diff --git a/paddle/fluid/framework/parallel_executor.cc b/paddle/fluid/framework/parallel_executor.cc
index f9fc35d8ce3ecc042739ed54a5595c590fb94de7..21a19cb5b274f9e746a8c9f10cb7b5f86f8ffa7e 100644
--- a/paddle/fluid/framework/parallel_executor.cc
+++ b/paddle/fluid/framework/parallel_executor.cc
@@ -402,10 +402,13 @@ struct NCCLAllReduceOpHandle : public OpHandle {
         platform::dynload::ncclAllReduce(
             buffer, buffer, numel, static_cast<ncclDataType_t>(dtype), ncclSum,
             nccl_ctx.comm, nccl_ctx.stream());
-        PADDLE_ENFORCE(cudaEventRecord(events_[dev_id], nccl_ctx.stream()));
       }
-
       platform::dynload::ncclGroupEnd();
+
+      for (auto &ev : events_) {
+        PADDLE_ENFORCE(cudaEventRecord(
+            ev.second, member_->communication_streams_.at(ev.first).stream()));
+      }
     }
   }