mindstudio-probe 1.1.0__py3-none-any.whl → 1.2.1__py3-none-any.whl

msprobe/mindspore/common/const.py

@@ -1,3 +1,18 @@
+# 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.
+
 import numpy as np
 import mindspore as ms
 
@@ -23,23 +38,18 @@ class Const:
     ASCEND_910A = "ascend910"
 
     OPS_PREFIX = "mindspore.ops."
-    Tensor_PREFIX = "mindspore.Tensor."
+    TENSOR_PREFIX = "mindspore.Tensor."
     MINT_PREFIX = "mindspore.mint."
     MINT_NN_FUNC_PREFIX = "mindspore.mint.nn.functional."
-    COMM_PREFIX = "mindspore.communication.comm_func."
-    COMMUNICATION_API_LIST = [
-        "mindspore.communication.comm_func.all_gather_into_tensor",
-        "mindspore.communication.comm_func.gather_into_tensor",
-        "mindspore.communication.comm_func.all_reduce",
-        "mindspore.communication.comm_func.reduce",
-        "mindspore.communication.comm_func.reduce_scatter_tensor"
-        ]
+
     TENSOR_DATA_PREFIX = "Tensor."
     STUB_TENSOR_DATA_PREFIX = "Tensor."
     OPS_DATA_PREFIX = "Functional."
     MINT_DATA_PREFIX = "Mint."
     MINT_NN_FUNC_DATA_PREFIX = "MintFunctional."
     DISTRIBUTED_DATA_PREFIX = "Distributed."
+    TORCH_DATA_PREFIX = "Torch."
+    TORCH_NPU_DATA_PREFIX = "NPU."
 
     SUPPORTED_API_LIST_FILE = "support_wrap_ops.yaml"
     SUPPORTED_TENSOR_LIST_KEY = "tensor"
@@ -50,6 +60,15 @@ class Const:
 
     DROPOUT_API_NAME_PREFIX = "dropout"
 
+    GRAPH_DATA_MODE_LIST = [CoreConst.ALL, CoreConst.INPUT, CoreConst.OUTPUT]
+
+    HOOK_MS_PREFIX_DICT = {
+        OPS_DATA_PREFIX: OPS_PREFIX,
+        TENSOR_DATA_PREFIX: TENSOR_PREFIX,
+        MINT_DATA_PREFIX: MINT_PREFIX,
+        MINT_NN_FUNC_DATA_PREFIX: MINT_NN_FUNC_PREFIX
+    }
+
 
 class FreeBenchmarkConst:
     ADD_NOISE = "add_noise"
@@ -65,19 +84,21 @@ class FreeBenchmarkConst:
     DEFAULT_PERT_TYPE = IMPROVE_PRECISION
     DEFAULT_HANDLER_TYPE = CHECK
     DEVICE_LIST = [DEFAULT_DEVICE]
-    STAGE_LIST = [CoreConst.FORWARD]
+    STAGE_LIST = [CoreConst.FORWARD, CoreConst.BACKWARD]
     DUMP_LEVEL_LIST = [DEFAULT_DUMP_LEVEL]
     PERT_TYPE_LIST = [IMPROVE_PRECISION, ADD_NOISE, BIT_NOISE, NO_CHANGE, EXCHANGE_VALUE]
     HANDLER_TYPE_LIST = [CHECK, FIX]
     NO_CHANGE_ERROR_THRESHOLD = 1.0
     SYMBOL_FLIPPING_RATIO = 8.0
 
+    SUPPORTED_CHECK_API_FILE = "support_wrap_ops.yaml"
+    CHECK_RESULT_FILE = "free_benchmark.csv"
+
     API_PREFIX_DICT = {
         "ops": Const.OPS_PREFIX,
-        "Tensor": Const.Tensor_PREFIX,
+        "Tensor": Const.TENSOR_PREFIX,
         "mint": Const.MINT_PREFIX,
-        "mint.nn.functional": Const.MINT_NN_FUNC_PREFIX,
-        "communication": Const.COMM_PREFIX
+        "mint.nn.functional": Const.MINT_NN_FUNC_PREFIX
     }
 
     PERT_VALUE_DICT = {
@@ -88,6 +109,7 @@ class FreeBenchmarkConst:
     }
 
     ERROR_THRESHOLD = {
+        ms.bfloat16: 1.004,
         ms.float16: 1.002,
         ms.float32: 1.0002
     }

msprobe/mindspore/common/log.py

@@ -1,4 +1,5 @@
-# Copyright 2024 Huawei Technologies Co., Ltd
+# Copyright (c) 2024-2024, 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.
@@ -11,15 +12,10 @@
 # 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 os
-import time
-import sys
-
-from msprobe.mindspore.common.utils import get_rank_if_initialized
-from msprobe.core.common.log import BaseLogger
 from msprobe.core.common.exceptions import DistributedNotInitializedError
+from msprobe.core.common.log import BaseLogger
+from msprobe.mindspore.common.utils import get_rank_if_initialized
 
 
 class MindsporeLogger(BaseLogger):
@@ -35,4 +31,4 @@ class MindsporeLogger(BaseLogger):
         return current_rank
 
 
-logger = MindsporeLogger()
+logger = MindsporeLogger()

msprobe/mindspore/common/utils.py

@@ -1,4 +1,5 @@
-# Copyright 2024 Huawei Technologies Co., Ltd
+# 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.
@@ -11,13 +12,15 @@
 # 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 os
 import random
 
 import mindspore as ms
 
+from mindspore import ops
+from mindspore.mint import nn
+
 from msprobe.core.common.exceptions import DistributedNotInitializedError
 from msprobe.core.common.file_utils import path_len_exceeds_limit, check_path_exists, save_npy
 from msprobe.core.common.log import logger
@@ -54,6 +57,11 @@ def convert_to_int(value):
         return -1
 
 
+def clean_input_kwargs(cell):
+    if hasattr(cell, 'input_kwargs'):
+        del cell.input_kwargs
+
+
 def list_lowest_level_directories(root_dir):
     check_path_exists(root_dir)
     lowest_level_dirs = []
@@ -74,13 +82,15 @@ def list_lowest_level_directories(root_dir):
     return lowest_level_dirs
 
 
-def seed_all(seed=1234, mode=False):
-    check_seed_all(seed, mode)
+def seed_all(seed=1234, mode=False, rm_dropout=True):
+    check_seed_all(seed, mode, rm_dropout)
     os.environ['PYTHONHASHSEED'] = str(seed)
     ms.set_seed(seed)
     random.seed(seed)
     ms.set_context(deterministic="ON" if mode else "OFF")
     os.environ['HCCL_DETERMINISTIC'] = str(mode)
+    if rm_dropout:
+        remove_dropout()
 
 
 class MsprobeStep(ms.train.Callback):
@@ -95,3 +105,77 @@ class MsprobeStep(ms.train.Callback):
     def on_train_step_end(self, run_context):
         self.debugger.stop()
         self.debugger.step()
+
+
+class Dropout(ops.Dropout):
+    def __init__(self, keep_prob=0.5, seed0=0, seed1=1):
+        super().__init__(1., seed0, seed1)
+
+
+class Dropout2D(ops.Dropout2D):
+    def __init__(self, keep_prob=0.5):
+        super().__init__(1.)
+
+
+class Dropout3D(ops.Dropout3D):
+    def __init__(self, keep_prob=0.5):
+        super().__init__(1.)
+
+
+class DropoutExt(nn.Dropout):
+    def __init__(self, p=0.5):
+        super().__init__(0)
+
+
+def dropout_ext(input_tensor, p=0.5, training=True):
+    return input_tensor
+
+
+def remove_dropout():
+    ops.Dropout = Dropout
+    ops.operations.Dropout = Dropout
+    ops.Dropout2D = Dropout2D
+    ops.operations.Dropout2D = Dropout2D
+    ops.Dropout3D = Dropout3D
+    ops.operations.Dropout3D = Dropout3D
+    nn.Dropout = DropoutExt
+    nn.functional.dropout = dropout_ext
+
+
+mindtorch_check_result = None
+
+
+def is_mindtorch():
+    global mindtorch_check_result
+    if mindtorch_check_result is None:
+        mindtorch_check_result = False
+        try:
+            import torch
+            from mindspore._c_expression import Tensor
+        except ImportError:
+            return mindtorch_check_result
+        tensor = torch.tensor(0.0)
+        if isinstance(tensor, Tensor):
+            mindtorch_check_result = True
+    return mindtorch_check_result
+
+
+register_backward_hook_functions = {}
+
+
+def set_register_backward_hook_functions():
+    global register_backward_hook_functions
+    if is_mindtorch():
+        import torch
+        from msprobe.mindspore.mindtorch import (_call_impl,
+                                                 register_full_backward_pre_hook,
+                                                 register_full_backward_hook)
+        if not hasattr(torch, "register_full_backward_hook"):
+            setattr(torch.nn.Module, "_call_impl", _call_impl)
+            setattr(torch.nn.Module, "register_full_backward_pre_hook", register_full_backward_pre_hook)
+            setattr(torch.nn.Module, "register_full_backward_hook", register_full_backward_hook)
+        register_backward_hook_functions["pre"] = torch.nn.Module.register_full_backward_pre_hook
+        register_backward_hook_functions["full"] = torch.nn.Module.register_full_backward_hook
+    else:
+        register_backward_hook_functions["pre"] = ms.nn.Cell.register_backward_pre_hook
+        register_backward_hook_functions["full"] = ms.nn.Cell.register_backward_hook

msprobe/mindspore/compare/distributed_compare.py

@@ -1,10 +1,24 @@
+# Copyright (c) 2024-2024, 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.
+
 import os
-from msprobe.core.common.utils import CompareException, check_compare_param, \
-    check_configuration_param, task_dumppath_get
+from msprobe.core.common.utils import CompareException
 from msprobe.core.common.file_utils import create_directory
 from msprobe.core.common.exceptions import FileCheckException
 from msprobe.mindspore.common.log import logger
-from msprobe.mindspore.compare.ms_compare import MSComparator
+from msprobe.mindspore.compare.ms_compare import ms_compare
 from msprobe.core.compare.utils import check_and_return_dir_contents, extract_json
 from msprobe.mindspore.compare.ms_graph_compare import GraphMSComparator
 
@@ -13,43 +27,27 @@ def ms_compare_distributed(npu_dump_dir, bench_dump_dir, output_path, **kwargs):
     if kwargs.get('suffix'):
         logger.error("Argument 'suffix' is not supported for compare_distributed.")
         raise CompareException(CompareException.INVALID_PARAM_ERROR)
-    stack_mode = kwargs.get('stack_mode', False)
-    auto_analyze = kwargs.get('auto_analyze', True)
-    fuzzy_match = kwargs.get('fuzzy_match', False)
+    is_print_compare_log = kwargs.get('is_print_compare_log', True)
     # get the ranks and match by order
     npu_ranks = sorted(check_and_return_dir_contents(npu_dump_dir, 'rank'))
     bench_ranks = sorted(check_and_return_dir_contents(bench_dump_dir, 'rank'))
     if len(npu_ranks) != len(bench_ranks):
         logger.error('The number of ranks in the two runs are different. '
-                        'Unable to match the ranks. Please use another folder to compare '
-                        'or use compare() api and manually match the ranks.')
+                     'Unable to match the ranks. Please use another folder to compare '
+                     'or use compare() api and manually match the ranks.')
         raise CompareException(CompareException.INVALID_PATH_ERROR)
     for nr, br in zip(npu_ranks, bench_ranks):
         npu_data_dir = os.path.join(npu_dump_dir, nr)
         bench_data_dir = os.path.join(bench_dump_dir, br)
         npu_path = extract_json(npu_data_dir, stack_json=False)
         bench_path = extract_json(bench_data_dir, stack_json=False)
-        stack_path = extract_json(npu_data_dir, stack_json=True)
 
         dump_result_param = {
             'npu_json_path': npu_path,
             'bench_json_path': bench_path,
-            'stack_json_path': stack_path,
-            'is_print_compare_log': True
+            'is_print_compare_log': is_print_compare_log
         }
-        try:
-            summary_compare, md5_compare = task_dumppath_get(dump_result_param)
-            check_configuration_param(stack_mode, auto_analyze, fuzzy_match, 
-                                      dump_result_param.get('is_print_compare_log', True))
-            create_directory(output_path)
-            check_compare_param(dump_result_param, output_path, 
-                                summary_compare=summary_compare, md5_compare=md5_compare)
-        except (CompareException, FileCheckException) as error:
-            logger.error('Compare failed. Please check the arguments and do it again!')
-            raise CompareException(error.code) from error
-        ms_comparator = MSComparator()
-        ms_comparator.compare_core(dump_result_param, output_path, suffix=f'_{nr}-{br}', 
-                                   summary_compare=summary_compare, md5_compare=md5_compare, **kwargs)
+        ms_compare(input_param=dump_result_param, output_path=output_path, suffix=f'_{nr}-{br}', **kwargs)
 
 
 def ms_graph_compare(inputs, outputs):