mindstudio-probe 1.2.1__py3-none-any.whl → 1.3.0__py3-none-any.whl

msprobe/README.md

@@ -15,7 +15,7 @@ debugger = PrecisionDebugger(config_path='./config.json')
 ...
 debugger.start() # 一般在训练循环开头启动工具
 ... # 循环体
-debugger.stop() # 一般在训练循环末尾结束工具
+debugger.stop() # 一般在训练循环末尾结束工具。必须调用，否则可能导致精度数据落盘不全
 debugger.step() # 在训练循环的最后需要重置工具，非循环场景不需要
 ```
 
@@ -44,6 +44,7 @@ export MSPROBE_LOG_LEVEL={x}
 
 - 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。
 
 
@@ -69,35 +70,37 @@ export MSPROBE_LOG_LEVEL={x}
 
 ### 1 数据采集
 
-msprobe 通过在训练脚本中添加 PrecisionDebugger 接口的方式对 API 执行精度数据 dump 操作，对应 config.json 中的 task 为 statistics 或 tensor。
+msprobe 通过在训练脚本中添加 PrecisionDebugger 接口的方式对 API 执行精度数据 dump 操作。对应 config.json 中的 "statistics" 或 "tensor" task。
 
 [PyTorch 场景的数据采集](./docs/05.data_dump_PyTorch.md)
 
 [MindSpore 场景的数据采集](./docs/06.data_dump_MindSpore.md)
 
+[MSAdapter 场景的数据采集](./docs/29.data_dump_MSAdapter.md)
+
 ### 2 精度预检
 
-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 task 为 run_ut。
+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)和[在线预检](./docs/08.accuracy_checker_online_PyTorch.md)
 
 MindSpore 动态图场景的[离线预检](./docs/09.accuracy_checker_MindSpore.md)
 
-### 3 精度比对
+### 3 分级可视化构图比对
 
-该功能进行 PyTorch 整网 API 粒度的数据 dump、精度比对，进而定位训练场景下的精度问题。
+该功能将msprobe工具dump的精度数据进行解析，还原模型图结构，实现模型各个层级的精度数据比对，方便用户理解模型结构、分析精度问题。
 
-[PyTorch 场景的精度比对](./docs/10.accuracy_compare_PyTorch.md)
+[PyTorch 场景的分级可视化构图比对](./docs/21.visualization_PyTorch.md)
 
-[MindSpore 场景的精度比对](./docs/11.accuracy_compare_MindSpore.md)
+[MindSpore 场景的分级可视化构图比对](./docs/22.visualization_MindSpore.md)
 
-### 4 溢出检测与解析
+### 4 精度比对
 
-溢出检测与解析是在执行精度数据 dump 时，判断是否存在输入正常但输出存在溢出的 API，从而判断是否为正常溢出。对应 config.json 中的 overflow_check。
+该功能进行 PyTorch 整网 API 粒度的数据 dump、精度比对，进而定位训练场景下的精度问题。
 
-[PyTorch 场景的溢出检测与解析](./docs/12.overflow_check_PyTorch.md)
+[PyTorch 场景的精度比对](./docs/10.accuracy_compare_PyTorch.md)
 
-[MindSpore 场景的溢出检测与解析](./docs/13.overflow_check_MindSpore.md)
+[MindSpore 场景的精度比对](./docs/11.accuracy_compare_MindSpore.md)
 
 ### 5 数据解析
 
@@ -127,28 +130,30 @@ MindSpore 动态图场景的[离线预检](./docs/09.accuracy_checker_MindSpore.
 
 该功能收集和聚合模型训练过程中的网络层，优化器， 通信算子的中间值，帮助诊断模型训练过程中计算， 通信，优化器各部分出现的异常情况。
 
-[PyTorch 场景的训练状态监控](./docs/19.monitor.md)
+[兼容 PyTorch 和 MindSpore 框架的训练状态监控](./docs/19.monitor.md)
 
-### 10 分级可视化构图比对
+### 10 单算子API自动生成脚本
 
-该功能将msprobe工具dump的精度数据进行解析，还原模型图结构，实现模型各个层级的精度数据比对，方便用户理解模型结构、分析精度问题。
+该功能将msprobe工具dump的精度数据进行解析，自动生成单API脚本，用于复现整网中出现的算子问题，降低用户复现问题的成本，供开发分析算子问题。
 
-[PyTorch 场景的分级可视化构图比对](./docs/21.visualization_PyTorch.md)
+[PyTorch 单算子API自动生成脚本](./docs/23.generate_operator_PyTorch.md)
 
-[MindSpore 场景的分级可视化构图比对](./docs/22.visualization_MindSpore.md)
+### 11 数码关联
 
+该功能只支持 MindSpore 静态图场景，用于将IR图与dump数据进行关联，获取dump数据和代码调用栈的关联关系。
 
-### 11 单算子API自动生成脚本
+[MindSpore 场景的数码关联](./docs/24.code_mapping_Mindspore.md)
 
-该功能将msprobe工具dump的精度数据进行解析，自动生成单API脚本，用于复现整网中出现的算子问题，降低用户复现问题的成本，供开发分析算子问题。
+### 12 溢出检测与解析
 
-[PyTorch 单算子API自动生成脚本](./docs/23.generate_operator_PyTorch.md)
+溢出检测用于采集溢出 API 或 模块的精度数据，而溢出解析则是通过对溢出数据的分析，进一步判断是否为正常溢出。对应 config.json 中的 "overflow_check" task。 
+推荐直接使用[数据采集](#1-数据采集)功能采集统计量信息,检测溢出问题。
 
-### 12 数码关联
+[PyTorch 场景的溢出检测与解析](./docs/12.overflow_check_PyTorch.md)
 
-该功能只支持 MindSpore 静态图场景，用于将IR图与dump数据进行关联，获取dump数据和代码调用栈的关联关系。
+[MindSpore 场景的溢出检测](./docs/13.overflow_check_MindSpore.md)
 
-[MindSpore 场景的数码关联](./docs/24.code_mapping_Mindspore.md)
+[MSAdapter 场景的溢出检测](./docs/30.overflow_check_MSAdapter.md)
 
 ## 📑 补充材料
 

msprobe/core/common/const.py

@@ -51,7 +51,10 @@ class Const:
     FOUR_SEGMENT = 4
     SIX_SEGMENT = 6
     SEVEN_SEGMENT = 7
+
     MAX_DEPTH = 10
+    CPU_QUARTER = 4
+    DUMP_MAX_DEPTH = 50
 
     # dump mode
     ALL = "all"
@@ -103,14 +106,16 @@ class Const:
     FREE_BENCHMARK = "free_benchmark"
     RUN_UT = "run_ut"
     GRAD_PROBE = "grad_probe"
-    TASK_LIST = [TENSOR, STATISTICS, OVERFLOW_CHECK, FREE_BENCHMARK, RUN_UT, GRAD_PROBE]
-    DUMP_DATA_COLLECTION_LIST = [STATISTICS, TENSOR]
+    STRUCTURE = "structure"
+    TASK_LIST = [TENSOR, STATISTICS, OVERFLOW_CHECK, FREE_BENCHMARK, RUN_UT, GRAD_PROBE, STRUCTURE]
+    DUMP_DATA_COLLECTION_LIST = [STATISTICS, TENSOR, STRUCTURE]
     DUMP_DATA_MODE_LIST = [ALL, INPUT, OUTPUT, FORWARD, BACKWARD]
     LEVEL_L0 = "L0"
     LEVEL_L1 = "L1"
     LEVEL_L2 = "L2"
     LEVEL_MIX = "mix"
-    LEVEL_LIST = [LEVEL_L0, LEVEL_L1, LEVEL_L2, LEVEL_MIX]
+    LEVEL_DEBUG = "debug"
+    LEVEL_LIST = [LEVEL_L0, LEVEL_L1, LEVEL_L2, LEVEL_MIX, LEVEL_DEBUG]
     ATTR_NAME_PREFIX = "wrap_"
     ATTR_NAME_PREFIX_LEN = len(ATTR_NAME_PREFIX)
     KERNEL_DUMP = "kernel_dump"
@@ -228,6 +233,92 @@ class Const:
 
     TENSOR_STAT_LEN = 2
 
+    SUPPORT_API_FILE_NAME = "support_wrap_ops.yaml"
+
+    PT_API_TYPE_FUNCTIONAL = "functional"
+    PT_API_TYPE_TENSOR = "tensor"
+    PT_API_TYPE_TORCH = "torch"
+    PT_API_TYPE_VF = "_VF"
+    PT_API_TYPE_NPU = "torch_npu"
+    PT_API_TYPE_ATEN = "aten"
+    PT_API_TYPE_DIST = "distributed"
+    PT_API_TYPE_NPU_DIST = "npu_distributed"
+
+    MS_API_TYPE_OPS = "ops"
+    MS_API_TYPE_TENSOR = "tensor"
+    MS_API_TYPE_STUB_TENSOR = "stubtensor"
+    MS_API_TYPE_MINT = "mint.ops"
+    MS_API_TYPE_MINT_FUNC = "mint.nn.functional"
+    MS_API_TYPE_COM = "communication.comm_func"
+
+    FUNCTIONAL_API_TYPE_PREFIX = "Functional"
+    TENSOR_API_TYPE_PREFIX = "Tensor"
+    DIST_API_TYPE_PREFIX = "Distributed"
+
+    TORCH_API_TYPE_PREFIX = "Torch"
+    NPU_API_TYPE_PREFIX = "NPU"
+    ATEN_API_TYPE_PREFIX = "Aten"
+    VF_API_TYPE_PREFIX = "VF"
+
+    MINT_API_TYPE_PREFIX = "Mint"
+    MINT_FUNC_API_TYPE_PREFIX = "MintFunctional"
+
+    SUPPORT_API_DICT_KEY_MAP = {
+        PT_FRAMEWORK: {
+            PT_API_TYPE_FUNCTIONAL: PT_API_TYPE_FUNCTIONAL,
+            PT_API_TYPE_TENSOR: PT_API_TYPE_TENSOR,
+            PT_API_TYPE_TORCH: PT_API_TYPE_TORCH,
+            PT_API_TYPE_VF: PT_API_TYPE_VF,
+            PT_API_TYPE_NPU: PT_API_TYPE_NPU,
+            PT_API_TYPE_ATEN: PT_API_TYPE_ATEN,
+            PT_API_TYPE_DIST: PT_API_TYPE_DIST,
+            PT_API_TYPE_NPU_DIST: PT_API_TYPE_NPU_DIST
+        },
+        MS_FRAMEWORK: {
+            MS_API_TYPE_OPS: MS_API_TYPE_OPS,
+            MS_API_TYPE_TENSOR: MS_API_TYPE_TENSOR,
+            MS_API_TYPE_STUB_TENSOR: MS_API_TYPE_TENSOR,
+            MS_API_TYPE_MINT: MS_API_TYPE_MINT,
+            MS_API_TYPE_MINT_FUNC: MS_API_TYPE_MINT_FUNC,
+            MS_API_TYPE_COM: MS_API_TYPE_COM
+        },
+        MT_FRAMEWORK: {
+            PT_API_TYPE_FUNCTIONAL: PT_API_TYPE_FUNCTIONAL,
+            PT_API_TYPE_TENSOR: PT_API_TYPE_TENSOR,
+            PT_API_TYPE_TORCH: PT_API_TYPE_TORCH,
+            PT_API_TYPE_NPU: PT_API_TYPE_NPU,
+            PT_API_TYPE_DIST: PT_API_TYPE_DIST
+        }
+    }
+
+    API_DATA_PREFIX = {
+        PT_FRAMEWORK: {
+            PT_API_TYPE_FUNCTIONAL: FUNCTIONAL_API_TYPE_PREFIX,
+            PT_API_TYPE_TENSOR: TENSOR_API_TYPE_PREFIX,
+            PT_API_TYPE_TORCH: TORCH_API_TYPE_PREFIX,
+            PT_API_TYPE_VF: VF_API_TYPE_PREFIX,
+            PT_API_TYPE_NPU: NPU_API_TYPE_PREFIX,
+            PT_API_TYPE_ATEN: ATEN_API_TYPE_PREFIX,
+            PT_API_TYPE_DIST: DIST_API_TYPE_PREFIX,
+            PT_API_TYPE_NPU_DIST: DIST_API_TYPE_PREFIX
+        },
+        MS_FRAMEWORK: {
+            MS_API_TYPE_OPS: FUNCTIONAL_API_TYPE_PREFIX,
+            MS_API_TYPE_TENSOR: TENSOR_API_TYPE_PREFIX,
+            MS_API_TYPE_STUB_TENSOR: TENSOR_API_TYPE_PREFIX,
+            MS_API_TYPE_MINT: MINT_API_TYPE_PREFIX,
+            MS_API_TYPE_MINT_FUNC: MINT_FUNC_API_TYPE_PREFIX,
+            MS_API_TYPE_COM: DIST_API_TYPE_PREFIX
+        },
+        MT_FRAMEWORK: {
+            PT_API_TYPE_FUNCTIONAL: FUNCTIONAL_API_TYPE_PREFIX,
+            PT_API_TYPE_TENSOR: TENSOR_API_TYPE_PREFIX,
+            PT_API_TYPE_TORCH: TORCH_API_TYPE_PREFIX,
+            PT_API_TYPE_NPU: NPU_API_TYPE_PREFIX,
+            PT_API_TYPE_DIST: DIST_API_TYPE_PREFIX
+        }
+    }
+
 
 class CompareConst:
     """
@@ -254,6 +345,7 @@ class CompareConst:
     MEAN_DIFF = "Mean diff"
     NORM_DIFF = "L2norm diff"
     COSINE = "Cosine"
+    EUC_DIST = "EucDist"
     MAX_ABS_ERR = "MaxAbsErr"
     MAX_RELATIVE_ERR = "MaxRelativeErr"
     MIN_RELATIVE_ERR = "MinRelativeErr"
@@ -328,8 +420,8 @@ class CompareConst:
     ULP_ERR_STATUS = "ulp_err_status"
 
     COMPARE_RESULT_HEADER = [
-        NPU_NAME, BENCH_NAME, NPU_DTYPE, BENCH_DTYPE, NPU_SHAPE, BENCH_SHAPE, COSINE, MAX_ABS_ERR, MAX_RELATIVE_ERR,
-        ONE_THOUSANDTH_ERR_RATIO, FIVE_THOUSANDTHS_ERR_RATIO,
+        NPU_NAME, BENCH_NAME, NPU_DTYPE, BENCH_DTYPE, NPU_SHAPE, BENCH_SHAPE, COSINE, EUC_DIST,
+        MAX_ABS_ERR, MAX_RELATIVE_ERR, ONE_THOUSANDTH_ERR_RATIO, FIVE_THOUSANDTHS_ERR_RATIO,
         NPU_MAX, NPU_MIN, NPU_MEAN, NPU_NORM, BENCH_MAX, BENCH_MIN, BENCH_MEAN, BENCH_NORM, ACCURACY, ERROR_MESSAGE
     ]
 
@@ -355,18 +447,16 @@ class CompareConst:
         Const.MD5: MD5_COMPARE_RESULT_HEADER
     }
 
-    ALL_COMPARE_INDEX = [COSINE, MAX_ABS_ERR, MAX_RELATIVE_ERR, ONE_THOUSANDTH_ERR_RATIO, FIVE_THOUSANDTHS_ERR_RATIO]
+    ALL_COMPARE_INDEX = [COSINE, EUC_DIST, MAX_ABS_ERR, MAX_RELATIVE_ERR, ONE_THOUSANDTH_ERR_RATIO,
+                         FIVE_THOUSANDTHS_ERR_RATIO]
     SUMMARY_COMPARE_INDEX = [MAX_DIFF, MIN_DIFF, MEAN_DIFF, NORM_DIFF,
                              MAX_RELATIVE_ERR, MIN_RELATIVE_ERR, MEAN_RELATIVE_ERR, NORM_RELATIVE_ERR]
 
     # dtype match
-    MS_TYPE = [
-        [Const.FLOAT16, Const.FLOAT32], [Const.FLOAT32, Const.FLOAT16],
-        [Const.FLOAT16, Const.BFLOAT16], [Const.BFLOAT16, Const.FLOAT16]
-    ]
-    TORCH_TYPE = [
-        [Const.TORCH_FLOAT16, Const.TORCH_FLOAT32], [Const.TORCH_FLOAT32, Const.TORCH_FLOAT16],
-        [Const.TORCH_FLOAT16, Const.TORCH_BFLOAT16], [Const.TORCH_BFLOAT16, Const.TORCH_FLOAT16]
+
+    DTYPE_MATCH_GROUPS = [
+        {Const.FLOAT16, Const.FLOAT32, Const.BFLOAT16},
+        {Const.TORCH_FLOAT16, Const.TORCH_FLOAT32, Const.TORCH_BFLOAT16}
     ]
 
     # read_op
@@ -465,7 +555,7 @@ class CompareConst:
         BENCH_MEAN: None, BENCH_NORM: None, ACCURACY: '', ERROR_MESSAGE: ''
     }
     MS_GRAPH_NPY = {
-        COSINE: None, MAX_ABS_ERR: None, MAX_RELATIVE_ERR: None, ONE_THOUSANDTH_ERR_RATIO: None,
+        COSINE: None, EUC_DIST: None, MAX_ABS_ERR: None, MAX_RELATIVE_ERR: None, ONE_THOUSANDTH_ERR_RATIO: None,
         FIVE_THOUSANDTHS_ERR_RATIO: None
     }
     MS_GRAPH_STATISTIC = {
@@ -536,46 +626,6 @@ class OverflowConst:
     OVERFLOW_DEBUG_MODE = 1
 
 
-class MsCompareConst:
-    # api_info field
-    MINT = "Mint"
-    MINT_FUNCTIONAL = "MintFunctional"
-    TENSOR_API = "Tensor"
-
-    API_NAME_STR_LENGTH = 4
-
-    TASK_FIELD = "task"
-    STATISTICS_TASK = "statistics"
-    TENSOR_TASK = "tensor"
-    DUMP_DATA_DIR_FIELD = "dump_data_dir"
-    DATA_FIELD = "data"
-
-    # supported api yaml
-    SUPPORTED_API_LIST_FILE = "checker_support_api.yaml"
-    SUPPORTED_TENSOR_LIST_KEY = "tensor"
-
-    # detail_csv
-    DETAIL_CSV_API_NAME = "API Name"
-    DETAIL_CSV_BENCH_DTYPE = "Bench Dtype"
-    DETAIL_CSV_TESTED_DTYPE = "Tested Dtype"
-    DETAIL_CSV_SHAPE = "Shape"
-    DETAIL_CSV_PASS_STATUS = "Status"
-    DETAIL_CSV_MESSAGE = "Message"
-    DETAIL_CSV_FILE_NAME = "accuracy_checking_details"
-
-    # result_csv
-    RESULT_CSV_FORWARD_TEST_SUCCESS = "Forward Test Success"
-    RESULT_CSV_BACKWARD_TEST_SUCCESS = "Backward Test Success"
-    RESULT_CSV_FILE_NAME = "accuracy_checking_result"
-
-    EPSILON = 1e-8
-
-    class ProcessStatus:
-        SUCCESS = "success"
-        API_NOT_FOUND = "api_not_found"
-        EXCEPTION_SKIP = "exception_skip"
-
-
 class MsgConst:
     """
     Class for log messages const
@@ -612,6 +662,15 @@ class MonitorConst:
     """
     Class for monitor const
     """
+
+    # monitor config set default values
+    DEFAULT_GRAD_ACC_STEPS = 1
+    DEFAULT_START_ITERATION = 0
+    DEFAULT_START_STEP = 0
+    DEFAULT_MAX_COLLECT_TIMES = 1e8
+    DEFAULT_MIN_COLLECT_TIMES = 0
+    DEFAULT_STEP_INTERVAL = 1
+
     OP_LIST = ["norm", "min", "max", "zeros", "nans", "id", "mean"]
     MONITOR_OUTPUT_DIR = "MONITOR_OUTPUT_DIR"
     DEFAULT_MONITOR_OUTPUT_DIR = "./monitor_output"
@@ -623,29 +682,39 @@ class MonitorConst:
         "DeepSpeedZeroOptimizer_Stage1_or_2",
         "DeepSpeedZeroOptimizer_Stage3"
     )
+    DEEPSPEED_ZERO_OPT_FILTER = "DeepSpeedZeroOptimizer"
     RULE_NAME = ['AnomalyTurbulence']
 
     SLICE_SIZE = 20480
+    # used for name
     DOT = "."
-    VPP_SEP = ":"
+    NAME_SEP = ":"
+    INPUT_GRAD = "input_grad"
+    OUTPUT_GRAD = "output_grad"
     ACTV_IN = "input"
     ACTV_OUT = "output"
     ACTVGRAD_IN = "input_grad"
     ACTVGRAD_OUT = "output_grad"
+    # used for tasks
+    ACTV = "actv"
+    ACTVGRAD = "actv_grad"
     POST_GRAD = "post_grad"
     PRE_GRAD = "pre_grad"
     ACC_GRAD = "acc_grad"
     PREFIX_POST = "post"
     PREFIX_PRE = "pre"
-    OUTPUT_DIR_PATTERN = r"([\w-]{0,20})-rank(\d{1,5})-"
-
     EXP_AVG = "exp_avg"
-    EFXP_AVG_SQ = "efxp_avg_sq"
+    EXP_AVG_SQ = "exp_avg_sq"
+    PARAM = "param"
 
+    CSV_HEADER = ["vpp_stage", "name", "step"]
+    CSV_HEADER_XY = ["vpp_stage", "name", "step", "micro_step"]
+    OUTPUT_DIR_PATTERN = r"([\w-]{0,20})-rank(\d{1,5})-"
     ANOMALY_JSON = "anomaly.json"
     ANALYSE_JSON = "anomaly_analyse.json"
     TENSORBOARD = "tensorboard"
     CSV = "csv"
     API = "api"
-    OPS_START_INDEX = 3
-    HEADER_NAME_INDEX = 1
+    HEADER_NAME = 'name'
+
+    MAX_NDIGITS = 20

msprobe/core/common/decorator.py

@@ -0,0 +1,50 @@
+# 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 collections import defaultdict
+from functools import wraps
+
+from msprobe.core.common.const import Const
+from msprobe.core.common.exceptions import MsprobeException
+from msprobe.core.common.log import logger
+
+# 记录工具函数递归的深度
+recursion_depth = defaultdict(int)
+
+
+def recursion_depth_decorator(func_info, max_depth=Const.MAX_DEPTH):
+    """装饰一个函数，当函数递归调用超过限制时，抛出异常并打印函数信息。"""
+    def decorator(func):
+        @wraps(func)
+        def wrapper(*args, **kwargs):
+            func_id = id(func)
+            recursion_depth[func_id] += 1
+            if recursion_depth[func_id] > max_depth:
+                msg = f"call {func_info} exceeds the recursion limit."
+                logger.error_log_with_exp(
+                    msg,
+                    MsprobeException(
+                        MsprobeException.RECURSION_LIMIT_ERROR, msg
+                    ),
+                )
+            try:
+                result = func(*args, **kwargs)
+            finally:
+                recursion_depth[func_id] -= 1
+            return result
+
+        return wrapper
+
+    return decorator

msprobe/core/common/exceptions.py

@@ -28,12 +28,14 @@ class MsprobeException(CodedException):
     OVERFLOW_NUMS_ERROR = 1
     RECURSION_LIMIT_ERROR = 2
     INTERFACE_USAGE_ERROR = 3
+    UNSUPPORTED_TYPE_ERROR = 4
 
     err_strs = {
         INVALID_PARAM_ERROR: "[msprobe] 无效参数：",
         OVERFLOW_NUMS_ERROR: "[msprobe] 超过预设溢出次数 当前溢出次数：",
         RECURSION_LIMIT_ERROR: "[msprobe] 递归调用超过限制：",
-        INTERFACE_USAGE_ERROR: "[msprobe] Invalid interface usage: "
+        INTERFACE_USAGE_ERROR: "[msprobe] Invalid interface usage: ",
+        UNSUPPORTED_TYPE_ERROR: "[msprobe] Unsupported type: "
     }
 
 

msprobe/core/common/file_utils.py

@@ -26,6 +26,7 @@ import yaml
 import numpy as np
 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
@@ -266,6 +267,7 @@ def make_dir(dir_path):
     file_check.common_check()
 
 
+@recursion_depth_decorator('msprobe.core.common.file_utils.create_directory', max_depth=16)
 def create_directory(dir_path):
     """
     Function Description:
@@ -332,6 +334,23 @@ def change_mode(path, mode):
                                  'Failed to change {} authority. {}'.format(path, str(ex))) from ex
 
 
+@recursion_depth_decorator('msprobe.core.common.file_utils.recursive_chmod')
+def recursive_chmod(path):
+    """
+    递归地修改目录及其子目录和文件的权限，文件修改为640，路径修改为750
+
+    :param path: 要修改权限的目录路径
+    """
+    for _, dirs, files in os.walk(path):
+        for file_name in files:
+            file_path = os.path.join(path, file_name)
+            change_mode(file_path, FileCheckConst.DATA_FILE_AUTHORITY)
+        for dir_name in dirs:
+            dir_path = os.path.join(path, dir_name)
+            change_mode(dir_path, FileCheckConst.DATA_DIR_AUTHORITY)
+            recursive_chmod(dir_path)
+
+
 def path_len_exceeds_limit(file_path):
     return len(os.path.realpath(file_path)) > FileCheckConst.DIRECTORY_LENGTH or \
         len(os.path.basename(file_path)) > FileCheckConst.FILE_NAME_LENGTH
@@ -632,7 +651,7 @@ def os_walk_for_files(path, depth):
     return res
 
 
-def check_crt_valid(pem_path):
+def check_crt_valid(pem_path, is_public_key=False):
     """
     Check the validity of the SSL certificate.
 
@@ -641,6 +660,7 @@ def check_crt_valid(pem_path):
 
     Parameters:
     pem_path (str): The file path of the SSL certificate.
+    is_public_key (bool): The file is public key or not.
 
     Raises:
     RuntimeError: If the SSL certificate is invalid or expired.
@@ -649,7 +669,10 @@ def check_crt_valid(pem_path):
     try:
         with FileOpen(pem_path, "r") as f:
             pem_data = f.read()
-        cert = OpenSSL.crypto.load_certificate(OpenSSL.crypto.FILETYPE_PEM, pem_data)
+        if is_public_key:
+            cert = OpenSSL.crypto.load_publickey(OpenSSL.crypto.FILETYPE_PEM, pem_data)
+        else:
+            cert = OpenSSL.crypto.load_certificate(OpenSSL.crypto.FILETYPE_PEM, pem_data)
         pem_start = parser.parse(cert.get_notBefore().decode("UTF-8"))
         pem_end = parser.parse(cert.get_notAfter().decode("UTF-8"))
         logger.info(f"The SSL certificate passes the verification and the validity period "

msprobe/core/common/inplace_ops.yaml

@@ -250,5 +250,6 @@ inplace_distributed_op:
   - all_to_all
   - all_gather_into_tensor
   - reduce_scatter_tensor
+  - batch_isend_irecv
 
 

msprobe/core/common/utils.py

@@ -18,9 +18,7 @@ import os
 import re
 import subprocess
 import time
-from collections import defaultdict
 from datetime import datetime, timezone
-from functools import wraps
 
 import numpy as np
 
@@ -75,6 +73,7 @@ class MsprobeBaseException(Exception):
     MERGE_COMPARE_RESULT_ERROR = 33
     NAMES_STRUCTS_MATCH_ERROR = 34
     INVALID_STATE_ERROR = 35
+    INVALID_API_NAME_ERROR = 36
 
     def __init__(self, code, error_info: str = ""):
         super(MsprobeBaseException, self).__init__()
@@ -239,12 +238,18 @@ def md5_find(data):
                 for data_detail in data[key_op][api_info]:
                     if data_detail and 'md5' in data_detail:
                         return True
+            if isinstance(data[key_op][api_info], bool):
+                continue
             elif data[key_op][api_info] and 'md5' in data[key_op][api_info]:
                 return True
     return False
 
 
 def detect_framework_by_dump_json(file_path):
+    json_data = load_json(file_path)
+    framework = json_data.get("framework", None)
+    if framework in [Const.PT_FRAMEWORK, Const.MS_FRAMEWORK]:
+        return framework
     pattern_ms = r'"type":\s*"mindspore'
     pattern_pt = r'"type":\s*"torch'
     with FileOpen(file_path, 'r') as file:
@@ -277,7 +282,7 @@ def set_dump_path(input_param):
     npu_path_valid = npu_path is not None and npu_path.endswith("dump.json")
     bench_path_valid = bench_path is not None and bench_path.endswith("dump.json")
     if not npu_path_valid or not bench_path_valid:
-        logger.error(f"Please check the json path is valid. npu_path: {npu_path}, bench_path: {bench_path}")
+        logger.error(f"Please check the json path is valid and ensure that neither npu_path nor bench_path is None.")
         raise CompareException(CompareException.INVALID_PATH_ERROR)
     input_param['npu_dump_data_dir'] = os.path.join(os.path.dirname(npu_path), Const.DUMP_TENSOR_DATA)
     input_param['bench_dump_data_dir'] = os.path.join(os.path.dirname(bench_path), Const.DUMP_TENSOR_DATA)
@@ -303,6 +308,9 @@ def get_dump_mode(input_param):
     if npu_task == Const.TENSOR:
         return Const.ALL
 
+    if npu_task == Const.STRUCTURE:
+        return Const.STRUCTURE
+
     if npu_task == Const.STATISTICS:
         npu_md5_compare = md5_find(npu_json_data['data'])
         bench_md5_compare = md5_find(bench_json_data['data'])
@@ -419,6 +427,15 @@ def get_real_step_or_rank(step_or_rank_input, obj):
     return real_step_or_rank
 
 
+def check_init_step(step):
+    if not is_int(step):
+        raise MsprobeException(MsprobeException.INVALID_PARAM_ERROR,
+                        f"{step} must be an integer")
+    if not step >= 0:
+        raise MsprobeException(MsprobeException.INVALID_PARAM_ERROR,
+                f"{step} must be greater than or equal to 0")
+
+
 def check_seed_all(seed, mode, rm_dropout):
     if is_int(seed):
         if seed < 0 or seed > Const.MAX_SEED_VALUE:
@@ -462,37 +479,30 @@ def safe_get_value(container, index, container_name, key=None):
         raise MsprobeBaseException(MsprobeBaseException.INVALID_OBJECT_TYPE_ERROR) from e
 
 
-# 记录工具函数递归的深度
-recursion_depth = defaultdict(int)
-
-
-# 装饰一个函数，当函数递归调用超过限制时，抛出异常并打印函数信息。
-def recursion_depth_decorator(func_info):
-    def decorator(func):
-        @wraps(func)
-        def wrapper(*args, **kwargs):
-            func_id = id(func)
-            recursion_depth[func_id] += 1
-            if recursion_depth[func_id] > Const.MAX_DEPTH:
-                msg = f"call {func_info} exceeds the recursion limit."
-                logger.error_log_with_exp(
-                    msg,
-                    MsprobeException(
-                        MsprobeException.RECURSION_LIMIT_ERROR, msg
-                    ),
-                )
-            try:
-                result = func(*args, **kwargs)
-            finally:
-                recursion_depth[func_id] -= 1
-            return result
-
-        return wrapper
-
-    return decorator
-
-
 def check_str_param(param):
     if not re.match(Const.REGEX_PREFIX_PATTERN, param):
         logger.error('The parameter {} contains special characters.'.format(param))
         raise MsprobeBaseException(MsprobeBaseException.INVALID_CHAR_ERROR)
+
+
+class DumpPathAggregation:
+    dump_file_path = None
+    stack_file_path = None
+    construct_file_path = None
+    dump_tensor_data_dir = None
+    free_benchmark_file_path = None
+    debug_file_path = None
+
+
+def is_save_variable_valid(variable, valid_special_types, depth=0):
+    if depth > Const.DUMP_MAX_DEPTH:
+        return False
+    if isinstance(variable, valid_special_types):
+        return True
+    elif isinstance(variable, (list, tuple)):
+        return all(is_save_variable_valid(item, valid_special_types, depth + 1) for item in variable)
+    elif isinstance(variable, dict):
+        return all(isinstance(key, str) and is_save_variable_valid(value, valid_special_types, depth + 1)
+                   for key, value in variable.items())
+    else:
+        return False