diff --git a/paddle/fluid/eager/auto_code_generator/eager_generator.cc b/paddle/fluid/eager/auto_code_generator/eager_generator.cc
index db1dbaa2aa7e1e1ad772c7bdd606b614f8a62153..da13b08b59f80aeb64002d0d2d38cda93e4c366c 100644
--- a/paddle/fluid/eager/auto_code_generator/eager_generator.cc
+++ b/paddle/fluid/eager/auto_code_generator/eager_generator.cc
@@ -1107,10 +1107,12 @@ static std::string GenerateGradNodeCreationContent(
   size_t bwd_in_slot_num = out_vars.size();
   size_t bwd_out_slot_num = in_vars.size();
   const char* GRAD_OP_NODE_TEMPLATE =
-      "      auto grad_node = std::make_shared<GradNode%s>(%d, %d);\n";
+      "      auto grad_node = std::shared_ptr<GradNode%s>(new GradNode%s(%d, "
+      "%d));\n";
   grad_node_creation_str += "    // Create GradOpNode\n";
-  grad_node_creation_str += paddle::string::Sprintf(
-      GRAD_OP_NODE_TEMPLATE, op_type, bwd_in_slot_num, bwd_out_slot_num);
+  grad_node_creation_str +=
+      paddle::string::Sprintf(GRAD_OP_NODE_TEMPLATE, op_type, op_type,
+                              bwd_in_slot_num, bwd_out_slot_num);
   grad_node_creation_str += "\n";
 
   VLOG(6) << "Generated GradOpNode construction";
diff --git a/paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py b/paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py
index 7339f3581ac1fd5ec50a8fd161a2723d49ec78a1..8d061c8929ae6b040b5a723097d2eb5d906b8d3f 100644
--- a/paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py
+++ b/paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py
@@ -650,8 +650,12 @@ class DygraphFunctionGeneratorBase(FunctionGeneratorBase):
         num_backward_inputs = len(forward_outputs_position_map.keys())
         num_backward_outputs = len(forward_inputs_position_map.keys())
         grad_node_name = GetGradNodeName(forward_api_name)
-
-        node_construction_str = f"            auto grad_node = std::make_shared<{grad_node_name}>({num_backward_inputs}, {num_backward_outputs});"
+        # NOTE(Aurelius74): DO NOT use make_shared here. Because some Node contains experimental::Scalar
+        # which contains "complex128" as data. "complex128" is memory-aligned manually. But make_shared
+        # request MEMALIGN for allocation (Maybe).
+        # See https://stackoverflow.com/questions/31228656/how-can-shared-ptr-disrupt-alignment
+        # and https://github.com/MRtrix3/mrtrix3/issues/957
+        node_construction_str = f"            auto grad_node = std::shared_ptr<{grad_node_name}>(new {grad_node_name}({num_backward_inputs}, {num_backward_outputs}));"
 
         # SetAttributes
         set_attributes_list = []
diff --git a/python/paddle/utils/code_gen/api.yaml b/python/paddle/utils/code_gen/api.yaml
index 09bfe746271092e5bd5777671893b5930ff76bdb..9a6059c53a7e57932aa9304d1018b5d98ae8509b 100644
--- a/python/paddle/utils/code_gen/api.yaml
+++ b/python/paddle/utils/code_gen/api.yaml
@@ -1373,7 +1373,15 @@
   backward : tile_grad
   # no_need_buffer : x
 
-# trace
+- api : top_k
+  args : (Tensor x, Scalar k, int axis = -1, bool largest = true, bool sorted = true)
+  output : Tensor(out), Tensor(indices)
+  infer_meta :
+    func : TopKInferMeta
+  kernel :
+    func : top_k
+  backward : top_k_grad
+
 - api : trace
   args : (Tensor x, int offset, int axis1, int axis2)
   output : Tensor
diff --git a/python/paddle/utils/code_gen/backward.yaml b/python/paddle/utils/code_gen/backward.yaml
index 43f512540ec4b24629b02dd0130831fb5387b97e..4f2b76db1a27f3428ad8a2d9f3172d02fae20657 100644
--- a/python/paddle/utils/code_gen/backward.yaml
+++ b/python/paddle/utils/code_gen/backward.yaml
@@ -866,6 +866,16 @@
   kernel :
     func : tile_grad
 
+- backward_api : top_k_grad
+  forward : top_k (Tensor x, Scalar k, int axis = -1, bool largest = true, bool sorted = true) -> Tensor(out), Tensor(indices)
+  args : (Tensor x, Tensor indices, Tensor out_grad, Scalar k = -1, int axis = -1, bool largest = true, bool sorted = true)
+  output : Tensor(x_grad)
+  infer_meta :
+    func : UnchangedInferMeta
+    param : [x]
+  kernel :
+    func : top_k_grad
+
 - backward_api : trace_grad
   forward : trace (Tensor x, int offset, int axis1, int axis2) -> Tensor(out)
   args : (Tensor x, Tensor out_grad, int offset, int axis1, int axis2)