mindstudio-probe 8.2.1__tar.gz → 8.3.3__tar.gz

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mindstudio-probe
-Version: 8.2.1
+Version: 8.3.3
 Summary: Ascend Probe Utils
 Home-page: https://gitcode.com/Ascend/mstt/tree/master/debug/accuracy_tools/msprobe
 Author: Ascend Team

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/mindstudio_probe.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mindstudio-probe
-Version: 8.2.1
+Version: 8.3.3
 Summary: Ascend Probe Utils
 Home-page: https://gitcode.com/Ascend/mstt/tree/master/debug/accuracy_tools/msprobe
 Author: Ascend Team

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/mindstudio_probe.egg-info/SOURCES.txt

@@ -46,7 +46,6 @@ msprobe/core/compare/utils.py
 msprobe/core/compare/diff_analyze/__init__.py
 msprobe/core/compare/diff_analyze/diff_analyze_threshold.yaml
 msprobe/core/compare/diff_analyze/first_diff_analyze.py
-msprobe/core/compare/diff_analyze/ignore_op_list.yaml
 msprobe/core/compare/find_first/__init__.py
 msprobe/core/compare/find_first/analyzer.py
 msprobe/core/compare/find_first/data_processor.py

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/msprobe/README.md

@@ -50,9 +50,14 @@ export MSPROBE_LOG_LEVEL={x}
 
 ## 🚨 工具限制与注意事项
 
-**1. Pytorch 框架下，工具暂不支持 Fully Sharded Data Parallel(FSDP)。**
+**1. 工具读写的所有路径，如config_path、dump_path等，只允许包含大小写字母、数字、下划线、斜杠、点和短横线。**
+
+**2. 出于安全性及权限最小化角度考虑，msprobe工具不应使用root等高权限账户使用，建议使用普通用户权限安装执行。**
+
+**3. 使用msprobe工具前请确保执行用户的umask值大于等于0027，否则可能会导致工具生成的精度数据文件和目录权限过大。**
+
+**4. 用户须自行保证使用最小权限原则，如给工具输入的文件要求other用户不可写，在一些对安全要求更严格的功能场景下还需确保输入的文件group用户不可写。**
 
-**2. 工具读写的所有路径，如config_path、dump_path等，只允许包含大小写字母、数字、下划线、斜杠、点和短横线。**
 
 ## ⚙️ 安装
 

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/msprobe/core/common/const.py

@@ -691,9 +691,23 @@ class FileCheckConst:
     """
     Class for file check const
     """
-    READ_ABLE = "read"
-    WRITE_ABLE = "write"
-    READ_WRITE_ABLE = "read and write"
+    READ_ABLE = "r"
+    WRITE_ABLE = "w"
+    EXECUTE_ABLE = "x"
+    READ_WRITE_ABLE = "rw"
+    READ_EXECUTE_ABLE = "rx"
+    WRITE_EXECUTE_ABLE = "wx"
+    READ_WRITE_EXECUTE_ABLE = "rwx"
+    PERM_OPTIONS = [
+        READ_ABLE,
+        WRITE_ABLE,
+        EXECUTE_ABLE,
+        READ_WRITE_ABLE,
+        READ_EXECUTE_ABLE,
+        WRITE_EXECUTE_ABLE,
+        READ_WRITE_EXECUTE_ABLE
+    ]
+
     DIRECTORY_LENGTH = 4096
     FILE_NAME_LENGTH = 255
     FILE_VALID_PATTERN = r"^[a-zA-Z0-9_.:/-]+$"

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/msprobe/core/common/file_utils.py

@@ -12,29 +12,31 @@
 # 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 atexit
 import csv
 import fcntl
 import io
+import json
+import multiprocessing
 import os
 import pickle
-from multiprocessing import shared_memory
-import stat
-import json
 import re
 import shutil
+import stat
 import sys
 import zipfile
-import multiprocessing
-import yaml
+from multiprocessing import shared_memory
+
 import numpy as np
 import pandas as pd
+import yaml
 
+from msprobe.core.common.const import FileCheckConst, CompareConst, Const
 from msprobe.core.common.decorator import recursion_depth_decorator
-from msprobe.core.common.log import logger
 from msprobe.core.common.exceptions import FileCheckException
-from msprobe.core.common.const import FileCheckConst, CompareConst, Const
 from msprobe.core.common.global_lock import global_lock, is_main_process
+from msprobe.core.common.log import logger
 
 proc_lock = multiprocessing.Lock()
 
@@ -44,18 +46,23 @@ class FileChecker:
     The class for check file.
 
     Attributes:
-        file_path: The file or dictionary path to be verified.
-        path_type: file or dictionary
-        ability(str): FileCheckConst.WRITE_ABLE or FileCheckConst.READ_ABLE to set file has writability or readability
+        file_path: The file or dictionary path to be verified
+        path_type: file or dictionary type
+        ability(str): one of ["r", "w", "x", "rw", "rx", "wx", "rwx"], r: read, w: write, x: execute
         file_type(str): The correct file type for file
     """
 
-    def __init__(self, file_path, path_type, ability=None, file_type=None, is_script=True):
+    def __init__(
+        self,
+        file_path,
+        path_type,
+        ability=None,
+        file_type=None
+    ):
         self.file_path = file_path
         self.path_type = self._check_path_type(path_type)
-        self.ability = ability
+        self.ability = self._check_ability_type(ability)
         self.file_type = file_type
-        self.is_script = is_script
 
     @staticmethod
     def _check_path_type(path_type):
@@ -64,9 +71,16 @@ class FileChecker:
             raise FileCheckException(FileCheckException.ILLEGAL_PARAM_ERROR)
         return path_type
 
+    @staticmethod
+    def _check_ability_type(ability):
+        if ability and ability not in FileCheckConst.PERM_OPTIONS:
+            logger.error(f'The ability must be one of {FileCheckConst.PERM_OPTIONS}.')
+            raise FileCheckException(FileCheckException.ILLEGAL_PARAM_ERROR)
+        return ability
+
     def common_check(self):
         """
-        功能：用户校验基本文件权限：软连接、文件长度、是否存在、读写权限、文件属组、文件特殊字符
+        功能：基本文件权限校验，包括文件存在性、软连接、文件长度、文件类型、文件读写权限、文件属组、文件路径特殊字符、文件后缀名等
         注意：文件后缀的合法性，非通用操作，可使用其他独立接口实现
         """
         check_path_exists(self.file_path)
@@ -75,23 +89,25 @@ class FileChecker:
         check_path_length(self.file_path)
         check_path_type(self.file_path, self.path_type)
         self.check_path_ability()
-        if self.is_script:
-            check_path_owner_consistent(self.file_path)
+        check_path_owner_consistent(self.file_path)
         check_path_pattern_valid(self.file_path)
         check_common_file_size(self.file_path)
         check_file_suffix(self.file_path, self.file_type)
+        check_path_no_others_write(self.file_path)
         if self.path_type == FileCheckConst.FILE:
-            check_dirpath_before_read(self.file_path)
+            check_dirpath_permission(self.file_path)
         return self.file_path
 
     def check_path_ability(self):
-        if self.ability == FileCheckConst.WRITE_ABLE:
-            check_path_writability(self.file_path)
-        if self.ability == FileCheckConst.READ_ABLE:
-            check_path_readability(self.file_path)
-        if self.ability == FileCheckConst.READ_WRITE_ABLE:
+        if not self.ability:
+            return
+
+        if FileCheckConst.READ_ABLE in self.ability:
             check_path_readability(self.file_path)
+        if FileCheckConst.WRITE_ABLE in self.ability:
             check_path_writability(self.file_path)
+        if FileCheckConst.EXECUTE_ABLE in self.ability:
+            check_path_executable(self.file_path)
 
 
 class FileOpen:
@@ -137,7 +153,8 @@ class FileOpen:
         check_path_pattern_valid(self.file_path)
         if os.path.exists(self.file_path):
             check_common_file_size(self.file_path)
-            check_dirpath_before_read(self.file_path)
+            check_path_no_others_write(self.file_path)
+            check_dirpath_permission(self.file_path)
 
     def check_ability_and_owner(self):
         if self.mode in self.SUPPORT_READ_MODE:
@@ -256,12 +273,15 @@ def check_path_type(file_path, file_type):
             raise FileCheckException(FileCheckException.INVALID_FILE_ERROR)
 
 
-def check_others_writable(directory):
-    dir_stat = os.stat(directory)
-    is_writable = (
-            bool(dir_stat.st_mode & stat.S_IWGRP) or  # 组可写
-            bool(dir_stat.st_mode & stat.S_IWOTH)  # 其他用户可写
-    )
+def check_group_writable(file_path):
+    path_stat = os.stat(file_path)
+    is_writable = bool(path_stat.st_mode & stat.S_IWGRP)
+    return is_writable
+
+
+def check_others_writable(file_path):
+    path_stat = os.stat(file_path)
+    is_writable = bool(path_stat.st_mode & stat.S_IWOTH)
     return is_writable
 
 
@@ -309,7 +329,7 @@ def check_path_before_create(path):
                                  'The file path {} contains special characters.'.format(path))
 
 
-def check_dirpath_before_read(path):
+def check_dirpath_permission(path):
     path = os.path.realpath(path)
     dirpath = os.path.dirname(path)
     if dedup_log('check_dirpath_before_read', dirpath):
@@ -321,13 +341,14 @@ def check_dirpath_before_read(path):
             logger.warning(f"The directory {dirpath} is not yours.")
 
 
-def check_file_or_directory_path(path, isdir=False):
+def check_file_or_directory_path(path, isdir=False, is_strict=False):
     """
     Function Description:
         check whether the path is valid
     Parameter:
         path: the path to check
         isdir: the path is dir or file
+        is_strict: whether to perform stricter validation (e.g., verify group cannot write to path)
     Exception Description:
         when invalid data throw exception
     """
@@ -337,6 +358,33 @@ def check_file_or_directory_path(path, isdir=False):
         path_checker = FileChecker(path, FileCheckConst.FILE, FileCheckConst.READ_ABLE)
     path_checker.common_check()
 
+    if is_strict:
+        if check_group_writable(path):
+            raise FileCheckException(
+                FileCheckException.FILE_PERMISSION_ERROR,
+                f"The directory/file must not allow write access to group. Directory/File path: {path}"
+            )
+
+
+def check_path_no_others_write(file_path):
+    if dedup_log('check_path_no_others_write', file_path):
+        if check_group_writable(file_path):
+            logger.warning(f"The directory/file path is writable by group: {file_path}.")
+
+    if check_others_writable(file_path):
+        raise FileCheckException(
+            FileCheckException.FILE_PERMISSION_ERROR,
+            f"The directory/file must not allow write access to others. Directory/File path: {file_path}"
+        )
+
+
+def check_path_no_group_others_write(file_path):
+    if check_group_writable(file_path) or check_others_writable(file_path):
+        raise FileCheckException(
+            FileCheckException.FILE_PERMISSION_ERROR,
+            f"The directory/file must not allow write access to group or others. Directory/File path: {file_path}"
+        )
+
 
 def change_mode(path, mode):
     if not os.path.exists(path) or os.path.islink(path):
@@ -388,6 +436,14 @@ def check_file_type(path):
         raise FileCheckException(FileCheckException.INVALID_FILE_ERROR)
 
 
+def root_privilege_warning():
+    if os.getuid() == 0:
+        logger.warning(
+            "msprobe is being run as root. "
+            "To avoid security risks, it is recommended to switch to a regular user to run it."
+        )
+
+
 def load_yaml(yaml_path):
     path_checker = FileChecker(yaml_path, FileCheckConst.FILE, FileCheckConst.READ_ABLE, FileCheckConst.YAML_SUFFIX)
     checked_path = path_checker.common_check()
@@ -968,7 +1024,13 @@ class SharedDict:
     def _safe_load(self):
         with io.BytesIO(self._shm.buf[:]) as buff:
             try:
-                self._dict = SafeUnpickler(buff).load()
+                data = SafeUnpickler(buff).load()
+                if not isinstance(data, dict):
+                    logger.debug(f"Data from shared memory is '{type(data)}' type, expected 'dict'.")
+                    self._dict = {}
+                    self._changed = True
+                else:
+                    self._dict = data
             except Exception as e:
                 logger.debug(f'shared dict is unreadable, reason: {e}, create new dict.')
                 self._dict = {}
@@ -985,4 +1047,48 @@ class SafeUnpickler(pickle.Unpickler):
         raise pickle.PicklingError(f'Unpickling {module}.{name} is illegal!')
 
 
+class DeserializationScanner:
+    """反序列化风险扫描器"""
+
+    DANGEROUS_METHODS = {
+        '__reduce__', '__reduce_ex__', '__setstate__', '__getstate__',
+        '__new__', '__init__', '__del__',
+        '__call__', '__enter__', '__exit__',
+        'eval', 'exec', 'compile', '__import__', 'open'
+        'os.system', 'os.popen', 'subprocess.call', 'subprocess.Popen',
+    }
+
+    DANGEROUS_MODULES = {
+        'os', 'sys', 'subprocess', 'shutil', 'socket',
+        'requests', 'urllib', 'ftplib', 'smtplib',
+    }
+
+    @classmethod
+    def scan_pickle_content(cls, filepath: str) -> bool:
+        with FileOpen(filepath, 'rb') as f:
+            content = f.read()
+
+        try:
+            text_content = content.decode('latin-1')
+        except Exception as e:
+            text_content = str(content)
+
+        for method in cls.DANGEROUS_METHODS:
+            if re.fullmatch(method, text_content):
+                logger.warning(f"Insecure method found: {method}")
+                return False
+
+        for module in cls.DANGEROUS_MODULES:
+            patterns = [
+                f"import {module}",
+                f"from {module} import",
+                f"{module}.",
+            ]
+            for pattern in patterns:
+                if pattern in text_content:
+                    logger.warning(f"Insecure module found: {module}")
+                    return False
+        return True
+
+
 atexit.register(SharedDict.destroy_shared_memory)

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/msprobe/core/common/framework_adapter.py

@@ -12,10 +12,13 @@
 # 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 functools
+
 import functools
+
 from msprobe.core.common.const import Const
-from msprobe.core.common.file_utils import check_file_or_directory_path
-from msprobe.core.common.file_utils import save_npy
+from msprobe.core.common.file_utils import check_file_or_directory_path, save_npy, DeserializationScanner
+from msprobe.core.common.log import logger
+from msprobe.core.common.utils import confirm
 
 
 class FrameworkDescriptor:
@@ -103,7 +106,7 @@ class FmkAdp:
     @classmethod
     def tensor_norm(cls, tensor):
         return cls.process_tensor(tensor, lambda x: x.norm())
-    
+
     @classmethod
     def save_tensor(cls, tensor, filepath):
         if cls.fmk == Const.PT_FRAMEWORK:
@@ -151,9 +154,16 @@ class FmkAdp:
 
     @classmethod
     def load_checkpoint(cls, path, to_cpu=True, weights_only=True):
-        check_file_or_directory_path(path)
+        check_file_or_directory_path(path, is_strict=not weights_only)
         if cls.fmk == Const.PT_FRAMEWORK:
             try:
+                if not weights_only:
+                    if not DeserializationScanner.scan_pickle_content(path):
+                        if not confirm(
+                                f"Some insecure methods or modules are detected in {path}, "
+                                f"input yes to ignore and continue, otherwise exit", False):
+                            logger.error("Insecure risks found and exit!")
+                            raise Exception("Insecure risks found and exit!")
                 if to_cpu:
                     return cls.framework.load(path, map_location=cls.framework.device("cpu"), weights_only=weights_only)
                 else:
@@ -161,9 +171,9 @@ class FmkAdp:
             except Exception as e:
                 raise RuntimeError(f"load pt file {path} failed: {e}") from e
         return mindspore.load_checkpoint(path)
-    
+
     @classmethod
     def asnumpy(cls, tensor):
         if cls.fmk == Const.PT_FRAMEWORK:
             return tensor.float().numpy()
-        return tensor.float().asnumpy()
+        return tensor.float().asnumpy()

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/msprobe/core/common/utils.py

@@ -708,3 +708,20 @@ def check_process_num(process_num):
         raise ValueError(f"process_num({process_num}) is not a positive integer")
     if process_num > Const.MAX_PROCESS_NUM:
         raise ValueError(f"The maximum supported process_num is {Const.MAX_PROCESS_NUM}, current value: {process_num}.")
+
+
+def confirm(prompt, default=False):
+    if default is True:
+        prompt_suffix = " [Y/n] "
+    elif default is False:
+        prompt_suffix = " [y/N] "
+    else:
+        prompt_suffix = " [y/n] "
+
+    full_prompt = prompt + prompt_suffix
+
+    user_input = input(full_prompt).strip().lower()
+    if user_input in ['y', 'yes']:
+        return True
+    else:
+        return default

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/msprobe/core/compare/diff_analyze/first_diff_analyze.py

@@ -26,8 +26,6 @@ from msprobe.core.compare.utils import gen_api_batches
 
 cur_dir = os.path.dirname(os.path.realpath(__file__))
 diff_threshold_yaml_path = os.path.join(cur_dir, 'diff_analyze_threshold.yaml')
-ignore_op_list_yaml_path = os.path.join(cur_dir, 'ignore_op_list.yaml')
-ignore_list = load_yaml(ignore_op_list_yaml_path)
 thresholds = load_yaml(diff_threshold_yaml_path)
 cmp_metrics = thresholds.get('compare_metrics')
 
@@ -53,7 +51,7 @@ class FirstDiffAnalyze:
                 return True
         return False
 
-    def single_api_check(self, result_slice, header, api_name=None):
+    def single_api_check(self, result_slice, header):
         """
         单个api差异检查
 
@@ -67,18 +65,14 @@ class FirstDiffAnalyze:
         }
 
         column_indices = {name: idx for idx, name in enumerate(header)}
-        output_idx = -1
+
         for line in result_slice:
             op_item = {
                 column_name: line[column_indices[column_name]]
                 for column_name in header
             }
             single_check_result['op_items'].append(op_item)
-            if op_item['state'] != 'output':
-                continue
-            output_idx += 1
-            if output_idx in ignore_list.get(api_name, []):
-                continue
+
             # set is_same
             if self.mode_config.dump_mode == Const.MD5:
                 if line[column_indices[CompareConst.RESULT]] == CompareConst.DIFF:
@@ -123,13 +117,7 @@ class FirstDiffAnalyze:
         with tqdm(total=len(api_batches), desc=bar_desc_add_rank, unit="api/module", ncols=100) as progress_bar:
             for api_batch in api_batches:
                 result_slice = result[api_batch.start: api_batch.params_grad_end_index]
-                api_compo = api_batch.api_name.split('.')
-                # suppose name is Tensor.MatMul.0.forward
-                if len(api_compo) < 4:
-                    continue
-                # get MatMul as api_name
-                api_name = api_compo[-3]
-                check_result[api_batch.api_name] = self.single_api_check(result_slice, header, api_name)
+                check_result[api_batch.api_name] = self.single_api_check(result_slice, header)
                 progress_bar.update(1)
 
         return check_result

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/msprobe/core/compare/find_first/utils.py

@@ -182,7 +182,7 @@ def analyze_diff_in_group(nodes_group):
     input_diff_nodes = list(filter(lambda node: node.is_diff, src_list))
     # 如果有异常回溯计算节点找到异常来源
     # 使用cpu模拟节点进行计算，查看结果是否有问题。需要对所有计算节点录入/映射，暂不实现。
-    get_compute_ops_from_comm_nodes(nodes_group)
+    get_compute_ops_from_comm_nodes(input_diff_nodes)
     # 筛选入参没问题但出参有问题的通信节点
     output_diff_nodes = list(filter(lambda node: node.data.is_diff, nodes_group))
     get_comm_ops(output_diff_nodes)

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/msprobe/core/compare/utils.py

@@ -789,23 +789,43 @@ def compare_distributed_inner(npu_dump_dir, bench_dump_dir, output_path, compare
 
     npu_ranks, bench_ranks = get_sorted_ranks(npu_dump_dir, bench_dump_dir)
 
-    # 统计量、md5比对
-    pre_check_dump_path = os.path.join(npu_dump_dir, npu_ranks[0], 'dump.json') if npu_ranks else ''
+    # ------------------预载rank0的json用于判断是什么类型dump数据------------------
+    # 判断是否存在dump.json或debug.json
+    if npu_ranks:
+        dir_path = os.path.join(npu_dump_dir, npu_ranks[0])
+        dump_file = os.path.join(dir_path, 'dump.json')
+        debug_file = os.path.join(dir_path, 'debug.json')
+
+        # 确定pre_check_dump_path
+        if os.path.exists(dump_file):
+            pre_check_dump_path = dump_file
+        elif os.path.exists(debug_file):
+            pre_check_dump_path = debug_file
+        else:
+            pre_check_dump_path = ''
+    else:
+        pre_check_dump_path = ''
+
+    # 如果pre_check_dump_path为空，直接返回
     if not pre_check_dump_path:
         return
+
     dump_data = load_json(pre_check_dump_path)
+
+    # ------------------统计量、md5比对------------------
     if dump_data.get('task') == Const.STATISTICS:
         # dump数据为统计量或md5时，多进程加速比对
         input_param_nr_list = []
         for nr, br in zip(npu_ranks, bench_ranks):
-            input_param, skip = extract_compare_param(Const.DUMP_JSON_FILE)
-            if not skip:
-                input_param_nr_list.append((input_param, nr))
+            for file_type in [Const.DUMP_JSON_FILE, Const.DEBUG_JSON_FILE]:
+                input_param, skip = extract_compare_param(file_type)
+                if not skip:
+                    input_param_nr_list.append((input_param, nr))
         func_args = (compare_func, input_param_nr_list, output_path, kwargs)
         multi_statistics_compare(multi_ranks_compare, func_args)
         return
 
-    # 真实数据比对
+    # ------------------真实数据比对------------------
     for nr, br in zip(npu_ranks, bench_ranks):
         for file_type in [Const.DUMP_JSON_FILE, Const.DEBUG_JSON_FILE]:
             input_param, skip = extract_compare_param(file_type)

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/msprobe/core/config_check/ckpt_compare/ckpt_comparator.py

@@ -19,11 +19,11 @@ from tqdm import tqdm
 from msprobe.core.common.file_utils import save_json, check_path_before_create, check_path_not_exists, \
     check_file_or_directory_path
 from msprobe.core.common.log import logger
+from msprobe.core.common.utils import confirm
 from msprobe.core.config_check.ckpt_compare.megatron_loader import load_megatron_weights
 from msprobe.core.config_check.ckpt_compare.metrics import METRIC_FUNC
 
 
-
 def compare_checkpoints(ckpt_path1, ckpt_path2, output_path) -> Dict:
     """Compare weights between two checkpoints using cosine similarity and L2 distance.
     
@@ -45,6 +45,11 @@ def compare_checkpoints(ckpt_path1, ckpt_path2, output_path) -> Dict:
     """
 
     # Load both checkpoints
+    if not confirm("You are using torch.load with weights_only is False, it may cause arbitrary code "
+                   "execution. Do it only if you get the file from a trusted source. Input yes to continue, "
+                   "otherwise exit", False):
+        logger.error("Insecure risks found and exit!")
+        raise Exception("Insecure risks found and exit!")
     check_file_or_directory_path(ckpt_path1, isdir=True)
     check_file_or_directory_path(ckpt_path2, isdir=True)
     check_path_before_create(output_path)

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/msprobe/core/hook_manager.py

@@ -63,7 +63,6 @@ class BaseHookManager(ABC):
     def reset_status():
         BaseHookManager.inner_switch = defaultdict(bool)
         BaseHookManager.inner_api_count = defaultdict(int)
-        BaseHookManager.hook_handle_dict.clear()
         BaseHookManager.params_grad_info.clear()
 
     @staticmethod

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/msprobe/docs/01.installation.md

@@ -16,6 +16,8 @@ pip install mindstudio-probe
 
 |  版本   |    发布日期    |支持 PyTorch 版本|支持 MindSpore 版本|                                                                下载链接                                                                |校验码|
 |:-----:|:----------:|:--:|:--:|:----------------------------------------------------------------------------------------------------------------------------------:|:--:|
+| 8.3.0 | 2025.10.30 |1.11/2.0/2.1/2.2/2.5/2.6/2.7|2.4.0/2.5.0/2.6.0| [mindstudio_probe-8.3.0-py3-none-any.whl](https://ptdbg.obs.myhuaweicloud.com/msprobe/8.3/mindstudio_probe-8.3.0-py3-none-any.whl) |e933657b8ceb20774f924865d1f47978bd49cd1d9cec5fb60dec4f18802afbcc|
+| 8.2.1 | 2025.9.29  |1.11/2.0/2.1/2.2|2.4.0/2.5.0/2.6.0| [mindstudio_probe-8.2.1-py3-none-any.whl](https://ptdbg.obs.myhuaweicloud.com/msprobe/8.2/mindstudio_probe-8.2.1-py3-none-any.whl) |2152fadefec3d70148a910a54f2cfb6fccfa60b5ccd59f835f20289a7a5f4764|
 | 8.2.0 | 2025.9.03  |1.11/2.0/2.1/2.2|2.4.0/2.5.0/2.6.0| [mindstudio_probe-8.2.0-py3-none-any.whl](https://ptdbg.obs.myhuaweicloud.com/msprobe/8.2/mindstudio_probe-8.2.0-py3-none-any.whl) |bbc1577d76754adf987069308177d3e0a04e36de9c7f22e75c34cf4ad0ce1af2|
 | 8.1.2 | 2025.8.01  |1.11/2.0/2.1/2.2|2.4.0/2.5.0/2.6.0| [mindstudio_probe-8.1.2-py3-none-any.whl](https://ptdbg.obs.myhuaweicloud.com/msprobe/8.1/mindstudio_probe-8.1.2-py3-none-any.whl) |ff07bb81fddd3b8f3096d119ca1481bde8fdb24f10644def5250caad727448ab|
 | 8.1.1 | 2025.6.20  |1.11/2.0/2.1/2.2|2.4.0/2.5.0/2.6.0| [mindstudio_probe-8.1.1-py3-none-any.whl](https://ptdbg.obs.myhuaweicloud.com/msprobe/8.1/mindstudio_probe-8.1.1-py3-none-any.whl) |2aad10a243575544d7feef552caf4d06aa93028488ebd0bbc9aa350379da859d|

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/msprobe/docs/02.config_introduction.md

@@ -145,7 +145,7 @@ PyTorch、MSAdapter 以及 MindSpore 动态图场景下，"level"须为"L0"或"L
     <tr><td>pert_mode</td><td>无标杆扰动因子，str 类型。可选参数：<br/>  "improve_precision"：对输入做升精度，默认值；<br/>        "add_noise"：对输入增加噪声；<br/>        "no_change"：不加扰动直接二次执行；<br/>        "bit_noise"：输入的末位比特翻转，MindSpore 场景不支持 BF16 类型的向量；<br/>        "change_value"：输入的张量首尾值调换；<br/>        "to_cpu"：在 CPU 等价执行（仅 PyTorch 场景支持）。<br/><b>配置示例</b>："pert_mode": "improve_precision"。</td><td>否</td></tr>
     <tr><td>handler_type</td><td>处理类型，可选参数：<br/> "check"：进行无标杆比对检查，默认值；<br/> "fix"：将扰动后的 API 输出结果覆盖原始 API 输出结果，尝试将 Loss 曲线恢复正常，该模式下不支持预热功能与反向过程，且仅支持"improve_precision"、"to_cpu"（ PyTorch 场景）两种扰动因子。<br/> <b>配置示例</b>："handler_type": "check"。</td><td>否</td></tr>
     <tr><td>fuzz_level</td><td>无标杆数据 dump 级别，即选择比对结果文件应输出的表头属性，当前仅支持取值为："L1"。输出结果详见 <a href="#161-无标杆比对数据存盘格式">1.6.1 无标杆比对数据存盘格式</a>。</td><td>否</td></tr>
-    <tr><td>fuzz_stage</td><td>比对过程，选择对 API 前向或反向进行无标杆比对，可选参数：<br/> "forward"：前向，默认值；<br/>  "backward"：反向。当 fuzz_stage 为 "backward" 时，handler_type 只能为 "check"。<br/>  <b>配置示例</b>："fuzz_stage": "backward"。</td><td>否</td></tr>
+    <tr><td>fuzz_stage</td><td>比对过程，选择对 API 前向或反向进行无标杆比对，可选参数：<br/> "forward"：前向，默认值；<br/>  "backward"：反向。当 fuzz_stage 为 "backward" 时，handler_type 只能为 "check"。pytorch场景下，当 fuzz_stage 为 "backward" 时， list 参数不能为空。<br/>  <b>配置示例</b>："fuzz_stage": "backward"。</td><td>否</td></tr>
     <tr><td>if_preheat</td><td>预热功能（仅 PyTorch 场景支持），bool 类型。开启功能后工具可以根据每次迭代的输出调整精度算法的阈值，从而更准确地找出存在精度问题的 API。当"handler_type": "fix"时，不支持预热。可选参数：<br/>  true（开启）或 false（关闭），默认关闭。<br/>  <b>配置示例</b>："if_preheat": "true"。</td><td>否</td></tr>
     <tr><td>preheat_step</td><td>开启预热的迭代数量（仅 PyTorch 场景支持），int 类型，默认值为 15。须配置 "if_preheat": "true"。</td><td>否</td></tr>
     <tr><td>max_sample</td><td>每个算子预热的采样次数的最大阈值（仅 PyTorch 场景支持），int 类型，默认值为 20。须配置 "if_preheat": "true"。</td><td>否</td></tr>

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.3}/msprobe/docs/05.data_dump_PyTorch.md

@@ -445,7 +445,7 @@ seed_all()
 debugger = PrecisionDebugger(config_path="./config.json", dump_path="./dump_path")
 # 模型定义及初始化等操作
 prompts = ["Hello, my name is"]
-sampling_params = SamplingParams(temprature=0.8, top_p=0.95)
+sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
 llm = LLM(model='...')
 model = llm.llm_engine.model_executor.driver_worker.worker.model_runner.get_model()
 # 开启数据dump, 指定采集推理模型逐字符循环推理中的第1~3次