nicolas
/
ik_llama_opt
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Revert "Fix race in the CUDA DeepSeek FA kernel (#406)" 

This reverts commit 36e6e888b75ae93fb5aac212bb0e147d8379ae23.
I should have tested. We get NaNs.
This commit is contained in:
Kawrakow 2025-05-11 12:22:19 +03:00
parent 0abcf0749e
commit 8f7bd74afb
1 changed files with 0 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-cuda/fattn-new-mma.cu
View File

@ -898,8 +898,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
KQ_crs += __shfl_xor_sync(0xFFFFFFFF, KQ_crs, offset, WARP_SIZE); KQ_crs += __shfl_xor_sync(0xFFFFFFFF, KQ_crs, offset, WARP_SIZE);
} }
__syncthreads();
// Write back combined meta data: // Write back combined meta data:
#pragma unroll #pragma unroll
for (int imeta = 0; imeta < nmeta; ++imeta) { for (int imeta = 0; imeta < nmeta; ++imeta) {