CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-15 11:17:31 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: optimize FA for GQA + large batches (#12014)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-02-22 12:20:17 +01:00
committed by GitHub
parent 335eb04a91
commit 5fa07c2f93
32 changed files with 940 additions and 411 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/ggml-cuda/fattn-tile-f16.cu
View File

@@ -302,14 +302,14 @@ void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
constexpr int nwarps = 8;
constexpr size_t nbytes_shared = 0;
fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
launch_fattn<D, cols_per_block, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
} break;
case 128: {
constexpr int D = 128;
constexpr int nwarps = 8;
constexpr size_t nbytes_shared = 0;
fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
launch_fattn<D, cols_per_block, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
launch_fattn<D, cols_per_block, 1, parallel_blocks, -1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, true, true);
} break;
default: {
GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");