update by comment

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -46,11 +46,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
 #endif
 
 void AllReduceOpHandle::RunImpl() {
-  if (dev_ctxes_.size() > 0UL) {
   platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
-  } else {
-    platform::RecordEvent record_event(Name(), nullptr);
-  }
 
   if (NoDummyInputSize() == 1) {
     return;  // No need to all reduce when GPU count = 1;

paddle/fluid/framework/details/broadcast_op_handle.cc

@@ -22,11 +22,7 @@ namespace framework {
 namespace details {
 
 void BroadcastOpHandle::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;
 

paddle/fluid/framework/details/data_balance_op_handle.cc

@@ -87,12 +87,6 @@ std::vector<std::array<int, 3>> DataBalanceOpHandle::GetBalancePlan(
 }
 
 void DataBalanceOpHandle::RunImpl() {
-  if (dev_ctxes_.size() > 0UL) {
-    platform::RecordEvent record_event(Name(), dev_ctxes_.begin()->second);
-  } else {
-    platform::RecordEvent record_event(Name(), nullptr);
-  }
-
   PADDLE_ENFORCE_GT(places_.size(), 1,
                     "Data balance can only be enabled when the number of "
                     "places to run larger than 1.");

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -431,10 +431,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
                     CreateReduceOp(&result, g_name, cur_device_id);
                     graph->Get<ShardedVarDevice>(kShardedVarDevice)
                         .emplace(g_name, cur_device_id);
-                    if (!is_dist_train) {
-                      // will send gradients directly when distributed training
                     bcast_var_name_set[cur_device_id].emplace(p_name);
-                    }
                     break;
                   case BuildStrategy::ReduceStrategy::kAllReduce:
                     if (IsSparseGradient(g_name)) {
@@ -461,9 +458,8 @@ 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];