CUDA: fix race conditions FlashAttention kernels (#13438)

2025-05-10 22:22:48 +02:00
parent d2a4ef05c6
commit 0208355f42
2 changed files with 3 additions and 0 deletions
--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
@@ -874,6 +874,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
            }
        }

+        __syncthreads();
+
        // Write back combined meta data:
 #pragma unroll
        for (int imeta = 0; imeta < nmeta; ++imeta) {