Skip to content

HIP: fattn-mma: use fp32 accumulation on MFMA devices #28576

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
IMbackK merged 2 commits into from
Sep 14, 2026
Merged

HIP: fattn-mma: use fp32 accumulation on MFMA devices #28576

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
65a133b
ggml: hip: use fp32 accumulators in fattn-mma on CDNA
IMbackK Sep 7, 2026
ca29eaf
Retune DKQ=64
IMbackK Sep 7, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 5 additions & 3 deletions ggml/src/ggml-cuda/fattn-mma-f16.cuh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -181,7 +181,7 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64, 64, 8, 128, 1, 64, 32, 32, 32, 1, true);
GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64, 64, 16, 256, 2, 64, 32, 32, 32, 1, true);
GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64, 64, 32, 256, 2, 64, 32, 32, 32, 1, true);
GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64, 64, 64, 256, 4, 64, 32, 32, 32, 1, true);
GGML_CUDA_FATTN_MMA_CONFIG_CASE( 64, 64, 64, 256, 3, 64, 32, 32, 32, 1, true);

GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80, 80, 8, 256, 2, 64, 40, 40, 40, 1, true);
GGML_CUDA_FATTN_MMA_CONFIG_CASE( 80, 80, 16, 256, 2, 64, 40, 40, 40, 1, true);
Expand Down Expand Up @@ -1141,7 +1141,7 @@ template<int DV, int ncols> struct mma_tile_sizes {
using T_C_KQ = tile<16, 16, float>; // column-major
using T_A_VKQ = tile<16, 8, half2>; // row-major
using T_B_VKQ = tile<16, 8, half2>; // column-major
using T_C_VKQ = tile<16, 8, half2>; // column-major
using T_C_VKQ = tile<16, 16, float>; // column-major
};
#else // Volta
template<int DV, int ncols> struct mma_tile_sizes {
Expand Down Expand Up @@ -1227,7 +1227,9 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
T_C_VKQ VKQ_C[cols_per_warp == 8 ? DV/T_C_VKQ::I : DV/(2*T_C_VKQ::J)];
#elif defined(AMD_WMMA_AVAILABLE) && defined(RDNA3)
T_C_VKQ VKQ_C[DV % 32 != 0 ? DV/T_C_VKQ::J : DV/(2*T_C_VKQ::J)];
#elif defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
#elif defined(AMD_MFMA_AVAILABLE)
T_C_VKQ VKQ_C[ DV/T_C_VKQ::J];
#elif defined(AMD_WMMA_AVAILABLE)
T_C_VKQ VKQ_C[ DV/(2*T_C_VKQ::J)];
#else // Volta
T_C_VKQ VKQ_C[ DV/(2*T_C_VKQ::J)];
Expand Down
Loading