PaddlePaddle · luotao1 · Mar 1, 2024 · Feb 11, 2024 · Feb 19, 2024 · Feb 20, 2024
diff --git a/paddle/fluid/distributed/fleet_executor/compute_interceptor.cc b/paddle/fluid/distributed/fleet_executor/compute_interceptor.cc
@@ -176,7 +176,7 @@ bool ComputeInterceptor::IsInputReady() {
       flag = flag && (ready_size_map.at(i) != 0);
     }
     if (flag) {
-      if (scope_id_to_finish_flag.empty()) {
+      if (scope_id_to_finish_flag.empty()) {  // NOLINT
         cur_scope_id_ = i;
         return true;
       } else if (scope_id_to_finish_flag.find(i) !=
@@ -303,7 +303,7 @@ void ComputeInterceptor::RunOps() {
                           cur_scope_id_));
   }
 
-  if (!cores_.empty()) {
+  if (!cores_.empty()) {  // NOLINT
     cores_[cur_scope_id_]->Run(/*feed_names=*/{}, /*need_fetch=*/false);
   } else {
     for (auto op : node_->ops()) {

diff --git a/paddle/fluid/distributed/fleet_executor/dist_model.cc b/paddle/fluid/distributed/fleet_executor/dist_model.cc
@@ -215,7 +215,7 @@ bool DistModel::Init() {
 }
 
 bool DistModel::PreparePlace() {
-  if (config_.place == "GPU") {
+  if (config_.place == "GPU") {  // NOLINT
     place_ = paddle::platform::CUDAPlace(config_.device_id);
   } else if (config_.place == "CPU") {
     place_ = paddle::platform::CPUPlace();

diff --git a/paddle/fluid/eager/custom_operator/custom_operator_utils.cc b/paddle/fluid/eager/custom_operator/custom_operator_utils.cc
@@ -558,7 +558,7 @@ std::vector<std::vector<phi::DDim>> RunInferShapeFn(
     out_dims =
         RunInferShapeFunc(ctx, infer_shape_func, inputs, outputs, inplace_map);
   } else {
-    if (is_forward) {
+    if (is_forward) {  // NOLINT
       out_dims = RunDefaultInferShapeFunc(ctx, inputs, outputs, inplace_map);
     } else {
       out_dims =
@@ -592,7 +592,7 @@ std::vector<std::vector<phi::DataType>> RunInferDtypeFn(
     out_dtypes =
         RunInferDtypeFunc(ctx, infer_dtype_func, inputs, outputs, inplace_map);
   } else {
-    if (is_forward) {
+    if (is_forward) {  // NOLINT
       out_dtypes = RunDefaultInferDtypeFunc(ctx, inputs, outputs, inplace_map);
     } else {
       out_dtypes =

@@ -79,7 +79,7 @@ void GradTensorHolder::CopyValueFromTensor(size_t slot_id,
     // Create new tensor->impl and fill it with 1.0
     if (t.defined()) {
       // Fill 1.0, use full to support complex, one_like don't support it.
-      if (t.is_dense_tensor()) {
+      if (t.is_dense_tensor()) {  // NOLINT
         buffer_[slot_id][rank] =
             paddle::experimental::full(t.shape(), 1, t.dtype(), t.place());
       } else if (t.is_sparse_csr_tensor() || t.is_sparse_coo_tensor()) {

diff --git a/paddle/fluid/framework/data_feed.cc b/paddle/fluid/framework/data_feed.cc
@@ -1813,7 +1813,7 @@ int PaddleBoxDataFeed::Next() {
     this->batch_size_ = index;
     VLOG(3) << "pv_batch_size_=" << this->batch_size_
             << ", thread_id=" << thread_id_;
-    if (this->batch_size_ != 0) {
+    if (this->batch_size_ != 0) {  // NOLINT
       PutToFeedVec(pv_vec);
     } else {
       VLOG(3) << "finish reading, output_pv_channel_ size="
@@ -2113,7 +2113,7 @@ void SlotRecordInMemoryDataFeed::Init(const DataFeedDesc& data_feed_desc) {
   finish_init_ = true;
   input_type_ = data_feed_desc.input_type();
   size_t pos = pipe_command_.find(".so");
-  if (pos != std::string::npos) {
+  if (pos != std::string::npos) {  // NOLINT
     pos = pipe_command_.rfind('|');
     if (pos == std::string::npos) {
       so_parser_name_ = pipe_command_;
@@ -2129,7 +2129,7 @@ void SlotRecordInMemoryDataFeed::Init(const DataFeedDesc& data_feed_desc) {
 #if defined(PADDLE_WITH_GPU_GRAPH) && defined(PADDLE_WITH_HETERPS)
   gpu_graph_data_generator_.SetConfig(data_feed_desc);
 #endif
-  if (gpu_graph_mode_) {
+  if (gpu_graph_mode_) {  // NOLINT
     train_mode_ = true;
   } else {
     train_mode_ = data_feed_desc.graph_config().gpu_graph_training();
@@ -2780,7 +2780,7 @@ int SlotRecordInMemoryDataFeed::Next() {
     this->batch_size_ = batch.second;
     VLOG(3) << "batch_size_=" << this->batch_size_
             << ", thread_id=" << thread_id_;
-    if (this->batch_size_ != 0) {
+    if (this->batch_size_ != 0) {  // NOLINT
       PutToFeedVec(&records_[batch.first], this->batch_size_);
     } else {
       VLOG(3) << "finish reading for heterps, batch size zero, thread_id="

diff --git a/paddle/fluid/framework/data_set.cc b/paddle/fluid/framework/data_set.cc
@@ -966,7 +966,7 @@ void DatasetImpl<T>::DynamicAdjustChannelNum(int channel_num,
     CHECK(output_channels_data_size == 0);  // NOLINT
     cur_channel = 1;
   }
-  if (cur_channel == 0) {
+  if (cur_channel == 0) {  // NOLINT
     origin_channels = &multi_output_channel_;
     other_channels = &multi_consume_channel_;
     origin_pv_channels = &multi_pv_output_;
@@ -1111,8 +1111,8 @@ void DatasetImpl<T>::CreateReaders() {
     if (input_pv_channel_ != nullptr) {
       readers_[i]->SetInputPvChannel(input_pv_channel_.get());
     }
-    if (cur_channel_ == 0 &&
-        static_cast<size_t>(channel_idx) < multi_output_channel_.size()) {
+    if (cur_channel_ == 0 && static_cast<size_t>(channel_idx) <
+                                 multi_output_channel_.size()) {  // NOLINT
       readers_[i]->SetOutputChannel(multi_output_channel_[channel_idx].get());
       readers_[i]->SetConsumeChannel(multi_consume_channel_[channel_idx].get());
       readers_[i]->SetOutputPvChannel(multi_pv_output_[channel_idx].get());
@@ -1722,7 +1722,7 @@ void MultiSlotDataset::PreprocessChannel(
     const std::set<std::string>& slots_to_replace,
     std::unordered_set<uint16_t>& index_slots) {  // NOLINT
   int out_channel_size = 0;
-  if (cur_channel_ == 0) {
+  if (cur_channel_ == 0) {  // NOLINT
     for (auto& item : multi_output_channel_) {
       out_channel_size += static_cast<int>(item->Size());
     }
@@ -1757,7 +1757,7 @@ void MultiSlotDataset::PreprocessChannel(
       input_channel_->ReadAll(slots_shuffle_original_data_);
     } else {
       CHECK(out_channel_size > 0);  // NOLINT
-      if (cur_channel_ == 0) {
+      if (cur_channel_ == 0) {      // NOLINT
         for (auto& item : multi_output_channel_) {
           std::vector<Record> vec_data;
           item->Close();
@@ -1792,7 +1792,7 @@ void MultiSlotDataset::PreprocessChannel(
   } else {
     // if already have original data for slots shuffle, clear channel
     input_channel_->Clear();
-    if (cur_channel_ == 0) {
+    if (cur_channel_ == 0) {  // NOLINT
       for (auto& item : multi_output_channel_) {
         if (!item) {
           continue;
@@ -1809,7 +1809,7 @@ void MultiSlotDataset::PreprocessChannel(
     }
   }
   int end_size = 0;
-  if (cur_channel_ == 0) {
+  if (cur_channel_ == 0) {  // NOLINT
     for (auto& item : multi_output_channel_) {
       if (!item) {
         continue;

diff --git a/paddle/fluid/framework/details/nan_inf_utils_detail.cc b/paddle/fluid/framework/details/nan_inf_utils_detail.cc
@@ -264,7 +264,7 @@ void CheckOpHasNanOrInf(const framework::OperatorBase& op,
 
   if (IsSkipOp(op)) return;
 
-  if (op_var_nan_inf_white_list().count(op.Type()) == 0) {
+  if (op_var_nan_inf_white_list().count(op.Type()) == 0) {  // NOLINT
     // NOTE. vname may destruct in the end of this func.
     for (auto& vname : op.OutputVars(true)) {
       auto* var = exec_scope.FindVar(vname);

diff --git a/paddle/fluid/framework/dist_multi_trainer.cc b/paddle/fluid/framework/dist_multi_trainer.cc
@@ -157,7 +157,7 @@ void DistMultiTrainer::Run() {
   std::vector<std::future<void>> wait_futures;
   CHECK_EQ(static_cast<int>(pool.size()), thread_num_);
   for (int i = 0; i < thread_num_; ++i) {
-    if (!debug_) {
+    if (!debug_) {  // NOLINT
       wait_futures.emplace_back(
           pool[i]->Run([this, i]() { workers_[i]->TrainFiles(); }));
     } else {

diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc
@@ -99,7 +99,7 @@ void Executor::CreateVariables(const ProgramDesc& pdesc,
   while (ancestor_scope->parent()) {
     ancestor_scope = ancestor_scope->parent();
   }
-  if (ancestor_scope != scope) {
+  if (ancestor_scope != scope) {  // NOLINT
     for (auto& var : global_block.AllVars()) {
       if (var->Name() == framework::kEmptyVarName) {
         continue;

diff --git a/paddle/fluid/framework/heter_section_worker.cc b/paddle/fluid/framework/heter_section_worker.cc
@@ -126,7 +126,7 @@ void HeterSectionWorker::Initialize(const TrainerDesc& desc) {
   bool is_first_stage = (pipeline_stage_ == 0);
   bool is_last_stage = (pipeline_stage_ + 1 == num_pipeline_stages_);
 
-  if (is_first_stage) {
+  if (is_first_stage) {  // NOLINT
     for (auto& op_desc : program_->Block(0).AllOps()) {
       auto op = std::move(OpRegistry::CreateOp(*op_desc));
       auto op_type = op->Type();

diff --git a/paddle/fluid/framework/infershape_utils.cc b/paddle/fluid/framework/infershape_utils.cc
@@ -658,7 +658,7 @@ CompatInferMetaContext BuildInferMetaContext(InferShapeContext* ctx,
         if (attr_ptr && !is_attr_var) {
           auto& attr = *attr_ptr;
           switch (AttrTypeID(attr)) {
-            case framework::proto::AttrType::INTS:
+            case framework::proto::AttrType::INTS:  // NOLINT
               infer_meta_context.EmplaceBackAttr(std::move(
                   phi::IntArray(PADDLE_GET_CONST(std::vector<int32_t>, attr))));
               break;
@@ -836,7 +836,7 @@ CompatInferMetaContext BuildInferMetaContext(InferShapeContext* ctx,
                       attr_names[i]));
               }
               break;
-            case phi::AttributeType::FLOAT32S:
+            case phi::AttributeType::FLOAT32S:  // NOLINT
               infer_meta_context.EmplaceBackAttr(
                   PADDLE_GET_CONST(std::vector<float>, attr));
               break;

diff --git a/paddle/fluid/framework/ir/coalesce_grad_tensor_pass.cc b/paddle/fluid/framework/ir/coalesce_grad_tensor_pass.cc
@@ -134,7 +134,7 @@ class CoalesceGradTensorPass : public ir::Pass {
 
     auto &pinned_var_set =
         graph->GetOrInit<details::PinnedVars>(details::kPinnedVars);
-    if (IsUnifiedDtype(p_g_dense_grad, vars_info)) {
+    if (IsUnifiedDtype(p_g_dense_grad, vars_info)) {  // NOLINT
       RecordGradients(p_g_dense_grad, vars_info, &pinned_var_set);
       CoalesceTensors(vars_info, p_g_dense_grad, &result);
     } else {

diff --git a/paddle/fluid/framework/ir/generate_pass_tester.cc b/paddle/fluid/framework/ir/generate_pass_tester.cc
@@ -25,7 +25,7 @@ REGISTER_GENERATE_PASS(generate_fc_fuse) {
       VLOG(3) << "exec lambda func.";
       auto mul = OP_(mul)({{"X", x}, {"Y", y}}).Out("Out");
       auto ewadd = OP_(elementwise_add)({{"X", mul}, {"Y", z}}).Out("Out");
-      if (with_relu) {
+      if (with_relu) {  // NOLINT
         return OP_(relu)({"X", ewadd}).Out("Out");
       } else {
         return ewadd;

diff --git a/paddle/fluid/framework/ir/identity_op_clean_pass.cc b/paddle/fluid/framework/ir/identity_op_clean_pass.cc
@@ -70,7 +70,7 @@ FindUselessOpPattern::FindUselessOpPattern(PDPattern* pattern,
               auto in_dtype = x->Op()->GetAttrIfExists<int>("in_dtype");
               auto out_dtype = x->Op()->GetAttrIfExists<int>("out_dtype");
               return in_dtype == out_dtype;
-            } else if (op_type == "c_identity") {
+            } else if (op_type == "c_identity") {  // NOLINT
               return true;
             } else if (op_type == "assign") {
               const auto& in_name = x->Op()->Input("X")[0];

diff --git a/paddle/fluid/framework/ir/mkldnn/compute_propagate_scales_mkldnn_pass_tester.cc b/paddle/fluid/framework/ir/mkldnn/compute_propagate_scales_mkldnn_pass_tester.cc
@@ -161,7 +161,7 @@ class ComputePropagateScalesMkldnnPassTest : public testing::Test {
           begin(wh[i]),
           end(wh[i]),
           wh_tensor->mutable_data<float>(phi::CPUPlace()) + i * wh[0].size());
-    if (type == "gru") {
+    if (type == "gru") {  // NOLINT
       ComputeGruWeightScales(
           graph, &scope, wx_name, wh_name, &var_quant_scales);
     } else {

diff --git a/paddle/fluid/framework/ir/mkldnn/cpu_quantize_pass_tester.cc b/paddle/fluid/framework/ir/mkldnn/cpu_quantize_pass_tester.cc
@@ -61,7 +61,7 @@ void SetOp(ProgramDesc* prog,
     op->SetOutput("Output", {outputs[0]});
   } else if (type == "pool2d" || type == "fused_transpose" ||
              type == "reshape2" || type == "nearest_interp" ||
-             type == "nearest_interp_v2") {
+             type == "nearest_interp_v2" || type == "dropout") {
     op->SetInput("X", {inputs[0]});
     op->SetOutput("Out", {outputs[0]});
   } else if (type == "slice") {
@@ -70,9 +70,6 @@ void SetOp(ProgramDesc* prog,
   } else if (type == "split") {
     op->SetInput("X", {inputs[0]});
     op->SetOutput("Out", {outputs});
-  } else if (type == "dropout") {
-    op->SetInput("X", {inputs[0]});
-    op->SetOutput("Out", {outputs[0]});
   } else if (type == "fc") {
     op->SetInput("Input", {inputs[0]});
     if (inputs.size() > 1) op->SetInput("W", {inputs[1]});

diff --git a/paddle/fluid/framework/ir/mkldnn/cpu_quantize_squash_pass_tester.cc b/paddle/fluid/framework/ir/mkldnn/cpu_quantize_squash_pass_tester.cc
@@ -41,7 +41,7 @@ void SetOp(ProgramDesc* prog,
   if (type != "dropout" && type != "quantize" && type != "dequantize") {
     op->SetAttr("mkldnn_data_type", mkldnn_data_type);
   }
-  if (type == "pool2d") {
+  if (type == "pool2d") {  // NOLINT
     op->SetInput("X", {inputs[0]});
     op->SetOutput("Out", {outputs[0]});
     if (!scale.empty()) op->SetAttr("Scale_in", scale[0]);

diff --git a/paddle/fluid/framework/ir/mkldnn/int8_scale_calculation_mkldnn_pass_tester.cc b/paddle/fluid/framework/ir/mkldnn/int8_scale_calculation_mkldnn_pass_tester.cc
@@ -70,14 +70,7 @@ ProgramDesc BuildProgramDesc(bool convWithExistingBias,
     }
   }
 
-  if (convWithExistingBias) {
-    SetOp(&prog,
-          "conv2d",
-          "conv",
-          std::vector<std::string>({"c", "weights", "conv_bias"}),
-          std::vector<std::string>({"f"}),
-          scale_weights);
-  } else if (scale_weights.size() > 1) {
+  if (convWithExistingBias || scale_weights.size() > 1) {
     SetOp(&prog,
           "conv2d",
           "conv",

diff --git a/paddle/fluid/framework/ir/multi_devices_graph_pass/multi_devices_graph_pass.cc b/paddle/fluid/framework/ir/multi_devices_graph_pass/multi_devices_graph_pass.cc
@@ -933,7 +933,7 @@ bool ReduceSSAGraphBuilder::DealWithSpecialOp(ir::Graph *result,
 
 void ReduceSSAGraphBuilder::InsertPostprocessOps(ir::Graph *result) const {
   if (UseGPU()) {
-    if (strategy_.fuse_broadcast_ops_ == true) {
+    if (strategy_.fuse_broadcast_ops_ == true) {  // NOLINT
       CreateFusedBroadcastOp(result, bcast_var_name_set_);
     } else {
       for (size_t dev_id = 0; dev_id < bcast_var_name_set_.size(); ++dev_id) {
@@ -1193,7 +1193,7 @@ int DistSSAGraphBuilder::CreateRPCOp(ir::Graph *result, ir::Node *node) const {
                                  node->Op()->Type()));
   // Create fetch_barrier op handle to enable output on all devices.
   // **NOTE** fetch_barrier should output variables list same as recv op does.
-  if (node->Op()->Type() == "fetch_barrier") {
+  if (node->Op()->Type() == "fetch_barrier") {  // NOLINT
     result->Get<GraphOps>(kGraphOps).emplace_back(
         new details::FetchBarrierOpHandle(
             result->CreateOpNode(node->Op()), local_scopes_, places_));
@@ -1354,7 +1354,7 @@ void DistSSAGraphBuilder::InsertPostprocessOps(ir::Graph *result) const {
         strategy_.reduce_ == details::BuildStrategy::ReduceStrategy::kReduce) {
       return;
     }
-    if (strategy_.fuse_broadcast_ops_ == true) {
+    if (strategy_.fuse_broadcast_ops_ == true) {  // NOLINT
       CreateFusedBroadcastOp(result, bcast_var_name_set_);
     } else {
       for (size_t dev_id = 0; dev_id < bcast_var_name_set_.size(); ++dev_id) {

diff --git a/paddle/fluid/framework/ir/transfer_layout_elim_pass.cc b/paddle/fluid/framework/ir/transfer_layout_elim_pass.cc
@@ -239,7 +239,7 @@ void TransferLayoutElimPass::ApplyImpl(ir::Graph *graph) const {
   FusePassBase::Init(pattern_name, graph);
 
   auto transfer_format = [&](std::string data_format) -> std::string {
-    if (data_format == "NCHW") {
+    if (data_format == "NCHW") {  // NOLINT
       return "NHWC";
     } else if (data_format == "NHWC") {
       return "NCHW";

diff --git a/paddle/fluid/framework/new_executor/garbage_collector/garbage_collector.cc b/paddle/fluid/framework/new_executor/garbage_collector/garbage_collector.cc
@@ -32,14 +32,14 @@ CreateInterpreterCoreGarbageCollector(
     const platform::Place& place,
     const std::vector<std::unique_ptr<InstructionBase>>& vec_instruction) {
   if (platform::is_gpu_place(place)) {
-    if (IsInterpretercoreFastGCEnabled()) {
+    if (IsInterpretercoreFastGCEnabled()) {  // NOLINT
       return std::unique_ptr<InterpreterCoreGarbageCollector>(
           new InterpreterCoreFastGarbageCollector());
     } else {
       return std::unique_ptr<InterpreterCoreGarbageCollector>(
           new InterpreterCoreEventGarbageCollector(vec_instruction));
     }
-  } else if (platform::is_xpu_place(place)) {
+  } else if (platform::is_xpu_place(place)) {  // NOLINT
     // Because there is no multi-stream on XPU device, fast GC can
     // be used.
     // Previously, XPU used no_event GC. But `Wait` in no_event GC
@@ -62,14 +62,14 @@ CreateInterpreterCoreGarbageCollector(
     const platform::Place& place,
     const std::vector<Instruction>& vec_instruction) {
   if (platform::is_gpu_place(place)) {
-    if (IsInterpretercoreFastGCEnabled()) {
+    if (IsInterpretercoreFastGCEnabled()) {  // NOLINT
       return std::unique_ptr<InterpreterCoreGarbageCollector>(
           new InterpreterCoreFastGarbageCollector());
     } else {
       return std::unique_ptr<InterpreterCoreGarbageCollector>(
           new InterpreterCoreEventGarbageCollector(vec_instruction));
     }
-  } else if (platform::is_xpu_place(place)) {
+  } else if (platform::is_xpu_place(place)) {  // NOLINT
     // Because there is no multi-stream on XPU device, fast GC can
     // be used.
     // Previously, XPU used no_event GC. But `Wait` in no_event GC

diff --git a/paddle/fluid/framework/new_executor/garbage_collector/no_event_garbage_collector.cc b/paddle/fluid/framework/new_executor/garbage_collector/no_event_garbage_collector.cc
@@ -49,9 +49,10 @@ void InterpreterCoreNoEventGarbageCollector::Add(
 
   if (var->IsType<phi::DenseTensor>()) {
     Add(var->GetMutable<phi::DenseTensor>()->MoveMemoryHolder(), ctx);
-  } else if (var->IsType<
-                 operators::reader::
-                     OrderedMultiDeviceLoDTensorBlockingQueueHolder>()) {
+  } else if (
+      var->IsType<
+          operators::reader::
+              OrderedMultiDeviceLoDTensorBlockingQueueHolder>()) {  // NOLINT
     // TODO(xiongkun03) in old executor, this type of variable is not support
     // eager deletion. so we just leave it here ?
   } else if (var->IsType<LoDRankTable>()) {

diff --git a/paddle/fluid/framework/new_executor/new_executor_defs.cc b/paddle/fluid/framework/new_executor/new_executor_defs.cc
@@ -96,7 +96,7 @@ void VariableScope::AddVar(const std::string& name,
     auto id = VarSize();
     name2id_[name] = static_cast<int>(id);
     vec_meta_info_.emplace_back(0, var_desc);
-    if (local_scope_ != nullptr) {
+    if (local_scope_ != nullptr) {  // NOLINT
       var_list_.push_back(local_scope_->FindVar(name));
     } else {
       var_list_.push_back(scope_->FindVar(name));

diff --git a/paddle/fluid/framework/new_executor/pir_interpreter.cc b/paddle/fluid/framework/new_executor/pir_interpreter.cc
@@ -701,7 +701,7 @@ void PirInterpreter::BuildInstruction() {
         continue;
       }
     } else if (op.dialect()->name() == "pd_op") {
-      if (op.isa<paddle::dialect::IfOp>()) {
+      if (op.isa<paddle::dialect::IfOp>()) {  // NOLINT
         vec_instruction_base_.emplace_back(std::make_unique<IfInstruction>(
             op_idx++, place_, &op, value_exe_info_.get(), execution_config_));
         sub_blocks_.insert(
@@ -742,7 +742,7 @@ void PirInterpreter::BuildInstruction() {
       }
       VLOG(6) << "process " << op_name;
 
-      if (op.isa<paddle::dialect::LegacyKernelOp>()) {
+      if (op.isa<paddle::dialect::LegacyKernelOp>()) {  // NOLINT
         CREATE_INSTR(LegacyKernelInstruction);
       } else {
         CREATE_INSTR(PhiKernelInstruction);

diff --git a/paddle/fluid/framework/new_executor/standalone_executor.cc b/paddle/fluid/framework/new_executor/standalone_executor.cc
@@ -57,7 +57,7 @@ StandaloneExecutor::StandaloneExecutor(const platform::Place& place,
     const std::string& job_type = job->Type();
     std::shared_ptr<ProgramDesc> program = nullptr;
     std::shared_ptr<::pir::Program> ir_program = nullptr;
-    if (FLAGS_enable_pir_api || FLAGS_enable_pir_in_executor) {
+    if (FLAGS_enable_pir_api || FLAGS_enable_pir_in_executor) {  // NOLINT
       ir_program = plan_.IrProgram(job_type);
     } else {
       // NOTE (liuchenghao): std::make_shared will duplicate ProgramDesc object,