diff --git a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
index d9b855503b4b3b3df40b0192e569b40731a94953..501e1dfad7644de274ff540458bb9417760dedaf 100644
--- a/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
+++ b/paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
@@ -208,6 +208,11 @@ void ThreadedSSAGraphExecutor::RunOp(
     try {
       VLOG(10) << op->DebugString();
       op->Run(use_event_);
+
+      for (auto &dev_ctx : op->dev_ctx_) {
+        dev_ctx.second->Wait();  // Sync error
+      }
+
       for (auto *ready : *ready_buffer) {
         ready->store(true, std::memory_order_release);
       }