EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 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

2
ggml/src/ggml-cuda/fattn-wmma-f16.cu
View File

@@ -490,7 +490,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggm
fattn_kernel = flash_attn_ext_f16<
D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), KQ_acc_t, use_logit_softcap>;
}
launch_fattn<D, cols_per_block, 1, -1>(ctx, dst, fattn_kernel, nwarps, 0, FATTN_KQ_STRIDE, true, true, false, warp_size);
launch_fattn<D, cols_per_block, 1>(ctx, dst, fattn_kernel, nwarps, 0, FATTN_KQ_STRIDE, true, true, false, warp_size);
}
void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {