mindstudio-probe 1.0.4__py3-none-any.whl → 1.1.0__py3-none-any.whl

msprobe/core/data_dump/data_processor/factory.py

@@ -1,3 +1,18 @@
+# Copyright (c) 2024-2024, Huawei Technologies Co., Ltd.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0  (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from msprobe.core.common.const import Const
 
 
@@ -34,14 +49,14 @@ class DataProcessorFactory:
     @classmethod
     def register_processors(cls, framework):
         if framework == Const.PT_FRAMEWORK:
-            from .pytorch_processor import (
+            from msprobe.core.data_dump.data_processor.pytorch_processor import (
                 StatisticsDataProcessor as PytorchStatisticsDataProcessor,
                 TensorDataProcessor as PytorchTensorDataProcessor,
                 OverflowCheckDataProcessor as PytorchOverflowCheckDataProcessor,
                 FreeBenchmarkDataProcessor as PytorchFreeBenchmarkDataProcessor,
                 KernelDumpDataProcessor as PytorchKernelDumpDataProcessor
             )
-            from ....pytorch.module_processer import ModuleProcesser
+            from msprobe.pytorch.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)
@@ -49,11 +64,13 @@ class DataProcessorFactory:
             cls.register_processor(Const.PT_FRAMEWORK, Const.KERNEL_DUMP, PytorchKernelDumpDataProcessor)
             cls.register_module_processor(Const.PT_FRAMEWORK, ModuleProcesser)
         elif framework == Const.MS_FRAMEWORK:
-            from .mindspore_processor import (
+            from msprobe.core.data_dump.data_processor.mindspore_processor import (
                 StatisticsDataProcessor as MindsporeStatisticsDataProcessor,
                 TensorDataProcessor as MindsporeTensorDataProcessor,
                 OverflowCheckDataProcessor as MindsporeOverflowCheckDataProcessor
             )
+            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_module_processor(Const.MS_FRAMEWORK, CellProcessor)

msprobe/core/data_dump/data_processor/mindspore_processor.py

@@ -17,6 +17,7 @@ import zlib
 
 import mindspore as ms
 from mindspore import mint, ops
+from mindspore._c_expression.typing import Number
 import numpy as np
 
 from msprobe.core.common.const import Const
@@ -29,7 +30,7 @@ from msprobe.mindspore.dump.hook_cell.api_registry import api_register
 
 
 class MindsporeDataProcessor(BaseDataProcessor):
-    mindspore_special_type = tuple([ms.Tensor])
+    mindspore_special_type = tuple([ms.Tensor, Number])
 
     def __init__(self, config, data_writer):
         super().__init__(config, data_writer)
@@ -69,13 +70,16 @@ class MindsporeDataProcessor(BaseDataProcessor):
             tensor_stat.mean = np.mean(data_abs).item()
             tensor_stat.norm = np.linalg.norm(data_abs).item()
         else:
-            if data.dtype == ms.bfloat16 or not ops.is_floating_point(data):
+            if not ops.is_floating_point(data):
                 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)
-            get_norm_value = api_register.functional_ori_attr.get("norm", ops.norm)
+            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.max = get_max_value(data).item()
             tensor_stat.min = get_min_value(data).item()
             tensor_stat.mean = get_mean_value(data).item()
@@ -90,9 +94,10 @@ class MindsporeDataProcessor(BaseDataProcessor):
         converted_numpy, numpy_type = self._convert_numpy_to_builtin(element)
         if converted_numpy is not element:
             return self._analyze_numpy(converted_numpy, numpy_type)
+        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))
-
         if isinstance(element, (bool, int, float, str, slice, type(Ellipsis))):
             return self._analyze_builtin(element)
         return {}
@@ -163,7 +168,8 @@ class OverflowCheckDataProcessor(MindsporeDataProcessor):
                 save_tensor_as_npy(tensor, file_path)
             self.real_overflow_nums += 1
             if self.overflow_nums != -1 and self.real_overflow_nums >= self.overflow_nums:
-                logger.info(f"[{Const.TOOL_NAME}] 超过预设溢出次数 当前溢出次数: {self.real_overflow_nums}")
+                logger.info(f"[{Const.TOOL_NAME}] Reached the preset overflow times, "
+                            f"current overflow times: {self.real_overflow_nums}.")
         self.cached_tensors_and_file_paths = {}
 
     def _analyze_maybe_overflow_tensor(self, tensor_json):

msprobe/core/data_dump/data_processor/pytorch_processor.py

@@ -1,20 +1,35 @@
+# Copyright (c) 2024-2024, Huawei Technologies Co., Ltd.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0  (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import zlib
 from dataclasses import asdict
 from typing import List
 
 import numpy as np
 import torch
-from msprobe.core.common.file_utils import path_len_exceeds_limit, change_mode
+from msprobe.core.common.const import Const
+from msprobe.core.common.file_utils import path_len_exceeds_limit
 from msprobe.core.common.log import logger
-from msprobe.core.common.const import Const, OverflowConst, FileCheckConst
 from msprobe.core.data_dump.data_processor.base import BaseDataProcessor, ModuleBackwardInputsOutputs, \
     ModuleForwardInputsOutputs, TensorStatInfo
-from msprobe.pytorch.free_benchmark import FreeBenchmarkCheck, UnequalRow
 from msprobe.pytorch.common.utils import save_pt, load_pt
+from msprobe.pytorch.free_benchmark import FreeBenchmarkCheck, UnequalRow
 
+is_gpu = False
 try:
     import torch_npu
-    is_gpu = False
 except ImportError:
     is_gpu = True
 
@@ -153,7 +168,7 @@ class StatisticsDataProcessor(PytorchDataProcessor):
 class TensorDataProcessor(PytorchDataProcessor):
     def _analyze_tensor(self, tensor, suffix):
         dump_data_name, file_path = self.get_save_file_path(suffix)
-        saved_tensor = tensor.contiguous().detach()
+        saved_tensor = tensor.clone().contiguous().detach()
         save_pt(saved_tensor, file_path)
         single_arg = super()._analyze_tensor(tensor, suffix)
         single_arg.update({"data_name": dump_data_name})
@@ -178,7 +193,6 @@ class OverflowCheckDataProcessor(PytorchDataProcessor):
         if self.overflow_nums == -1:
             return False
         if self.real_overflow_nums >= self.overflow_nums:
-            logger.info(f"[msprobe] 超过预设溢出次数 当前溢出次数: {self.real_overflow_nums}")
             return True
         return False
 
@@ -219,6 +233,9 @@ class OverflowCheckDataProcessor(PytorchDataProcessor):
             for file_path, tensor in self.cached_tensors_and_file_paths.items():
                 save_pt(tensor, file_path)
             self.real_overflow_nums += 1
+            if self.overflow_nums != -1 and self.real_overflow_nums >= self.overflow_nums:
+                logger.info(f"[{Const.TOOL_NAME}] Reached the preset overflow times, "
+                            f"current overflow times: {self.real_overflow_nums}.")
         self.cached_tensors_and_file_paths = {}
 
     def _is_support_inf_nan(self):
@@ -243,7 +260,7 @@ class OverflowCheckDataProcessor(PytorchDataProcessor):
         if tensor_json['Max'] is None or tensor_json['Min'] is None:
             return
         self.has_overflow = np.isinf(tensor_json['Max']) or np.isnan(tensor_json['Max']) or \
-            np.isinf(tensor_json['Min']) or np.isnan(tensor_json['Min'])
+                            np.isinf(tensor_json['Min']) or np.isnan(tensor_json['Min'])
 
     def _analyze_tensor(self, tensor, suffix):
         dump_data_name, file_path = self.get_save_file_path(suffix)

msprobe/core/data_dump/json_writer.py

@@ -1,24 +1,36 @@
-import os
+# Copyright (c) 2024-2024, Huawei Technologies Co., Ltd.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0  (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import csv
+import os
 
-from msprobe.core.common.file_utils import change_mode, FileOpen
-from msprobe.core.common.log import logger
 from msprobe.core.common.const import Const, FileCheckConst
-from msprobe.core.common.file_utils import remove_path, load_json, save_json
+from msprobe.core.common.file_utils import change_mode, FileOpen, save_json
+from msprobe.core.common.log import logger
 
 
 class DataWriter:
 
-    def __init__(self, init_json=None) -> None:
-        self.dump_count = 0
-        self.init_json = init_json
-        self.dump_file_path = None  # os.path.join(dump_dir, DataWriter.dump_json_name)
-        self.stack_file_path = None  # os.path.join(dump_dir, DataWriter.stack_json_name)
-        self.construct_file_path = None  # os.path.join(dump_dir, DataWriter.construct_json_name)
+    def __init__(self) -> None:
+        self.dump_file_path = None
+        self.stack_file_path = None
+        self.construct_file_path = None
         self.free_benchmark_file_path = None
         self.dump_tensor_data_dir = None
-        self.buffer_size = 1000
-        self.cache_data = {Const.DATA: {}}
+        self.flush_size = 1000
+        self.cache_data = {}
         self.cache_stack = {}
         self.cache_construct = {}
 
@@ -37,18 +49,22 @@ class DataWriter:
         if is_new_file:
             change_mode(file_path, FileCheckConst.DATA_FILE_AUTHORITY)
 
-    def initialize_json_file(self, **kwargs):
-        kwargs.update({"dump_data_dir": self.dump_tensor_data_dir, Const.DATA: {}})
-        save_json(self.dump_file_path, kwargs)
-
-        empty_dict = {}
-        remove_path(self.stack_file_path)
-        save_json(self.stack_file_path, empty_dict)
-
-        remove_path(self.construct_file_path)
-        save_json(self.construct_file_path, empty_dict)
+    def reset_cache(self):
+        self.cache_data = {}
+        self.cache_stack = {}
+        self.cache_construct = {}
 
-    def update_dump_paths(self, dump_file_path, stack_file_path, construct_file_path, dump_data_dir, 
+    def initialize_json_file(self, **kwargs):
+        if not self.cache_data:
+            kwargs.update({"dump_data_dir": self.dump_tensor_data_dir, Const.DATA: {}})
+            self.cache_data = kwargs
+            save_json(self.dump_file_path, self.cache_data, indent=1)
+        if not self.cache_stack:
+            save_json(self.stack_file_path, self.cache_stack, indent=1)
+        if not self.cache_construct:
+            save_json(self.construct_file_path, self.cache_construct, indent=1)
+
+    def update_dump_paths(self, dump_file_path, stack_file_path, construct_file_path, dump_data_dir,
                           free_benchmark_file_path):
         self.dump_file_path = dump_file_path
         self.stack_file_path = stack_file_path
@@ -56,16 +72,25 @@ class DataWriter:
         self.dump_tensor_data_dir = dump_data_dir
         self.free_benchmark_file_path = free_benchmark_file_path
 
+    def flush_data_periodically(self):
+        dump_data = self.cache_data.get(Const.DATA)
+        if dump_data and isinstance(dump_data, dict) and len(dump_data) % self.flush_size == 0:
+            self.write_json()
+
     def update_data(self, new_data):
-        key = next(iter(new_data.keys()))  # assert len(new_data.keys()) == 1
-        if key in self.cache_data[Const.DATA]:
-            self.cache_data[Const.DATA][key].update(new_data[key])
-        else:
-            self.cache_data[Const.DATA].update(new_data)
+        if not isinstance(new_data, dict) or len(new_data.keys()) != 1:
+            logger.warning(f"The data info({new_data}) should be a dict with only one outer key.")
+            return
+        dump_data = self.cache_data.get(Const.DATA)
+        if not isinstance(dump_data, dict):
+            logger.warning(f"The dump data({dump_data}) should be a dict.")
+            return
 
-    def flush_data_when_buffer_is_full(self):
-        if len(self.cache_data[Const.DATA]) >= self.buffer_size:
-            self.write_data_json(self.dump_file_path)
+        key = next(iter(new_data.keys()))
+        if key in dump_data:
+            dump_data.get(key).update(new_data.get(key))
+        else:
+            dump_data.update(new_data)
 
     def update_stack(self, new_data):
         self.cache_stack.update(new_data)
@@ -75,14 +100,7 @@ class DataWriter:
 
     def write_data_json(self, file_path):
         logger.info(f"dump.json is at {os.path.dirname(os.path.dirname(file_path))}. ")
-        if os.path.exists(file_path) and os.path.getsize(file_path) > 0:
-            data_to_write = load_json(file_path)
-        else:
-            self.init_json['data_path'] = self.dump_tensor_data_dir
-            data_to_write = self.init_json
-        data_to_write[Const.DATA].update(self.cache_data[Const.DATA])
-        save_json(file_path, data_to_write, indent=1)
-        self.cache_data[Const.DATA].clear()
+        save_json(file_path, self.cache_data, indent=1)
 
     def write_stack_info_json(self, file_path):
         save_json(file_path, self.cache_stack, indent=1)
@@ -91,6 +109,9 @@ class DataWriter:
         save_json(file_path, self.cache_construct, indent=1)
 
     def write_json(self):
-        self.write_data_json(self.dump_file_path)
-        self.write_stack_info_json(self.stack_file_path)
-        self.write_construct_info_json(self.construct_file_path)
+        if self.cache_data:
+            self.write_data_json(self.dump_file_path)
+        if self.cache_stack:
+            self.write_stack_info_json(self.stack_file_path)
+        if self.cache_construct:
+            self.write_construct_info_json(self.construct_file_path)

msprobe/core/data_dump/scope.py

@@ -1,6 +1,22 @@
+# Copyright (c) 2024-2024, Huawei Technologies Co., Ltd.
+# All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0  (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 from abc import ABC, abstractmethod
-from msprobe.core.common.exceptions import ScopeException
+
 from msprobe.core.common.const import Const
+from msprobe.core.common.exceptions import ScopeException
 
 
 def build_scope(scope_class, scope=None, api_list=None):
@@ -33,6 +49,7 @@ def build_range_scope_according_to_scope_name(scope, api_list):
 class BaseScope(ABC):
     Module_Type_Module = "Module"
     Module_Type_API = "api"
+    module_type = ["Module", "Cell"]
 
     def __init__(self, scope, api_list):
         scope, api_list = self.rectify_args(scope, api_list)
@@ -81,9 +98,9 @@ class ListScope(BaseScope):
                 f"scope和api_list不可以同时配置，实际配置为scope={scope}, api_list={api_list}.")
         return super(ListScope, ListScope).rectify_args(scope, api_list)
 
-    def check(self, module_name):
-        if not self.scope or module_name in self.scope:
-            return self.check_api_list(module_name)
+    def check(self, name):
+        if not self.scope or name in self.scope:
+            return self.check_api_list(name)
         return False
 
 
@@ -94,7 +111,6 @@ class RangeScope(BaseScope, ABC):
         self.in_scope = False
         self.is_valid = self.check_scope_is_valid()
 
-
     @staticmethod
     def rectify_args(scope, api_list):
         scope, api_list = super(RangeScope, RangeScope).rectify_args(scope, api_list)
@@ -104,7 +120,6 @@ class RangeScope(BaseScope, ABC):
             elif len(scope) > 2:
                 raise ScopeException(ScopeException.InvalidScope,
                     f"scope参数指定区间断点，须传入长度为1或2的列表，实际长度为{len(scope)}.")
-
         return scope, api_list
 
     @abstractmethod
@@ -123,23 +138,23 @@ class APIRangeScope(RangeScope):
         if not self.scope:
             return True
         scope_start_type = self.scope[0].split(Const.SEP)[0]
-        if scope_start_type == BaseScope.Module_Type_Module:
+        if scope_start_type in BaseScope.module_type:
             return False
         scope_stop_type = self.scope[1].split(Const.SEP)[0]
-        if scope_stop_type == BaseScope.Module_Type_Module:
+        if scope_stop_type in BaseScope.module_type:
             return False
         return True
 
-    def check(self, api_name):
-        if self.scope and api_name == self.scope[0]:
+    def check(self, name):
+        if self.scope and name == self.scope[0]:
             self.in_scope = True
 
         if not self.scope or self.in_scope:
-            result = self.check_api_list(api_name)
+            result = self.check_api_list(name)
         else:
             result = False
 
-        if self.scope and api_name == self.scope[1]:
+        if self.scope and name == self.scope[1]:
             self.in_scope = False
         return result
 
@@ -150,13 +165,14 @@ class ModuleRangeScope(RangeScope):
         需要用pre_hook和full_backward_hook来精确控制module的开始和结束，
         在这些hook触发时调用begin_module和end_module做区间控制
     """
+
     def check_scope_is_valid(self):
         if not self.scope:
             return True
         scope_start_type = self.scope[0].split(Const.SEP)[0]
         scope_stop_type = self.scope[1].split(Const.SEP)[0]
-        if scope_start_type == BaseScope.Module_Type_Module and \
-                scope_stop_type == BaseScope.Module_Type_Module:
+        if scope_start_type in BaseScope.module_type and \
+                scope_stop_type in BaseScope.module_type:
             return True
         return False
 
@@ -172,7 +188,7 @@ class ModuleRangeScope(RangeScope):
         if module_name == self.scope[1]:
             self.in_scope = False
 
-    def check(self, module_name):
+    def check(self, name):
         if not self.scope or self.in_scope:
-            return self.check_api_list(module_name)
+            return self.check_api_list(name)
         return False

msprobe/core/grad_probe/constant.py

@@ -33,6 +33,10 @@ class GradConst:
     # direction suffix
     DIR_SUFFIX = "dir.npy"
 
+    # bounds safety
+    BOUNDS_MINIMUM = -2**63
+    BOUNDS_MAXIMUM = 2**63 - 1
+
     # file safty
     DATA_DIR_AUTHORITY = 0o750
     DATA_FILE_AUTHORITY = 0o640

msprobe/core/grad_probe/grad_compare.py

@@ -2,12 +2,11 @@ import os
 from typing import List
 
 from tqdm import tqdm
-import pandas as pd
 import matplotlib.pyplot as plt
 
 from msprobe.core.common.file_utils import create_directory, check_path_before_create, check_file_or_directory_path
 from msprobe.core.common.log import logger
-from msprobe.core.common.file_utils import remove_path, load_npy, write_csv
+from msprobe.core.common.file_utils import remove_path, load_npy, write_csv, read_csv
 from msprobe.core.grad_probe.constant import GradConst
 from msprobe.core.grad_probe.utils import plt_savefig
 
@@ -21,7 +20,7 @@ class GradComparator:
                 continue
             if not os.path.exists(os.path.join(path2, summary_file)):
                 continue
-            summary_csv = pd.read_csv(os.path.join(path1, summary_file))
+            summary_csv = read_csv(os.path.join(path1, summary_file))
             return summary_csv["param_name"]
         raise RuntimeError("no matched grad_summary.csv for comparison, please dump data in same configuration")
     

msprobe/core/grad_probe/utils.py

@@ -20,12 +20,25 @@ def check_numeral_list_ascend(lst):
 def check_param(param_name):
     if not re.match(GradConst.PARAM_VALID_PATTERN, param_name):
         raise RuntimeError("The parameter name contains special characters.")
-    
+
 
 def check_str(string, variable_name):
     if not isinstance(string, str):
         raise ValueError(f'The variable: "{variable_name}" is not a string.')
-    
+
+def check_bounds_element(bound):
+    return GradConst.BOUNDS_MINIMUM <= bound and bound <= GradConst.BOUNDS_MAXIMUM
+
+def check_bounds(bounds):
+    prev = GradConst.BOUNDS_MINIMUM - 1
+    for element in bounds:
+        if not isinstance(element, (int, float)):
+            raise Exception("bounds element is not int or float")
+        if not check_bounds_element(element):
+            raise Exception("bounds element is out of int64 range")
+        if prev >= element:
+            raise Exception("bounds list is not ascending")
+        prev = element
 
 class ListCache(list):
     threshold = 1000
@@ -50,7 +63,7 @@ class ListCache(list):
         list.append(self, data)
         if len(self) >= ListCache.threshold:
             self.flush()
-    
+
     def set_output_file(self, output_file):
         self._output_file = output_file
 

msprobe/docs/01.installation.md

@@ -3,19 +3,20 @@
 推荐使用 [miniconda](https://docs.anaconda.com/miniconda/) 管理环境依赖。
 
 ```bash
-conda create -n msprobe python=3.8
+conda create -n msprobe python
 conda activate msprobe
 ```
 
-## 1. 从 PyPI 安装
+## 1 从 PyPI 安装
 ```shell
-pip install mindstudio-probe[==版本号]
+pip install mindstudio-probe
 ```
 
-## 2. 下载 whl 包安装
+## 2 下载 whl 包安装
 
 |版本|发布日期|支持 PyTorch 版本|支持 MindSpore 版本|下载链接|校验码|
 |:--:|:--:|:--:|:--:|:--:|:--:|
+|1.0.4|2024.09.09|1.11/2.0/2.1/2.2|2.4.0|[mindstudio_probe-1.0.4-py3-none-any.whl](https://ptdbg.obs.myhuaweicloud.com/msprobe/1.0/mindstudio_probe-1.0.4-py3-none-any.whl)|4e1909566a71a855b356597750c20ee43d964a22b2c2b02ac08312a5def75fd6|
 | 1.0.3 | 2024.08.23 | 1.11/2.0/2.1/2.2 | 2.4.0 | [mindstudio_probe-1.0.3-py3-none-any.whl](https://ptdbg.obs.myhuaweicloud.com/msprobe/1.0/mindstudio_probe-1.0.3-py3-none-any.whl) | 7060cc141a5b98ef770cd9220995d299393f32a61938261e632c7e8b5160bef2 |
 | 1.0.2 | 2024.08.09 | 1.11/2.0/2.1/2.2 | 2.4.0 | [mindstudio_probe-1.0.2-py3-none-any.whl](https://ptdbg.obs.myhuaweicloud.com/msprobe/1.0/mindstudio_probe-1.0.2-py3-none-any.whl) | e4a980e5d98c426ce5ce9842520d9bc031d3b3de621c74b3d59414cc6e238e0e |
 | 1.0.1 | 2024.07.25 | 2.0/2.1/2.2 | 2.4.0 | [mindstudio_probe-1.0.1-py3-none-any.whl](https://ptdbg.obs.myhuaweicloud.com/msprobe/1.0/mindstudio_probe-1.0.1-py3-none-any.whl) | b699e224e4d4e3bcf9412c54fa858a1ee370f0d7a2bc69cb3f1273ac14a6dc82 |
@@ -31,7 +32,7 @@ pip install ./mindstudio_probe-{version}-py3-none-any.whl # 安装whl包
 
 若覆盖安装，请在命令行末尾添加 `--force-reinstall` 参数。
 
-## 3. 从源码安装
+## 3 从源码安装
 
 ```shell
 git clone https://gitee.com/ascend/mstt.git
@@ -40,9 +41,18 @@ cd mstt/debug/accuracy_tools
 pip install setuptools wheel
 
 python setup.py bdist_wheel
-pip install ./dist/mindstudio_probe*.whl
+cd ./dist
+pip install ./mindstudio_probe*.whl
 ```
 
+# 历史版本特性
+
+<table>
+    <tr><th>版本</th><th>特性</th></tr>
+    <tr><td rowspan="2">1.0.3</td><td>【精度预检】</br>1. 落盘数据小；</br>2. 支持随机生成模式和真实数据模式；</br>3. 单 API 测试，排除整网中的累计误差问题。</td></tr>
+    <tr><td>【梯度检测】</br>1. 使用便捷，无需在训练流程里插入代码。</br>2. 可以精准定位问题出现的 step。</td></tr>
+</table>
+
 # 查看 msprobe 工具信息
 
 ```bash
@@ -59,7 +69,7 @@ Home-page: https://gitee.com/ascend/mstt/tree/master/debug/accuracy_tools/msprob
 Author: Ascend Team
 Author-email: pmail_mindstudio@huawei.com
 License: Apache License 2.0
-Location: /home/xxx/miniconda3/envs/xxx/lib/python3.8/site-packages/mindstudio_probe-1.0.0-py3.8.egg
+Location: /home/xxx/miniconda3/envs/xxx/lib/python3.x/site-packages/mindstudio_probe-1.0.x-py3.x.egg
 Requires: einops, matplotlib, numpy, openpyxl, pandas, pyOpenSSL, pyyaml, rich, tqdm, twisted, wheel
 Required-by: 
 ```
@@ -68,11 +78,11 @@ Required-by:
 
 ## 1 安装 CANN 包
 
-1.1 根据 CPU 架构和 NPU 型号选择 toolkit 和 kernal 包，可以参考 [CANN 软件安装指南](https://gitee.com/link?target=https%3A%2F%2Fwww.hiascend.com%2Fdocument%2Fdetail%2Fzh%2Fcanncommercial%2F700%2Fenvdeployment%2Finstg%2Finstg_0001.html)和[昇腾社区](https://www.hiascend.cn/developer/download/community/result?module=cann)。
+1.1 根据 CPU 架构和 NPU 型号选择 toolkit 和 kernel，可以参考 [CANN 软件安装指南](https://gitee.com/link?target=https%3A%2F%2Fwww.hiascend.com%2Fdocument%2Fdetail%2Fzh%2Fcanncommercial%2F700%2Fenvdeployment%2Finstg%2Finstg_0001.html)和[昇腾社区](https://www.hiascend.cn/developer/download/community/result?module=cann)。
 
 1.2 运行示例
 ```bash
-Ascend-cann-toolkit_x.x.x_linux-aarch64.run --full --install-path={cann_path}
+Ascend-cann-toolkit_x.x.x_linux-xxxx.run --full --install-path={cann_path}
 Ascend-cann-kernels_x.x.x_linux.run --install --install-path={cann_path}
 ```