fix halfattention

ztxz16 · Jul 12, 2024 · 1952607 · 1952607
1 parent f98f3b9
commit 1952607
Showing 1 changed file with 2 additions and 1 deletion.
diff --git a/src/devices/cuda/fastllm-cuda.cu b/src/devices/cuda/fastllm-cuda.cu
@@ -830,7 +830,8 @@ __device__ void FastllmSoftmaxKernelInner1Func(half *input, half *output, int ch
     }
     __syncthreads();
     float maxV = sdata[0];
-
+    __syncthreads();
+
     // 4. 求和
     float sum = 0;
     for (int i = tid; i < channels; i += THREAD_PER_BLOCK) {