diff --git a/paddle/operators/fetch_op.cc b/paddle/operators/fetch_op.cc
index c35d7d49e31f6ca11e2b37a455af430aac50a232..f1086e3dc774a5e57f1abb5d4f91f859fc0e64aa 100644
--- a/paddle/operators/fetch_op.cc
+++ b/paddle/operators/fetch_op.cc
@@ -52,6 +52,7 @@ class FetchOp : public framework::OperatorBase {
     // FIXME(yuyang18): Should we assume the fetch operator always generate
     // CPU outputs?
     dst_item.CopyFrom(src_item, platform::CPUPlace(), dev_ctx);
+    dev_ctx.Wait();
     dst_item.set_lod(src_item.lod());
 
     VLOG(3) << "Fetch variable " << fetch_var_name << " to " << out_name;