From: Johannes Gäßler <redacted>
Date: Sat, 10 May 2025 20:22:48 +0000 (+0200)
Subject: CUDA: fix race conditions FlashAttention kernels (#13438)
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=0208355f42bdab88a08507ead4a6302790a08323;p=pkg%2Fggml%2Fsources%2Fllama.cpp

CUDA: fix race conditions FlashAttention kernels (#13438)
---

diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
index b2f95fa3..9873ea75 100644
--- 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) {
diff --git a/ggml/src/ggml-cuda/fattn-vec-f16.cuh b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
index ef0addc1..d96e3921 100644
--- a/ggml/src/ggml-cuda/fattn-vec-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-vec-f16.cuh
@@ -168,6 +168,7 @@ static __global__ void flash_attn_vec_ext_f16(
     for (int j = 0; j < ncols; ++j) {
         KQ[j*D + tid] = -HALF_MAX_HALF;
     }
+    __syncthreads();
 
     half2 VKQ[ncols] = {{0.0f, 0.0f}};