Debug

paddle/fluid/framework/parallel_executor.cc

@@ -345,8 +345,9 @@ struct NCCLAllReduceOpHandle : public OpHandle {
   }
 
   void Wait(platform::DeviceContext *waited_dev) override {
-    VLOG(3) << "Wait NCCL AllReduce";
+    for (auto &pair : member_->communication_streams_) {
-    this->dev_ctx_.at(waited_dev->GetPlace())->Wait();
+      pair.second.ctx_->Wait();
+    }
   }
 };