Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

CUDA: add gqa_ratio 4 for GLM 4.7 flash (#18953)

Browse Source
This commit is contained in:
Aman Gupta
2026-01-22 18:51:53 +08:00
committed by GitHub
parent 5516b9c16a
commit b70d251076
8 changed files with 47 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/fattn-mma-f16.cuh
+23 -8
View File
@@ -432,7 +432,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
constexpr int ncols = ncols1 * ncols2;
constexpr int cols_per_warp = T_B_KQ::I;
constexpr int cols_per_thread = get_cols_per_thread();
constexpr int np = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
constexpr int np = cols_per_warp > ncols ? nwarps : nwarps * cols_per_warp/ncols; // Number of parallel CUDA warps per Q column.
constexpr int nbatch_fa = ggml_cuda_fattn_mma_get_nbatch_fa(DKQ, DV, ncols);
constexpr int nbatch_K2 = ggml_cuda_fattn_mma_get_nbatch_K2(DKQ, DV, ncols);
constexpr int nbatch_V2 = ggml_cuda_fattn_mma_get_nbatch_V2(DKQ, DV, ncols);
@@ -510,7 +510,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
}
}
} else {
static_assert(cols_per_warp != 8, "cols_per_warp == 8 not implemented");
#pragma unroll
for (int k_KQ_0 = k0_start; k_KQ_0 < k0_stop; k_KQ_0 += T_A_KQ::J) {
load_ldmatrix(Q_B[0], tile_Q + (threadIdx.y / np)*(T_B_KQ::I*stride_tile_Q) + k_KQ_0, stride_tile_Q);
@@ -522,14 +521,18 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
T_A_KQ K_A;
load_ldmatrix(K_A, tile_K + i_KQ_0*stride_tile_K + (k_KQ_0 - k0_start), stride_tile_K);
// Wide version of KQ_C is column-major
if constexpr (cols_per_warp == 8) {
mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
} else {
// Wide version of KQ_C is column-major
#if defined(AMD_WMMA_AVAILABLE)
// RDNA matrix C is column-major.
mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
// RDNA matrix C is column-major.
mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
#else
// swap A and B for CUDA.
mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
// swap A and B for CUDA.
mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
#endif // defined(AMD_WMMA_AVAILABLE)
}
}
}
}
@@ -953,7 +956,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
constexpr int cols_per_warp = T_B_KQ::I;
constexpr int cols_per_thread = get_cols_per_thread();
constexpr int np = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
constexpr int np = cols_per_warp > ncols ? nwarps : nwarps * cols_per_warp/ncols; // Number of parallel CUDA warps per Q column.
constexpr int nbatch_fa = ggml_cuda_fattn_mma_get_nbatch_fa (DKQ, DV, ncols);
constexpr int nbatch_K2 = ggml_cuda_fattn_mma_get_nbatch_K2 (DKQ, DV, ncols);
constexpr int nbatch_V2 = ggml_cuda_fattn_mma_get_nbatch_V2 (DKQ, DV, ncols);
@@ -1484,6 +1487,13 @@ static __global__ void flash_attn_ext_f16(
NO_DEVICE_CODE;
return;
}
#ifdef VOLTA_MMA_AVAILABLE
if (ncols1*ncols2 < 32) {
NO_DEVICE_CODE;
return;
}
#endif // VOLTA_MMA_AVAILABLE
#if __CUDA_ARCH__ == GGML_CUDA_CC_TURING
if (ncols1*ncols2 > 32) {
NO_DEVICE_CODE;
@@ -1728,3 +1738,8 @@ DECL_FATTN_MMA_F16_CASE_ALL_NCOLS2(256, 256, 64)
extern DECL_FATTN_MMA_F16_CASE(576, 512, 1, 16);
extern DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16);
extern DECL_FATTN_MMA_F16_CASE(576, 512, 4, 16);
// For GLM 4.7 Flash
extern DECL_FATTN_MMA_F16_CASE(576, 512, 4, 4);
extern DECL_FATTN_MMA_F16_CASE(576, 512, 8, 4);
extern DECL_FATTN_MMA_F16_CASE(576, 512, 16, 4);