diff --git a/paddle/fluid/framework/details/op_handle_base.cc b/paddle/fluid/framework/details/op_handle_base.cc
index b2fa31f73b9d96ef7fe56dd59ca9b4b18f114c95..cbaaa91d1ea7dd171c21552e571ae919ef6f44eb 100644
--- a/paddle/fluid/framework/details/op_handle_base.cc
+++ b/paddle/fluid/framework/details/op_handle_base.cc
@@ -188,26 +188,16 @@ bool OpHandleBase::NeedWait(VarHandleBase *in_var) {
 }
 
 void OpHandleBase::RunAndRecordEvent(const std::function<void()> &callback) {
+  callback();
 #ifdef PADDLE_WITH_CUDA
   if (!events_.empty()) {  // Use event
-    std::function<void()> method = callback;
     for (auto &p : dev_ctxes_) {
-      method = [method, p, this]() {
-        VLOG(10) << "cudadevicecontext:"
-                 << static_cast<platform::CUDADeviceContext *>(p.second)
-                 << ", dev_id:"
-                 << boost::get<platform::CUDAPlace>(p.first).device;
-
-        static_cast<platform::CUDADeviceContext *>(p.second)->RecordEvent(
-            events_.at(boost::get<platform::CUDAPlace>(p.first).device),
-            method);
-      };
+      auto dev_id = boost::get<platform::CUDAPlace>(p.first).device;
+      auto *cuda_dev_ctx = static_cast<platform::CUDADeviceContext *>(p.second);
+      VLOG(10) << "cudadevicecontext:" << cuda_dev_ctx << ", dev_id:" << dev_id;
+      PADDLE_ENFORCE_CUDA_SUCCESS(
+          cudaEventRecord(events_.at(dev_id), cuda_dev_ctx->stream()));
     }
-    method();
-  } else {
-#endif
-    callback();
-#ifdef PADDLE_WITH_CUDA
   }
 #endif
 }