diff --git a/paddle/fluid/framework/data_device_transform.cc b/paddle/fluid/framework/data_device_transform.cc
index 85dbb39e6fba735471446b5e5e71a612282c498a..a876725ac0f17838458065c4b4753a03e2812801 100644
--- a/paddle/fluid/framework/data_device_transform.cc
+++ b/paddle/fluid/framework/data_device_transform.cc
@@ -36,9 +36,11 @@ void TransDataDevice(const Tensor& in, const platform::Place& dst_place,
   VLOG(3) << "DeviceTransform in, src_place " << in.place()
           << " dst_place: " << dst_place;
   auto* dev_ctx = GetDeviceContext(in.place(), dst_place);
-  dev_ctx->Wait();
+
   TensorCopy(in, dst_place, *dev_ctx, out);
-  dev_ctx->Wait();
+  if (platform::is_gpu_place(in.place()) && platform::is_cpu_place(dst_place)) {
+    dev_ctx->Wait();
+  }
 }
 
 }  // namespace framework