warp_size / threads_per_row; @@ -2340,13 +2504,13 @@ template static __device__ __forceinline__ void loa const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl); const int k0 = kbx * (2 * QI4_NL) + kqsx; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 0] = v.x; x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + QI4_NL] = v.y; #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + QI4_NL] = v.y; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } constexpr int blocks_per_tile_x_row = MMQ_TILE_NE_K / QI4_NL; @@ -2363,11 +2527,11 @@ template static __device__ __forceinline__ void loa const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbxd; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = __half2float(bxi->d); #else x_df[i*(MMQ_TILE_NE_K/QI4_NL) + i/QI4_NL + kbxd] = __half2float(bxi->d); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -2376,14 +2540,14 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); #else constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_XXS, mmq_y); int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XXS)) / 2; constexpr int nrows = warp_size / threads_per_row; @@ -2414,22 +2578,22 @@ template static __device__ __forceinline__ void loa const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000); const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid0; x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 1)] = grid1; #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid0; x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } const int ls = aux32 >> 28; const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = (ls*d + d/2)/4; #else x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = (ls*d + d/2)/4; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -2438,14 +2602,14 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); #else constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16; int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_XS)) / 2; constexpr int nrows = warp_size / threads_per_row; @@ -2472,24 +2636,24 @@ template static __device__ __forceinline__ void loa const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]); const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l; x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 1)] = grid_h; #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } const int ls = bxi->scales[kqsx]; const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; #else x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -2498,15 +2662,14 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); #else constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_S, mmq_y); int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) - +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR2_S)) / 2; constexpr int nrows = warp_size / threads_per_row; const int kqsx = threadIdx.x % threads_per_row; @@ -2539,24 +2702,24 @@ template static __device__ __forceinline__ void loa const int grid_l = __vsub4(grid_pos[0] ^ signs0, signs0); const int grid_h = __vsub4(grid_pos[1] ^ signs1, signs1); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l; x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 1)] = grid_h; #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } const int ls = bxi->scales[kqsx]; const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; x_df[i*MMQ_MMA_TILE_X_K_Q3_K + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; #else x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls & 0x0F)*d + d/2)/4; x_df[i*(2*MMQ_TILE_NE_K*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >> 4)*d + d/2)/4; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -2565,14 +2728,14 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); #else constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_XXS, mmq_y); int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_XXS)) / 2; constexpr int nrows = warp_size / threads_per_row; @@ -2601,22 +2764,22 @@ template static __device__ __forceinline__ void loa const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]); const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid_l; x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 1)] = grid_h; #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 0)] = grid_l; x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l + 1)] = grid_h; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } const int ls = aux32 >> 28; const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = (ls*d + d/2)/2; #else x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = (ls*d + d/2)/2; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -2625,14 +2788,14 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); #else constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, mmq_y); int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = (MMQ_ITER_K / (4 * QR3_S)) / 2; constexpr int nrows = warp_size / threads_per_row; @@ -2668,22 +2831,22 @@ template static __device__ __forceinline__ void loa const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0); const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l+0)] = grid_l; x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l+1)] = grid_h; #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid_l; x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid_h; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } const int ls = 1 + 2*((bxi->scales[kqsx/2] >> (((2*kqsx) << 1) & 0x04)) & 0x0F); const float d = bxi->d; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = ls*d; #else x_df[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = ls*d; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -2692,14 +2855,14 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; half2 * x_ds = (half2 *) (x_qs + MMQ_TILE_NE_K*2); #else constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, mmq_y); int * x_qs = (int *) x_tile; half2 * x_ds = (half2 *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = MMQ_ITER_K / (4 * QR1_S); constexpr int nrows = warp_size / threads_per_row; @@ -2727,23 +2890,23 @@ template static __device__ __forceinline__ void loa const int grid0 = (grid >> 0) & 0x0F0F0F0F; const int grid1 = (grid >> 4) & 0x0F0F0F0F; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 8*kqsx + (2*l+0)] = grid0; x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 8*kqsx + (2*l+1)] = grid1; #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+0)] = grid0; x_qs[i*(2*MMQ_TILE_NE_K + 1) + 8*kqsx + (2*l+1)] = grid1; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } const float d1q = __half2float(bxi->d) * (((qh >> 11) & 0x0E) + 1); const float delta = -1.0f + IQ1S_DELTA - (qh & 0x8000) * (2.0f*IQ1S_DELTA/0x8000); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_ds[i*MMQ_MMA_TILE_X_K_Q8_1 + kqsx] = make_half2(d1q, d1q*delta); #else x_ds[i*(MMQ_TILE_NE_K/4) + i/4 + kqsx] = make_half2(d1q, d1q*delta); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -2752,14 +2915,14 @@ template static __device__ __forceinline__ void loa constexpr int nwarps = mmq_get_nwarps_device(); constexpr int warp_size = ggml_cuda_get_physical_warp_size(); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + MMQ_TILE_NE_K*2); #else constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_XS, mmq_y); int * x_qs = (int *) x_tile; float * x_df = (float *) (x_qs + txs.qs); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int threads_per_row = MMQ_ITER_K / (4 * QR4_XS); constexpr int nrows = warp_size / threads_per_row; @@ -2779,13 +2942,13 @@ template static __device__ __forceinline__ void loa const int2 v = get_int_from_table_16(aux_q4, kvalues_iq4nl); const int k0 = 8 * (kqsx / 4) + kqsx % 4; -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 0] = v.x; x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 4] = v.y; #else x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 0] = v.x; x_qs[i*(2*MMQ_TILE_NE_K + 1) + k0 + 4] = v.y; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } constexpr int rows_per_warp = warp_size / 8; @@ -2804,11 +2967,11 @@ template static __device__ __forceinline__ void loa const int ls = ((bxi->scales_l[(threadIdx.x % 8)/2] >> (4*(threadIdx.x % 2))) & 0x0F) | (((bxi->scales_h >> (2*(threadIdx.x % 8))) & 0x03) << 4); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + threadIdx.x % 8] = d * (ls - 32); #else x_df[i*(MMQ_TILE_NE_K/4) + i/4 + threadIdx.x % 8] = d * (ls - 32); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) } } @@ -2848,7 +3011,7 @@ static __device__ __forceinline__ void mmq_write_back_mma( constexpr int granularity = mmq_get_granularity_device(mmq_x); constexpr int nwarps = mmq_get_nwarps_device(); -#if defined(AMD_MFMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int tileC_IJ = mmq_get_granularity_device(0); typedef tile tile_C; constexpr int rows_per_warp = granularity; @@ -2859,11 +3022,11 @@ static __device__ __forceinline__ void mmq_write_back_mma( constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp. const int i0 = (threadIdx.y / ntx) * (ntx*tile_C::I); -#if defined(TURING_MMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE) +#if defined(TURING_MMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) static_assert(nwarps*tile_C::I == mmq_y, "nwarps*tile_C::I != mmq_y"); #else GGML_UNUSED(nwarps); -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) #pragma unroll for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) { @@ -3063,13 +3226,13 @@ static __device__ __forceinline__ void mul_mat_q_process_tile( int * tile_y = data_mul_mat_q + mmq_x; int * tile_x = tile_y + GGML_PAD(mmq_x*MMQ_TILE_Y_K, nwarps*warp_size); -#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr vec_dot_mmq_t vec_dot = mmq_type_traits::vec_dot_mma; constexpr mmq_write_back_t write_back = mmq_write_back_mma; #else constexpr vec_dot_mmq_t vec_dot = mmq_type_traits::vec_dot_dp4a; constexpr mmq_write_back_t write_back = mmq_write_back_dp4a; -#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) +#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) constexpr int blocks_per_iter = MMQ_ITER_K / qk; From 134e6940caf5c64071b7f3b7bc6c2f32f1b3a5a4 Mon Sep 17 00:00:00 2001 From: Daniel Bevenius Date: Mon, 24 Nov 2025 21:06:17 +0100 Subject: [PATCH 03/18] llama : skip output reordering for single token batches (#17466) This commit adds a check to skip the output reordering logic when n_outputs == 1. With a single output token, the data is trivially sorted and the reordering code is currently doing unnecessary work (resetting and rebuilding output_ids to the same values). The motivation for this change is improved code clarity and avoiding confusion when debugging. While the performance impact is probably negligible, this unnecessary work happens on every decode call in llama-server when processing batches with single-token outputs. --- src/llama-context.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/src/llama-context.cpp b/src/llama-context.cpp index 70a3ec62df..2aa6d52a24 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -1248,7 +1248,7 @@ int llama_context::decode(const llama_batch & batch_inp) { // make the outputs have the same order they had in the user-provided batch // note: this is mostly relevant for recurrent models atm - if (!sorted_output) { + if (!sorted_output && n_outputs > 1) { GGML_ASSERT((size_t) n_outputs == out_ids.size()); // TODO: is there something more efficient which also minimizes swaps? From 3d07caa99bff9213411202b4063aa2f44e919654 Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Mon, 24 Nov 2025 15:25:24 -0600 Subject: [PATCH 04/18] vulkan: more FA details in vk_perf_logger (#17443) --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 16 ++++++++++++++++ 1 file changed, 16 insertions(+) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index bc8d3cdcb5..d78c727e53 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -1629,6 +1629,22 @@ class vk_perf_logger { timings[name].push_back(time); return; } + if (node->op == GGML_OP_FLASH_ATTN_EXT) { + const ggml_tensor * dst = node; + const ggml_tensor * q = node->src[0]; + const ggml_tensor * k = node->src[1]; + const ggml_tensor * v = node->src[2]; + const ggml_tensor * m = node->src[3]; + std::stringstream name; + name << ggml_op_name(node->op) << + " dst(" << dst->ne[0] << "," << dst->ne[1] << "," << dst->ne[2] << "," << dst->ne[3] << "), " << + " q(" << q->ne[0] << "," << q->ne[1] << "," << q->ne[2] << "," << q->ne[3] << "), " << + " k(" << k->ne[0] << "," << k->ne[1] << "," << k->ne[2] << "," << k->ne[3] << "), " << + " v(" << v->ne[0] << "," << v->ne[1] << "," << v->ne[2] << "," << v->ne[3] << "), " << + " m(" << (m?m->ne[0]:0) << "," << (m?m->ne[1]:0) << "," << (m?m->ne[2]:0) << "," << (m?m->ne[3]:0) << ")"; + timings[name.str()].push_back(time); + return; + } timings[ggml_op_name(node->op)].push_back(time); } private: From 877566d512d02a3a288ea38c1f7b1c25ccd6c082 Mon Sep 17 00:00:00 2001 From: Aaron Teo Date: Tue, 25 Nov 2025 09:56:07 +0800 Subject: [PATCH 05/18] llama: introduce support for model-embedded sampling parameters (#17120) --- common/arg.cpp | 12 ++++++ common/common.cpp | 55 ++++++++++++++++++++++++ common/common.h | 18 ++++++++ gguf-py/gguf/constants.py | 14 ++++++ gguf-py/gguf/gguf_writer.py | 36 ++++++++++++++++ gguf-py/gguf/metadata.py | 85 +++++++++++++++++++++++++++++++++++++ include/llama.h | 18 ++++++++ src/llama-arch.cpp | 38 +++++++++++------ src/llama-arch.h | 12 ++++++ src/llama-model.cpp | 18 ++++++++ 10 files changed, 293 insertions(+), 13 deletions(-) diff --git a/common/arg.cpp b/common/arg.cpp index 430ab45dfe..dd787290d2 100644 --- a/common/arg.cpp +++ b/common/arg.cpp @@ -1232,6 +1232,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params, const std::string & value) { const auto sampler_names = string_split(value, ';'); params.sampling.samplers = common_sampler_types_from_names(sampler_names, true); + params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS; } ).set_sparam()); add_opt(common_arg( @@ -1261,6 +1262,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex [](common_params & params, const std::string & value) { params.sampling.temp = std::stof(value); params.sampling.temp = std::max(params.sampling.temp, 0.0f); + params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP; } ).set_sparam()); add_opt(common_arg( @@ -1268,6 +1270,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex string_format("top-k sampling (default: %d, 0 = disabled)", params.sampling.top_k), [](common_params & params, int value) { params.sampling.top_k = value; + params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K; } ).set_sparam()); add_opt(common_arg( @@ -1275,6 +1278,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex string_format("top-p sampling (default: %.1f, 1.0 = disabled)", (double)params.sampling.top_p), [](common_params & params, const std::string & value) { params.sampling.top_p = std::stof(value); + params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P; } ).set_sparam()); add_opt(common_arg( @@ -1282,6 +1286,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex string_format("min-p sampling (default: %.1f, 0.0 = disabled)", (double)params.sampling.min_p), [](common_params & params, const std::string & value) { params.sampling.min_p = std::stof(value); + params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P; } ).set_sparam()); add_opt(common_arg( @@ -1296,6 +1301,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sampling.xtc_probability), [](common_params & params, const std::string & value) { params.sampling.xtc_probability = std::stof(value); + params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY; } ).set_sparam()); add_opt(common_arg( @@ -1303,6 +1309,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex string_format("xtc threshold (default: %.1f, 1.0 = disabled)", (double)params.sampling.xtc_threshold), [](common_params & params, const std::string & value) { params.sampling.xtc_threshold = std::stof(value); + params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD; } ).set_sparam()); add_opt(common_arg( @@ -1321,6 +1328,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex } params.sampling.penalty_last_n = value; params.sampling.n_prev = std::max(params.sampling.n_prev, params.sampling.penalty_last_n); + params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N; } ).set_sparam()); add_opt(common_arg( @@ -1328,6 +1336,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex string_format("penalize repeat sequence of tokens (default: %.1f, 1.0 = disabled)", (double)params.sampling.penalty_repeat), [](common_params & params, const std::string & value) { params.sampling.penalty_repeat = std::stof(value); + params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT; } ).set_sparam()); add_opt(common_arg( @@ -1425,6 +1434,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex "(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", params.sampling.mirostat), [](common_params & params, int value) { params.sampling.mirostat = value; + params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT; } ).set_sparam()); add_opt(common_arg( @@ -1432,6 +1442,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex string_format("Mirostat learning rate, parameter eta (default: %.1f)", (double)params.sampling.mirostat_eta), [](common_params & params, const std::string & value) { params.sampling.mirostat_eta = std::stof(value); + params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA; } ).set_sparam()); add_opt(common_arg( @@ -1439,6 +1450,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex string_format("Mirostat target entropy, parameter tau (default: %.1f)", (double)params.sampling.mirostat_tau), [](common_params & params, const std::string & value) { params.sampling.mirostat_tau = std::stof(value); + params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU; } ).set_sparam()); add_opt(common_arg( diff --git a/common/common.cpp b/common/common.cpp index f3cc55247e..0d7fd9a937 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -8,6 +8,7 @@ #include "common.h" #include "log.h" #include "llama.h" +#include "sampling.h" #include #include @@ -949,6 +950,58 @@ std::vector fs_list_files(const std::string & path) { // Model utils // +static inline void common_init_sampler_from_model( + const llama_model * model, + common_params_sampling & sparams) { + + const uint64_t config = sparams.user_sampling_config; + + auto get_int32 = [&](const char * key, int32_t & dst, uint64_t user_config) { + if (config & user_config) return; + + char buf[64] = {0}; + if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) { + char * end = nullptr; + int32_t v = strtol(buf, &end, 10); + if (end && end != buf) dst = v; + } + }; + + auto get_float = [&](const char * key, float & dst, uint64_t user_config) { + if (config & user_config) return; + + char buf[128] = {0}; + if (llama_model_meta_val_str(model, key, buf, sizeof(buf)) > 0) { + char * end = nullptr; + float v = strtof(buf, &end); + if (end && end != buf) dst = v; + } + }; + + // Sampling sequence + if (!(config & common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS)) { + char buf[512] = {0}; + if (llama_model_meta_val_str(model, llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE), buf, sizeof(buf)) > 0) { + const std::vector sampler_names = string_split(std::string(buf), ';'); + if (!sampler_names.empty()) { + sparams.samplers = common_sampler_types_from_names(sampler_names, true); + } + } + } + + get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_K), sparams.top_k, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_K); + get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TOP_P), sparams.top_p, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TOP_P); + get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIN_P), sparams.min_p, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIN_P); + get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY), sparams.xtc_probability, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY); + get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD), sparams.xtc_threshold, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD); + get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_TEMP), sparams.temp, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_TEMP); + get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N), sparams.penalty_last_n, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N); + get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT), sparams.penalty_repeat, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT); + get_int32(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT), sparams.mirostat, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT); + get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU), sparams.mirostat_tau, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU); + get_float(llama_model_meta_key_str(LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA), sparams.mirostat_eta, common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA); +} + struct common_init_result common_init_from_params(common_params & params) { common_init_result iparams; auto mparams = common_model_params_to_llama(params); @@ -960,6 +1013,8 @@ struct common_init_result common_init_from_params(common_params & params) { return iparams; } + common_init_sampler_from_model(model, params.sampling); + const llama_vocab * vocab = llama_model_get_vocab(model); auto cparams = common_context_params_to_llama(params); diff --git a/common/common.h b/common/common.h index de5b404dd8..2f23d0baa8 100644 --- a/common/common.h +++ b/common/common.h @@ -140,6 +140,22 @@ struct common_grammar_trigger { llama_token token = LLAMA_TOKEN_NULL; }; +enum common_params_sampling_config : uint64_t { + COMMON_PARAMS_SAMPLING_CONFIG_SAMPLERS = 1 << 0, + COMMON_PARAMS_SAMPLING_CONFIG_TOP_K = 1 << 1, + COMMON_PARAMS_SAMPLING_CONFIG_TOP_P = 1 << 2, + COMMON_PARAMS_SAMPLING_CONFIG_MIN_P = 1 << 3, + COMMON_PARAMS_SAMPLING_CONFIG_XTC_PROBABILITY = 1 << 4, + COMMON_PARAMS_SAMPLING_CONFIG_XTC_THRESHOLD = 1 << 5, + COMMON_PARAMS_SAMPLING_CONFIG_TEMP = 1 << 6, + COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_LAST_N = 1 << 7, + COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT = 1 << 8, + COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT = 1 << 9, + COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_TAU = 1 << 10, + COMMON_PARAMS_SAMPLING_CONFIG_MIROSTAT_ETA = 1 << 11, +}; + + // sampling parameters struct common_params_sampling { uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler @@ -172,6 +188,8 @@ struct common_params_sampling { bool no_perf = false; // disable performance metrics bool timing_per_token = false; + uint64_t user_sampling_config = 0; // bitfield to track user-specified samplers + std::vector dry_sequence_breakers = {"\n", ":", "\"", "*"}; // default sequence breakers for DRY diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 8bc558fe4b..6f5a742e04 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -25,6 +25,20 @@ class Keys: ALIGNMENT = "general.alignment" FILE_TYPE = "general.file_type" + # Recommended Sampler Parameters + SAMPLING_SEQUENCE = "general.sampling.sequence" + SAMPLING_TOP_K = "general.sampling.top_k" + SAMPLING_TOP_P = "general.sampling.top_p" + SAMPLING_MIN_P = "general.sampling.min_p" + SAMPLING_XTC_PROBABILITY = "general.sampling.xtc_probability" + SAMPLING_XTC_THRESHOLD = "general.sampling.xtc_threshold" + SAMPLING_TEMP = "general.sampling.temp" + SAMPLING_PENALTY_LAST_N = "general.sampling.penalty_last_n" + SAMPLING_PENALTY_REPEAT = "general.sampling.penalty_repeat" + SAMPLING_MIROSTAT = "general.sampling.mirostat" + SAMPLING_MIROSTAT_TAU = "general.sampling.mirostat_tau" + SAMPLING_MIROSTAT_ETA = "general.sampling.mirostat_eta" + # Authorship Metadata NAME = "general.name" AUTHOR = "general.author" diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index a051daeeb1..642ae2ae59 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -496,6 +496,42 @@ class GGUFWriter: def add_file_type(self, ftype: int) -> None: self.add_uint32(Keys.General.FILE_TYPE, ftype) + def add_sampling_sequence(self, sequence: str) -> None: + self.add_string(Keys.General.SAMPLING_SEQUENCE, sequence) + + def add_sampling_top_k(self, top_k: int) -> None: + self.add_int32(Keys.General.SAMPLING_TOP_K, top_k) + + def add_sampling_top_p(self, top_p: float) -> None: + self.add_float32(Keys.General.SAMPLING_TOP_P, top_p) + + def add_sampling_min_p(self, min_p: float) -> None: + self.add_float32(Keys.General.SAMPLING_MIN_P, min_p) + + def add_sampling_xtc_probability(self, xtc_probability: float) -> None: + self.add_float32(Keys.General.SAMPLING_XTC_PROBABILITY, xtc_probability) + + def add_sampling_xtc_threshold(self, xtc_threshold: float) -> None: + self.add_float32(Keys.General.SAMPLING_XTC_THRESHOLD, xtc_threshold) + + def add_sampling_temp(self, temp: float) -> None: + self.add_float32(Keys.General.SAMPLING_TEMP, temp) + + def add_sampling_penalty_last_n(self, penalty_last_n: int) -> None: + self.add_int32(Keys.General.SAMPLING_PENALTY_LAST_N, penalty_last_n) + + def add_sampling_penalty_repeat(self, penalty_repeat: float) -> None: + self.add_float32(Keys.General.SAMPLING_PENALTY_REPEAT, penalty_repeat) + + def add_sampling_mirostat(self, mirostat: int) -> None: + self.add_int32(Keys.General.SAMPLING_MIROSTAT, mirostat) + + def add_sampling_mirostat_tau(self, mirostat_tau: float) -> None: + self.add_float32(Keys.General.SAMPLING_MIROSTAT_TAU, mirostat_tau) + + def add_sampling_mirostat_eta(self, mirostat_eta: float) -> None: + self.add_float32(Keys.General.SAMPLING_MIROSTAT_ETA, mirostat_eta) + def add_name(self, name: str) -> None: self.add_string(Keys.General.NAME, name) diff --git a/gguf-py/gguf/metadata.py b/gguf-py/gguf/metadata.py index 67efedbdbc..e0d478ce95 100644 --- a/gguf-py/gguf/metadata.py +++ b/gguf-py/gguf/metadata.py @@ -17,6 +17,20 @@ logger = logging.getLogger("metadata") @dataclass class Metadata: + # Recommended Sampler Parameters to be written to GGUF KV Store + sampling_sequence: Optional[str] = None + sampling_top_k: Optional[int] = None + sampling_top_p: Optional[float] = None + sampling_min_p: Optional[float] = None + sampling_xtc_probability: Optional[float] = None + sampling_xtc_threshold: Optional[float] = None + sampling_temp: Optional[float] = None + sampling_penalty_last_n: Optional[int] = None + sampling_penalty_repeat: Optional[float] = None + sampling_mirostat: Optional[int] = None + sampling_mirostat_tau: Optional[float] = None + sampling_mirostat_eta: Optional[float] = None + # Authorship Metadata to be written to GGUF KV Store name: Optional[str] = None author: Optional[str] = None @@ -54,15 +68,43 @@ class Metadata: model_card = Metadata.load_model_card(model_path) hf_params = Metadata.load_hf_parameters(model_path) + gen_config = Metadata.load_generation_config(model_path) # TODO: load adapter_config.json when possible, it usually contains the base model of the LoRA adapter # heuristics metadata = Metadata.apply_metadata_heuristic(metadata, model_card, hf_params, model_path, total_params) + if gen_config: + metadata.sampling_sequence = gen_config.get("sequence", metadata.sampling_sequence) + metadata.sampling_top_k = gen_config.get("top_k", metadata.sampling_top_k) + metadata.sampling_top_p = gen_config.get("top_p", metadata.sampling_top_p) + metadata.sampling_min_p = gen_config.get("min_p", metadata.sampling_min_p) + metadata.sampling_xtc_probability = gen_config.get("xtc_probability", metadata.sampling_xtc_probability) + metadata.sampling_xtc_threshold = gen_config.get("xtc_threshold", metadata.sampling_xtc_threshold) + metadata.sampling_temp = gen_config.get("temperature", metadata.sampling_temp) + metadata.sampling_penalty_last_n = gen_config.get("penalty_last_n", metadata.sampling_penalty_last_n) + metadata.sampling_penalty_repeat = gen_config.get("penalty_repeat", metadata.sampling_penalty_repeat) + metadata.sampling_mirostat = gen_config.get("mirostat", metadata.sampling_mirostat) + metadata.sampling_mirostat_tau = gen_config.get("mirostat_tau", metadata.sampling_mirostat_tau) + metadata.sampling_mirostat_eta = gen_config.get("mirostat_eta", metadata.sampling_mirostat_eta) + # Metadata Override File Provided # This is based on LLM_KV_NAMES mapping in llama.cpp metadata_override = Metadata.load_metadata_override(metadata_override_path) + metadata.sampling_sequence = metadata_override.get(Keys.General.SAMPLING_SEQUENCE, metadata.sampling_sequence) + metadata.sampling_top_k = metadata_override.get(Keys.General.SAMPLING_TOP_K, metadata.sampling_top_k) + metadata.sampling_top_p = metadata_override.get(Keys.General.SAMPLING_TOP_P, metadata.sampling_top_p) + metadata.sampling_min_p = metadata_override.get(Keys.General.SAMPLING_MIN_P, metadata.sampling_min_p) + metadata.sampling_xtc_probability = metadata_override.get(Keys.General.SAMPLING_XTC_PROBABILITY, metadata.sampling_xtc_probability) + metadata.sampling_xtc_threshold = metadata_override.get(Keys.General.SAMPLING_XTC_THRESHOLD, metadata.sampling_xtc_threshold) + metadata.sampling_temp = metadata_override.get(Keys.General.SAMPLING_TEMP, metadata.sampling_temp) + metadata.sampling_penalty_last_n = metadata_override.get(Keys.General.SAMPLING_PENALTY_LAST_N, metadata.sampling_penalty_last_n) + metadata.sampling_penalty_repeat = metadata_override.get(Keys.General.SAMPLING_PENALTY_REPEAT, metadata.sampling_penalty_repeat) + metadata.sampling_mirostat = metadata_override.get(Keys.General.SAMPLING_MIROSTAT, metadata.sampling_mirostat) + metadata.sampling_mirostat_tau = metadata_override.get(Keys.General.SAMPLING_MIROSTAT_TAU, metadata.sampling_mirostat_tau) + metadata.sampling_mirostat_eta = metadata_override.get(Keys.General.SAMPLING_MIROSTAT_ETA, metadata.sampling_mirostat_eta) + metadata.name = metadata_override.get(Keys.General.NAME, metadata.name) metadata.author = metadata_override.get(Keys.General.AUTHOR, metadata.author) metadata.version = metadata_override.get(Keys.General.VERSION, metadata.version) @@ -172,6 +214,23 @@ class Metadata: with open(config_path, "r", encoding="utf-8") as f: return json.load(f) + @staticmethod + def load_generation_config(model_path: Optional[Path] = None) -> dict[str, Any]: + if model_path is None or not model_path.is_dir(): + return {} + + generation_config_path = model_path / "generation_config.json" + + if not generation_config_path.is_file(): + return {} + + try: + with open(generation_config_path, "r", encoding="utf-8") as f: + return json.load(f) + except (json.JSONDecodeError, IOError): + # not all models have valid generation_config.json + return {} + @staticmethod def id_to_title(string): # Convert capitalization into title form unless acronym or version number @@ -546,6 +605,32 @@ class Metadata: def set_gguf_meta_model(self, gguf_writer: gguf.GGUFWriter): assert self.name is not None + + if self.sampling_sequence is not None: + gguf_writer.add_sampling_sequence(self.sampling_sequence) + if self.sampling_top_k is not None: + gguf_writer.add_sampling_top_k(self.sampling_top_k) + if self.sampling_top_p is not None: + gguf_writer.add_sampling_top_p(self.sampling_top_p) + if self.sampling_min_p is not None: + gguf_writer.add_sampling_min_p(self.sampling_min_p) + if self.sampling_xtc_probability is not None: + gguf_writer.add_sampling_xtc_probability(self.sampling_xtc_probability) + if self.sampling_xtc_threshold is not None: + gguf_writer.add_sampling_xtc_threshold(self.sampling_xtc_threshold) + if self.sampling_temp is not None: + gguf_writer.add_sampling_temp(self.sampling_temp) + if self.sampling_penalty_last_n is not None: + gguf_writer.add_sampling_penalty_last_n(self.sampling_penalty_last_n) + if self.sampling_penalty_repeat is not None: + gguf_writer.add_sampling_penalty_repeat(self.sampling_penalty_repeat) + if self.sampling_mirostat is not None: + gguf_writer.add_sampling_mirostat(self.sampling_mirostat) + if self.sampling_mirostat_tau is not None: + gguf_writer.add_sampling_mirostat_tau(self.sampling_mirostat_tau) + if self.sampling_mirostat_eta is not None: + gguf_writer.add_sampling_mirostat_eta(self.sampling_mirostat_eta) + gguf_writer.add_name(self.name) if self.author is not None: diff --git a/include/llama.h b/include/llama.h index 8547226ff2..b52eaacfa7 100644 --- a/include/llama.h +++ b/include/llama.h @@ -246,6 +246,21 @@ extern "C" { LLAMA_KV_OVERRIDE_TYPE_STR, }; + enum llama_model_meta_key { + LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE, + LLAMA_MODEL_META_KEY_SAMPLING_TOP_K, + LLAMA_MODEL_META_KEY_SAMPLING_TOP_P, + LLAMA_MODEL_META_KEY_SAMPLING_MIN_P, + LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY, + LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD, + LLAMA_MODEL_META_KEY_SAMPLING_TEMP, + LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N, + LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT, + LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT, + LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU, + LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA, + }; + struct llama_model_kv_override { enum llama_model_kv_override_type tag; @@ -518,6 +533,9 @@ extern "C" { // Get the number of metadata key/value pairs LLAMA_API int32_t llama_model_meta_count(const struct llama_model * model); + // Get sampling metadata key name. Returns nullptr if the key is invalid + LLAMA_API const char * llama_model_meta_key_str(enum llama_model_meta_key key); + // Get metadata key name by index LLAMA_API int32_t llama_model_meta_key_by_index(const struct llama_model * model, int32_t i, char * buf, size_t buf_size); diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index fc6cddc92f..7ef87acf1b 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -114,19 +114,31 @@ static const std::map LLM_ARCH_NAMES = { }; static const std::map LLM_KV_NAMES = { - { LLM_KV_GENERAL_TYPE, "general.type" }, - { LLM_KV_GENERAL_ARCHITECTURE, "general.architecture" }, - { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version" }, - { LLM_KV_GENERAL_ALIGNMENT, "general.alignment" }, - { LLM_KV_GENERAL_FILE_TYPE, "general.file_type" }, - { LLM_KV_GENERAL_NAME, "general.name" }, - { LLM_KV_GENERAL_AUTHOR, "general.author" }, - { LLM_KV_GENERAL_VERSION, "general.version" }, - { LLM_KV_GENERAL_URL, "general.url" }, - { LLM_KV_GENERAL_DESCRIPTION, "general.description" }, - { LLM_KV_GENERAL_LICENSE, "general.license" }, - { LLM_KV_GENERAL_SOURCE_URL, "general.source.url" }, - { LLM_KV_GENERAL_SOURCE_HF_REPO, "general.source.huggingface.repository" }, + { LLM_KV_GENERAL_TYPE, "general.type" }, + { LLM_KV_GENERAL_ARCHITECTURE, "general.architecture" }, + { LLM_KV_GENERAL_QUANTIZATION_VERSION, "general.quantization_version" }, + { LLM_KV_GENERAL_ALIGNMENT, "general.alignment" }, + { LLM_KV_GENERAL_FILE_TYPE, "general.file_type" }, + { LLM_KV_GENERAL_SAMPLING_SEQUENCE, "general.sampling.sequence" }, + { LLM_KV_GENERAL_SAMPLING_TOP_K, "general.sampling.top_k" }, + { LLM_KV_GENERAL_SAMPLING_TOP_P, "general.sampling.top_p" }, + { LLM_KV_GENERAL_SAMPLING_MIN_P, "general.sampling.min_p" }, + { LLM_KV_GENERAL_SAMPLING_XTC_PROBABILITY, "general.sampling.xtc_probability" }, + { LLM_KV_GENERAL_SAMPLING_XTC_THRESHOLD, "general.sampling.xtc_threshold" }, + { LLM_KV_GENERAL_SAMPLING_TEMP, "general.sampling.temp" }, + { LLM_KV_GENERAL_SAMPLING_PENALTY_LAST_N, "general.sampling.penalty_last_n" }, + { LLM_KV_GENERAL_SAMPLING_PENALTY_REPEAT, "general.sampling.penalty_repeat" }, + { LLM_KV_GENERAL_SAMPLING_MIROSTAT, "general.sampling.mirostat" }, + { LLM_KV_GENERAL_SAMPLING_MIROSTAT_TAU, "general.sampling.mirostat_tau" }, + { LLM_KV_GENERAL_SAMPLING_MIROSTAT_ETA, "general.sampling.mirostat_eta" }, + { LLM_KV_GENERAL_NAME, "general.name" }, + { LLM_KV_GENERAL_AUTHOR, "general.author" }, + { LLM_KV_GENERAL_VERSION, "general.version" }, + { LLM_KV_GENERAL_URL, "general.url" }, + { LLM_KV_GENERAL_DESCRIPTION, "general.description" }, + { LLM_KV_GENERAL_LICENSE, "general.license" }, + { LLM_KV_GENERAL_SOURCE_URL, "general.source.url" }, + { LLM_KV_GENERAL_SOURCE_HF_REPO, "general.source.huggingface.repository" }, { LLM_KV_VOCAB_SIZE, "%s.vocab_size" }, { LLM_KV_CONTEXT_LENGTH, "%s.context_length" }, diff --git a/src/llama-arch.h b/src/llama-arch.h index 02a1c2dc25..9ad3157bf6 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -123,6 +123,18 @@ enum llm_kv { LLM_KV_GENERAL_QUANTIZATION_VERSION, LLM_KV_GENERAL_ALIGNMENT, LLM_KV_GENERAL_FILE_TYPE, + LLM_KV_GENERAL_SAMPLING_SEQUENCE, + LLM_KV_GENERAL_SAMPLING_TOP_K, + LLM_KV_GENERAL_SAMPLING_TOP_P, + LLM_KV_GENERAL_SAMPLING_MIN_P, + LLM_KV_GENERAL_SAMPLING_XTC_PROBABILITY, + LLM_KV_GENERAL_SAMPLING_XTC_THRESHOLD, + LLM_KV_GENERAL_SAMPLING_TEMP, + LLM_KV_GENERAL_SAMPLING_PENALTY_LAST_N, + LLM_KV_GENERAL_SAMPLING_PENALTY_REPEAT, + LLM_KV_GENERAL_SAMPLING_MIROSTAT, + LLM_KV_GENERAL_SAMPLING_MIROSTAT_TAU, + LLM_KV_GENERAL_SAMPLING_MIROSTAT_ETA, LLM_KV_GENERAL_NAME, LLM_KV_GENERAL_AUTHOR, LLM_KV_GENERAL_VERSION, diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 35179a98e0..a042ea9632 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -7687,6 +7687,24 @@ int32_t llama_model_meta_count(const llama_model * model) { return (int)model->gguf_kv.size(); } +const char * llama_model_meta_key_str(llama_model_meta_key key) { + switch (key) { + case LLAMA_MODEL_META_KEY_SAMPLING_SEQUENCE: return "general.sampling.sequence"; + case LLAMA_MODEL_META_KEY_SAMPLING_TOP_K: return "general.sampling.top_k"; + case LLAMA_MODEL_META_KEY_SAMPLING_TOP_P: return "general.sampling.top_p"; + case LLAMA_MODEL_META_KEY_SAMPLING_MIN_P: return "general.sampling.min_p"; + case LLAMA_MODEL_META_KEY_SAMPLING_XTC_PROBABILITY: return "general.sampling.xtc_probability"; + case LLAMA_MODEL_META_KEY_SAMPLING_XTC_THRESHOLD: return "general.sampling.xtc_threshold"; + case LLAMA_MODEL_META_KEY_SAMPLING_TEMP: return "general.sampling.temp"; + case LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_LAST_N: return "general.sampling.penalty_last_n"; + case LLAMA_MODEL_META_KEY_SAMPLING_PENALTY_REPEAT: return "general.sampling.penalty_repeat"; + case LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT: return "general.sampling.mirostat"; + case LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_TAU: return "general.sampling.mirostat_tau"; + case LLAMA_MODEL_META_KEY_SAMPLING_MIROSTAT_ETA: return "general.sampling.mirostat_eta"; + default: return nullptr; + } +} + int32_t llama_model_meta_key_by_index(const llama_model * model, int i, char * buf, size_t buf_size) { if (i < 0 || i >= (int)model->gguf_kv.size()) { if (buf_size > 0) { From d414db02d3ab3744402bc57a7b3fce7de66e3d5a Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Tue, 25 Nov 2025 00:11:27 -0600 Subject: [PATCH 06/18] vulkan: Use fewer rows for scalar FA when HS is not a multiple of 16 (#17455) --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 12 +++++++----- tests/test-backend-ops.cpp | 3 +++ 2 files changed, 10 insertions(+), 5 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index d78c727e53..6cf15b43bb 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -2501,9 +2501,11 @@ static void ggml_vk_wait_events(vk_context& ctx, std::vector&& events static constexpr uint32_t flash_attention_num_small_rows = 32; static constexpr uint32_t scalar_flash_attention_num_small_rows = 1; -static uint32_t get_fa_scalar_num_large_rows(uint32_t hsv) { +static uint32_t get_fa_scalar_num_large_rows(uint32_t hsk, uint32_t hsv) { if (hsv >= 192) { return 2; + } else if ((hsv | hsk) & 8) { + return 4; } else { return 8; } @@ -2535,9 +2537,9 @@ static std::array fa_rows_cols(FaCodePath path, uint32_t hsk, uint3 if ((hsv | hsk) & 8) { // HSV/HSK not being a multiple of 16 makes D_split smaller, which makes cols_per_iter // larger, and Bc needs to be >= cols_per_thread. 64 is large enough, 32 is not. - return {get_fa_scalar_num_large_rows(hsv), 64}; + return {get_fa_scalar_num_large_rows(hsk, hsv), 64}; } else { - return {get_fa_scalar_num_large_rows(hsv), 32}; + return {get_fa_scalar_num_large_rows(hsk, hsv), 32}; } } } @@ -7740,7 +7742,7 @@ static bool ggml_vk_flash_attn_scalar_shmem_support(const vk_device& device, con // Needs to be kept up to date on shader changes GGML_UNUSED(hsv); const uint32_t wg_size = scalar_flash_attention_workgroup_size; - const uint32_t Br = get_fa_scalar_num_large_rows(hsv); + const uint32_t Br = get_fa_scalar_num_large_rows(hsk, hsv); const uint32_t Bc = scalar_flash_attention_Bc; const uint32_t tmpsh = wg_size * sizeof(float); @@ -7871,7 +7873,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx case FA_SCALAR: case FA_COOPMAT1: // We may switch from coopmat1 to scalar, so use the scalar limit for both - max_gqa = get_fa_scalar_num_large_rows(HSV); + max_gqa = get_fa_scalar_num_large_rows(HSK, HSV); break; case FA_COOPMAT2: max_gqa = get_fa_num_small_rows(FA_COOPMAT2); diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp index ce8c068d7a..fd48d25475 100644 --- a/tests/test-backend-ops.cpp +++ b/tests/test-backend-ops.cpp @@ -7859,6 +7859,9 @@ static std::vector> make_test_cases_perf() { } } + // Qwen3-VL-8B https://github.com/ggml-org/llama.cpp/issues/17012 + test_cases.emplace_back(new test_flash_attn_ext(72, 72, 16, {1, 1}, 5776, 5776, false, false, 0, 0, GGML_PREC_F32, GGML_TYPE_F16)); + for (int kv : { 4096, 8192, 16384, }) { for (int hs : { 64, 128, }) { for (int nr : { 1, 4, }) { From b1846f1c8ecd97ee08593e9498ef3244d43c1ad6 Mon Sep 17 00:00:00 2001 From: Pascal Date: Tue, 25 Nov 2025 08:01:02 +0100 Subject: [PATCH 07/18] webui: add rehype plugin to restore HTML in Markdown table cells (#17477) * webui: add rehype plugin to restore HTML in Markdown table cells The remark/rehype pipeline neutralizes inline HTML as literal text (remarkLiteralHtml) so that XML/HTML snippets in LLM responses display as-is instead of being rendered. This causes
and