mindstudio-probe 8.2.1__tar.gz → 8.3.0__tar.gz

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mindstudio-probe
-Version: 8.2.1
+Version: 8.3.0
 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-8.3.0}/mindstudio_probe.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mindstudio-probe
-Version: 8.2.1
+Version: 8.3.0
 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-8.3.0}/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
@@ -115,6 +114,7 @@ msprobe/docs/04.kernel_dump_PyTorch.md
 msprobe/docs/05.data_dump_PyTorch.md
 msprobe/docs/06.data_dump_MindSpore.md
 msprobe/docs/07.accuracy_checker_PyTorch.md
+msprobe/docs/08.accuracy_checker_online_PyTorch.md
 msprobe/docs/09.accuracy_checker_MindSpore.md
 msprobe/docs/10.accuracy_compare_PyTorch.md
 msprobe/docs/11.accuracy_compare_MindSpore.md
@@ -333,6 +333,7 @@ msprobe/nan_analyze/analyzer.py
 msprobe/nan_analyze/graph.py
 msprobe/nan_analyze/utils.py
 msprobe/pytorch/__init__.py
+msprobe/pytorch/attl_manager.py
 msprobe/pytorch/function_factory.py
 msprobe/pytorch/pt_config.py
 msprobe/pytorch/pytorch_service.py
@@ -372,6 +373,14 @@ msprobe/pytorch/api_accuracy_checker/run_ut/run_overflow_check.py
 msprobe/pytorch/api_accuracy_checker/run_ut/run_ut.py
 msprobe/pytorch/api_accuracy_checker/run_ut/run_ut_utils.py
 msprobe/pytorch/api_accuracy_checker/run_ut/torch_ut_setting.json
+msprobe/pytorch/api_accuracy_checker/tensor_transport_layer/__init__.py
+msprobe/pytorch/api_accuracy_checker/tensor_transport_layer/attl.py
+msprobe/pytorch/api_accuracy_checker/tensor_transport_layer/client.py
+msprobe/pytorch/api_accuracy_checker/tensor_transport_layer/device_dispatch.py
+msprobe/pytorch/api_accuracy_checker/tensor_transport_layer/dump_dispatch.py
+msprobe/pytorch/api_accuracy_checker/tensor_transport_layer/server.py
+msprobe/pytorch/api_accuracy_checker/tensor_transport_layer/torch_ops_config.yaml
+msprobe/pytorch/api_accuracy_checker/tensor_transport_layer/utils.py
 msprobe/pytorch/bench_functions/__init__.py
 msprobe/pytorch/bench_functions/apply_adam.py
 msprobe/pytorch/bench_functions/apply_adam_w.py

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.0}/msprobe/README.md

@@ -84,7 +84,9 @@ msprobe 通过在训练脚本中添加 PrecisionDebugger 接口的方式对 API
 
 精度预检旨在昇腾 NPU 上扫描训练模型中的所有 API 进行 API 复现，给出精度情况的诊断和分析。对应 config.json 中的 "run_ut" task。
 
-PyTorch 场景的[离线预检](./docs/07.accuracy_checker_PyTorch.md)
+PyTorch 场景的[离线预检](./docs/07.accuracy_checker_PyTorch.md)和[在线预检](./docs/08.accuracy_checker_online_PyTorch.md)
+
+在线预检计划在 Mindstudio 8.3.0 版本正式下线。
 
 MindSpore 动态图场景的[离线预检](./docs/09.accuracy_checker_MindSpore.md)
 

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.0}/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()
 
@@ -46,16 +48,15 @@ class FileChecker:
     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
+        ability(str): one of [FileCheckConst.READ_ABLE, FileCheckConst.WRITE_ABLE, FileCheckConst.READ_WRITE_ABLE]
         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 +65,17 @@ class FileChecker:
             raise FileCheckException(FileCheckException.ILLEGAL_PARAM_ERROR)
         return path_type
 
+    @staticmethod
+    def _check_ability_type(ability):
+        ability_list = [FileCheckConst.READ_ABLE, FileCheckConst.WRITE_ABLE, FileCheckConst.READ_WRITE_ABLE]
+        if ability and ability not in ability_list:
+            logger.error(f'The ability must be one of {ability_list}.')
+            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)
@@ -75,13 +84,13 @@ 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):
@@ -137,7 +146,8 @@ 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 +266,15 @@ 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 +322,7 @@ 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 +334,14 @@ 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 +351,33 @@ 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 +429,14 @@ 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 +1017,13 @@ 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 = {}

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.0}/msprobe/core/common/framework_adapter.py

@@ -12,10 +12,11 @@
 # 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
 
 
 class FrameworkDescriptor:
@@ -103,7 +104,7 @@ 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:
@@ -151,7 +152,7 @@ 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 to_cpu:
@@ -161,9 +162,9 @@ 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-8.3.0}/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-8.3.0}/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-8.3.0}/msprobe/core/hook_manager.py

@@ -46,9 +46,10 @@ class BaseHookManager(ABC):
     hook_handle_dict = {}
     params_grad_info = {}
 
-    def __init__(self, data_collector, config):
+    def __init__(self, data_collector, config, attl_manager=None):
         self.data_collector = data_collector
         self.config = config
+        self.attl_manager = attl_manager
 
     @property
     def _pid(self):
@@ -63,7 +64,6 @@ 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
@@ -216,6 +216,10 @@ class BaseHookManager(ABC):
 
                 args = self._register_backward_hook(module, full_backward_name, args)
                 with self._no_grad_context():
+                    if getattr(self.config, "online_run_ut", False):
+                        BaseHookManager.inner_switch[tid] = False
+                        ThreadSafe.release()
+                        return
                     self.data_collector.update_api_or_module_name(full_forward_name)
                     self.data_collector.forward_input_data_collect(
                         full_forward_name,
@@ -261,6 +265,13 @@ class BaseHookManager(ABC):
                     output = self._register_backward_pre_hook(module, full_backward_name, output)
 
                 with self._no_grad_context():
+                    if getattr(self.config, "online_run_ut", False):
+                        if self.data_collector.scope and not self.data_collector.scope.check(full_name):
+                            return None
+                        if self.attl_manager:
+                            self.attl_manager.attl_send(full_name, args, kwargs, output)
+                        BaseHookManager.inner_switch[tid] = False
+                        return None
                     if hook_type == Const.MODULE:
                         params_dict = self._get_params_dict(module)
                         setattr(module_input_output, Const.PARAMS, params_dict)
@@ -309,7 +320,9 @@ class BaseHookManager(ABC):
                 original_state = self.ensure_gc_enabled()
                 BaseHookManager.inner_switch[tid] = True
                 self.data_collector.update_api_or_module_name(full_name)
-
+                if getattr(self.config, "online_run_ut", False):
+                    BaseHookManager.inner_switch[tid] = False
+                    return
                 need_exchange = self._need_exchange(module) if hook_type == Const.MODULE else True
                 if need_exchange:
                     module_input_output = ModuleBackwardInputsOutputs(grad_input=grad_output, grad_output=grad_input)

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.0}/msprobe/core/service.py

@@ -35,6 +35,7 @@ class BaseService(ABC):
         self.config.level = getattr(config, 'level_ori', config.level)  # 兼容MindSpore配置
         self.model = None
         self.data_collector = build_data_collector(self.config)
+        self.attl_manager = None
         self.current_iter = 0
         self.loop = 0
         self.init_step = 0
@@ -90,6 +91,10 @@ class BaseService(ABC):
             self.config.task in self.data_collector.tasks_need_tensor_data or
             (self.config.task == Const.STATISTICS and self.config.tensor_list)
         )
+    
+    @property
+    def _is_online_run_ut(self):
+        return getattr(self.config, "online_run_ut", False)
 
     @property
     @abstractmethod
@@ -141,9 +146,11 @@ class BaseService(ABC):
             self.primitive_switch = True
             self._change_jit_switch(True)
             self.logger.info(f"Dump switch is turned on at step {self.current_iter}. ")
-
-        self.create_dirs()
-        self.logger.info(f"Dump data will be saved in {self.dump_iter_dir}.")
+        if self._is_online_run_ut:
+            self._run_ut_dispatch(True)
+        else:
+            self.create_dirs()
+            self.logger.info(f"Dump data will be saved in {self.dump_iter_dir}.")
 
     def stop(self):
         """通用stop模板"""
@@ -158,7 +165,8 @@ class BaseService(ABC):
         self._change_jit_switch(False)
         if self._is_l2_level:
             return
-
+        if self._is_online_run_ut:
+            self._run_ut_dispatch(False)
         self._process_async_dump()
         self.data_collector.write_json()
 
@@ -258,6 +266,8 @@ class BaseService(ABC):
         end_service = self.config.step and self.current_iter > max(self.config.step) or \
             self.data_collector and self.data_collector.data_processor.is_terminated
         if end_service:
+            if self._is_online_run_ut and self.attl_manager:
+                self.attl_manager.attl_stop()
             self.primitive_switch = False
             self._change_jit_switch(False)
             Runtime.is_running = False
@@ -300,7 +310,8 @@ class BaseService(ABC):
         if root_model and isinstance(root_model, list):
             root_model = root_model[0]
             self.logger.warning("Infer model can only input one to support token_range, choose the first one.")
-
+        if self._is_online_run_ut:
+            return
         root_model.register_forward_pre_hook(infer_hook)
 
     def _create_l2_dirs(self, cur_rank):

{mindstudio_probe-8.2.1 → mindstudio_probe-8.3.0}/msprobe/docs/02.config_introduction.md

@@ -73,7 +73,14 @@
 | data_mode      | 与[ 1.2 task 配置为 statistics ](#12-task-配置为-statistics)中的解释相同 | 否       |
 | file_format    | tensor 数据的保存格式，str 类型，仅支持 MindSpore 静态图场景的 L2 级别配置该字段，其他场景不生效。可选参数：<br/> "bin"：dump 的 tensor 文件为二进制格式；<br/>"npy"：dump 的 tensor 文件后缀为 .npy，默认值。 | 否       |
 | summary_mode  | 控制 dump 文件输出的模式，str 类型，支持 PyTorch、MSAdapter、MindSpore 动态图。可选参数：<br/> md5：dump 输出包含 CRC-32 值以及 API 统计信息的 dump.json 文件，用于验证数据的完整性；<br/> statistics：dump 仅输出包含 API 统计信息的 dump.json 文件，默认值。| 否 |
+| online_run_ut<sup>a</sup>  | 在线预检模式开关，bool 类型，可选参数 true（开启）、false（关闭），默认未配置，表示关闭。配置为 true 表示开启在线预检。| 否 |
+| nfs_path<sup>a</sup> | 在线预检模式共享存储目录路径，str 类型，用于 GPU 设备和 NPU 设备间进行通信。仅在 online_run_ut 字段配置为 true 时生效，配置该参数后 host 和 port 不生效。 | 否 |
+| host<sup>a</sup> | 在线预检模式局域网场景信息接收端 IP，str 类型，用于 GPU 设备和 NPU 设备间进行通信，NPU 侧须配置为 GPU 侧的局域网 IP 地址。仅在 online_run_ut 字段配置为 true 时生效，局域网场景时，不能配置 nfs_path 参数，否则局域网场景不生效。 | 否 |
+| port<sup>a</sup> | 在线预检模式局域网场景信息接收端端口号，int 类型，用于 GPU 设备和 NPU 设备间进行通信，NPU 侧须配置为 GPU 侧的端口号。仅在 online_run_ut 字段配置为 true 时生效，局域网场景时，不能配置 nfs_path 参数，否则局域网场景不生效。| 否 |
 
+**说明**：
+
+1. online_run_ut、nfs_path、host、port 等字段仅在线预检场景 NPU 机器生效。
 
 **示例**：
  - [PyTorch场景](03.config_examples.md#12-task-配置为-tensor)
@@ -88,11 +95,17 @@
 | white_list<sup>a</sup>      | API dump 白名单，仅对指定的 API 进行 dump。<br/>**配置示例**："white_list": ["conv1d", "conv2d"]。默认未配置白名单，即 dump 全量 API 数据。 | 否       |
 | black_list<sup>a</sup>      | API dump 黑名单，被指定的 API 不进行 dump。<br/>**配置示例**："black_list": ["conv1d", "conv2d"]。默认未配置黑名单，即 dump 全量 API 数据。 | 否       |
 | error_data_path | 配置保存精度未达标的 API 输入输出数据路径，默认为当前路径。<br/>**配置示例**："error_data_path": "./"。 | 否       |
+| is_online<sup>b</sup>       | 在线预检模式开关，bool 类型，可选参数 true（开启）、false（关闭），默认关闭。 | 否 |
+| nfs_path<sup>b</sup>        | 在线预检模式共享存储目录路径，str 类型，用于 GPU 设备和 NPU 设备间进行通信。配置该参数后 host 和 port 不生效，仅在 is_online 字段配置为 true 时生效。 | 否    |
+| host<sup>b</sup>            | 在线预检模式局域网场景信息接收端 IP，str 类型，用于 GPU 设备和 NPU 设备间进行通信，GPU 侧配置为本机地址 127.0.0.1 或本机局域网 IP。局域网场景时，不能配置 nfs_path 参数，否则局域网场景不生效。仅在 is_online 字段配置为 true 时生效。 | 否 |
+| port<sup>b</sup>            | 在线预检模式局域网场景信息接收端端口号，int 类型，用于 GPU 设备和 NPU 设备间进行通信，GPU 侧配置为本机可用端口。局域网场景时，不能配置 nfs_path 参数，否则局域网场景不生效。仅在 is_online 字段配置为 true 时生效。| 否 |
+| rank_list<sup>b</sup>       | 指定在线预检的 Rank ID，默认值为 [0]，list[int] 类型，应配置为大于等于 0 的整数，且须根据实际卡的 Rank ID 配置，若所配置的值大于实际训练所运行的卡的 Rank ID，则在线预检输出数据为空。GPU 和 NPU 须配置一致。仅在 is_online 字段配置为 true 时生效。 | 否    |
 
 **说明**：
 
 1. white_list 和 black_list 同时配置时，二者配置的 API 名单若无交集，则白名单生效，若 API 名单存在交集，则白名单排除的部分以及交集的 API 不进行 dump。
 
+2. is_online、nfs_path、host、port、rank_list 等字段仅在线预检场景 GPU 机器生效。
 
 **示例**：
 ```json
@@ -145,7 +158,7 @@ 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>