Adjusted python-level trace_op to accomodate final state Eager Dygraph (#39319)

* Removed debug info

* Added automatic code generation for final state Eager Dygraph

* Modified backward yaml

* Added EagerUtils helper functions for final state CodeGen

* Adjusted CMakeFiles to support compilation for final state auto generated codes

* Added python-c code generation for final state Eager Dygraph

* Fixed minor issue

* Fixed yaml.load() method failure

* Fixed minor issues

* Refactored Python-C Attributes Parsing Functions

* Fixed minor issue with Python-C AddFunctions

* Adjusted python-level trace_op to accomodate final state Eager Dygraph

* Added Logs for final state Eager Dygraph

* Fixed merge issues

* Fixed minor issue

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

@@ -17,6 +17,12 @@ import re
 import argparse
 import os
 
+# For API dispatch used at python-level
+# { op_name : [arg_name, ...] }
+core_ops_returns_info = {}
+core_ops_args_info = {}
+core_ops_args_type_info = {}
+
 
 def ParseArguments():
     parser = argparse.ArgumentParser(
@@ -130,17 +136,16 @@ def ParseYamlArgs(string):
     attrs_list = []
 
     args = [x.strip() for x in string.strip().split(",")]
-
     atype = r'((const )?\S+) '
-    aname = r'(\S+)'
+    aname = r'(.*)'
     pattern = f'{atype}{aname}'
     for i in range(len(args)):
         arg = args[i]
         m = re.search(pattern, arg)
-        arg_type = m.group(1)
-        arg_name = m.group(3).split("=")[0]
-        default_value = m.group(3).split("=")[1] if len(m.group(3).split(
-            "=")) > 1 else None
+        arg_type = m.group(1).strip()
+        arg_name = m.group(3).split("=")[0].strip()
+        default_value = m.group(3).split("=")[1].strip() if len(
+            m.group(3).split("=")) > 1 else None
         if "Tensor" in arg_type:
             assert default_value is None
             inputs_list.append([arg_name, arg_type, i])
@@ -262,7 +267,6 @@ def ForwardsValidationCheck(forward_inputs_list, forward_attrs_list,
         forward_attr_type = forward_attrs_list[i][1]
         forward_attr_default = forward_attrs_list[i][2]
         forward_attr_pos = forward_attrs_list[i][3]
-
         assert orig_attr_type == forward_attr_type
         assert orig_attr_default == forward_attr_default
         assert orig_attr_pos == forward_attr_pos
@@ -741,26 +745,34 @@ def GenerateForwardDefinition(fwd_api_name, bwd_api_name,
     # Get Function Args
     num_inputs = len(forward_attrs_list) + len(forward_inputs_position_map.keys(
     ))
-    inputs_args_list = ["" for i in range(num_inputs)]
+    inputs_args_definition_list = ["" for i in range(num_inputs)]
+    inputs_args_declaration_list = ["" for i in range(num_inputs)]
     inputs_call_list = ["" for i in range(num_inputs)]
     for name, (ttype, pos) in forward_inputs_position_map.items():
         inputs_call_list[pos] = f"{name}"
         if IsPlainTensorType(ttype):
-            inputs_args_list[
+            inputs_args_definition_list[
+                pos] = f"const paddle::experimental::Tensor& {name}"
+            inputs_args_declaration_list[
                 pos] = f"const paddle::experimental::Tensor& {name}"
         else:
             assert IsVectorTensorType(ttype)
-            inputs_args_list[
+            inputs_args_definition_list[
+                pos] = f"const std::vector<paddle::experimental::Tensor>& {name}"
+            inputs_args_declaration_list[
                 pos] = f"const std::vector<paddle::experimental::Tensor>& {name}"
 
     for name, atype, default_val, pos in forward_attrs_list:
         inputs_call_list[pos] = name
         if default_val is not None:
-            inputs_args_list[pos] = f"{atype} {name} = {default_val}"
+            inputs_args_declaration_list[
+                pos] = f"{atype} {name} = {default_val}"
         else:
-            inputs_args_list[pos] = f"{atype} {name}"
+            inputs_args_declaration_list[pos] = f"{atype} {name}"
+        inputs_args_definition_list[pos] = f"{atype} {name}"
 
-    inputs_args_str = ", ".join(inputs_args_list)
+    inputs_args_declaration_str = ", ".join(inputs_args_declaration_list)
+    inputs_args_definition_str = ", ".join(inputs_args_definition_list)
     inputs_call_args_str = ", ".join(inputs_call_list)
 
     # Forward Full Logic
@@ -812,13 +824,95 @@ def GenerateForwardDefinition(fwd_api_name, bwd_api_name,
 
     forward_function_name = GetForwardFunctionName(fwd_api_name)
     forward_function_str = FORWARD_FUNCTION_TEMPLATE.format(
-        returns_type_str, forward_function_name, inputs_args_str,
+        returns_type_str, forward_function_name, inputs_args_definition_str,
         forward_call_str, node_creation_str, returns_str)
-    forward_function_declaration_str = f"{returns_type_str} {forward_function_name}({inputs_args_str});"
+    forward_function_declaration_str = f"{returns_type_str} {forward_function_name}({inputs_args_declaration_str});"
 
     return forward_function_str, forward_function_declaration_str
 
 
+def CollectCoreOpsInformation(fwd_api_name, forward_inputs_position_map,
+                              forward_outputs_position_map, forward_attrs_list):
+    # fwd_api_name : ""
+    # forward_inputs_position_map = { "name" : [type, fwd_position] }
+    # forward_outputs_position_map = { "name" : [type, fwd_position] }
+    # forward_attrs_list = [ [attr_name, attr_type, default_value, orig_position], ...]
+    num_args = len(forward_inputs_position_map.keys()) + len(forward_attrs_list)
+    num_returns = len(forward_outputs_position_map.keys())
+
+    final_state_fwd_api_name = "final_state_" + fwd_api_name
+    core_ops_returns_info[
+        final_state_fwd_api_name] = ["" for i in range(num_returns)]
+    core_ops_args_info[final_state_fwd_api_name] = ["" for i in range(num_args)]
+    core_ops_args_type_info[
+        final_state_fwd_api_name] = ["" for i in range(num_args)]
+    for name, (ttype, pos) in forward_inputs_position_map.items():
+        core_ops_args_info[final_state_fwd_api_name][pos] = name
+        if IsPlainTensorType(ttype):
+            core_ops_args_type_info[final_state_fwd_api_name][pos] = "tensor"
+        else:
+            assert IsVectorTensorType(ttype)
+            core_ops_args_type_info[final_state_fwd_api_name][pos] = "list"
+
+    for name, _, _, pos in forward_attrs_list:
+        core_ops_args_info[final_state_fwd_api_name][pos] = name
+
+    for name, (ttype, pos) in forward_outputs_position_map.items():
+        core_ops_returns_info[final_state_fwd_api_name][pos] = name
+
+
+def GenerateCoreOpInfoDeclaration():
+    core_ops_declaration_str = """
+    extern std::unordered_map<std::string, std::vector<std::string>> core_ops_final_state_args_info;
+    extern std::unordered_map<std::string, std::vector<std::string>> core_ops_final_state_args_type_info;
+    extern std::unordered_map<std::string, std::vector<std::string>> core_ops_final_state_returns_info;
+
+"""
+    return core_ops_declaration_str
+
+
+def GenerateCoreOpInfoDefinition():
+
+    CORE_OPS_INFO_TEMPLATE = """
+std::unordered_map<std::string, std::vector<std::string>> core_ops_final_state_args_info = {{
+    {}
+}};
+std::unordered_map<std::string, std::vector<std::string>> core_ops_final_state_args_type_info = {{
+    {}
+}};
+std::unordered_map<std::string, std::vector<std::string>> core_ops_final_state_returns_info = {{
+    {}
+}};
+
+"""
+    op_args_info_list = []
+    for op_name, arg_list in core_ops_args_info.items():
+        arg_str = ",".join(["\"" + v + "\"" for v in arg_list])
+        op_args_info = f"{{ \"{op_name}\", {{ {arg_str} }} }},"
+        op_args_info_list.append(op_args_info)
+
+    op_types_info_list = []
+    for op_name, type_list in core_ops_args_type_info.items():
+        type_str = ",".join(["\"" + v + "\"" for v in type_list])
+        op_types_info = f"{{ \"{op_name}\", {{ {type_str} }} }},"
+        op_types_info_list.append(op_types_info)
+
+    op_returns_info_list = []
+    for op_name, return_list in core_ops_returns_info.items():
+        return_str = ",".join(["\"" + v + "\"" for v in return_list])
+        return_types_info = f"{{ \"{op_name}\", {{ {return_str} }} }},"
+        op_returns_info_list.append(return_types_info)
+
+    op_args_info_str = "\n".join(op_args_info_list)
+    op_types_info_str = "\n".join(op_types_info_list)
+    op_returns_info_str = "\n".join(op_returns_info_list)
+
+    core_ops_info_definition_str = CORE_OPS_INFO_TEMPLATE.format(
+        op_args_info_str, op_types_info_str, op_returns_info_str)
+
+    return core_ops_info_definition_str
+
+
 def GenerateNodeCCFile(filepath, node_definition_str):
     file_contents = """
 #include "glog/logging.h"
@@ -856,6 +950,8 @@ def GenerateForwardCCFile(filepath, forward_definition_str):
 #include "paddle/fluid/eager/api/utils/global_utils.h"
 
 """
+
+    file_contents += GenerateCoreOpInfoDefinition()
     file_contents += forward_definition_str
     with open(filepath, 'a') as f:
         f.write(file_contents)
@@ -871,6 +967,7 @@ def GenerateForwardHFile(filepath, forward_function_declaration_str):
 #include "paddle/fluid/framework/op_registry.h"
 
 """
+    file_contents += GenerateCoreOpInfoDeclaration()
     file_contents += forward_function_declaration_str
     with open(filepath, 'a') as f:
         f.write(file_contents)
@@ -985,6 +1082,11 @@ if __name__ == "__main__":
         forward_definition_str += definition_declaration_pair[0]
         forward_declaration_str += definition_declaration_pair[1]
 
+        # For python-level API dispatch
+        CollectCoreOpsInformation(fwd_api_name, forward_inputs_position_map,
+                                  forward_outputs_position_map,
+                                  forward_attrs_list)
+
     # Generate Files
     nodes_h_path = args.nodes_h_path
     nodes_cc_path = args.nodes_cc_path

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

@@ -104,6 +104,8 @@ static PyObject * eager_final_state_api_{}(PyObject *self, PyObject *args, PyObj
   PyThreadState *tstate = nullptr;
   try
   {{
+    VLOG(6) << "Running Eager Final State API: {}";
+
     // Get EagerTensors from args
 {}
 
@@ -129,16 +131,87 @@ static PyObject * eager_final_state_api_{}(PyObject *self, PyObject *args, PyObj
 
 """
     python_c_function_str = PYTHON_C_FUNCTION_TEMPLATE.format(
-        fwd_api_name, get_eager_tensor_str, parse_attributes_str,
+        fwd_api_name, fwd_api_name, get_eager_tensor_str, parse_attributes_str,
         GetForwardFunctionName(fwd_api_name), dygraph_function_call_str)
 
-    python_c_function_reg_str = f"{{\"final_state_{fwd_api_name}\", (PyCFunction)(void(*)(void))eager_final_state_api_{fwd_api_name}, METH_VARARGS | METH_KEYWORDS, \"C++ interface function for {fwd_api_name} in dygraph.\"}}"
+    python_c_function_reg_str = f"{{\"final_state_{fwd_api_name}\", (PyCFunction)(void(*)(void))eager_final_state_api_{fwd_api_name}, METH_VARARGS | METH_KEYWORDS, \"C++ interface function for {fwd_api_name} in dygraph.\"}},\n"
 
     return python_c_function_str, python_c_function_reg_str
 
 
+def GenerateCoreOpsInfoMap():
+    result = """
+static PyObject * eager_get_final_state_core_ops_args_info(PyObject *self) {
+    PyThreadState *tstate = nullptr;
+    try
+    {
+      return ToPyObject(core_ops_final_state_args_info);
+    }
+    catch(...) {
+      if (tstate) {
+        PyEval_RestoreThread(tstate);
+      }
+      ThrowExceptionToPython(std::current_exception());
+      return nullptr;
+    }
+}
+
+static PyObject * eager_get_final_state_core_ops_args_type_info(PyObject *self) {
+    PyThreadState *tstate = nullptr;
+    try
+    {
+      return ToPyObject(core_ops_final_state_args_type_info);
+    }
+    catch(...) {
+      if (tstate) {
+        PyEval_RestoreThread(tstate);
+      }
+      ThrowExceptionToPython(std::current_exception());
+      return nullptr;
+    }
+}
+
+static PyObject * eager_get_final_state_core_ops_returns_info(PyObject *self) {
+    PyThreadState *tstate = nullptr;
+    try
+    {
+      return ToPyObject(core_ops_final_state_returns_info);
+    }
+    catch(...) {
+      if (tstate) {
+        PyEval_RestoreThread(tstate);
+      }
+      ThrowExceptionToPython(std::current_exception());
+      return nullptr;
+    }
+}
+    """
+
+    core_ops_infos_registry = """
+    {\"get_final_state_core_ops_args_info\",
+    (PyCFunction)(void(*)(void))eager_get_final_state_core_ops_args_info, METH_NOARGS,
+    \"C++ interface function for eager_get_final_state_core_ops_args_info.\"},
+    {\"get_final_state_core_ops_args_type_info\",
+    (PyCFunction)(void(*)(void))eager_get_final_state_core_ops_args_type_info,
+    METH_NOARGS,
+    \"C++ interface function for eager_get_final_state_core_ops_args_type_info.\"},
+    {\"get_final_state_core_ops_returns_info\",
+    (PyCFunction)(void(*)(void))eager_get_final_state_core_ops_returns_info,
+    METH_NOARGS, \"C++ interface function for eager_get_final_state_core_ops_returns_info.\"},
+"""
+
+    return result, core_ops_infos_registry
+
+
 def GeneratePythonCWrappers(python_c_function_str, python_c_function_reg_str):
 
+    core_ops_infos_definition, core_ops_infos_registry = GenerateCoreOpsInfoMap(
+    )
+
+    python_c_function_str += core_ops_infos_definition
+    python_c_function_reg_str += core_ops_infos_registry
+    python_c_function_reg_str += "\n {nullptr,nullptr,0,nullptr}"
+
     PYTHON_C_WRAPPER_TEMPLATE = """
 #pragma once
 
@@ -215,12 +288,12 @@ if __name__ == "__main__":
         python_c_function_reg_list.append(python_c_function_reg_str)
         print("Generated Python-C Function: ", python_c_function_str)
 
-    python_c_function_reg_list.append("{nullptr,nullptr,0,nullptr}")
     python_c_functions_str = "\n".join(python_c_function_list)
     python_c_functions_reg_str = ",\n".join(python_c_function_reg_list)
 
     python_c_str = GeneratePythonCWrappers(python_c_functions_str,
                                            python_c_functions_reg_str)
+
     print("Generated Python-C Codes: ", python_c_str)
 
     output_path = args.output_path

python/paddle/fluid/dygraph/tracer.py

@@ -21,6 +21,17 @@ from paddle.fluid import core
 from paddle.fluid import framework
 from paddle import _C_ops
 
+final_state_name_mapping = {
+    "matmul_v2": {
+        "final_op_name": "final_state_matmul",
+        "transpose_x": "trans_x",
+        "transpose_y": "trans_y",
+        "x": "X",
+        "y": "Y",
+        "out": "Out",
+    }
+}
+
 
 class Tracer(core.Tracer):
     """
@@ -40,6 +51,169 @@ class Tracer(core.Tracer):
 
         self._train_mode = True
 
+    def eager_trace_op(self,
+                       type,
+                       inputs,
+                       outputs,
+                       attrs,
+                       stop_gradient=False,
+                       inplace_map=None):
+        function_ptr = _C_ops.__dict__[type]
+
+        core_ops_args_info = _C_ops.get_core_ops_args_info()
+        core_ops_args_type_info = _C_ops.get_core_ops_args_type_info()
+        core_ops_returns_info = _C_ops.get_core_ops_returns_info()
+
+        op_args = core_ops_args_info[type]
+        op_args_type = core_ops_args_type_info[type]
+        op_returns = core_ops_returns_info[type]
+
+        arg_list = []
+        for i in range(len(op_args)):
+            arg_name = op_args[i]
+            arg_type = op_args_type[i]
+            if arg_name in inputs.keys():
+                arg_to_append = inputs[arg_name]
+            elif arg_name in outputs.keys():
+                arg_to_append = outputs[arg_name]
+            else:
+                if "Num" in arg_name:
+                    # Remove "Num" suffix to get out_name
+                    out_name = arg_name[:-3]
+                    assert out_name in outputs.keys()
+                    num_outs = len(outputs[out_name])
+                    arg_to_append = num_outs
+                else:
+                    arg_to_append = None
+
+            if arg_to_append is None:
+                arg_list.append(arg_to_append)
+            elif arg_type == "tensor":
+                if isinstance(arg_to_append, list):
+                    arg_list.append(arg_to_append[0])
+                else:
+                    arg_list.append(arg_to_append)
+            elif arg_type == "list":
+                assert isinstance(arg_to_append, list)
+                arg_list.append(arg_to_append)
+            else:
+                assert arg_type == "int"
+                assert isinstance(arg_to_append, int)
+                arg_list.append(arg_to_append)
+
+        attrs_list = []
+        for k, v in attrs.items():
+            attrs_list.append(k)
+            attrs_list.append(v)
+        returns = function_ptr(*arg_list, *attrs_list)
+
+        if isinstance(returns, tuple):
+            for i in range(len(op_returns)):
+                retname = op_returns[i]
+                if retname in outputs.keys():
+                    # Replaced outputs by function returns
+                    if isinstance(returns[i], list):
+                        for j in range(len(returns[i])):
+                            outputs[retname][j].reconstruct_from_(returns[i][j],
+                                                                  False)
+                    else:
+                        outputs[retname][0].reconstruct_from_(returns[i], False)
+        elif isinstance(returns, list):
+            assert len(outputs.keys()) == 1
+            key = list(outputs.keys())[0]
+            for j in range(len(returns)):
+                outputs[key][j].reconstruct_from_(returns[j], False)
+        else:
+            assert len(outputs.keys()) == 1
+            key = list(outputs.keys())[0]
+            if isinstance(outputs[key], list):
+                outputs[key][0].reconstruct_from_(returns, False)
+            else:
+                outputs[key].reconstruct_from_(returns, False)
+
+    def eager_final_state_trace_op(self,
+                                   type,
+                                   inputs,
+                                   outputs,
+                                   attrs,
+                                   stop_gradient=False,
+                                   inplace_map=None):
+        assert type in final_state_name_mapping.keys()
+
+        final_state_type = final_state_name_mapping[type]["final_op_name"]
+        function_ptr = _C_ops.__dict__[final_state_type]
+
+        core_ops_args_info = _C_ops.get_final_state_core_ops_args_info()
+        core_ops_args_type_info = _C_ops.get_final_state_core_ops_args_type_info(
+        )
+        core_ops_returns_info = _C_ops.get_final_state_core_ops_returns_info()
+
+        op_args = core_ops_args_info[final_state_type]
+        op_args_type = core_ops_args_type_info[final_state_type]
+        op_returns = core_ops_returns_info[final_state_type]
+
+        arg_list = []
+        for i in range(len(op_args)):
+            eager_arg_name = op_args[i]
+            arg_type = op_args_type[i]
+
+            assert eager_arg_name in final_state_name_mapping[type].keys()
+            arg_name = final_state_name_mapping[type][eager_arg_name]
+
+            if arg_name in inputs.keys():
+                arg_to_append = inputs[arg_name]
+            elif arg_name in outputs.keys():
+                arg_to_append = outputs[arg_name]
+            elif arg_name in attrs.keys() and arg_type == "":
+                arg_to_append = attrs[arg_name]
+            else:
+                # dispensable
+                arg_to_append = None
+
+            if arg_type == "":
+                # attribute
+                arg_list.append(arg_to_append)
+            elif arg_type == "tensor":
+                if isinstance(arg_to_append, list):
+                    arg_list.append(arg_to_append[0])
+                else:
+                    arg_list.append(arg_to_append)
+            elif arg_type == "list":
+                assert isinstance(arg_to_append, list)
+                arg_list.append(arg_to_append)
+            else:
+                assert arg_to_append is None
+                arg_list.append(arg_to_append)
+
+        returns = function_ptr(*arg_list)
+
+        if isinstance(returns, tuple):
+            for i in range(len(op_returns)):
+                eager_retname = op_returns[i]
+
+                assert eager_retname in final_state_name_mapping[type].keys()
+                retname = final_state_name_mapping[type][eager_retname]
+                if retname in outputs.keys():
+                    # Replaced outputs by function returns
+                    if isinstance(returns[i], list):
+                        for j in range(len(returns[i])):
+                            outputs[retname][j].reconstruct_from_(returns[i][j],
+                                                                  False)
+                    else:
+                        outputs[retname][0].reconstruct_from_(returns[i], False)
+        elif isinstance(returns, list):
+            assert len(outputs.keys()) == 1
+            key = list(outputs.keys())[0]
+            for j in range(len(returns)):
+                outputs[key][j].reconstruct_from_(returns[j], False)
+        else:
+            assert len(outputs.keys()) == 1
+            key = list(outputs.keys())[0]
+            if isinstance(outputs[key], list):
+                outputs[key][0].reconstruct_from_(returns, False)
+            else:
+                outputs[key].reconstruct_from_(returns, False)
+
     def trace_op(self,
                  type,
                  inputs,
@@ -51,78 +225,16 @@ class Tracer(core.Tracer):
             # inputs : {"sum": [tensor], ...}
             # outputs : {"sum": [tensor], ...}
 
-            function_ptr = _C_ops.__dict__[type]
-
-            core_ops_args_info = _C_ops.get_core_ops_args_info()
-            core_ops_args_type_info = _C_ops.get_core_ops_args_type_info()
-            core_ops_returns_info = _C_ops.get_core_ops_returns_info()
-
-            op_args = core_ops_args_info[type]
-            op_args_type = core_ops_args_type_info[type]
-            op_returns = core_ops_returns_info[type]
-
-            arg_list = []
-            for i in range(len(op_args)):
-                arg_name = op_args[i]
-                arg_type = op_args_type[i]
-                if arg_name in inputs.keys():
-                    arg_to_append = inputs[arg_name]
-                elif arg_name in outputs.keys():
-                    arg_to_append = outputs[arg_name]
-                else:
-                    if "Num" in arg_name:
-                        # Remove "Num" suffix to get out_name
-                        out_name = arg_name[:-3]
-                        assert out_name in outputs.keys()
-                        num_outs = len(outputs[out_name])
-                        arg_to_append = num_outs
-                    else:
-                        arg_to_append = None
+            if type in final_state_name_mapping.keys():
+                final_state_type = final_state_name_mapping[type][
+                    "final_op_name"]
 
-                if arg_to_append is None:
-                    arg_list.append(arg_to_append)
-                elif arg_type == "tensor":
-                    if isinstance(arg_to_append, list):
-                        arg_list.append(arg_to_append[0])
-                    else:
-                        arg_list.append(arg_to_append)
-                elif arg_type == "list":
-                    assert isinstance(arg_to_append, list)
-                    arg_list.append(arg_to_append)
-                else:
-                    assert arg_type == "int"
-                    assert isinstance(arg_to_append, int)
-                    arg_list.append(arg_to_append)
-
-            attrs_list = []
-            for k, v in attrs.items():
-                attrs_list.append(k)
-                attrs_list.append(v)
-            returns = function_ptr(*arg_list, *attrs_list)
-
-            if isinstance(returns, tuple):
-                for i in range(len(op_returns)):
-                    retname = op_returns[i]
-                    if retname in outputs.keys():
-                        # Replaced outputs by function returns
-                        if isinstance(returns[i], list):
-                            for j in range(len(returns[i])):
-                                outputs[retname][j].reconstruct_from_(returns[i]
-                                                                      [j])
-                        else:
-                            outputs[retname][0].reconstruct_from_(returns[i])
-            elif isinstance(returns, list):
-                assert len(outputs.keys()) == 1
-                key = list(outputs.keys())[0]
-                for j in range(len(returns)):
-                    outputs[key][j].reconstruct_from_(returns[j])
+                assert final_state_type in _C_ops.__dict__
+                self.eager_final_state_trace_op(type, inputs, outputs, attrs,
+                                                stop_gradient, inplace_map)
             else:
-                assert len(outputs.keys()) == 1
-                key = list(outputs.keys())[0]
-                if isinstance(outputs[key], list):
-                    outputs[key][0].reconstruct_from_(returns)
-                else:
-                    outputs[key].reconstruct_from_(returns)
+                self.eager_trace_op(type, inputs, outputs, attrs, stop_gradient,
+                                    inplace_map)
         else:
             self.trace(type, inputs, outputs, attrs,
                        framework._current_expected_place(), self._has_grad and

python/paddle/utils/code_gen/backward.yaml

 - backward_api : matmul_grad
-  forward : matmul (const Tensor& x, const Tensor& y, bool transpose_x, bool transpose_y) -> Tensor(out)
+  forward : matmul (const Tensor& x, const Tensor& y, bool transpose_x=false, bool transpose_y=false) -> Tensor(out)
   args : (const Tensor& x, const Tensor& y, const Tensor& out_grad, bool transpose_x=false, bool transpose_y=false)
   output : Tensor(x_grad), Tensor(y_grad)
   infer_meta :