diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc
index cfd6609a4b14fe852fe38ec6300f790b1ad8fae4..8236773672562e04bc525f6e320f6c1784bee228 100644
--- a/paddle/fluid/framework/parallel_executor.cc
+++ b/paddle/fluid/framework/parallel_executor.cc
@@ -217,6 +217,11 @@ ParallelExecutor::ParallelExecutor(
     }
   }
 
+  if (build_strategy.async_mode_) {
+    PADDLE_ENFORCE(!member_->use_cuda_,
+                   "gpu mode does not support async_mode_ now!");
+  }
+
   // FIXME(Yancey1989): parallel graph mode get better performance
   // in GPU allreduce distributed training. Need an elegant way to
   // choice the execution strategy.