Fixed performance issue regarding BackwardRun using add_final_state_dygraph (#41912) (#41991)

Co-authored-by: NZhanlue Yang <jim19930609@gmail.com>

paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py

@@ -1471,9 +1471,6 @@ class DygraphNodeGenerator(DygraphFunctionGeneratorBase):
 
         grad_node_name = GetGradNodeName(forward_api_name)
 
-        if len(grad_node_creation_str) == 0:
-            grad_node_creation_str = f"if(create_graph) VLOG(3) << \"Higher order grad node for {grad_node_name} has not been implemented yet.\";"
-
         self.node_definition_str = GRAD_FUNCTION_TEMPLATE.format(
             grad_node_name, fill_zero_str, get_grad_in_args_str, grad_node_name,
             grad_function_call_str, get_outputs_str, inputs_autograd_meta_str,

paddle/fluid/eager/backward.cc

@@ -766,7 +766,8 @@ std::vector<paddle::experimental::Tensor> RunBackward(
                 << ", rank: " << edge_rank.second;
 
         node_input_buffers_dict[next_node]->add(
-            edge_rank.first, edge_rank.second, grad_output_tensor);
+            edge_rank.first, edge_rank.second, grad_output_tensor,
+            create_graph);
 
         // Update queue
         node_in_degree_map[next_node]--;

paddle/fluid/eager/grad_tensor_holder.cc

@@ -72,7 +72,8 @@ void GradTensorHolder::CopyValueFromTensor(
 }
 
 void GradTensorHolder::add(size_t slot_id, size_t rank,
-                           const paddle::experimental::Tensor& t) {
+                           const paddle::experimental::Tensor& t,
+                           bool create_graph) {
   // TODO(jiabin): We need to deal with empty input_buffer with slot size not
   // empty;
   PADDLE_ENFORCE(slot_id < buffer_.size(),
@@ -113,8 +114,12 @@ void GradTensorHolder::add(size_t slot_id, size_t rank,
 
     if (t.is_dense_tensor()) {
       if (buffer_tensor.is_dense_tensor()) {
-        buffer_tensor = add_final_state_dygraph_function(t, buffer_tensor);
-
+        if (create_graph) {
+          buffer_tensor = add_final_state_dygraph_function(t, buffer_tensor);
+        } else {
+          paddle::imperative::TensorAdd<paddle::experimental::Tensor>(
+              t, &buffer_tensor);
+        }
       } else {
         // TODO(jiabin): Support Other TensorBase later
         // TODO(zhanlve): Replace SelectedRowsAddTensor with

paddle/fluid/eager/grad_tensor_holder.h

@@ -45,7 +45,8 @@ class GradTensorHolder {
   GradTensorHolder& operator=(const GradTensorHolder& other) = default;
 
   // Create new tensor and copy tensor->impl
-  void add(size_t slot_id, size_t rank, const paddle::experimental::Tensor& t);
+  void add(size_t slot_id, size_t rank, const paddle::experimental::Tensor& t,
+           bool create_graph = false);
   void CopyValueFromTensor(size_t slot_id, size_t rank,
                            const paddle::experimental::Tensor& t,
                            bool fill_one = false);