fix performance

47740ace · Yancey1989 · 73edf137 · 47740ace
显示空白变更内容
内联并排

Showing with 1 addition and 0 deletion

paddle/fluid/framework/details/all_reduce_op_handle.cc paddle/fluid/framework/details/all_reduce_op_handle.cc +1 -0

未找到文件。
--- a/paddle/fluid/framework/details/all_reduce_op_handle.cc
+++ b/paddle/fluid/framework/details/all_reduce_op_handle.cc
@@ -107,6 +107,7 @@ void AllReduceOpHandle::RunImpl() {
          PADDLE_ENFORCE(platform::dynload::ncclAllReduce(
              buffer, buffer, numel, static_cast<ncclDataType_t>(dtype),
              ncclSum, comm, stream));
+          if (!nccl_ctxs_->need_group_call_) cudaStreamSynchronize(stream);
        });
      }