mindstudio-probe 8.2.0__tar.gz → 8.2.1__tar.gz

{mindstudio_probe-8.2.0 → mindstudio_probe-8.2.1}/PKG-INFO

@@ -1,8 +1,8 @@
 Metadata-Version: 2.1
 Name: mindstudio-probe
-Version: 8.2.0
+Version: 8.2.1
 Summary: Ascend Probe Utils
-Home-page: https://gitee.com/ascend/mstt/tree/master/debug/accuracy_tools/msprobe
+Home-page: https://gitcode.com/Ascend/mstt/tree/master/debug/accuracy_tools/msprobe
 Author: Ascend Team
 Author-email: pmail_mindstudio@huawei.com
 License: Apache License 2.0

{mindstudio_probe-8.2.0 → mindstudio_probe-8.2.1}/mindstudio_probe.egg-info/PKG-INFO

@@ -1,8 +1,8 @@
 Metadata-Version: 2.1
 Name: mindstudio-probe
-Version: 8.2.0
+Version: 8.2.1
 Summary: Ascend Probe Utils
-Home-page: https://gitee.com/ascend/mstt/tree/master/debug/accuracy_tools/msprobe
+Home-page: https://gitcode.com/Ascend/mstt/tree/master/debug/accuracy_tools/msprobe
 Author: Ascend Team
 Author-email: pmail_mindstudio@huawei.com
 License: Apache License 2.0

{mindstudio_probe-8.2.0 → mindstudio_probe-8.2.1}/mindstudio_probe.egg-info/SOURCES.txt

@@ -30,6 +30,7 @@ msprobe/core/common/global_lock.py
 msprobe/core/common/inplace_op_checker.py
 msprobe/core/common/inplace_ops.yaml
 msprobe/core/common/log.py
+msprobe/core/common/megatron_utils.py
 msprobe/core/common/parallel_state.py
 msprobe/core/common/runtime.py
 msprobe/core/common/utils.py
@@ -45,6 +46,7 @@ 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-8.2.0 → mindstudio_probe-8.2.1}/msprobe/README.md

@@ -35,17 +35,17 @@ export MSPROBE_LOG_LEVEL={x}
 
 ## 环境和依赖
 
-- 硬件环境请参见《[昇腾产品形态说明](https://gitee.com/link?target=https%3A%2F%2Fwww.hiascend.com%2Fdocument%2Fdetail%2Fzh%2Fcanncommercial%2F80RC22%2Fquickstart%2Fquickstart%2Fquickstart_18_0002.html)》。
-- 软件环境请参见《[CANN 软件安装指南](https://gitee.com/link?target=https%3A%2F%2Fwww.hiascend.com%2Fdocument%2Fdetail%2Fzh%2Fcanncommercial%2F80RC22%2Fsoftwareinst%2Finstg%2Finstg_0000.html%3FMode%3DPmIns%26OS%3DUbuntu%26Software%3DcannToolKit)》安装昇腾设备开发或运行环境，即toolkit软件包。
+- 硬件环境请参见《[昇腾产品形态说明](https://www.hiascend.com/document/detail/zh/canncommercial/80RC22/quickstart/quickstart/quickstart_18_0002.html)》。
+- 软件环境请参见《[CANN 软件安装指南](https://www.hiascend.com/document/detail/zh/canncommercial/80RC22/softwareinst/instg/instg_0000.html?Mode=PmIns&OS=Ubuntu&Software=cannToolKit)》安装昇腾设备开发或运行环境，即toolkit软件包。
 
 以上环境依赖请根据实际环境选择适配的版本。
 
 ## 版本配套说明
 
-- msprobe支持AscendPyTorch 1.11.0或更高版本，支持的PyTorch和CANN以及PyTorch和python软件版本配套关系请参见《[Ascend Extension for PyTorch插件](https://gitee.com/ascend/pytorch)》。
+- msprobe支持AscendPyTorch 1.11.0或更高版本，支持的PyTorch和CANN以及PyTorch和python软件版本配套关系请参见《[Ascend Extension for PyTorch插件](https://gitcode.com/Ascend/pytorch)》。
 - msprobe支持MindSpore 2.4.0或更高版本，支持的MindSpore和CANN以及MindSpore和python软件版本配套关系请参见《[MindSpore版本发布列表](https://www.mindspore.cn/versions)》。
 - msprobe支持MSAdapter 2.1.0。
-- msprobe支持的固件驱动版本与配套CANN软件支持的固件驱动版本相同，开发者可通过“[昇腾社区-固件与驱动](https://gitee.com/link?target=https%3A%2F%2Fwww.hiascend.com%2Fhardware%2Ffirmware-drivers%2Fcommunity%3Fproduct%3D2%26model%3D28%26cann%3D8.0.RC3.alpha003%26driver%3D1.0.25.alpha)”页面根据产品型号与CANN软件版本获取配套的固件与驱动。
+- msprobe支持的固件驱动版本与配套CANN软件支持的固件驱动版本相同，开发者可通过“[昇腾社区-固件与驱动](https://www.hiascend.com/hardware/firmware-drivers/community?product=2&model=28&cann=8.0.RC3.alpha003&driver=1.0.25.alpha)”页面根据产品型号与CANN软件版本获取配套的固件与驱动。
 
 
 ## 🚨 工具限制与注意事项

{mindstudio_probe-8.2.0 → mindstudio_probe-8.2.1}/msprobe/core/common/const.py

@@ -274,6 +274,7 @@ class Const:
     TENSOR_TYPE = "torch.Tensor"
     DTENSOR_TYPE = "torch.distributed.tensor.DTensor"
     FAKE_TENSOR_TYPE = "torch._subclasses.fake_tensor.FakeTensor"
+    AC_TENSOR_TYPE = "torch.distributed._functional_collectives.AsyncCollectiveTensor"
 
     SUPPORT_API_FILE_NAME = "support_wrap_ops.yaml"
 
@@ -405,6 +406,8 @@ class Const:
 
     MIX_DUMP_NAMES = {'graph', 'pynative'}
 
+    MEGATRON_MICRO_STEP_NUMBER = 'megatron_micro_step_number'
+
 
 class CompareConst:
     """
@@ -425,6 +428,8 @@ class CompareConst:
     NPU_MIN = "NPU min"
     NPU_MEAN = "NPU mean"
     NPU_NORM = "NPU l2norm"
+    NPU_P2POP_PEER = "NPU P2POp peer"
+
     BENCH_MAX = "Bench max"
     BENCH_MIN = "Bench min"
     BENCH_MEAN = "Bench mean"
@@ -823,6 +828,7 @@ class MonitorConst:
     ACTV_OUT = "output"
     ACTVGRAD_IN = "input_grad"
     ACTVGRAD_OUT = "output_grad"
+    FSDP_FLAT_SEP = "_fsdp_wrapped_module."
     # used for tasks
     ACTV = "actv"
     ACTVGRAD = "actv_grad"

{mindstudio_probe-8.2.0 → mindstudio_probe-8.2.1}/msprobe/core/common/db_manager.py

@@ -12,6 +12,7 @@
 # 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 re
 import sqlite3
 from typing import List, Tuple, Dict, Any
 from functools import wraps
@@ -20,6 +21,14 @@ from msprobe.pytorch.common.log import logger
 from msprobe.core.common.file_utils import check_path_before_create, change_mode
 from msprobe.core.common.const import FileCheckConst
 
+SAFE_SQL_PATTERN = re.compile(r'^[a-zA-Z0-9_]+$')
+
+
+def check_identifier_safety(name):
+    """验证标识符是否安全（防止SQL注入）"""
+    if not isinstance(name, str) or SAFE_SQL_PATTERN.match(name) is None:
+        raise ValueError(f"Invalid SQL identifier: {name}, potential SQL injection risk!")
+
 
 def _db_operation(func):
     """数据库操作装饰器，自动管理连接"""
@@ -68,6 +77,7 @@ class DBManager:
         where_values = []
         if where_list:
             for col, val in where_list.items():
+                check_identifier_safety(col)
                 where_clauses.append(f"{col} = ?")
                 where_values.append(val)
             if where_clauses:
@@ -84,13 +94,22 @@ class DBManager:
         :param batch_size: 每批插入的大小
         :return: 插入的行数
         """
+        check_identifier_safety(table_name)
+
         if not data:
             return 0
         columns = len(data[0])
-        if key_list and columns != len(key_list):
-            raise ValueError(
-                f"When inserting into table {table_name}, the length of key list ({key_list})"
-                f"does not match the data({columns}).")
+        if key_list:
+            if not isinstance(key_list, list):
+                raise TypeError(
+                    f"key_list must be a list, got {type(key_list)}"
+                )
+            if columns != len(key_list):
+                raise ValueError(
+                    f"When inserting into table {table_name}, the length of key list ({key_list})"
+                    f"does not match the data({columns}).")
+            for key in key_list:
+                check_identifier_safety(key)
 
         batch_size = self.DEFAULT_INSERT_SIZE
         placeholders = ", ".join(["?"] * columns)
@@ -121,12 +140,16 @@ class DBManager:
         :param where: WHERE条件
         :return: 查询结果列表(字典形式)
         """
+        check_identifier_safety(table_name)
 
         if not columns:
             raise ValueError("columns parameter cannot be empty, specify columns to select (e.g. ['id', 'name'])")
         if not isinstance(columns, list) or not all(isinstance(col, str) for col in columns):
             raise TypeError("columns must be a list of strings (e.g. ['id', 'name'])")
         
+        for col in columns:
+            check_identifier_safety(col)
+        
         cols = ", ".join(columns)
         sql = f"SELECT {cols} FROM {table_name}"
 
@@ -147,6 +170,14 @@ class DBManager:
         :param where_params: WHERE条件参数
         :return: 影响的行数
         """
+        check_identifier_safety(table_name)
+        if not updates:
+            raise ValueError("columns parameter cannot be empty, specify it to update (e.g. {'name': 'xxx'}")
+        if not isinstance(updates, dict):
+            raise TypeError(f"updates must be a dictionary, got: {type(updates)}")
+        for key in updates.keys():
+            check_identifier_safety(key)
+
         set_clause = ", ".join([f"{k} = ?" for k in updates.keys()])
         sql = f"UPDATE {table_name} SET {set_clause}"
 

{mindstudio_probe-8.2.0 → mindstudio_probe-8.2.1}/msprobe/core/common/file_utils.py

@@ -33,7 +33,7 @@ import pandas as pd
 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
+from msprobe.core.common.const import FileCheckConst, CompareConst, Const
 from msprobe.core.common.global_lock import global_lock, is_main_process
 
 proc_lock = multiprocessing.Lock()
@@ -172,7 +172,7 @@ def check_path_exists(path):
     if not os.path.exists(path):
         logger.error('The file path %s does not exist.' % path)
         raise FileCheckException(FileCheckException.ILLEGAL_PATH_ERROR)
-    
+
 
 def check_path_not_exists(path):
     if os.path.exists(path):
@@ -259,8 +259,8 @@ def check_path_type(file_path, file_type):
 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
+            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
     )
     return is_writable
 
@@ -319,7 +319,7 @@ def check_dirpath_before_read(path):
             check_path_owner_consistent(dirpath)
         except FileCheckException:
             logger.warning(f"The directory {dirpath} is not yours.")
-    
+
 
 def check_file_or_directory_path(path, isdir=False):
     """
@@ -422,6 +422,26 @@ def load_json(json_path):
     return data
 
 
+def load_construct_json(json_path):
+    construct_dict_o = load_json(json_path)
+    if Const.MEGATRON_MICRO_STEP_NUMBER in construct_dict_o:
+        construct_dict = {}
+        micro_step_dict = {Const.MEGATRON_MICRO_STEP_NUMBER: construct_dict_o.get(Const.MEGATRON_MICRO_STEP_NUMBER)}
+        del construct_dict_o[Const.MEGATRON_MICRO_STEP_NUMBER]
+        for key, value in construct_dict_o.items():
+            if isinstance(value, list):
+                if len(value) != 2:
+                    logger.error(f'Parse construct json file "{os.path.basename(json_path)}" failed.')
+                    raise RuntimeError()
+                construct_dict[key] = value[0]
+                micro_step_dict[key] = value[1]
+            else:
+                construct_dict[key] = value
+                micro_step_dict[key] = 0
+        return construct_dict, micro_step_dict
+    return construct_dict_o, {}
+
+
 def save_json(json_path, data, indent=None, mode="w"):
     check_path_before_create(json_path)
     json_path = os.path.realpath(json_path)
@@ -520,6 +540,9 @@ def move_directory(src_path, dst_path):
     check_file_or_directory_path(src_path, isdir=True)
     check_path_before_create(dst_path)
     try:
+        if os.path.exists(dst_path):
+            logger.warning(f"The destination directory {dst_path} already exists, it will be removed.")
+            shutil.rmtree(dst_path)
         shutil.move(src_path, dst_path)
     except Exception as e:
         logger.error(f"move directory {src_path} to {dst_path} failed")

mindstudio_probe-8.2.1/msprobe/core/common/megatron_utils.py

@@ -0,0 +1,59 @@
+# Copyright (c) 2024-2025, 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 functools import wraps
+
+
+class MegatronStepInfo:
+    is_megatron = False
+    is_forward = False
+    is_backward = False
+    forward_micro_step = -1
+    backward_micro_step = -1
+
+    @classmethod
+    def reset(cls):
+        """重置所有类属性到初始状态"""
+        cls.is_megatron = False
+        cls.is_forward = False
+        cls.is_backward = False
+        cls.forward_micro_step = -1
+        cls.backward_micro_step = -1
+
+
+def wrap_megatron_step(func, is_forward=True):
+    @wraps(func)
+    def wrapped_func(*args, **kwargs):
+        if not MegatronStepInfo.is_megatron:
+            MegatronStepInfo.is_megatron = True
+        if is_forward:
+            MegatronStepInfo.is_forward = True
+            MegatronStepInfo.is_backward = False
+            MegatronStepInfo.forward_micro_step += 1
+        else:
+            MegatronStepInfo.is_forward = False
+            MegatronStepInfo.is_backward = True
+            MegatronStepInfo.backward_micro_step += 1
+        return func(*args, **kwargs)
+
+    return wrapped_func
+
+
+def get_micro_step():
+    return MegatronStepInfo.forward_micro_step if MegatronStepInfo.is_forward else MegatronStepInfo.backward_micro_step
+
+
+def is_megatron():
+    return MegatronStepInfo.is_megatron

{mindstudio_probe-8.2.0 → mindstudio_probe-8.2.1}/msprobe/core/common/utils.py

@@ -28,7 +28,7 @@ import numpy as np
 from msprobe.core.common.const import Const, CompareConst
 from msprobe.core.common.decorator import recursion_depth_decorator
 from msprobe.core.common.exceptions import MsprobeException
-from msprobe.core.common.file_utils import (FileOpen, check_file_or_directory_path, load_json)
+from msprobe.core.common.file_utils import (FileOpen, check_file_or_directory_path, load_json, load_construct_json)
 from msprobe.core.common.log import logger
 
 device = collections.namedtuple('device', ['type', 'index'])
@@ -83,7 +83,8 @@ class MsprobeBaseException(Exception):
     INVALID_API_NAME_ERROR = 36
     CROSS_FRAME_ERROR = 37
     MISSING_THRESHOLD_ERROR = 38
-    WRONG_THRESHOLD_ERROR = 38
+    WRONG_THRESHOLD_ERROR = 39
+    MULTIPROCESS_ERROR = 40
 
     def __init__(self, code, error_info: str = ""):
         super(MsprobeBaseException, self).__init__()
@@ -348,8 +349,18 @@ def get_stack_construct_by_dump_json_path(dump_json_path):
     stack_json = os.path.join(directory, "stack.json")
     construct_json = os.path.join(directory, "construct.json")
 
+    stack_json_exist = os.path.exists(stack_json)
+    construct_json_exist = os.path.exists(construct_json)
+
+    if not stack_json_exist and not construct_json_exist:
+        logger.info("stack.json and construct.json not found")
+        return {}, {}
+    if not stack_json_exist or not construct_json_exist:
+        logger.error("stack.json or construct.json not found, please check.")
+        raise CompareException(CompareException.INVALID_PATH_ERROR)
+
     stack = load_json(stack_json)
-    construct = load_json(construct_json)
+    construct, _ = load_construct_json(construct_json)
     return stack, construct
 
 

{mindstudio_probe-8.2.0 → mindstudio_probe-8.2.1}/msprobe/core/compare/diff_analyze/first_diff_analyze.py

@@ -26,6 +26,8 @@ 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')
 
@@ -51,7 +53,7 @@ class FirstDiffAnalyze:
                 return True
         return False
 
-    def single_api_check(self, result_slice, header):
+    def single_api_check(self, result_slice, header, api_name=None):
         """
         单个api差异检查
 
@@ -65,14 +67,18 @@ 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:
@@ -117,7 +123,13 @@ 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]
-                check_result[api_batch.api_name] = self.single_api_check(result_slice, header)
+                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)
                 progress_bar.update(1)
 
         return check_result

mindstudio_probe-8.2.1/msprobe/core/compare/diff_analyze/ignore_op_list.yaml

@@ -0,0 +1,3 @@
+npu_fusion_attention:
+  - 4
+  - 5

{mindstudio_probe-8.2.0 → mindstudio_probe-8.2.1}/msprobe/core/compare/find_first/analyzer.py

@@ -47,7 +47,6 @@ class DiffAnalyzer:
             analyze_func()
             if self._diff_nodes:
                 self._gen_analyze_info()
-                self._post_process()
                 return
         logger.info('Cannot find any diff node, no need to generate analyze file.')
 
@@ -56,12 +55,6 @@ class DiffAnalyzer:
         self._resolve_input_path(self._output_path)
         logger.info("Pre Process completed.")
 
-    def _post_process(self):
-        for rank_path in self._paths.values():
-            dump_path = rank_path.dump_path
-            logger.debug(f"Remove {dump_path} success")
-        logger.info("Post Process completed.")
-
     """
     这里需要生成stack，但是直接用dict中自带就行，在op_items.NPU_Stack_Info中
     """
@@ -105,6 +98,8 @@ class DiffAnalyzer:
                 logger.warning(f'Rank {path.rank} has no dump data!')
                 continue
             for op_name, op_data in dump_data.items():
+                if is_ignore_op(op_name):
+                    continue
                 if is_communication_op(op_name):
                     self._first_comm_nodes[path.rank] = op_name
                     break
@@ -131,10 +126,16 @@ class DiffAnalyzer:
         for rank, nodes in list(self._rank_comm_nodes_dict.items())[:-1]:
             searched_ranks.add(rank)
             seen_nodes = set()
+            last_node = None
             for cur_node in nodes.values():
+                is_overflow = last_node and hasattr(last_node, 'layer') and hasattr(cur_node, 'layer') and \
+                last_node.layer >= cur_node.layer
+                if is_overflow:
+                    cur_node.layer = last_node.layer + 1
                 conn_info = cur_node.find_connected_nodes()
                 if not conn_info.get('ranks'):
                     conn_info['ranks'] = self._rank_comm_nodes_dict.keys()
+                last_node = cur_node
                 if not self._find_connection(conn_info, cur_node, searched_ranks, seen_nodes):
                     logger.debug(f'Cannot find connected communication node for "{cur_node.node_id}".')
 

{mindstudio_probe-8.2.0 → mindstudio_probe-8.2.1}/msprobe/core/compare/find_first/graph.py

@@ -52,19 +52,25 @@ class DataNode:
         metrics = {}
         for cmp_data in self.op_data:
             name = cmp_data.get(CompareConst.NPU_NAME)
+            # 构建度量指标字典
+            metrics = {}
+            
             if CompareConst.NPU_MAX in cmp_data:
                 metrics = {CompareConst.NPU_MAX: cmp_data.get(CompareConst.NPU_MAX),
                         CompareConst.NPU_MIN: cmp_data.get(CompareConst.NPU_MIN),
                         CompareConst.NPU_MEAN: cmp_data.get(CompareConst.NPU_MEAN),
                         CompareConst.NPU_NORM: cmp_data.get(CompareConst.NPU_NORM)}
             elif CompareConst.NPU_MD5 in cmp_data:
-                metrics = {CompareConst.NPU_MD5: cmp_data.get(CompareConst.NPU_MD5)}
+                metrics[CompareConst.NPU_MD5] = cmp_data.get(CompareConst.NPU_MD5)
+                
+            if CompareConst.NPU_P2POP_PEER in cmp_data:
+                metrics[CompareConst.NPU_P2POP_PEER] = cmp_data.get(CompareConst.NPU_P2POP_PEER)
 
             if cmp_data.get(CompareConst.STACK) != CompareConst.N_A and not self.stack:
                 self.stack = cmp_data.get(CompareConst.STACK)
-            if Const.INPUT in name:
+            if cmp_data.get('state') == "input":
                 self.inputs[name] = metrics
-            elif Const.OUTPUT in name:
+            elif cmp_data.get('state') == "output":
                 self.outputs[name] = metrics
 
     def gen_node_info(self, path: RankPath):
@@ -161,6 +167,8 @@ class CommunicationNode:
                 if val and val.startswith('[') and val.endswith(']'):
                     val = [int(part) for part in val.strip('[]').split(',')]
                     ranks.update(val)
+            elif v.get(CompareConst.NPU_P2POP_PEER) != "None":
+                ranks.add(v.get(CompareConst.NPU_P2POP_PEER))
 
         return {'ranks': ranks, 'api': f'Distributed.{tar_api}',
                 'type': DiffAnalyseConst.OPPOSITE_DIR.get(self.type, DiffAnalyseConst.LINK)}

{mindstudio_probe-8.2.0 → mindstudio_probe-8.2.1}/msprobe/core/compare/find_first/utils.py

@@ -120,7 +120,8 @@ def is_communication_op(op_name):
 def is_ignore_op(op_name):
     ignore_keywords = [
         'Torch.empty',
-        'Torch.fill'
+        'Torch.fill',
+        'Tensor.__setitem__'
     ]
     return any(keyword in op_name for keyword in ignore_keywords)
 
@@ -181,7 +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(input_diff_nodes)
+    get_compute_ops_from_comm_nodes(nodes_group)
     # 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-8.2.0 → mindstudio_probe-8.2.1}/msprobe/core/compare/highlight.py

@@ -26,7 +26,7 @@ from tqdm import tqdm
 from msprobe.core.common.const import CompareConst, Const
 from msprobe.core.common.file_utils import save_workbook
 from msprobe.core.common.log import logger
-from msprobe.core.common.utils import get_header_index
+from msprobe.core.common.utils import get_header_index, CompareException
 from msprobe.core.compare.utils import table_value_is_valid, gen_api_batches
 from msprobe.core.compare.config import ModeConfig
 
@@ -359,18 +359,25 @@ class HighLight:
 
         def err_call(args):
             logger.error("Multiprocessing malicious value check failed! Reason: {}".format(args))
-            try:
-                pool.close()
-            except OSError:
-                logger.error("Pool terminate failed")
 
         result_df_columns = result_df.columns.tolist()
         for column in result_df_columns:
             self.value_check(column)
+        async_results = []
         for df_chunk in chunks:
-            pool.apply_async(func, args=(df_chunk, result_df_columns,), error_callback=err_call)
+            result = pool.apply_async(func, args=(df_chunk, result_df_columns,), error_callback=err_call)
+            async_results.append(result)
 
         pool.close()
+
+        for ar in async_results:
+            try:
+                ar.get(timeout=3600)
+            except Exception as e:
+                logger.error(f"Task failed with exception: {e}")
+                pool.terminate()
+                raise CompareException(CompareException.MULTIPROCESS_ERROR) from e
+
         pool.join()
 
     def df_malicious_value_check(self, result_df):

{mindstudio_probe-8.2.0 → mindstudio_probe-8.2.1}/msprobe/core/compare/multiprocessing_compute.py

@@ -52,16 +52,20 @@ def _ms_graph_handle_multi_process(func, result_df, mode):
 
     def err_call(args):
         logger.error('multiprocess compare failed! Reason: {}'.format(args))
-        try:
-            pool.close()
-        except OSError as e:
-            logger.error(f'pool terminate failed: {str(e)}')
 
     for df_chunk in df_chunks:
         result = pool.apply_async(func, args=(df_chunk, mode), error_callback=err_call)
         results.append(result)
-    final_results = [r.get() for r in results]
+
     pool.close()
+
+    try:
+        final_results = [r.get(timeout=3600) for r in results]
+    except Exception as e:
+        logger.error(f"Task failed with exception: {e}")
+        pool.terminate()
+        raise CompareException(CompareException.MULTIPROCESS_ERROR) from e
+
     pool.join()
     return pd.concat(final_results, ignore_index=True)
 
@@ -277,10 +281,6 @@ class CompareRealData:
 
         def err_call(args):
             logger.error('multiprocess compare failed! Reason: {}'.format(args))
-            try:
-                pool.close()
-            except OSError:
-                logger.error("pool terminate failed")
 
         progress_bar = tqdm(total=len(result_df), desc="API/Module Item Compare Process", unit="row", ncols=100)
 
@@ -298,7 +298,14 @@ class CompareRealData:
                                       )
             results.append(result)
 
-        final_results = [r.get() for r in results]
         pool.close()
+
+        try:
+            final_results = [r.get(timeout=3600) for r in results]
+        except Exception as e:
+            logger.error(f"Task failed with exception: {e}")
+            pool.terminate()
+            raise CompareException(CompareException.MULTIPROCESS_ERROR) from e
+
         pool.join()
         return pd.concat(final_results, ignore_index=True)

{mindstudio_probe-8.2.0 → mindstudio_probe-8.2.1}/msprobe/core/compare/utils.py

@@ -695,10 +695,6 @@ def get_sorted_ranks(npu_dump_dir, bench_dump_dir):
 def multi_statistics_compare(func, func_args):
     def err_call(args):
         logger.error(f'Multiprocess statistics compare failed! Reason: {args}')
-        try:
-            pool.close()
-        except OSError:
-            logger.error("Pool terminate failed")
 
     compare_func, input_param_nr_list, output_path, kwargs = func_args
 
@@ -715,9 +711,22 @@ def multi_statistics_compare(func, func_args):
             chunks[i].append(input_param_nr_list[param_num - remainder + i])
 
     pool = multiprocessing.Pool(process_num)
+
+    async_results = []
     for chunk in chunks:
-        pool.apply_async(func, args=(compare_func, chunk, output_path, kwargs), error_callback=err_call)
+        result = pool.apply_async(func, args=(compare_func, chunk, output_path, kwargs), error_callback=err_call)
+        async_results.append(result)
+
     pool.close()
+
+    for ar in async_results:
+        try:
+            ar.get(timeout=3600)
+        except Exception as e:
+            logger.error(f"Task failed with exception: {e}")
+            pool.terminate()
+            raise CompareException(CompareException.MULTIPROCESS_ERROR) from e
+
     pool.join()