follow comment test=develop

paddle/fluid/framework/details/build_strategy.cc

@@ -133,15 +133,15 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
   void AppendMultiDevPass(const BuildStrategy &strategy) {
     ir::Pass *multi_devices_pass;
     if (strategy_.is_distribution_) {
-      VLOG(3) << "dist train mode";
+      VLOG(3) << "multi device dist train mode";
       multi_devices_pass = AppendPass("dist_multi_devices_pass").get();
     } else {
       if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
-        VLOG(3) << "allreduce mode";
+        VLOG(3) << "multi device allreduce mode";
         multi_devices_pass =
             AppendPass("allreduce_mode_multi_devices_pass").get();
       } else if (strategy.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
-        VLOG(3) << "reduce mode";
+        VLOG(3) << "multi device reduce mode";
         multi_devices_pass = AppendPass("reduce_mode_multi_devices_pass").get();
       } else {
         PADDLE_THROW("Unknown reduce strategy.");

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -731,7 +731,6 @@ bool DistSSAGraphBuilder::DealWithSpecialOp(ir::Graph *result,
       }
     }
     insert_op = true;
-    need_broadcast_var_ = true;
   } else if (OpHaveRole(*node, OpRole::kDist)) {
     int op_dev_id = CreateDistTrainOp(result, node);
     if (node->Op()->Type() == "concat") {
@@ -925,6 +924,7 @@ void DistSSAGraphBuilder::InsertCollectiveOp(ir::Graph *result,
 }
 
 void DistSSAGraphBuilder::InsertPostprocessOps(ir::Graph *result) const {
+  // only GPU reduce mode need to broadcast parameters to each device.
   if (UseGPU() && strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) {
     if (strategy_.fuse_broadcast_op_) {
       CreateFusedBroadcastOp(result, bcast_var_name_set_);