mindstudio-probe 1.2.2__py3-none-any.whl → 8.1.0__py3-none-any.whl

msprobe/README.md

@@ -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。
 
 
@@ -53,7 +54,9 @@ export MSPROBE_LOG_LEVEL={x}
 
 **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](./docs/01.installation.md)
+## ⚙️ 安装
+
+请参见[安装指导说明](./docs/01.installation.md)。
 
 ## 🌟 新版本特性
 
@@ -69,35 +72,37 @@ 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
 
@@ -129,26 +134,57 @@ MindSpore 动态图场景的[离线预检](./docs/09.accuracy_checker_MindSpore.
 
 [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)
+[MindSpore 单算子API自动生成脚本](./docs/33.generate_operator_MindSpore.md)
 
+### 11 数码关联
 
-### 11 单算子API自动生成脚本
+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。
 
-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。
+[MindSpore 场景的数码关联](./docs/24.code_mapping_Mindspore.md)
 
-[PyTorch 单算子API自动生成脚本](./docs/23.generate_operator_PyTorch.md)
+### 12 溢出检测与解析
 
-### 12 数码关联
+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。
 
-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。
+[PyTorch 场景的溢出检测与解析](./docs/12.overflow_check_PyTorch.md)
+
+[MindSpore 场景的溢出检测](./docs/13.overflow_check_MindSpore.md)
+
+[MSAdapter 场景的溢出检测](./docs/30.overflow_check_MSAdapter.md)
+
+### 13 训练检查
+
+该工具主要包括：
+
+训练前或精度比对前，对比两个环境下可能影响训练精度的配置差异。
+
+[PyTorch 训练前配置检查](./docs/31.config_check.md)
+
+训练过程中或结束后，比较两个不同的checkpoint，评估模型相似度。
+
+[checkpoint比对](./docs/32.ckpt_compare.md)
+
+### 14 强化学习数据采集
+
+主要能力：
+
+灵活采集强化学习中重要关键过程数据，并支持比对。
+
+[强化学习数据采集](./docs/34.RL_collect.md)
+
+### 15 整网首个溢出节点分析
+
+多rank场景下通过dump数据找到首个出现Nan或Inf的节点。
+
+[PyTorch 场景整网首个溢出节点分析](./docs/35.nan_analyze.md)
 
-[MindSpore 场景的数码关联](./docs/24.code_mapping_Mindspore.md)
 
 ## 📑 补充材料
 

msprobe/core/__init__.py

@@ -0,0 +1,17 @@
+# Copyright (c) 2025-2025, 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 msprobe.core.single_save.single_saver import SingleSave
+from msprobe.core.single_save.single_comparator import SingleComparator

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"
@@ -67,7 +70,7 @@ class Const:
     SUMMARY = "summary"
     MD5 = "md5"
     VALUE = "value"
-    SUMMARY_MODE = [ALL, SUMMARY, MD5]
+    SUMMARY_MODE = ["statistics", "md5"]
 
     WRITE_FLAGS = os.O_WRONLY | os.O_CREAT
     WRITE_MODES = stat.S_IWUSR | stat.S_IRUSR
@@ -77,6 +80,8 @@ class Const:
     NUMPY_SUFFIX = ".npy"
     NUMPY_PATTERN = "*.npy"
     PT_SUFFIX = ".pt"
+    PY_SUFFIX = ".py"
+    INIT_PY = "init.py"
     ONE_GB = 1073741824  # 1 * 1024 * 1024 * 1024
     TEN_GB = 10737418240  # 10 * 1024 * 1024 * 1024
     ONE_MB = 1048576  # 1 * 1024 * 1024
@@ -92,6 +97,7 @@ class Const:
     GRAD_OUTPUT = 'grad_output'
     PARAMS = 'parameters'
     PARAMS_GRAD = 'parameters_grad'
+    DEBUG = 'debug'
     START = "start"
     STOP = "stop"
     ENV_ENABLE = "1"
@@ -129,6 +135,7 @@ class Const:
     NPU = 'NPU'
     NPU_LOWERCASE = 'npu'
     CPU_LOWERCASE = 'cpu'
+    GPU_LOWERCASE = 'gpu'
     CUDA_LOWERCASE = 'cuda'
     DEVICE = 'device'
     DISTRIBUTED = 'Distributed'
@@ -137,6 +144,10 @@ class Const:
     MODULE_PREFIX = ["Module", "Cell"]
     FORWARD_NAME_SUFFIX = ".forward"
 
+    DUMP_JSON_FILE = "dump_json_file"
+    DEBUG_JSON_FILE = "debug_json_file"
+    STACK_JSON_FILE = "stack_json_file"
+
     # struct json param
     ORIGIN_DATA = "origin_data"
     SCOPE = "scope"
@@ -167,6 +178,10 @@ class Const:
     TOP_LAYER = "TopLayer"
     CELL = "Cell"
     MODULE = "Module"
+    API = "api"
+    PYNATIVE_MODE = "pynative"
+    PYNATIVE_GRAPH_MODE = "pynative_graph"
+
     FRAME_FILE_LIST = ["site-packages/torch", "package/torch", "site-packages/mindspore", "package/mindspore"]
     INPLACE_LIST = [
         "broadcast", "all_reduce", "reduce", "all_gather", "gather", "scatter", "reduce_scatter",
@@ -188,7 +203,11 @@ class Const:
 
     FILL_CHAR_NUMS = 50
     TOOL_ENDS_SUCCESSFULLY = f"{TOOL_NAME} ends successfully."
+
     WITHOUT_CALL_STACK = "The call stack retrieval failed."
+    STACK_FILTER_KEYWORDS = ["msprobe/core", "msprobe/pytorch", "msprobe/mindspore"]
+    CALL_STACK_FLAG = "data_dump/api_registry"
+    NEW_STACK_FLAG = "0"
 
     STEP = "step"
     RANK = "rank"
@@ -206,12 +225,16 @@ class Const:
     TORCH_FLOAT32 = "torch.float32"
     TORCH_BFLOAT16 = "torch.bfloat16"
 
+    TYPE = 'type'
     DTYPE = 'dtype'
     SHAPE = 'shape'
+    STACK_INFO = 'stack_info'
     MAX = 'Max'
     MIN = 'Min'
     MEAN = 'Mean'
     NORM = 'Norm'
+    DATA_NAME = 'data_name'
+    TENSOR_STAT_INDEX = 'tensor_stat_index'
 
     CODE_STACK = 'Code Stack'
     OP_NAME = 'Op Name'
@@ -223,6 +246,10 @@ class Const:
     # 分隔符常量
     SCOPE_SEPARATOR = "/"
     REPLACEMENT_CHARACTER = "_"
+    PIPE_SEPARATOR = "|"
+
+    FORWARD_PATTERN = SEP + FORWARD + SEP
+    BACKWARD_PATTERN = SEP + BACKWARD + SEP
 
     OPTIMIZER = "optimizer"
     CLIP_GRAD = "clip_grad"
@@ -230,12 +257,136 @@ 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"
+    PT_API_TYPE_MINDSPEED = "mindspeed"
+
+    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"
+    MS_API_TYPE_MINT_DIST = "mint.distributed"
+
+    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"
+    MINDSPEED_API_TYPE_PREFIX = "MindSpeed"
+
+    MINT_API_TYPE_PREFIX = "Mint"
+    MINT_FUNC_API_TYPE_PREFIX = "MintFunctional"
+    MINT_DIST_API_TYPE_PREFIX = "MintDistributed"
+
+    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,
+            PT_API_TYPE_MINDSPEED: PT_API_TYPE_MINDSPEED
+        },
+        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,
+            MS_API_TYPE_MINT_DIST: MS_API_TYPE_MINT_DIST
+        },
+        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,
+            PT_API_TYPE_MINDSPEED: MINDSPEED_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,
+            MS_API_TYPE_MINT_DIST: MINT_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
+        }
+    }
+
+    def _fused_adamw_(
+        self,
+        grads,
+        exp_avgs,
+        exp_avg_sqs,
+        max_exp_avg_sqs,
+        state_steps,
+        *,
+        lr,
+        beta1,
+        beta2,
+        weight_decay,
+        eps,
+        amsgrad,
+        maximize,
+        grad_scale=None,
+        found_inf=None
+    ):
+        pass
+
+    API_WITH_SELF_ARG = {
+        'Torch._fused_adamw_': _fused_adamw_
+    }
+
+    ASCEND = "ASCEND"
+    MATCH_MODE_NAME = "pure name"
+    MATCH_MODE_MAPPING = "mapping"
+    MATCH_MODE_SIMILARITY = "similarity"
+
 
 class CompareConst:
     """
     Class for compare module const
     """
     SPACE = " "
+    NAME = "Name"
     # compare result column name
     NPU_NAME = "NPU Name"
     BENCH_NAME = "Bench Name"
@@ -256,6 +407,7 @@ 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"
@@ -278,6 +430,7 @@ class CompareConst:
     OUTPUT_STRUCT = "output_struct"
     PARAMS_STRUCT = "params_struct"
     PARAMS_GRAD_STRUCT = "params_grad_struct"
+    DEBUG_STRUCT = "debug_struct"
     SUMMARY = "summary"
     COMPARE_RESULT = "compare_result"
     COMPARE_MESSAGE = "compare_message"
@@ -330,8 +483,8 @@ 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
     ]
 
@@ -357,18 +510,16 @@ 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
@@ -386,16 +537,10 @@ class CompareConst:
         Const.KWARGS: INPUT_STRUCT,
         Const.OUTPUT: OUTPUT_STRUCT,
         Const.PARAMS: PARAMS_STRUCT,
-        Const.PARAMS_GRAD: PARAMS_GRAD_STRUCT
+        Const.PARAMS_GRAD: PARAMS_GRAD_STRUCT,
+        Const.DEBUG: DEBUG_STRUCT
     }
 
-    STRUCT_COMPARE_KEY = [
-        INPUT_STRUCT,
-        OUTPUT_STRUCT,
-        PARAMS_STRUCT,
-        PARAMS_GRAD_STRUCT
-    ]
-
     # compare standard
     HUNDRED_RATIO_THRESHOLD = 0.01
     THOUSAND_RATIO_THRESHOLD = 0.001
@@ -467,22 +612,42 @@ 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 = {
         MAX_DIFF: None, MIN_DIFF: None, MEAN_DIFF: None, NORM_DIFF: None, MAX_RELATIVE_ERR: None,
         MIN_RELATIVE_ERR: None, MEAN_RELATIVE_ERR: None, NORM_RELATIVE_ERR: None
     }
+
+    API_MAPPING_KEYS_TO_COMPARE = [
+        ('ms_args', 'pt_args'),
+        ('ms_outputs', 'pt_outputs'),
+        ('ms_parameters', 'pt_parameters'),
+        ('ms_parameters_grad', 'pt_parameters_grad')
+    ]
+
     INPUT_PATTERN = Const.SEP + Const.INPUT + Const.SEP
     KWARGS_PATTERN = Const.SEP + Const.KWARGS + Const.SEP
     OUTPUT_PATTERN = Const.SEP + Const.OUTPUT + Const.SEP
     PARAMS_PATTERN = Const.SEP + Const.PARAMS + Const.SEP
     PARAMS_GRAD_PATTERN = Const.SEP + Const.PARAMS_GRAD + Const.SEP
-    COMPARE_KEY = 'compare_key'
-    COMPARE_SHAPE = 'compare_shape'
+
+    CMP_KEY = 'compare_key'
+    CMP_SHAPE = 'compare_shape'
+
+    OP_NAME_X = 'op_name_x'
+    MATCH_RESULT_COLUMNS = [
+        OP_NAME_X, 'dtype_x', 'shape_x', 'summary_x', 'stack_info_x', 'data_name_x',
+        CMP_KEY, CMP_SHAPE,
+        'op_name_y', 'dtype_y', 'shape_y', 'summary_y', 'stack_info_y', 'data_name_y',
+    ]
+
     INTERNAL_API_MAPPING_FILE = 'ms_to_pt_api.yaml'
     UNREADABLE = 'unreadable data'
+    NPU_DUMP_DATA_DIR = 'npu_dump_data_dir'
+    BENCH_DUMP_DATA_DIR = 'bench_dump_data_dir'
+    NO_REAL_DATA_FLAG = '-1'
 
 
 class FileCheckConst:
@@ -504,6 +669,8 @@ class FileCheckConst:
     XLSX_SUFFIX = ".xlsx"
     YAML_SUFFIX = ".yaml"
     IR_SUFFIX = ".ir"
+    ZIP_SUFFIX = ".zip"
+    SHELL_SUFFIX = ".sh"
     MAX_PKL_SIZE = 1073741824  # 1 * 1024 * 1024 * 1024
     MAX_NUMPY_SIZE = 10737418240  # 10 * 1024 * 1024 * 1024
     MAX_JSON_SIZE = 1073741824  # 1 * 1024 * 1024 * 1024
@@ -512,6 +679,8 @@ class FileCheckConst:
     MAX_XLSX_SIZE = 1073741824  # 1 * 1024 * 1024 * 1024
     MAX_YAML_SIZE = 1073741824  # 1 * 1024 * 1024 * 1024
     MAX_IR_SIZE = 1073741824  # 1 * 1024 * 1024 * 1024
+    MAX_ZIP_SIZE = 10737418240  # 10 * 1024 * 1024 * 1024
+    MAX_FILE_IN_ZIP_SIZE = 1073741824  # 1 * 1024 * 1024 * 1024
     COMMOM_FILE_SIZE = 1048576  # 1 * 1024 * 1024
     DIR = "dir"
     FILE = "file"
@@ -525,7 +694,8 @@ class FileCheckConst:
         CSV_SUFFIX: MAX_CSV_SIZE,
         XLSX_SUFFIX: MAX_XLSX_SIZE,
         YAML_SUFFIX: MAX_YAML_SIZE,
-        IR_SUFFIX: MAX_IR_SIZE
+        IR_SUFFIX: MAX_IR_SIZE,
+        ZIP_SUFFIX: MAX_ZIP_SIZE
     }
     CSV_BLACK_LIST = r'^[＋－＝％＠\+\-=%@]|;[＋－＝％＠\+\-=%@]'
 
@@ -538,61 +708,6 @@ class OverflowConst:
     OVERFLOW_DEBUG_MODE = 1
 
 
-class MsCompareConst:
-    # api_info field
-    MINT = "Mint"
-    MINT_FUNCTIONAL = "MintFunctional"
-    TENSOR_API = "Tensor"
-
-    API_NAME_STR_LENGTH = 4
-    MAX_RECURSION_DEPTH = 20
-
-    # Mindtorch api_info field
-    MINDTORCH_TENSOR = "Tensor"
-    MINDTORCH = "Torch"
-    MINDTORCH_FUNC = "Functional"
-    MINDTORCH_NPU = "NPU"
-    MINDTORCH_DIST = "Distributed"
-
-
-
-    MT_VALID_API_TYPES = [
-        MINDTORCH, MINDTORCH_FUNC, MINDTORCH_TENSOR
-    ]
-
-    TASK_FIELD = "task"
-    STATISTICS_TASK = "statistics"
-    FRAMEWORK = "framework"
-    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
@@ -629,7 +744,16 @@ class MonitorConst:
     """
     Class for monitor const
     """
-    OP_LIST = ["norm", "min", "max", "zeros", "nans", "id", "mean"]
+
+    # 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", "shape", "dtype"]
     MONITOR_OUTPUT_DIR = "MONITOR_OUTPUT_DIR"
     DEFAULT_MONITOR_OUTPUT_DIR = "./monitor_output"
     DATABASE = "database"
@@ -641,7 +765,7 @@ class MonitorConst:
         "DeepSpeedZeroOptimizer_Stage3"
     )
     DEEPSPEED_ZERO_OPT_FILTER = "DeepSpeedZeroOptimizer"
-    RULE_NAME = ['AnomalyTurbulence']
+    RULE_NAME = ['AnomalyTurbulence', 'AnomalyNan']
 
     SLICE_SIZE = 20480
     # used for name
@@ -658,15 +782,16 @@ class MonitorConst:
     ACTVGRAD = "actv_grad"
     POST_GRAD = "post_grad"
     PRE_GRAD = "pre_grad"
+    PRE_PARAM = "param_origin"
+    POST_PARAM = "param_updated"
     ACC_GRAD = "acc_grad"
     PREFIX_POST = "post"
     PREFIX_PRE = "pre"
     EXP_AVG = "exp_avg"
     EXP_AVG_SQ = "exp_avg_sq"
-    PARAM = "param"
 
     CSV_HEADER = ["vpp_stage", "name", "step"]
-    CSV_HEADER_XY = ["vpp_stage", "name", "step", "micro_step"]
+    CSV_HEADER_MICRO_STEP = ["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"
@@ -674,3 +799,20 @@ class MonitorConst:
     CSV = "csv"
     API = "api"
     HEADER_NAME = 'name'
+    MAX_NDIGITS = 20
+
+    DEFAULT_STAGE = -1
+    FORWARD_STAGE = 0
+    BACKWARD_STAGE = 1
+    OPTIMIZER_STAGE = 2
+    FORWARD_KEY = [ACTV]
+    BACKWARD_KEY = [ACTVGRAD, PRE_GRAD, POST_GRAD, ACC_GRAD]
+    OPTIMIZER_KEY = [EXP_AVG, EXP_AVG_SQ]
+
+    TRAIN_STAGE = {}
+    for key in FORWARD_KEY:
+        TRAIN_STAGE[key] = FORWARD_STAGE
+    for key in BACKWARD_KEY:
+        TRAIN_STAGE[key] = BACKWARD_STAGE
+    for key in OPTIMIZER_KEY:
+        TRAIN_STAGE[key] = OPTIMIZER_STAGE