diff --git a/paddle/fluid/framework/new_executor/interpretercore.cc b/paddle/fluid/framework/new_executor/interpretercore.cc
index f6157367cd4e2e53940ceada2e61b9f5b9fd0489..b26d213ddf7740f6de02d07e3ce7ed50ae64c646 100644
--- a/paddle/fluid/framework/new_executor/interpretercore.cc
+++ b/paddle/fluid/framework/new_executor/interpretercore.cc
@@ -118,6 +118,8 @@ void InterpreterCore::Convert() {
     temp_inst.input_index_ = vec_func_list_[i].input_index;
     temp_inst.output_index_ = vec_func_list_[i].output_index;
     temp_inst.type_ = vec_func_list_[i].type_;
+    temp_inst.no_data_transform_index_ =
+        vec_func_list_[i].no_data_transform_index;
 
     OpInOutInfo info;
 
diff --git a/paddle/fluid/framework/new_executor/interpretercore_util.cc b/paddle/fluid/framework/new_executor/interpretercore_util.cc
index 3438fc3bd4dcd168b33ba60c5cd497cd23a226b5..7bb0429c6228b2be2d14f2923b009367e7b4d5c3 100644
--- a/paddle/fluid/framework/new_executor/interpretercore_util.cc
+++ b/paddle/fluid/framework/new_executor/interpretercore_util.cc
@@ -278,6 +278,7 @@ void build_op_func_list(const platform::Place& place,
 
     // step 3. Insert memcpy_op if needed
     VariableValueMap& ins_map_temp = runtime_context.inputs;
+    std::unordered_set<int> no_data_transform_index;
     for (auto& var_name_item : ins_map_temp) {
       for (size_t i = 0; i < var_name_item.second.size(); ++i) {
         auto var = var_name_item.second[i];
@@ -289,8 +290,14 @@ void build_op_func_list(const platform::Place& place,
             static_cast<const framework::OperatorWithKernel*>(op_base)
                 ->GetKernelTypeForVar(var_name_item.first, *tensor_in,
                                       expected_kernel_key);
-        if (!platform::is_same_place(kernel_type_for_var.place_,
-                                     expected_kernel_key.place_)) {
+        if (platform::is_same_place(kernel_type_for_var.place_,
+                                    expected_kernel_key.place_)) {
+          // record no need data transformer input var_id
+          auto& var_name = inputs_names[var_name_item.first][i];
+          VLOG(3) << op->Type() << " found no data_transform var: " << var_name
+                  << " with id: " << var_scope->name2id[var_name];
+          no_data_transform_index.emplace(var_scope->name2id[var_name]);
+        } else {
           if (op_base->Type() == "fetch_v2") {
             op_base->SetAttr("deepcopy", false);
           }
@@ -385,6 +392,7 @@ void build_op_func_list(const platform::Place& place,
         }
       }
     }
+    op_func_node.no_data_transform_index = std::move(no_data_transform_index);
     // step 4. Run op kernel
     op_list->push_back(op_base);
     VLOG(3) << op_base->Type()
diff --git a/paddle/fluid/framework/new_executor/new_executor_defs.h b/paddle/fluid/framework/new_executor/new_executor_defs.h
index 19b7b6d5dc299fc0117b12f4a7fa56bc371dadb8..e6cff353a659d77f1cf0b12692e4e4cdccc9fda2 100644
--- a/paddle/fluid/framework/new_executor/new_executor_defs.h
+++ b/paddle/fluid/framework/new_executor/new_executor_defs.h
@@ -511,6 +511,8 @@ struct Instruction {
   std::map<std::string, std::vector<int>> input_index_;
   std::map<std::string, std::vector<int>> output_index_;
 
+  std::unordered_set<int> no_data_transform_index_;
+
   std::vector<size_t> gc_check_var_list;
   NextInstruction next_instruction_;
 
@@ -527,6 +529,7 @@ struct OpFuncNode {
   // int unsed;
   std::map<std::string, std::vector<int>> input_index;
   std::map<std::string, std::vector<int>> output_index;
+  std::unordered_set<int> no_data_transform_index;
 
   OpKernelComputeFunc kernel_func_;
   platform::DeviceContext* dev_ctx_;  // not owned
diff --git a/paddle/fluid/framework/new_executor/stream_analyzer.cc b/paddle/fluid/framework/new_executor/stream_analyzer.cc
index a9322d8fc88edb58af24a4ac560c635a94e29137..ffc2da499e1f7b927b5629b9194746d4fe671199 100644
--- a/paddle/fluid/framework/new_executor/stream_analyzer.cc
+++ b/paddle/fluid/framework/new_executor/stream_analyzer.cc
@@ -38,7 +38,8 @@ std::vector<size_t> StreamAnalyzer::ParseEventVarIds(
   std::vector<size_t> new_event_var_ids;
   for (auto& item : next_instr.input_index_) {
     for (auto var_id : item.second) {
-      if (unique_var_ids.count(var_id) > 0) {
+      if (unique_var_ids.count(var_id) > 0 &&
+          next_instr.no_data_transform_index_.count(var_id) == 0) {
         new_event_var_ids.push_back(var_id);
       }
     }