cuda : fix data race in soft max (llama/5853)

diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 9382d96aa5190d721525b388eac8aa5a04129d6a..7d027a30a81010ed00f432fb839d3f6f30c8631e 100644 (file)
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -6947,6 +6947,7 @@ 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;
         }