mindstudio-probe 8.3.2__py3-none-any.whl → 26.0.0a1__py3-none-any.whl

msprobe/mindspore/monitor/distributed/wrap_distributed.py

@@ -1,21 +1,24 @@
-# Copyright (c) 2024-2025, Huawei Technologies Co., Ltd.
-# All rights reserved.
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
 #
-# 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
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
 #
-# http://www.apache.org/licenses/LICENSE-2.0
+#          http://license.coscl.org.cn/MulanPSL2
 #
-# 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.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
+
 
 import inspect
 import os
 import re
+from dataclasses import dataclass
 
 import numpy as np
 
@@ -29,7 +32,7 @@ from msprobe.mindspore.monitor.utils import get_metrics, get_summary_writer_tag_
 
 enable_communication = True
 try:
-    from mindspore._c_expression import CommHandle as CommHandle_
+    from mindspore.communication.comm_func import CommHandle as CommHandle_
 except ImportError:
     enable_communication = False
 
@@ -42,6 +45,16 @@ WrapDistributedOps = load_yaml(OpsPath).get("communication.comm_func", [])
 StackBlackListPath = os.path.join(os.path.dirname(__file__), "stack_blacklist.yaml")
 StackBlackList = load_yaml(StackBlackListPath).get("stack", [])
 
+
+@dataclass(frozen=True)
+class CatchDataParams:
+    cc_context: object
+    cc_name: str
+    ops_list: list
+    args: tuple
+    prefix: str
+    call_idx: int = 0
+
 distributed_func = {}
 for f in dir(comm_func):
     distributed_func[f] = getattr(comm_func, f)
@@ -61,6 +74,7 @@ class DistributedOPTemplate(nn.Cell):
         super(DistributedOPTemplate, self).__init__()
         self.op_name_ = str(op_name)
         self.__name__ = self.op_name_
+        self.idx = 0
         self.cc_hooks = []
         for pre_hook in pre_hooks:
             handle = self.register_forward_pre_hook(pre_hook)
@@ -130,6 +144,11 @@ class ApiRegistry:
         self.set_api_attr(comm_func, self.distributed_attr_origin)
         setattr(CommHandle_, 'wait', ORIGIN_WAIT)
 
+    def reset_idx(self):
+        for op_template in self.distributed_attr_hooked.values():
+            if hasattr(op_template, "idx"):
+                op_template.idx = 0
+
     def initialize_hook(self, pre_hooks, post_hooks):
         self.store_ori_attr(comm_func, get_distributed_ops(), self.distributed_attr_origin)
         cc_hooks = []
@@ -208,27 +227,35 @@ def is_target_line(codeline):
 
 
 @_no_grad()
-def catch_data(cc_context, cc_name, ops_list, args, prefix):
+def catch_data(params: CatchDataParams):
     tensor_args = {}
-    for arg in args:
+    for arg in params.args:
         if isinstance(arg, Tensor):
-            key = get_summary_writer_tag_name(cc_name, f'{prefix}_{len(tensor_args)}', RANK)
+            key = get_summary_writer_tag_name(
+                params.cc_name,
+                f'{params.prefix}_{params.call_idx}_{len(tensor_args)}',
+                RANK,
+            )
             tensor_args[key] = arg
         elif isinstance(arg, list):
             if isinstance(arg[0], Tensor):
                 stacked_arg = ops.stack(arg)
             elif isinstance(arg[0], comm_func.P2POp):
                 stacked_arg = ops.stack([op.tensor for op in arg])
-            key = get_summary_writer_tag_name(cc_name, f'{prefix}_{len(tensor_args)}', RANK)
+            key = get_summary_writer_tag_name(
+                params.cc_name,
+                f'{params.prefix}_{params.call_idx}_{len(tensor_args)}',
+                RANK,
+            )
             tensor_args[key] = stacked_arg
 
-    new_data = get_metrics(ops_list, tensor_args, 1e-8)
-    cc_context.data = update_data(cc_context.data, new_data)
+    new_data = get_metrics(params.ops_list, tensor_args, 1e-8)
+    params.cc_context.data = update_data(params.cc_context.data, new_data)
 
 
-def create_async_callback_func(context, cc_name, ops_list, args, prefix):
+def create_async_callback_func(params: CatchDataParams):
     def store_data():
-        catch_data(context, cc_name, ops_list, args, prefix)
+        catch_data(params)
 
     return store_data
 
@@ -242,40 +269,67 @@ def create_hooks(context, monitor):
     def cc_pre_hook(module, inputs):
         if not is_target_line(monitor.cc_codeline):
             return
-        catch_data(context[module.op_name_], module.op_name_, monitor.ops, inputs, MonitorConst.PREFIX_PRE)
+        call_idx = getattr(module, "idx", 0)
+        catch_data(CatchDataParams(
+            context[module.op_name_],
+            module.op_name_,
+            monitor.ops,
+            inputs,
+            MonitorConst.PREFIX_PRE,
+            call_idx,
+        ))
         return
 
     def cc_hook(module, inputs, out=None):
         if not is_target_line(monitor.cc_codeline):
             return out
-        if out and enable_communication:  # async
+        call_idx = getattr(module, "idx", 0)
+        is_async_op = False
+        if enable_communication:
             if isinstance(out, CommHandle_):
-                PENDING_ASYNC_CC_BY_HANDLE[out] = create_async_callback_func(
+                PENDING_ASYNC_CC_BY_HANDLE[out] = create_async_callback_func(CatchDataParams(
                     context[module.op_name_],
                     module.op_name_,
-                    monitor.ops, inputs,
-                    MonitorConst.PREFIX_POST
-                )
+                    monitor.ops,
+                    inputs,
+                    MonitorConst.PREFIX_POST,
+                    call_idx,
+                ))
+                is_async_op = True
             elif isinstance(out, list):  # batch_isend_irecv
                 for out_element in out:
                     if isinstance(out_element, comm_func.P2POp):
-                        PENDING_ASYNC_CC_BY_HANDLE[out_element] = create_async_callback_func(
+                        PENDING_ASYNC_CC_BY_HANDLE[out_element] = create_async_callback_func(CatchDataParams(
                             context[module.op_name_],
                             module.op_name_,
-                            monitor.ops, inputs,
-                            MonitorConst.PREFIX_POST
-                        )
+                            monitor.ops,
+                            inputs,
+                            MonitorConst.PREFIX_POST,
+                            call_idx,
+                        ))
+                        is_async_op = True
             elif isinstance(out, tuple):
                 if len(out) == 2 and isinstance(out[1], CommHandle_):
-                    PENDING_ASYNC_CC_BY_HANDLE[out[1]] = create_async_callback_func(
-                            context[module.op_name_],
-                            module.op_name_,
-                            monitor.ops, inputs,
-                            MonitorConst.PREFIX_POST
-                    )
-
-            return out
-        catch_data(context[module.op_name_], module.op_name_, monitor.ops, inputs, MonitorConst.PREFIX_POST)
+                    PENDING_ASYNC_CC_BY_HANDLE[out[1]] = create_async_callback_func(CatchDataParams(
+                        context[module.op_name_],
+                        module.op_name_,
+                        monitor.ops,
+                        inputs,
+                        MonitorConst.PREFIX_POST,
+                        call_idx,
+                    ))
+                    is_async_op = True
+        if not is_async_op:
+            out_list = [ii for ii in out if isinstance(ii, Tensor)]
+            catch_data(CatchDataParams(
+                context[module.op_name_],
+                module.op_name_,
+                monitor.ops,
+                out_list,
+                MonitorConst.PREFIX_POST,
+                call_idx,
+            ))
+        module.idx = call_idx + 1
         return out
 
     global RANK

msprobe/mindspore/monitor/features.py

@@ -1,17 +1,18 @@
-# Copyright (c) 2024-2025, Huawei Technologies Co., Ltd.
-# All rights reserved.
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
 #
-# 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
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
 #
-# http://www.apache.org/licenses/LICENSE-2.0
+#          http://license.coscl.org.cn/MulanPSL2
 #
-# 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.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
 
 from mindspore import mint, ops, _no_grad
 from mindspore import Tensor

msprobe/mindspore/monitor/module_hook.py

@@ -1,17 +1,18 @@
-# Copyright (c) 2024-2025, Huawei Technologies Co., Ltd.
-# All rights reserved.
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
 #
-# 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
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
 #
-# http://www.apache.org/licenses/LICENSE-2.0
+#          http://license.coscl.org.cn/MulanPSL2
 #
-# 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.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
 
 from gzip import FEXTRA
 import os
@@ -29,7 +30,7 @@ from mindspore import nn, _no_grad
 
 from msprobe.core.common.log import logger
 from msprobe.core.common.const import MonitorConst, Const
-from msprobe.core.common.file_utils import load_json, save_json, make_dir
+from msprobe.core.common.file_utils import load_json, save_json, create_directory
 from msprobe.core.monitor.utils import validate_config, get_output_base_dir, get_target_output_dir
 from msprobe.core.monitor.anomaly_processor import AnomalyScanner, AnomalyDataFactory, AnomalyDataWriter
 from msprobe.mindspore.common.utils import is_mindtorch
@@ -198,7 +199,7 @@ class TrainerMon:
         self.process_group = process_group
         self.params_have_main_grad = params_have_main_grad
         self.is_mindtorch = is_mindtorch()
-        self.config_timestamp = 0  # 后面有校验时间戳, 首次监控无需为了更新config文件时间戳而去改, 可通过dynamic_on开关直接打开
+        self.config_timestamp = 0  # 后面有校验时间戳, 首次监测无需为了更新config文件时间戳而去改, 可通过dynamic_on开关直接打开
         self.config = load_json(config_file_path)
         validate_config(self.config)
 
@@ -232,7 +233,7 @@ class TrainerMon:
         self.micro_batch_number = 1
         self.optimizer_mon = None
 
-        # TYPE3: 会随着训练中途config配置更新或监控状态改变而重置的变量
+        # TYPE3: 会随着训练中途config配置更新或监测状态改变而重置的变量
         self.module_fwd_hook_context_by_module = defaultdict(ModuleHookContext)
         self.module_bwd_hook_context_by_module = defaultdict(ModuleHookContext)
         self.feature_hook_context_by_module = defaultdict(FeatureHookContext)
@@ -249,7 +250,6 @@ class TrainerMon:
         self.param_name_call_id = {}
         self.call_id = 0
         self.module_struct = defaultdict(dict)
-        self.grad_accs = []
         self.weight_hooked = False
         self.optimizer_hooked = False
         self.param_registered = False
@@ -762,7 +762,7 @@ class TrainerMon:
 
     def _save_module_struct(self):
         output_dir = os.path.join(get_output_base_dir(), 'module_struct', f'rank{self.rank}')
-        make_dir(output_dir)
+        create_directory(output_dir)
         module_struct_file = os.path.realpath(os.path.join(output_dir, 'module_struct.json'))
         save_json(module_struct_file, self.module_struct, indent=2)
         logger.info(f"> save module struct to {module_struct_file}")
@@ -981,6 +981,7 @@ class TrainerMon:
         self._hook_weights()
 
     def _hook_weights(self):
+        context = self.grad_context
 
         @_no_grad()
         def param_hook(grad, param, name):
@@ -990,11 +991,8 @@ class TrainerMon:
             key = get_summary_writer_tag_name(key, 'acc_grad', self.rank)
             self.register_param_call_id("param_hook", key)
             param.micro_step += 1
-            grad_dict = {}
             if self.monitor_mbs_grad or (param.micro_step == self.micro_batch_number):
-                grad_dict[key] = grad
-            get_metrics(self.ops, grad_dict, self.eps, self.grad_context.pre)
-
+                get_metrics(self.ops, {key: grad}, self.eps, self.grad_context.pre)
             if param.micro_step == self.micro_batch_number:
                 param.micro_step = 0
 
@@ -1004,7 +1002,7 @@ class TrainerMon:
 
             return wrapper
 
-        logger.info("hooking weights.")
+        logger.info("hooking weight grads.")
         for param, name in self.param2name.items():
             setattr(param, 'micro_step', 0)
             handle = param.register_hook(
@@ -1078,7 +1076,6 @@ class TrainerMon:
         self.duplicate_param.clear()
         self.name2tag.clear()
         self.module_struct.clear()
-        self.grad_accs.clear()
 
         # 关闭采集状态
         self.monitoring = False

msprobe/mindspore/monitor/optimizer_collect.py

@@ -1,17 +1,19 @@
-# Copyright (c) 2024-2025, Huawei Technologies Co., Ltd.
-# All rights reserved.
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
 #
-# 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
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
 #
-# http://www.apache.org/licenses/LICENSE-2.0
+#          http://license.coscl.org.cn/MulanPSL2
 #
-# 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.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
+
 from abc import abstractmethod
 
 from mindspore import mint, ops
@@ -28,7 +30,7 @@ class OptimizerMon(object):
 
     def narrow_from_flatten(self, param, flatten_state):
         return flatten_state
-    
+
     def get_state(self, optim):
         if hasattr(optim, 'chained_optimizers'):
             for opt in optim.chained_optimizers:
@@ -52,7 +54,7 @@ class OptimizerMon(object):
             if param.numel() != element_in_cur_partition:
                 if first_param:
                     grad = grad.flatten()[-element_in_cur_partition:]
-                else: # supposed to be the last one
+                else:  # supposed to be the last one
                     grad = grad.flatten()[:element_in_cur_partition]
             first_param = False
             if grad is None:
@@ -61,7 +63,7 @@ class OptimizerMon(object):
             monitor.register_param_call_id("hook_optimizer", tag)
             grad_dict[tag] = grad
         return grad_dict
-    
+
     def map_fp16_to_fp32_param(self, optim):
         pass
 
@@ -115,7 +117,7 @@ class OptimizerMon(object):
                     monitor.update_heatmap_visualizer[name].pre_cal(update_dict[name])
                     monitor.ratio_heatmap_visualizer[name].pre_cal(ratio_dict[name])
         return exp_avg_dict, exp_avg_sq_dict, update_dict, ratio_dict
-    
+
     def _get_single_state(self, optim):
         state = {}
         if hasattr(optim, 'param_to_cpu_states_map'):
@@ -129,9 +131,10 @@ class OptimizerMon(object):
 
 class MixPrecisionOptimizerMon(OptimizerMon):
     """
-    混合精度优化器监控类。在混合精度训练中监控和管理优化器。
+    混合精度优化器监测类。在混合精度训练中监测和管理优化器。
     混合精度训练通过适当降低某些计算的精度来加速训练过程并减少内存消耗。
     """
+
     def map_fp16_to_fp32_param(self, optim):
         for fp16_group, fp32_group in zip(optim.float16_groups, optim.fp32_from_float16_groups):
             for fp16_param, fp32_param in zip(fp16_group, fp32_group):
@@ -175,6 +178,7 @@ class DeepSpeedZeroOptimizerMon(OptimizerMon):
     - Handling gradient collection for different ZeRO stages
     - Managing optimizer state access for monitoring
     """
+
     def __init__(self, optim):
         super().__init__(optim)
         self.stage = ''
@@ -187,12 +191,12 @@ class DeepSpeedZeroOptimizerMon(OptimizerMon):
     @abstractmethod
     def get_grad_for_param(self, lp_param, group_idx, param_id):
         raise NotImplementedError
-    
+
     def param_not_in_partition(self, lp_param, group_idx):
         param_slice_mapping = self.optim.state_dict()['param_slice_mappings'][group_idx]
         hp_address = param_slice_mapping.get(self.optim.param_names.get(lp_param))
         return hp_address is None
-    
+
     def get_position(self, lp_param, group_idx):
         param_slice_mapping = self.optim.state_dict()['param_slice_mappings'][group_idx]
         hp_address = param_slice_mapping.get(self.optim.param_names.get(lp_param))
@@ -204,7 +208,7 @@ class DeepSpeedZeroOptimizerMon(OptimizerMon):
             for param in bit16_group:
                 param2group[param] = group_idx
         return param2group
-    
+
     def get_param_index(self, lp_param, group_idx):
         if not self.param2index:
             for group in self.bit16_groups:
@@ -212,9 +216,9 @@ class DeepSpeedZeroOptimizerMon(OptimizerMon):
                 for index, param in enumerate(group):
                     param2index[param] = index
                 self.param2index.append(param2index)
-                
+
         return self.param2index[group_idx][lp_param]
-    
+
     def narrow_from_flatten(self, param, flatten_state):
         if flatten_state is None:
             return flatten_state
@@ -223,7 +227,7 @@ class DeepSpeedZeroOptimizerMon(OptimizerMon):
             return None
         start, numel = self.get_position(param, group_idx)
         return flatten_state.narrow(0, start, numel)
-        
+
     def map_fp16_to_fp32_param(self, optim):
         for group_idx, group in enumerate(self.bit16_groups):
             for param in group:
@@ -253,7 +257,7 @@ class DeepSpeedZeroOptimizerStage0Mon(DeepSpeedZeroOptimizerMon):
         self.bit16_groups = optim.bf16_groups
         self.fp32_flat_groups = optim.fp32_groups_flat_partition
         self.param2group = self.get_group_index()
-            
+
     def get_grad_for_param(self, lp_param, group_idx, param_id):
         return self.optim.fp32_groups_gradient_dict[group_idx][param_id]
 
@@ -298,11 +302,11 @@ class DeepSpeedZeroOptimizerStage3Mon(DeepSpeedZeroOptimizerMon):
     def param_not_in_partition(self, lp_param, group_idx):
         """Each param partioned across all zero ranks"""
         return False
-    
+
     def get_position(self, lp_param, group_idx):
         param_id = self.optim.get_param_id(lp_param)
         return self.optim.grad_position[param_id][1:]
-    
+
     def get_grad_for_param(self, lp_param, group_idx, param_id):
         return self.optim.averaged_gradients[group_idx][param_id]
 

msprobe/mindspore/monitor/utils.py

@@ -1,17 +1,19 @@
-# Copyright (c) 2024-2025, Huawei Technologies Co., Ltd.
-# All rights reserved.
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
 #
-# 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
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
 #
-# http://www.apache.org/licenses/LICENSE-2.0
+#          http://license.coscl.org.cn/MulanPSL2
 #
-# 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.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
+
 from mindspore import dtype as mstype, Tensor
 
 from msprobe.mindspore.monitor.features import FUNC_MAP, cal_entropy, cal_stable_rank

msprobe/msaccucmp/advisor/__init__.py

@@ -0,0 +1,16 @@
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
+#
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
+#
+#          http://license.coscl.org.cn/MulanPSL2
+#
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
+

msprobe/msaccucmp/advisor/advisor_const.py

@@ -0,0 +1,65 @@
+# coding=utf-8
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
+#
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
+#
+#          http://license.coscl.org.cn/MulanPSL2
+#
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
+
+"""
+Function:
+This file mainly involves the const value.
+"""
+
+
+class AdvisorConst:
+    """
+    The class for advisor const
+    """
+    # column const
+    COSINE_SIMILARITY = "CosineSimilarity"
+    INDEX = "Index"
+    NPU_DUMP = "NPUDump"
+    OVERFLOW = "OverFlow"
+
+    # advisor summary key
+    DETECTION_TYPE = "Detection Type"
+    OPERATOR_INDEX = "Operator Index"
+    ADVISOR_SUGGEST = "Expert Advice"
+
+    # detection type
+    OVERFLOW_DETECTION = "FP16 Overflow"
+    INPUT_DETECTION = "Input Inconsistent"
+    CONSISTENCY_DETECTION = "Global Consistency"
+
+    # operator index
+    NO_ERROR_OP = "NA"
+
+    # advisor suggest
+    OVERFLOW_SUGGEST = "Float16 data overflow occurs. Rectify the fault and perform comparison again."
+    INPUT_SUGGEST = "The input data of NPUDump is inconsistent with that of GroundTruth. Use the same data " \
+                    "or check the data preprocessing process."
+    CONSISTENCY_SUGGEST = "All data in the comparison result meets the accuracy requirements. " \
+                          "If data accuracy of the model is still not up to standard in practical application, " \
+                          "please check the post-processing process of model outputs."
+    PROBLEM_SUGGEST = "The accuracy of some tensors is low, resulting in an unqualified final accuracy. " \
+                      "This may be caused by quantization. Calibrate the data or contact Huawei for further diagnosis. "
+    DEVIATION_SUGGEST = "The accuracy of some tensors is low, while the final accuracy is qualified. " \
+                        "This may be caused by Ascend internal optimization. " \
+                        "Ignore or contact Huawei for further diagnosis. "
+
+    # text symbol
+    NEW_LINE = "\n"
+    COLON = ": "
+
+    ACCURACY_THRESHOLD = 0.99
+