mindstudio-probe 1.1.1__py3-none-any.whl → 1.2.1__py3-none-any.whl

msprobe/mindspore/overflow_check/kernel_graph_overflow_check.py

@@ -46,6 +46,13 @@ class KernelGraphOverflowCheck:
             self.dump_json["common_dump_settings"]["op_debug_mode"] = 2
 
     def handle(self):
+        try:
+            from msprobe.lib import _msprobe_c
+            return
+        except ImportError:
+            # 如果没有_msprobe_ce_c走MindSpore老流程
+            logger.info("Module _msprobe_c has not been installed, use interface in mindspore instead.")
+
         if os.getenv("GRAPH_OP_RUN") == "1":
             raise Exception("Must run in graph mode, not kbk mode")
         json_path = self.dump_json["common_dump_settings"]["path"]

msprobe/mindspore/service.py

@@ -1,4 +1,4 @@
-# Copyright (c) 2024-2024, Huawei Technologies Co., Ltd.
+# Copyright (c) 2024-2025, Huawei Technologies Co., Ltd.
 # All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -20,6 +20,8 @@ from collections import defaultdict
 
 import mindspore as ms
 from mindspore import nn
+from mindspore.common.api import _no_grad
+from mindspore.ops.primitive import Primitive
 try:
     from mindspore.common._pijit_context import PIJitCaptureContext
 except ImportError:
@@ -27,19 +29,25 @@ except ImportError:
 else:
     pijit_label = True
 
-
 from msprobe.core.common.exceptions import DistributedNotInitializedError, MsprobeException
 from msprobe.core.common.file_utils import create_directory
 from msprobe.core.common.utils import Const, print_tools_ends_info
 from msprobe.core.data_dump.data_collector import build_data_collector
-from msprobe.core.data_dump.data_processor.base import ModuleBackwardInputsOutputs, ModuleForwardInputsOutputs
+from msprobe.core.data_dump.data_processor.base import (ModuleBackwardInputsOutputs, ModuleForwardInputsOutputs,
+                                                        ModuleBackwardInputs)
 from msprobe.core.data_dump.scope import BaseScope
 from msprobe.mindspore.cell_processor import CellProcessor
 from msprobe.mindspore.common.log import logger
-from msprobe.mindspore.common.utils import get_rank_if_initialized
+from msprobe.mindspore.common.utils import (get_rank_if_initialized, clean_input_kwargs,
+                                            is_mindtorch, register_backward_hook_functions)
 from msprobe.mindspore.dump.hook_cell.api_registry import api_register
 from msprobe.mindspore.dump.hook_cell.primitive_hooks import PrimitiveHookService
 from msprobe.mindspore.dump.jit_dump import JitDump
+from msprobe.mindspore.dump.hook_cell.hook_cell import HOOKCell
+from msprobe.mindspore.dump.kernel_dump.kernel_config import create_kernel_config_json
+
+if is_mindtorch():
+    import torch
 
 
 class Service:
@@ -51,54 +59,144 @@ class Service:
         self.cell_processor = CellProcessor(self.data_collector.scope)
         self.primitive_hook_service = PrimitiveHookService(self)
         self.switch = False
+        self.inner_switch = False
         self.primitive_switch = False
         self.current_iter = 0
         self.first_start = True
         self.current_rank = None
         self.dump_iter_dir = None
         self.start_call = False
-        self.check_level_valid()
         self.should_stop_service = False
+        self.params_grad_info = {}
+        # 提前注册，确保注册尽可能多的API hook
+        self.register_api_hook()
 
     @staticmethod
-    def check_model_valid(model):
-        if not model or isinstance(model, nn.Cell):
-            return model
-        raise MsprobeException(
-            MsprobeException.INVALID_PARAM_ERROR, "model 参数必须是 mindspore.nn.Cell 类型。"
-        )
+    def check_model_valid(models):
+        target_module_type = (torch.nn.Module, "torch.nn.Module") if is_mindtorch() else (nn.Cell, "mindspore.nn.Cell")
+        if models is None or isinstance(models, target_module_type[0]):
+            return models
+        error_model = None
+        if isinstance(models, (list, tuple)):
+            for model in models:
+                if not isinstance(model, target_module_type[0]):
+                    error_model = model
+                    break
+        else:
+            error_model = models
 
-    def check_level_valid(self):
-        if self.config.level == Const.LEVEL_L2:
+        if error_model is not None:
+            error_info = (f"The 'model' parameter must be a {target_module_type[1]} or list[{target_module_type[1]}] "
+                          f"type, currently there is a {type(error_model)} type.")
             raise MsprobeException(
-                MsprobeException.INVALID_PARAM_ERROR, "L2 level dump function is currently not supported."
-            )
+                MsprobeException.INVALID_PARAM_ERROR, error_info)
+        return models
+
+    @staticmethod
+    def prepare_module_input_output(target_type, cell, input_data, output):
+        if target_type == BaseScope.Module_Type_Module:
+            module_input_output = ModuleForwardInputsOutputs(args=input_data, kwargs={}, output=output)
+        else:
+            module_input_output = ModuleForwardInputsOutputs(args=input_data, kwargs=cell.input_kwargs, output=output)
+        return module_input_output
 
     def build_hook(self, target_type, name):
-        def forward_hook(api_or_cell_name, cell, input_data, output):
-            if not self.should_excute_hook():
-                if hasattr(cell, 'input_kwargs'):
-                    del cell.input_kwargs
+        def pre_hook(api_or_cell_name, cell, input_data):
+            if not self.should_execute_hook(target_type, cell, True):
+                clean_input_kwargs(cell)
                 return None
 
-            if target_type == BaseScope.Module_Type_Module:
-                api_or_cell_name = self.cell_processor.set_and_get_reserved_name(cell, api_or_cell_name)
-                module_input_output = ModuleForwardInputsOutputs(args=input_data, kwargs={}, output=output)
-            else:
-                module_input_output = ModuleForwardInputsOutputs(args=input_data, kwargs=cell.input_kwargs,
-                                                                 output=output)
+            with _no_grad():
+                self.inner_switch = True
+                if target_type == BaseScope.Module_Type_Module:
+                    api_or_cell_name = self.cell_processor.set_and_get_reserved_name(cell, api_or_cell_name)
+                else:
+                    cell.forward_data_collected = True
+                    HOOKCell.add_cell_count(name)
+                module_input_output = self.prepare_module_input_output(target_type, cell, input_data, None)
+                self.data_collector.update_api_or_module_name(api_or_cell_name)
+                self.data_collector.forward_input_data_collect(api_or_cell_name, cell, pid, module_input_output)
+                self.inner_switch = False
+                return input_data
+
+        def grad_hook(cell, ori_name, param_name):
+            def hook_fn(grad):
+                if not self.should_execute_hook(target_type, cell, False):
+                    return None
+                self.inner_switch = True
+                self.data_collector.params_data_collect(ori_name, param_name, pid, grad)
+                self.inner_switch = False
+                return None
 
-            self.data_collector.update_api_or_module_name(api_or_cell_name)
-            self.data_collector.forward_data_collect(api_or_cell_name, cell, pid, module_input_output)
-            if self.data_collector.if_return_forward_new_output():
-                return self.data_collector.get_forward_new_output()
-            if hasattr(cell, 'input_kwargs'):
-                del cell.input_kwargs
-            return output
+            return hook_fn
+
+        def register_param_hook(ori_name, cell, params_dict):
+            '''
+            注册参数hook
+            '''
+            # data_mode为forward时，不注册参数hook
+            if not (Const.FORWARD in self.config.data_mode and Const.BACKWARD not in self.config.data_mode):
+                for param_name, param in params_dict.items():
+                    if param.requires_grad:
+                        param.register_hook(grad_hook(cell, ori_name, param_name))
+
+        def init_params_grad_info(cell, params_dict):
+            '''
+            初始化参数梯度信息, 在前向hook结束后, 将参数梯度信息写入cache_data中用于占位
+            '''
+            if not params_dict:
+                return
+            if not (Const.FORWARD in self.config.data_mode and Const.BACKWARD not in self.config.data_mode):
+                grad_name = cell.params_grad_name if hasattr(cell, 'params_grad_name') else None
+                # 判断是否已经在cache_data中进行了占位, 若没有则先写入cache_data中
+                if not self.params_grad_info.get(grad_name):
+                    data_info = {grad_name: {key: [None] for key, value in params_dict.items() if value.requires_grad}}
+                    # 当模块中的参数有requires_grad属性为True时，才会进行梯度计算，此时才需要占位
+                    if data_info.get(grad_name):
+                        # 将grad_name的data_info先写入cache_data中, 梯度计算后再更新
+                        self.data_collector.handle_data(grad_name, data_info,
+                                                        flush=self.data_collector.data_processor.is_terminated)
+                    # 记录当前模块的参数梯度信息已占位
+                    self.params_grad_info[grad_name] = True
+
+        def forward_hook(api_or_cell_name, cell, input_data, output):
+            if not self.should_execute_hook(target_type, cell, True):
+                clean_input_kwargs(cell)
+                return None
+            with _no_grad():
+                self.inner_switch = True
+                module_input_output = self.prepare_module_input_output(target_type, cell, input_data, output)
+                if target_type == BaseScope.Module_Type_Module:
+                    api_or_cell_name = self.cell_processor.set_and_get_reserved_name(cell, api_or_cell_name)
+                    params_dict = {key.split(Const.SEP)[-1]: value for key, value in cell.parameters_dict(
+                        recurse=False).items()}
+                    setattr(module_input_output, Const.PARAMS, params_dict)
+                    # 判断是否需要注册参数hook
+                    if not hasattr(cell, 'params_grad_name') and params_dict:
+                        ori_name = api_or_cell_name.rsplit(Const.SEP, 2)[0]
+                        grad_name = ori_name + Const.SEP + Const.PARAMS_GRAD
+                        # 首次执行前向hook时，添加params_grad_name属性，并注册参数hook
+                        setattr(cell, 'params_grad_name', grad_name)
+                        register_param_hook(ori_name, cell, params_dict)
+                    self.data_collector.update_api_or_module_name(api_or_cell_name)
+                    self.data_collector.forward_data_collect(api_or_cell_name, cell, pid, module_input_output)
+                    init_params_grad_info(cell, params_dict)
+                else:
+                    self.data_collector.update_api_or_module_name(api_or_cell_name)
+                    self.data_collector.forward_output_data_collect(api_or_cell_name, cell, pid, module_input_output)
+
+                if self.data_collector.if_return_forward_new_output():
+                    forward_new_output = self.data_collector.get_forward_new_output()
+                    self.inner_switch = False
+                    return forward_new_output
+                clean_input_kwargs(cell)
+                self.inner_switch = False
+                return output
 
         def backward_hook(api_or_cell_name, cell, grad_input, grad_output):
-            if not self.should_excute_hook():
+            if not self.should_execute_hook(target_type, cell, False):
                 return
+            self.inner_switch = True
 
             need_exchange = True
             if target_type == BaseScope.Module_Type_Module:
@@ -114,12 +212,32 @@ class Service:
                 else:
                     module_input_output = ModuleBackwardInputsOutputs(grad_input=grad_input, grad_output=grad_output)
                 self.data_collector.backward_data_collect(api_or_cell_name, cell, pid, module_input_output)
+            self.inner_switch = False
+
+        def pre_backward_hook(api_or_cell_name, cell, grad_input):
+            if not self.should_execute_hook(target_type, cell, False):
+                return
+            self.inner_switch = True
+            module_input = ModuleBackwardInputs(grad_input=grad_input)
+            self.data_collector.update_api_or_module_name(api_or_cell_name)
+            self.data_collector.backward_input_data_collect(api_or_cell_name, cell, pid, module_input)
+
+            self.inner_switch = False
 
         pid = os.getpid()
-        forward_name_template = name + Const.FORWARD
-        backward_name_template = name + Const.BACKWARD
-        forward_hook = functools.partial(forward_hook, forward_name_template)
-        backward_hook = functools.partial(backward_hook, backward_name_template)
+        if target_type == BaseScope.Module_Type_Module:
+            full_forward_name = name + Const.FORWARD
+            full_backward_name = name + Const.BACKWARD
+        else:
+            full_forward_name = name + str(HOOKCell.get_cell_count(name)) + Const.SEP + Const.FORWARD
+            full_backward_name = name + str(HOOKCell.get_cell_count(name)) + Const.SEP + Const.BACKWARD
+        pre_forward_hook = functools.partial(pre_hook, full_forward_name)
+        forward_hook = functools.partial(forward_hook, full_forward_name)
+        backward_hook = functools.partial(backward_hook, full_backward_name)
+        pre_backward_hook = functools.partial(pre_backward_hook, full_backward_name)
+
+        def wrap_pre_forward_hook(cell, input_data):
+            return pre_forward_hook(cell, input_data)
 
         def wrap_forward_hook(cell, input_data, output_data):
             return forward_hook(cell, input_data, output_data)
@@ -127,7 +245,10 @@ class Service:
         def wrap_backward_hook(cell, grad_input, grad_output):
             return backward_hook(cell, grad_input, grad_output)
 
-        return wrap_forward_hook, wrap_backward_hook
+        def wrap_pre_backward_hook(cell, grad_input):
+            return pre_backward_hook(cell, grad_input)
+
+        return wrap_pre_forward_hook, wrap_forward_hook, wrap_backward_hook, wrap_pre_backward_hook
 
     def update_primitive_counters(self, primitive_name):
         if primitive_name not in self.primitive_counters:
@@ -135,33 +256,20 @@ class Service:
         else:
             self.primitive_counters[primitive_name] += 1
 
-    def register_primitive_hooks(self):
-        primitive_set = set()
-        for _, cell in self.model.cells_and_names():
-            for pname, primitive in cell._primitives.items():
-                primitive_set.add((pname, primitive))
-
-        for pname, primitive in primitive_set:
-            primitive_class_name = primitive.__class__.__name__
-            primitive_combined_name = pname + Const.SEP + primitive_class_name
-            new_primitive = type('NewPrimitive', (primitive.__class__,),
-                                 {'__call__': self.primitive_hook_service.wrap_primitive(primitive.__call__,
-                                                                                         primitive_combined_name)})
-            primitive.__class__ = new_primitive
-
     def step(self):
+        if self.config.async_dump:
+            self.data_collector.fill_stack_tensor_data()
+            self.data_collector.data_processor.dump_async_data()
+        self.data_collector.write_json()
         self.current_iter += 1
         self.data_collector.update_iter(self.current_iter)
-        self.primitive_hook_service.primitive_counters.clear()
-        self.data_collector.data_writer.reset_cache()
-        JitDump.jit_count = defaultdict(int)
+        self.reset_status()
 
     def start(self, model=None):
         self.start_call = True
         if self.should_stop_service:
             return
         if self.need_end_service():
-            api_register.api_set_ori_func()
             self.should_stop_service = True
             self.switch = False
             self.primitive_switch = False
@@ -181,7 +289,8 @@ class Service:
 
             if self.config.rank and self.current_rank not in self.config.rank:
                 return
-            self.register_hook_new()
+            self.register_primitive_hook()
+            self.register_cell_hook()
             if self.config.level in [Const.LEVEL_MIX, Const.LEVEL_L1]:
                 JitDump.set_config(self.config)
                 JitDump.set_data_collector(self.data_collector)
@@ -200,25 +309,6 @@ class Service:
         logger.info(f"Dump data will be saved in {self.dump_iter_dir}.")
         JitDump.jit_dump_switch = True
 
-    def forward_backward_dump_end(self):
-        if self.should_stop_service:
-            return
-        logger.info(f"{Const.TOOL_NAME}: debugger.forward_backward_dump_end() is set successfully. ")
-        if not self.start_call:
-            logger.error(f"{Const.TOOL_NAME}: debugger.start() is not set in the current scope.")
-            raise Exception("debugger.start() is not set in the current scope.")
-        if not self.switch:
-            logger.error(f"{Const.TOOL_NAME}: debugger.forward_backward_dump_end() should be called between "
-                         "debugger.start() and debugger.stop() ")
-            raise Exception("debugger.stop() is already called. ")
-        if self.config.step and self.current_iter not in self.config.step:
-            return
-        if self.config.rank and self.current_rank not in self.config.rank:
-            return
-        self.primitive_switch = False
-        api_register.api_set_ori_func()
-        JitDump.jit_dump_switch = False
-
     def stop(self):
         if self.should_stop_service:
             return
@@ -234,6 +324,9 @@ class Service:
         self.switch = False
         self.primitive_switch = False
         self.start_call = False
+        if self.config.async_dump:
+            self.data_collector.fill_stack_tensor_data()
+            self.data_collector.data_processor.dump_async_data()
         self.data_collector.write_json()
         JitDump.jit_dump_switch = False
 
@@ -244,8 +337,16 @@ class Service:
             return True
         return False
 
-    def should_excute_hook(self):
-        if not self.switch:
+    def should_execute_hook(self, hook_type, cell, is_forward):
+        is_cell_hook = hook_type == BaseScope.Module_Type_Module
+        if is_cell_hook and not self.switch:
+            return False
+        elif not is_cell_hook and is_forward and not self.switch:
+            return False
+        elif not is_cell_hook and not is_forward and not cell.forward_data_collected:
+            return False
+
+        if self.inner_switch:
             return False
         if not self.data_collector or self.data_collector.data_processor.is_terminated:
             return False
@@ -255,6 +356,12 @@ class Service:
         create_directory(self.config.dump_path)
         self.dump_iter_dir = os.path.join(self.config.dump_path, f"step{self.current_iter}")
         cur_rank = self.current_rank if self.current_rank is not None else ''
+        if self.config.level == Const.LEVEL_L2:
+            create_directory(self.dump_iter_dir)
+            kernel_config_path = create_kernel_config_json(self.dump_iter_dir, cur_rank)
+            self.config.kernel_config_path = kernel_config_path
+            return
+
         dump_dir = os.path.join(self.dump_iter_dir, f"rank{cur_rank}")
         create_directory(dump_dir)
         if self.config.task in self.data_collector.tasks_need_tensor_data:
@@ -267,37 +374,96 @@ class Service:
         stack_file_path = os.path.join(dump_dir, "stack.json")
         construct_file_path = os.path.join(dump_dir, "construct.json")
         self.data_collector.update_dump_paths(
-            dump_file_path, stack_file_path, construct_file_path, dump_data_dir, None)
+            dump_file_path, stack_file_path, construct_file_path, dump_data_dir, None
+        )
+        self.data_collector.initialize_json_file(
+            framework=Const.MT_FRAMEWORK if is_mindtorch() else Const.MS_FRAMEWORK
+        )
 
     def empty(self, *args, **kwargs):
         pass
 
-    def register_hook_new(self):
-        logger.info("The {} hook function is successfully mounted to the model.".format(self.config.task))
-        if self.config.level in [Const.LEVEL_MIX, Const.LEVEL_L1]:
+    def register_api_hook(self):
+        if self.config.level in [Const.LEVEL_MIX, Const.LEVEL_L1, Const.LEVEL_L2]:
+            logger.info(f"The api {self.config.task} hook function is successfully mounted to the model.")
             api_register.initialize_hook(functools.partial(self.build_hook, BaseScope.Module_Type_API))
             api_register.api_set_hook_func()
-            if self.model and self.config.task in Const.DUMP_DATA_COLLECTION_LIST:
-                self.register_primitive_hooks()
 
+    def get_cells_and_names(self):
+        cells_and_names_with_index = {}
+
+        def get_cell_or_module(model):
+            return model.named_modules() if is_mindtorch() else model.cells_and_names()
+        
+        if isinstance(self.model, (list, tuple)):
+            for index, model in enumerate(self.model):
+                cells_and_names_with_index[str(index)] = get_cell_or_module(model)
+        else:
+            cells_and_names_with_index["-1"] = get_cell_or_module(self.model)
+        return cells_and_names_with_index        
+
+    def register_primitive_hook(self):
+        if self.config.level not in [Const.LEVEL_MIX, Const.LEVEL_L1]:
+            return
+        if not self.model or self.config.task not in Const.DUMP_DATA_COLLECTION_LIST:
+            return
+
+        primitive_set = set()
+        cells_and_names_with_index = self.get_cells_and_names()
+        for cells_and_names in cells_and_names_with_index.values():
+            for _, cell in cells_and_names:
+                for attribute, value in vars(cell).items():
+                    if isinstance(value, Primitive):
+                        primitive_set.add((attribute, value))
+
+        for pname, primitive in primitive_set:
+            primitive_class_name = primitive.__class__.__name__
+            primitive_combined_name = pname + Const.SEP + primitive_class_name
+            new_primitive = type('NewPrimitive', (primitive.__class__,),
+                                 {'__call__': self.primitive_hook_service.wrap_primitive(primitive.__call__,
+                                                                                         primitive_combined_name)})
+            primitive.__class__ = new_primitive
+
+    def register_cell_hook(self):
         if self.config.level in [Const.LEVEL_MIX, Const.LEVEL_L0]:
+            logger.info(f"The cell {self.config.task} hook function is successfully mounted to the model.")
             if not self.model:
                 raise MsprobeException(MsprobeException.INVALID_PARAM_ERROR,
                                        f"The current level is {self.config.level}, the model cannot be None")
-            for name, cell in self.model.cells_and_names():
-                if cell == self.model:
-                    continue
-                prefix = 'Cell' + Const.SEP + name + Const.SEP + \
-                         cell.__class__.__name__ + Const.SEP
-                forward_hook, backward_hook = self.build_hook(BaseScope.Module_Type_Module, prefix)
-                cell.register_forward_hook(forward_hook)
-                cell.register_backward_hook(backward_hook)
-
-                cell.register_forward_pre_hook(
-                    self.cell_processor.node_hook(prefix + Const.FORWARD, Const.START))
-                cell.register_forward_hook(
-                    self.cell_processor.node_hook(prefix + Const.FORWARD, Const.STOP))
-                cell.register_backward_pre_hook(
-                    self.cell_processor.node_hook(prefix + Const.BACKWARD, Const.START))
-                cell.register_backward_hook(
-                    self.cell_processor.node_hook(prefix + Const.BACKWARD, Const.STOP))
+            model_type = Const.MODULE if is_mindtorch() else Const.CELL 
+            cells_and_names_with_index = self.get_cells_and_names()
+
+            for index, cells_and_names in cells_and_names_with_index.items():
+                model = self.model if index == "-1" else self.model[int(index)]
+                for name, cell in cells_and_names:
+                    if cell == model:
+                        continue
+                    cell_index = (index + Const.SEP) if index != "-1" else ""
+                    prefix = (model_type + Const.SEP + cell_index + name + 
+                              Const.SEP + cell.__class__.__name__ + Const.SEP)
+                    _, forward_hook, backward_hook, _ = self.build_hook(BaseScope.Module_Type_Module, prefix)
+                    cell.register_forward_hook(forward_hook)
+                    cell.register_forward_pre_hook(
+                        self.cell_processor.node_hook(prefix + Const.FORWARD, Const.START))
+                    cell.register_forward_hook(
+                        self.cell_processor.node_hook(prefix + Const.FORWARD, Const.STOP))
+
+                    register_backward_hook_functions["full"](cell, backward_hook)
+                    register_backward_hook_functions["pre"](
+                        cell, self.cell_processor.node_hook(prefix + Const.BACKWARD, Const.START))
+                    register_backward_hook_functions["full"](
+                        cell, self.cell_processor.node_hook(prefix + Const.BACKWARD, Const.STOP))
+
+    def reset_status(self):
+        self.primitive_hook_service.primitive_counters.clear()
+        self.data_collector.data_writer.reset_cache()
+        JitDump.jit_count = defaultdict(int)
+        self.params_grad_info.clear()
+
+        if self.config.level == Const.LEVEL_L2:
+            self.data_collector.data_processor.reset_status()
+            return
+        if self.config.step and self.current_iter not in self.config.step:
+            return
+        if self.config.rank and self.current_rank not in self.config.rank:
+            return

msprobe/msprobe.py

@@ -16,10 +16,12 @@
 import argparse
 import sys
 import importlib.util
-from msprobe.core.compare.utils import _compare_parser
+
+from msprobe.core.common.const import Const
 from msprobe.core.common.log import logger
+from msprobe.core.compare.utils import _compare_parser
 from msprobe.core.compare.compare_cli import compare_cli
-from msprobe.core.common.const import Const
+from msprobe.core.compare.merge_result.merge_result_cli import _merge_result_parser, merge_result_cli
 
 
 def is_module_available(module_name):
@@ -45,10 +47,15 @@ def main():
     multi_run_ut_cmd_parser = subparsers.add_parser('multi_run_ut')
     api_precision_compare_cmd_parser = subparsers.add_parser('api_precision_compare')
     run_overflow_check_cmd_parser = subparsers.add_parser('run_overflow_check')
+    code_mapping_cmd_parser = subparsers.add_parser('code_mapping')
     graph_service_cmd_parser = subparsers.add_parser('graph')
+    op_generate_cmd_parser = subparsers.add_parser('op_generate')
+    merge_result_parser = subparsers.add_parser('merge_result')
     _compare_parser(compare_cmd_parser)
+    _merge_result_parser(merge_result_parser)
+
     is_torch_available = is_module_available("torch")
-    is_mindspore_available = is_module_available("mindspore")
+
     if len(sys.argv) < 4:
         parser.print_help()
         sys.exit(0)
@@ -62,6 +69,8 @@ def main():
         from msprobe.pytorch.api_accuracy_checker.run_ut.run_overflow_check import _run_overflow_check_parser, \
             _run_overflow_check_command
         from msprobe.visualization.graph_service import _pt_graph_service_parser, _pt_graph_service_command
+        from msprobe.pytorch.api_accuracy_checker.generate_op_script.op_generator import _op_generator_parser, \
+            _run_operator_generate_commond
 
         _run_ut_parser(run_ut_cmd_parser)
         _run_ut_parser(multi_run_ut_cmd_parser)
@@ -70,12 +79,15 @@ def main():
         _api_precision_compare_parser(api_precision_compare_cmd_parser)
         _run_overflow_check_parser(run_overflow_check_cmd_parser)
         _pt_graph_service_parser(graph_service_cmd_parser)
+        _op_generator_parser(op_generate_cmd_parser)
     elif framework_args.framework == Const.MS_FRAMEWORK:
         from msprobe.mindspore.api_accuracy_checker.cmd_parser import add_api_accuracy_checker_argument
         from msprobe.visualization.graph_service import _ms_graph_service_parser, _ms_graph_service_command
         add_api_accuracy_checker_argument(run_ut_cmd_parser)
         from msprobe.mindspore.api_accuracy_checker.cmd_parser import multi_add_api_accuracy_checker_argument
         multi_add_api_accuracy_checker_argument(multi_run_ut_cmd_parser)
+        from msprobe.mindspore.code_mapping.cmd_parser import add_ir_parser_arguments
+        add_ir_parser_arguments(code_mapping_cmd_parser)
 
         _ms_graph_service_parser(graph_service_cmd_parser)
 
@@ -97,17 +109,23 @@ def main():
             _run_overflow_check_command(args)
         elif sys.argv[3] == "graph":
             _pt_graph_service_command(args)
+        elif sys.argv[3] == 'op_generate':
+            _run_operator_generate_commond(args)
         elif sys.argv[3] == "compare":
             if args.cell_mapping is not None or args.api_mapping is not None:
                 logger.error("Argument -cm or -am is not supported in PyTorch framework")
                 raise Exception("Argument -cm or -am is not supported in PyTorch framework")
             compare_cli(args)
+        elif sys.argv[3] == "merge_result":
+            merge_result_cli(args)
     else:
         if not is_module_available(Const.MS_FRAMEWORK):
             logger.error("MindSpore does not exist, please install MindSpore library")
             raise Exception("MindSpore does not exist, please install MindSpore library")
         if sys.argv[3] == "compare":
             compare_cli(args)
+        elif sys.argv[3] == "merge_result":
+            merge_result_cli(args)
         elif sys.argv[3] == "run_ut":
             from msprobe.mindspore.api_accuracy_checker.main import api_checker_main
             api_checker_main(args)
@@ -116,6 +134,9 @@ def main():
             mul_api_checker_main(args)
         elif sys.argv[3] == "graph":
             _ms_graph_service_command(args)
+        elif sys.argv[3] == "code_mapping":
+            from msprobe.mindspore.code_mapping.main import code_mapping_main
+            code_mapping_main(args)
 
 
 if __name__ == "__main__":

msprobe/pytorch/__init__.py

@@ -1,6 +1,4 @@
-#!/usr/bin/env python3
-# -*- coding: utf-8 -*-
-# Copyright (c) 2024-2024, Huawei Technologies Co., Ltd.
+# Copyright (c) 2024-2025, Huawei Technologies Co., Ltd.
 # All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0  (the "License");
@@ -16,9 +14,12 @@
 # limitations under the License.
 
 
-from msprobe.pytorch.monitor.module_hook import TrainerMon
+import torch
 from .compare.distributed_compare import compare_distributed
 from .compare.pt_compare import compare
 from .common.utils import seed_all
-from .debugger.precision_debugger import PrecisionDebugger
-from .functional.module_dump import module_dump, module_dump_end
+from .debugger.precision_debugger import PrecisionDebugger, module_dump, module_dump_end
+
+torch_version_above_or_equal_2 = torch.__version__.split('+')[0] >= '2.0'
+if torch_version_above_or_equal_2:
+    from msprobe.pytorch.monitor.module_hook import TrainerMon