mindstudio-probe 1.1.1__py3-none-any.whl → 1.2.2__py3-none-any.whl

msprobe/core/data_dump/data_processor/base.py

@@ -17,6 +17,9 @@ import inspect
 import os
 from dataclasses import dataclass, is_dataclass
 from typing import Tuple, Dict, Optional, Any
+from functools import partial
+import copy
+from typing import Union
 
 import numpy as np
 
@@ -39,9 +42,8 @@ class ModuleForwardInputsOutputs:
     def output_tuple(self):
         return convert_tuple(self.output)
 
-    def concat_args_and_kwargs(self):
-        args = self.args + tuple(self.kwargs.values())
-        return args
+    def update_output_with_args_and_kwargs(self):
+        self.output = self.args + tuple(self.kwargs.values())
 
 
 @dataclass
@@ -77,17 +79,18 @@ class ModuleBackwardOutputs:
 
 
 class TensorStatInfo:
-    def __init__(self, max_val=None, min_val=None, mean_val=None, norm_val=None):
+    def __init__(self, max_val=None, min_val=None, mean_val=None, norm_val=None, stack_tensor_stat=None):
         self.max = max_val
         self.min = min_val
         self.mean = mean_val
         self.norm = norm_val
+        self.stack_tensor_stat = stack_tensor_stat
 
 
 class BaseDataProcessor:
     _recursive_key_stack = []
     special_type = (
-        np.integer, np.floating, np.bool_, np.complexfloating, np.str_, np.byte, np.unicode_,
+        np.integer, np.floating, np.bool_, np.complexfloating, np.str_, np.byte, np.unicode_, np.ndarray,
         bool, int, float, str, slice,
         type(Ellipsis)
     )
@@ -102,6 +105,7 @@ class BaseDataProcessor:
         self.current_iter = 0
         self._return_forward_new_output = False
         self._forward_new_output = None
+        self.save_name = None
         if hasattr(config, "data_mode"):
             self.allowed_data_mode = self._get_allowed_data_mode(config.data_mode)
 
@@ -142,6 +146,37 @@ class BaseDataProcessor:
         else:
             return data
 
+    @staticmethod
+    def set_value_into_nested_structure(data_structure, indexes, value):
+        '''
+        Args:
+            data_structure: nested data structure
+            indexes: List
+            value: value to be set
+        '''
+        if not indexes:
+            raise ValueError("set_value_into_nested_structure failed: "
+                             "indexes need to be non empty when set value to nested data structure")
+        current_level = data_structure
+        for i, index in enumerate(indexes):
+            valid_for_list = isinstance(current_level, list) and isinstance(index, int) and len(current_level) > index
+            valid_for_dict = isinstance(current_level, dict) and index in current_level
+            is_last = i == len(indexes) - 1
+            if valid_for_dict or valid_for_list:
+                if is_last:
+                    try:
+                        current_level[index] = value
+                    except Exception as e:
+                        raise IndexError("set_value_into_nested_structure failed: passed indexes wrong") from e
+                else:
+                    try:
+                        current_level = current_level[index]
+                    except Exception as e:
+                        raise IndexError("set_value_into_nested_structure failed: passed indexes wrong") from e
+            else:
+                raise ValueError("set_value_into_nested_structure failed: "
+                                 "invalid data_structure type or invalid index")
+
     @staticmethod
     def _convert_numpy_to_builtin(arg):
         type_mapping = {
@@ -182,8 +217,22 @@ class BaseDataProcessor:
         return single_arg
 
     @staticmethod
-    def _analyze_numpy(value, numpy_type):
-        return {"type": numpy_type, "value": value}
+    def _analyze_numpy(ndarray, numpy_type):
+        ndarray_json = {}
+        ndarray_json.update({'type': 'numpy.ndarray'})
+        ndarray_json.update({'dtype': str(ndarray.dtype)})
+        ndarray_json.update({'shape': ndarray.shape})
+        if ndarray.size > 0:
+            ndarray_json.update({"Max": np.max(ndarray).item()})
+            ndarray_json.update({"Min": np.min(ndarray).item()})
+            ndarray_json.update({"Mean": np.mean(ndarray).item()})
+            ndarray_json.update({"Norm": np.linalg.norm(ndarray).item()})
+        else:
+            ndarray_json.update({"Max": None})
+            ndarray_json.update({"Min": None})
+            ndarray_json.update({"Mean": None})
+            ndarray_json.update({"Norm": None})
+        return ndarray_json
 
     @staticmethod
     def _get_allowed_data_mode(data_mode):
@@ -202,7 +251,7 @@ class BaseDataProcessor:
         return cls.special_type
 
     @classmethod
-    def recursive_apply_transform(cls, args, transform, depth=0):
+    def recursive_apply_transform(cls, args, transform, depth=0) -> Union[dict, list, None]:
         if depth > Const.MAX_DEPTH:
             logger.error(f"The maximum depth of recursive transform, {Const.MAX_DEPTH} is reached.")
             raise CompareException(CompareException.RECURSION_LIMIT_ERROR)
@@ -219,20 +268,20 @@ class BaseDataProcessor:
             return cls.apply_transform_dict(args_dict, transform, depth)
         elif isinstance(args, (list, tuple)):
             result_list = cls.apply_transform_list(args, transform, depth)
-            return type(args)(result_list)
+            return result_list
         elif isinstance(args, dict):
             return cls.apply_transform_dict(args, transform, depth)
         elif args is not None:
-            logger.warning(f"Data type {type(args)} is not supported.")
+            logger.debug(f"Data type {type(args)} is not supported.")
             return None
         else:
             return None
-    
+
     @classmethod
     def apply_transform_dict(cls, args, transform, depth):
         result_dict = {}
         for k, arg in args.items():
-            cls._recursive_key_stack.append(str(k))
+            cls._recursive_key_stack.append(k)
             result_dict[k] = cls.recursive_apply_transform(arg, transform, depth=depth + 1)
             cls._recursive_key_stack.pop()
         return result_dict
@@ -241,11 +290,21 @@ class BaseDataProcessor:
     def apply_transform_list(cls, args, transform, depth):
         result_list = []
         for i, arg in enumerate(args):
-            cls._recursive_key_stack.append(str(i))
+            cls._recursive_key_stack.append(i)
             result_list.append(cls.recursive_apply_transform(arg, transform, depth=depth + 1))
             cls._recursive_key_stack.pop()
         return result_list
 
+    @classmethod
+    def register_hook_single_element(cls, element, suffix_stack, hook_fn):
+        if cls.is_hookable_element(element):
+            indexes = copy.deepcopy(suffix_stack)
+            wrap_hook_fn = partial(hook_fn, indexes=indexes)
+
+            def real_hook_fn(grad):
+                return wrap_hook_fn(grad)
+            element.register_hook(real_hook_fn)
+
     def if_return_forward_new_output(self):
         return self._return_forward_new_output
 
@@ -273,13 +332,10 @@ class BaseDataProcessor:
         """
         return forward_backward in self.allowed_data_mode and input_output in self.allowed_data_mode
 
-    def analyze_pre_forward(self, name, module, module_input_output: ModuleForwardInputsOutputs):
-        pass
-
     def analyze_element(self, element):
         return self.recursive_apply_transform(element, self.analyze_single_element)
 
-    def analyze_forward(self, name, module, module_input_output: ModuleForwardInputsOutputs):
+    def analyze_forward_input(self, name, module, module_input_output: ModuleForwardInputsOutputs):
         api_info_struct = {}
         # check whether data_mode contains forward or input
         if self.is_dump_for_data_mode(Const.FORWARD, Const.INPUT):
@@ -291,16 +347,22 @@ class BaseDataProcessor:
             kwargs_info_list = self.analyze_element(module_input_output.kwargs)
             api_info_struct[name][Const.INPUT_KWARGS] = kwargs_info_list
 
-        # check whether data_mode contains forward or output
+        return api_info_struct
+
+    def analyze_forward_output(self, name, module, module_input_output: ModuleForwardInputsOutputs):
+        api_info_struct = {}
+        # check whether data_mode contains forward or input
         if self.is_dump_for_data_mode(Const.FORWARD, Const.OUTPUT):
-            api_info_struct[name] = api_info_struct.get(name, {})
+            api_info_struct[name] = {}
             self.api_data_category = Const.OUTPUT
             output_info_list = self.analyze_element(module_input_output.output_tuple)
             api_info_struct[name][Const.OUTPUT] = output_info_list
+
         return api_info_struct
 
-    def analyze_pre_forward_inplace(self, name, module_input_output: ModuleForwardInputsOutputs):
+    def analyze_forward(self, name, module, module_input_output: ModuleForwardInputsOutputs):
         api_info_struct = {}
+        # check whether data_mode contains forward or input
         if self.is_dump_for_data_mode(Const.FORWARD, Const.INPUT):
             api_info_struct[name] = {}
             self.api_data_category = Const.INPUT
@@ -309,16 +371,18 @@ class BaseDataProcessor:
             self.api_data_category = Const.KWARGS
             kwargs_info_list = self.analyze_element(module_input_output.kwargs)
             api_info_struct[name][Const.INPUT_KWARGS] = kwargs_info_list
-        return api_info_struct
 
-    def analyze_forward_inplace(self, name, module_input_output: ModuleForwardInputsOutputs):
-        concat_args = module_input_output.concat_args_and_kwargs()
-        api_info_struct = {}
+        # check whether data_mode contains forward or output
         if self.is_dump_for_data_mode(Const.FORWARD, Const.OUTPUT):
-            api_info_struct[name] = {}
+            api_info_struct[name] = api_info_struct.get(name, {})
             self.api_data_category = Const.OUTPUT
-            output_info_list = self.analyze_element(concat_args)
+            output_info_list = self.analyze_element(module_input_output.output_tuple)
             api_info_struct[name][Const.OUTPUT] = output_info_list
+
+        if name in api_info_struct and hasattr(module_input_output, Const.PARAMS):
+            self.api_data_category = Const.PARAMS
+            api_info_struct[name][Const.PARAMS] = self.analyze_element(getattr(module_input_output, Const.PARAMS))
+
         return api_info_struct
 
     def analyze_backward(self, name, module, module_input_output: ModuleBackwardInputsOutputs):
@@ -359,9 +423,47 @@ class BaseDataProcessor:
             api_info_struct[name][Const.OUTPUT] = output_info_list
         return api_info_struct
 
+    def analyze_params(self, name, param_name, grad):
+        api_info_struct = {}
+        self.save_name = name + Const.SEP + param_name
+        data_info = self.analyze_element(grad)
+        grad_info_dict = {param_name: [data_info]}
+        api_info_struct[name] = grad_info_dict
+        return api_info_struct
+
     def get_save_file_path(self, suffix):
         file_format = Const.PT_SUFFIX if self.config.framework == Const.PT_FRAMEWORK else Const.NUMPY_SUFFIX
-        dump_data_name = (self.current_api_or_module_name + Const.SEP + self.api_data_category + Const.SEP +
-                          suffix + file_format)
+        if self.save_name is not None:
+            dump_data_name = (self.save_name + file_format)
+            self.save_name = None
+        else:
+            dump_data_name = (self.current_api_or_module_name + Const.SEP + self.api_data_category + Const.SEP +
+                              suffix + file_format)
         file_path = os.path.join(self.data_writer.dump_tensor_data_dir, dump_data_name)
         return dump_data_name, file_path
+
+    def analyze_element_to_all_none(self, element):
+        return self.recursive_apply_transform(element, lambda element, stack: None)
+
+    def analyze_debug_forward(self, variable, name_with_count):
+        self.current_api_or_module_name = name_with_count
+        self.api_data_category = Const.TENSOR
+        # these two attributes are used to construct tensor file name {name_with_count}.tensor.{indexes}.npy/pt
+        data_info = self.analyze_element(variable)
+        return data_info
+
+    def analyze_debug_backward(self, variable, grad_name_with_count, nested_data_structure):
+        def hook_fn(grad, indexes):
+            suffix = Const.SEP.join([str(index) for index in indexes])
+            self.save_name = grad_name_with_count + Const.SEP + Const.TENSOR + Const.SEP + suffix
+            grad_data_info = self.analyze_element(grad)
+            self.save_name = None
+            full_index = [grad_name_with_count] + indexes
+            try:
+                self.set_value_into_nested_structure(nested_data_structure, full_index, grad_data_info)
+            except (ValueError, IndexError) as e:
+                logger.warning(f"error occured while recording statistics of {grad_name_with_count} variable, "
+                               f"skip current recording, detailed infomation: {e}")
+            return grad
+        wrap_register_hook_single_element = partial(self.register_hook_single_element, hook_fn=hook_fn)
+        self.recursive_apply_transform(variable, wrap_register_hook_single_element)

msprobe/core/data_dump/data_processor/factory.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");
@@ -14,6 +14,7 @@
 # limitations under the License.
 
 from msprobe.core.common.const import Const
+from msprobe.core.data_dump.data_processor.base import BaseDataProcessor
 
 
 class DataProcessorFactory:
@@ -56,21 +57,25 @@ class DataProcessorFactory:
                 FreeBenchmarkDataProcessor as PytorchFreeBenchmarkDataProcessor,
                 KernelDumpDataProcessor as PytorchKernelDumpDataProcessor
             )
-            from msprobe.pytorch.module_processer import ModuleProcesser
+            from msprobe.pytorch.dump.module_dump.module_processer import ModuleProcesser
             cls.register_processor(Const.PT_FRAMEWORK, Const.STATISTICS, PytorchStatisticsDataProcessor)
             cls.register_processor(Const.PT_FRAMEWORK, Const.TENSOR, PytorchTensorDataProcessor)
             cls.register_processor(Const.PT_FRAMEWORK, Const.OVERFLOW_CHECK, PytorchOverflowCheckDataProcessor)
             cls.register_processor(Const.PT_FRAMEWORK, Const.FREE_BENCHMARK, PytorchFreeBenchmarkDataProcessor)
             cls.register_processor(Const.PT_FRAMEWORK, Const.KERNEL_DUMP, PytorchKernelDumpDataProcessor)
+            cls.register_processor(Const.PT_FRAMEWORK, Const.STRUCTURE, BaseDataProcessor)
             cls.register_module_processor(Const.PT_FRAMEWORK, ModuleProcesser)
         elif framework == Const.MS_FRAMEWORK:
             from msprobe.core.data_dump.data_processor.mindspore_processor import (
                 StatisticsDataProcessor as MindsporeStatisticsDataProcessor,
                 TensorDataProcessor as MindsporeTensorDataProcessor,
-                OverflowCheckDataProcessor as MindsporeOverflowCheckDataProcessor
+                OverflowCheckDataProcessor as MindsporeOverflowCheckDataProcessor,
+                KernelDumpDataProcessor as MindsporeKernelDumpDataProcessor
             )
             from msprobe.mindspore.cell_processor import CellProcessor
             cls.register_processor(Const.MS_FRAMEWORK, Const.STATISTICS, MindsporeStatisticsDataProcessor)
             cls.register_processor(Const.MS_FRAMEWORK, Const.TENSOR, MindsporeTensorDataProcessor)
             cls.register_processor(Const.MS_FRAMEWORK, Const.OVERFLOW_CHECK, MindsporeOverflowCheckDataProcessor)
+            cls.register_processor(Const.MS_FRAMEWORK, Const.KERNEL_DUMP, MindsporeKernelDumpDataProcessor)
+            cls.register_processor(Const.MS_FRAMEWORK, Const.STRUCTURE, BaseDataProcessor)
             cls.register_module_processor(Const.MS_FRAMEWORK, CellProcessor)

msprobe/core/data_dump/data_processor/mindspore_processor.py

@@ -1,4 +1,4 @@
-# Copyright 2024 Huawei Technologies Co., Ltd
+# Copyright 2024-2025 Huawei Technologies Co., Ltd
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -16,18 +16,24 @@
 import zlib
 
 import mindspore as ms
-from mindspore import mint, ops
+from mindspore import mint, ops, hal
 from mindspore._c_expression.typing import Number
 import numpy as np
 
 from msprobe.core.common.const import Const
 from msprobe.core.data_dump.data_processor.base import (BaseDataProcessor, TensorStatInfo,
                                                         ModuleForwardInputsOutputs, ModuleBackwardInputsOutputs)
-from msprobe.core.common.file_utils import path_len_exceeds_limit
+from msprobe.core.common.file_utils import path_len_exceeds_limit, save_npy
 from msprobe.mindspore.common.utils import convert_bf16_to_fp32, save_tensor_as_npy
 from msprobe.mindspore.common.log import logger
 from msprobe.mindspore.dump.hook_cell.api_registry import api_register
 
+has_adump = True
+try:
+    from msprobe.lib import _msprobe_c
+except ImportError:
+    has_adump = False
+
 
 class MindsporeDataProcessor(BaseDataProcessor):
     mindspore_special_type = tuple([ms.Tensor, Number])
@@ -37,11 +43,12 @@ class MindsporeDataProcessor(BaseDataProcessor):
         self.mindspore_object_key = {
             "dtype": self.analyze_dtype_in_kwargs
         }
+        self._async_dump_cache = {}
 
     @staticmethod
     def get_md5_for_tensor(x):
         x = convert_bf16_to_fp32(x)
-        tensor_bytes = x.contiguous().asnumpy().tobytes()
+        tensor_bytes = x.asnumpy().tobytes()
         crc32_hash = zlib.crc32(tensor_bytes)
         return f"{crc32_hash:08x}"
 
@@ -49,22 +56,17 @@ class MindsporeDataProcessor(BaseDataProcessor):
     def analyze_dtype_in_kwargs(element):
         return {"type": "mindspore.dtype", "value": str(element)}
 
-    @classmethod
-    def get_special_types(cls):
-        return super().get_special_types() + cls.mindspore_special_type
-
-    def get_stat_info(self, data):
+    @staticmethod
+    def get_stat_info_sync(data):
         tensor_stat = TensorStatInfo()
-        if data.numel() == 0:
-            return tensor_stat
-        elif data.dtype == ms.bool_:
-            data_np = data.contiguous().asnumpy()
+        if data.dtype == ms.bool_:
+            data_np = data.asnumpy()
             tensor_stat.max = np.max(data_np).item()
             tensor_stat.min = np.min(data_np).item()
         elif not data.shape:
             tensor_stat.max = tensor_stat.min = tensor_stat.mean = tensor_stat.norm = data.item()
         elif data.dtype == ms.complex64 or data.dtype == ms.complex128:
-            data_abs = np.abs(data.contiguous().asnumpy())
+            data_abs = np.abs(data.asnumpy())
             tensor_stat.max = np.max(data_abs).item()
             tensor_stat.min = np.min(data_abs).item()
             tensor_stat.mean = np.mean(data_abs).item()
@@ -87,17 +89,64 @@ class MindsporeDataProcessor(BaseDataProcessor):
             api_register.norm_inner_op_set_hook_func()
         return tensor_stat
 
+    @staticmethod
+    def get_stat_info_async(data):
+        tensor_stat = TensorStatInfo()
+        stack_method = api_register.functional_ori_attr.get("stack", ms.ops.stack)
+        if data.dtype == ms.complex64 or data.dtype == ms.complex128:
+            logger.warning("Async dump do not support complex data!")
+            return tensor_stat
+        elif data.dtype == ms.bool_:
+            tensor_stat.stack_tensor_stat = (["Max", "Min"], stack_method([data.any(), data.all()]))
+        elif not data.shape:
+            tensor_stat.stack_tensor_stat = (["Max", "Min", "Mean", "Norm"], stack_method([data, data, data, data]))
+        else:
+            if not ops.is_floating_point(data) or data.dtype == ms.float64:
+                data = data.to(ms.float32)
+            api_register.norm_inner_op_set_ori_func()
+            get_max_value = api_register.mint_ops_ori_attr.get("max", mint.max)
+            get_min_value = api_register.mint_ops_ori_attr.get("min", mint.min)
+            get_mean_value = api_register.mint_ops_ori_attr.get("mean", mint.mean)
+            if hasattr(mint, "norm"):
+                get_norm_value = api_register.mint_ops_ori_attr.get("norm", mint.norm)
+            else:
+                get_norm_value = api_register.functional_ori_attr.get("norm", ops.norm)
+            tensor_stat.stack_tensor_stat = (["Max", "Min", "Mean", "Norm"], stack_method(
+                [get_max_value(data), get_min_value(data), get_mean_value(data), get_norm_value(data)]))
+            api_register.norm_inner_op_set_hook_func()
+        return tensor_stat
+
+    @staticmethod
+    def is_hookable_element(element):
+        return hasattr(element, "register_hook") and callable(element.register_hook)
+
+    @classmethod
+    def get_special_types(cls):
+        return super().get_special_types() + cls.mindspore_special_type
+
+    def get_stat_info(self, data):
+        tensor_stat = TensorStatInfo()
+        if data.numel() == 0:
+            return tensor_stat
+        else:
+            if self.config.async_dump:
+                return MindsporeDataProcessor.get_stat_info_async(data)
+            else:
+                return MindsporeDataProcessor.get_stat_info_sync(data)
+
     def analyze_single_element(self, element, suffix_stack):
         if suffix_stack and suffix_stack[-1] in self.mindspore_object_key:
             return self.mindspore_object_key[suffix_stack[-1]](element)
 
         converted_numpy, numpy_type = self._convert_numpy_to_builtin(element)
         if converted_numpy is not element:
-            return self._analyze_numpy(converted_numpy, numpy_type)
+            return {"type": numpy_type, "value": converted_numpy}
         if isinstance(element, Number):
             return self.analyze_dtype_in_kwargs(element)
         if isinstance(element, ms.Tensor):
-            return self._analyze_tensor(element, Const.SEP.join(suffix_stack))
+            return self._analyze_tensor(element, Const.SEP.join([str(suffix) for suffix in suffix_stack]))
+        if isinstance(element, np.ndarray):
+            return self._analyze_numpy(element, Const.SEP.join([str(suffix) for suffix in suffix_stack]))
         if isinstance(element, (bool, int, float, str, slice, type(Ellipsis))):
             return self._analyze_builtin(element)
         return {}
@@ -107,13 +156,17 @@ class MindsporeDataProcessor(BaseDataProcessor):
         tensor_json = {
             'type': 'mindspore.Tensor',
             'dtype': str(tensor.dtype),
-            'shape': tensor.shape,
-            'Max': self.transfer_type(tensor_stat.max),
-            'Min': self.transfer_type(tensor_stat.min),
-            'Mean': self.transfer_type(tensor_stat.mean),
-            'Norm': self.transfer_type(tensor_stat.norm),
+            'shape': tensor.shape
         }
-        if self.config.summary_mode == Const.MD5:
+
+        if tensor_stat.stack_tensor_stat is None:
+            tensor_json.update({'Max': self.transfer_type(tensor_stat.max)})
+            tensor_json.update({'Min': self.transfer_type(tensor_stat.min)})
+            tensor_json.update({'Mean': self.transfer_type(tensor_stat.mean)})
+            tensor_json.update({'Norm': self.transfer_type(tensor_stat.norm)})
+        else:
+            tensor_json.update({'tensor_stat': tensor_stat.stack_tensor_stat})
+        if self.config.summary_mode == Const.MD5 and not self.config.async_dump:
             tensor_md5 = self.get_md5_for_tensor(tensor)
             tensor_json.update({Const.MD5: tensor_md5})
         return tensor_json
@@ -124,12 +177,27 @@ class StatisticsDataProcessor(MindsporeDataProcessor):
 
 
 class TensorDataProcessor(MindsporeDataProcessor):
+    def dump_async_data(self):
+        for file_path, tensor in self._async_dump_cache.items():
+            save_tensor_as_npy(tensor, file_path)
+        self._async_dump_cache.clear()
+
     def _analyze_tensor(self, tensor, suffix):
         dump_data_name, file_path = self.get_save_file_path(suffix)
         single_arg = super()._analyze_tensor(tensor, suffix)
         single_arg.update({"data_name": dump_data_name})
-        save_tensor_as_npy(tensor, file_path)
+        if self.config.async_dump:
+            self._async_dump_cache[file_path] = tensor.copy()
+        else:
+            save_tensor_as_npy(tensor, file_path)
         return single_arg
+    
+    def _analyze_numpy(self, ndarray, suffix):
+        dump_data_name, file_path = self.get_save_file_path(suffix)
+        save_npy(ndarray, file_path)
+        ndarray_json = super()._analyze_numpy(ndarray, suffix)
+        ndarray_json.update({"data_name": dump_data_name})
+        return ndarray_json
 
 
 class OverflowCheckDataProcessor(MindsporeDataProcessor):
@@ -138,6 +206,7 @@ class OverflowCheckDataProcessor(MindsporeDataProcessor):
     def __init__(self, config, data_writer):
         super().__init__(config, data_writer)
         self.has_overflow = False
+        self.cached_api_info = {}
         self.cached_tensors_and_file_paths = {}
         self.real_overflow_nums = 0
         self.overflow_nums = config.overflow_nums
@@ -150,6 +219,20 @@ class OverflowCheckDataProcessor(MindsporeDataProcessor):
             return True
         return False
 
+    def analyze_forward_input(self, name, module, module_input_output: ModuleForwardInputsOutputs):
+        self.has_overflow = False
+        self.cached_api_info = super().analyze_forward_input(name, module, module_input_output)
+        return None
+
+    def analyze_forward_output(self, name, module, module_input_output: ModuleForwardInputsOutputs):
+        api_info_struct = super().analyze_forward_output(name, module, module_input_output)
+        if name in self.cached_api_info and name in api_info_struct:
+            self.cached_api_info[name].update(api_info_struct[name])
+        elif name in api_info_struct:
+            self.cached_api_info = api_info_struct
+        self.maybe_save_overflow_data()
+        return self.cached_api_info if self.has_overflow else None
+
     def analyze_forward(self, name, module, module_input_output: ModuleForwardInputsOutputs):
         self.has_overflow = False
         api_info_struct = super().analyze_forward(name, module, module_input_output)
@@ -162,6 +245,12 @@ class OverflowCheckDataProcessor(MindsporeDataProcessor):
         self.maybe_save_overflow_data()
         return api_info_struct if self.has_overflow else None
 
+    def analyze_params(self, name, param_name, grad):
+        self.has_overflow = False
+        api_info_struct = super().analyze_params(name, param_name, grad)
+        self.maybe_save_overflow_data()
+        return api_info_struct if self.has_overflow else None
+
     def maybe_save_overflow_data(self):
         if self.has_overflow:
             for file_path, tensor in self.cached_tensors_and_file_paths.items():
@@ -190,3 +279,61 @@ class OverflowCheckDataProcessor(MindsporeDataProcessor):
         self._analyze_maybe_overflow_tensor(single_arg)
         single_arg.update({"data_name": dump_data_name})
         return single_arg
+
+
+class KernelDumpDataProcessor(MindsporeDataProcessor):
+    def __init__(self, config, data_writer):
+        super().__init__(config, data_writer)
+        self.enable_kernel_dump = True
+
+    @staticmethod
+    def start_kernel_dump(config_path):
+        hal.synchronize()
+        _msprobe_c.init_dump()
+        _msprobe_c.set_dump(config_path)
+        hal.synchronize()
+
+    @staticmethod
+    def stop_kernel_dump():
+        hal.synchronize()
+        _msprobe_c.finalize_dump()
+        hal.synchronize()
+
+    @staticmethod
+    def _print_unsupported_log(api_name):
+        logger.warning(f"The kernel dump does not support the {api_name} API.")
+
+    def analyze_forward_input(self, name, module, module_input_output):
+        if not self.enable_kernel_dump:
+            return
+        if not has_adump:
+            logger.warning("The current msprobe package does not compile adump, and kernel dump cannot be used.")
+            self.enable_kernel_dump = False
+            return
+        self.start_kernel_dump(self.config.kernel_config_path)
+
+    def analyze_forward_output(self, name, module, module_input_output):
+        if not self.enable_kernel_dump:
+            return
+        self.enable_kernel_dump = False
+        self.stop_kernel_dump()
+        logger.info(f"The kernel data of {name} is dumped successfully.")
+
+    def analyze_backward_input(self, name, module, module_input_output):
+        if not self.enable_kernel_dump:
+            return
+        if not has_adump:
+            logger.warning("The current msprobe package does not compile adump, and kernel dump cannot be used.")
+            self.enable_kernel_dump = False
+            return
+        self.start_kernel_dump(self.config.kernel_config_path)
+
+    def analyze_backward(self, name, module, module_input_output):
+        if not self.enable_kernel_dump:
+            return
+        self.enable_kernel_dump = False
+        self.stop_kernel_dump()
+        logger.info(f"The kernel data of {name} is dumped successfully.")
+
+    def reset_status(self):
+        self.enable_kernel_dump = True