happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

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

This commit is contained in:
Aman Gupta 2026-01-22 18:51:53 +08:00 committed by GitHub
parent 5516b9c16a
commit b70d251076
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
8 changed files with 47 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
ggml/src/ggml-cuda/fattn-mma-f16.cuh
View File

@ -432,7 +432,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
constexpr int ncols = ncols1 * ncols2; constexpr int ncols = ncols1 * ncols2;
constexpr int cols_per_warp = T_B_KQ::I; constexpr int cols_per_warp = T_B_KQ::I;
constexpr int cols_per_thread = get_cols_per_thread(); 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_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_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); 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 { } else {
static_assert(cols_per_warp != 8, "cols_per_warp == 8 not implemented");
#pragma unroll #pragma unroll
for (int k_KQ_0 = k0_start; k_KQ_0 < k0_stop; k_KQ_0 += T_A_KQ::J) { 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); 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; 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); 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) #if defined(AMD_WMMA_AVAILABLE)
// RDNA matrix C is column-major. // RDNA matrix C is column-major.
mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]); mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
#else #else
// swap A and B for CUDA. // swap A and B for CUDA.
mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A); mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
#endif // defined(AMD_WMMA_AVAILABLE) #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_warp = T_B_KQ::I;
constexpr int cols_per_thread = get_cols_per_thread(); 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_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_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); 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; NO_DEVICE_CODE;
return; 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 __CUDA_ARCH__ == GGML_CUDA_CC_TURING
if (ncols1*ncols2 > 32) { if (ncols1*ncols2 > 32) {
NO_DEVICE_CODE; 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, 1, 16);
extern DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16); extern DECL_FATTN_MMA_F16_CASE(576, 512, 2, 16);
extern DECL_FATTN_MMA_F16_CASE(576, 512, 4, 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);

12
ggml/src/ggml-cuda/fattn-tile.cuh
View File

@ -68,6 +68,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 64, 64) GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 64, 64) GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 64, 64) GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 64, 64)
return 0; return 0;
@ -122,6 +124,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 32, 128) GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 32, 128)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 32, 64) GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 32, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 32, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 32, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 32, 64) GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 32, 64)
return 0; return 0;
@ -183,6 +187,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 32, 128) GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 2, 32, 128)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 32, 128) GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 2, 32, 128)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 64, 64) GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 512, 1, 128, 64) GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 512, 1, 128, 64)
@ -245,6 +251,8 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5, 32, 256) GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 16, 256, 5, 32, 256)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3, 64, 128) GGML_CUDA_FATTN_TILE_CONFIG_CASE(256, 256, 32, 256, 3, 64, 128)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 4, 128, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 8, 256, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4, 64, 64) GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 16, 256, 4, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2, 128, 64) GGML_CUDA_FATTN_TILE_CONFIG_CASE(576, 512, 32, 256, 2, 128, 64)
@ -1187,6 +1195,10 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst); launch_fattn_tile_switch_ncols1<DKQ, DV, 16, use_logit_softcap>(ctx, dst);
return; return;
} }
if (use_gqa_opt && gqa_ratio % 4 == 0) {
launch_fattn_tile_switch_ncols1<DKQ, DV, 4, use_logit_softcap>(ctx, dst);
return;
}
} }
if constexpr (DV <= 256) { if constexpr (DV <= 256) {

10
ggml/src/ggml-cuda/fattn.cu
View File

@ -121,8 +121,12 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
GGML_ASSERT(Q->ne[2] % K->ne[2] == 0); GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
const int gqa_ratio = Q->ne[2] / K->ne[2]; const int gqa_ratio = Q->ne[2] / K->ne[2];
GGML_ASSERT(gqa_ratio % 16 == 0); GGML_ASSERT(gqa_ratio % 4 == 0);
ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst); if (gqa_ratio % 16 == 0) {
ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
} else {
ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 4>(ctx, dst);
}
} break; } break;
default: default:
GGML_ABORT("fatal error"); GGML_ABORT("fatal error");
@ -262,7 +266,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
if (V->ne[0] != 512) { if (V->ne[0] != 512) {
return BEST_FATTN_KERNEL_NONE; return BEST_FATTN_KERNEL_NONE;
} }
if (!gqa_opt_applies || gqa_ratio % 16 != 0) { if (!gqa_opt_applies || gqa_ratio % 4 != 0) {
return BEST_FATTN_KERNEL_NONE; return BEST_FATTN_KERNEL_NONE;
} }
break; break;

1
ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_4.cu
View File

@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 16, 4);
DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4); DECL_FATTN_MMA_F16_CASE(112, 112, 16, 4);
DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4); DECL_FATTN_MMA_F16_CASE(128, 128, 16, 4);
DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4); DECL_FATTN_MMA_F16_CASE(256, 256, 16, 4);
DECL_FATTN_MMA_F16_CASE(576, 512, 16, 4);

1
ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_2-ncols2_4.cu
View File

@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 2, 4);
DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4); DECL_FATTN_MMA_F16_CASE(112, 112, 2, 4);
DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4); DECL_FATTN_MMA_F16_CASE(128, 128, 2, 4);
DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4); DECL_FATTN_MMA_F16_CASE(256, 256, 2, 4);
DECL_FATTN_MMA_F16_CASE(576, 512, 2, 4);

1
ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_4.cu
View File

@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 4, 4);
DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4); DECL_FATTN_MMA_F16_CASE(112, 112, 4, 4);
DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4); DECL_FATTN_MMA_F16_CASE(128, 128, 4, 4);
DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4); DECL_FATTN_MMA_F16_CASE(256, 256, 4, 4);
DECL_FATTN_MMA_F16_CASE(576, 512, 4, 4);

1
ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_4.cu
View File

@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 8, 4);
DECL_FATTN_MMA_F16_CASE(112, 112, 8, 4); DECL_FATTN_MMA_F16_CASE(112, 112, 8, 4);
DECL_FATTN_MMA_F16_CASE(128, 128, 8, 4); DECL_FATTN_MMA_F16_CASE(128, 128, 8, 4);
DECL_FATTN_MMA_F16_CASE(256, 256, 8, 4); DECL_FATTN_MMA_F16_CASE(256, 256, 8, 4);
DECL_FATTN_MMA_F16_CASE(576, 512, 8, 4);

2
ggml/src/ggml-cuda/template-instances/generate_cu_files.py
View File

@ -85,7 +85,7 @@ for ncols in [8, 16, 32, 64]:
continue continue
if head_size_kq != 576 and ncols2 == 16: if head_size_kq != 576 and ncols2 == 16:
continue continue
if head_size_kq == 576 and ncols2 != 16: if head_size_kq == 576 and ncols2 not in (4, 16):
continue continue
head_size_v = head_size_kq if head_size_kq != 576 else 512 head_size_v = head_size_kq if head_size_kq != 576 else 512
f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size_kq=head_size_kq, head_size_v=head_size_v)) f.write(SOURCE_FATTN_MMA_CASE.format(ncols1=ncols1, ncols2=ncols2, head_size_kq=head_size_kq, head_size_v=head_size_v))