diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
index 14e75e7b7b582d994b83d6c74ad9947135f6c449..3b7d61607301e685e67b5f4bc97fc837471e5722 100644
--- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
+++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
@@ -128,7 +128,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
     //
     // NOTE: DelayedOps have a lower priority. It will be scheduled after all
     // ready_ops have been performed.
-    if (ready_ops.empty() && allow_op_delay_) {
+    if (ready_ops.empty() && allow_op_delay_ && running_ops_ == 0) {
       run_all_ops(delayed_ops);
     } else {
       run_all_ops(ready_ops);