update by comment

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -458,8 +458,9 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
   use_gpu = nccl_ctxs_ != nullptr;
 #endif
 
-  if (use_gpu && strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce &&
-      !is_dist_train) {
+  if ((use_gpu &&
+       strategy_.reduce_ == BuildStrategy::ReduceStrategy::kReduce) ||
+      is_dist_train) {
     // Insert BCast Ops
     for (size_t dev_id = 0; dev_id < bcast_var_name_set.size(); ++dev_id) {
       auto &to_bcast_set = bcast_var_name_set[dev_id];

paddle/fluid/framework/details/reduce_op_handle.cc

@@ -27,11 +27,8 @@ namespace framework {
 namespace details {
 
 void ReduceOpHandle::RunImpl() {
-  if (dev_ctxes_.size() > 0UL) {
   platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
-  } else {
-    platform::RecordEvent record_event(Name(), nullptr);
-  }
+
   if (places_.size() == 1) return;
   // the input and output may have dummy var.
   auto in_var_handles = DynamicCast<VarHandle>(inputs_);