CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-03 09:22:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: FA support for Deepseek (Ampere or newer) (#13306)

Browse Source 
* CUDA: FA support for Deepseek (Ampere or newer)

* do loop unrolling via C++ template
This commit is contained in:
Johannes Gäßler
2025-05-09 13:34:58 +02:00
committed by GitHub
parent 27ebfcacba
commit 0cf6725e9f
33 changed files with 852 additions and 547 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -307,7 +307,7 @@ 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, use_logit_softcap>;
launch_fattn<D, cols_per_block, 1, -1>
launch_fattn<D, cols_per_block, 1>
(ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false);
} break;
case 128: {
@@ -315,7 +315,7 @@ 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, use_logit_softcap>;
launch_fattn<D, cols_per_block, 1, -1>
launch_fattn<D, cols_per_block, 1>
(ctx, dst, fattn_kernel, nwarps, nbytes_shared, FATTN_KQ_STRIDE_TILE_F16, true, true, false);
} break;
default: {