cuda : fix data race in soft max (#5853)

ggerganov · Mar 3, 2024 · 67be2ce · 67be2ce
1 parent 231ae28
commit 67be2ce
Showing 1 changed file with 1 addition and 0 deletions.
diff --git a/ggml-cuda.cu b/ggml-cuda.cu
@@ -6904,6 +6904,7 @@ static __global__ void soft_max_f32(const float * x, const float * mask, const f
     // find the sum of exps in the block
     tmp = warp_reduce_sum(tmp);
     if (block_size > WARP_SIZE) {
+        __syncthreads();
         if (warp_id == 0) {
             buf_iw[lane_id] = 0.0f;
         }