mindstudio-probe 1.0.4__py3-none-any.whl → 1.1.1__py3-none-any.whl

msprobe/pytorch/api_accuracy_checker/run_ut/multi_run_ut.py

@@ -1,3 +1,20 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# 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 subprocess
 import json
 import os
@@ -16,9 +33,10 @@ from msprobe.pytorch.api_accuracy_checker.compare.compare import Comparator
 from msprobe.pytorch.common import parse_json_info_forward_backward
 from msprobe.pytorch.common.log import logger
 from msprobe.core.common.file_utils import FileChecker, check_file_suffix, check_link, FileOpen, \
-    check_path_before_create, create_directory
+    create_directory, load_json, save_json
 from msprobe.core.common.file_utils import remove_path
-from msprobe.core.common.const import FileCheckConst
+from msprobe.core.common.const import FileCheckConst, Const
+from msprobe.core.common.utils import CompareException
 
 
 def split_json_file(input_file, num_splits, filter_api):
@@ -30,9 +48,11 @@ def split_json_file(input_file, num_splits, filter_api):
     for data_name in list(backward_data.keys()):
         backward_data[f"{data_name}.backward"] = backward_data.pop(data_name)
 
-    with FileOpen(input_file, 'r') as file:
-        input_data = json.load(file)
-        input_data.pop("data")
+    input_data = load_json(input_file)
+    if input_data.get("data") is None:
+        logger.error("Invalid input file, 'data' field is missing")
+        raise CompareException("Invalid input file, 'data' field is missing")
+    input_data.pop("data")
 
     items = list(forward_data.items())
     total_items = len(items)
@@ -52,8 +72,7 @@ def split_json_file(input_file, num_splits, filter_api):
             }
         }
         split_filename = f"temp_part{i}.json"
-        with FileOpen(split_filename, 'w') as split_file:
-            json.dump(temp_data, split_file)
+        save_json(split_filename, temp_data)
         split_files.append(split_filename)
 
     return split_files, total_items
@@ -105,7 +124,7 @@ def run_parallel_ut(config):
                 if output == '':
                     break
                 if '[ERROR]' in output:
-                    print(output, end='')
+                    logger.warning(output)
                     sys.stdout.flush()
         except ValueError as e:
             logger.warning(f"An error occurred while reading subprocess output: {e}")
@@ -119,7 +138,8 @@ def run_parallel_ut(config):
 
     for api_info in config.api_files:
         cmd = create_cmd(api_info, next(device_id_cycle))
-        process = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.DEVNULL, text=True, bufsize=1, shell=False)
+        process = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.DEVNULL, 
+                                   text=True, bufsize=1, shell=False)
         processes.append(process)
         threading.Thread(target=read_process_output, args=(process,), daemon=True).start()
 
@@ -150,7 +170,8 @@ def run_parallel_ut(config):
         logger.error(f"An unexpected error occurred: {e}")
     finally:
         if progress_bar.n < config.total_items:
-            logger.warning("The UT task has not been completed. The parameter '-csv_path' along with the path to the result CSV file will be utilized to resume the UT task.")
+            logger.warning("The UT task has not been completed. The parameter '-csv_path' along with the path to " \
+                           "the result CSV file will be utilized to resume the UT task.")
         clean_up()
         progress_bar_thread.join()
     try:
@@ -163,17 +184,21 @@ def run_parallel_ut(config):
 
 
 def prepare_config(args):
-    check_link(args.api_info_file)
-    api_info = os.path.realpath(args.api_info_file)
-    check_file_suffix(api_info, FileCheckConst.JSON_SUFFIX)
-    out_path = os.path.realpath(args.out_path) if args.out_path else "./"
-    check_path_before_create(out_path)
+    api_info_file_checker = FileChecker(file_path=args.api_info_file, path_type=FileCheckConst.FILE, 
+                                            ability=FileCheckConst.READ_ABLE, file_type=FileCheckConst.JSON_SUFFIX)
+    api_info = api_info_file_checker.common_check()
+    out_path = args.out_path if args.out_path else Const.DEFAULT_PATH
     create_directory(out_path)
     out_path_checker = FileChecker(out_path, FileCheckConst.DIR, ability=FileCheckConst.WRITE_ABLE)
     out_path = out_path_checker.common_check()
     split_files, total_items = split_json_file(api_info, args.num_splits, args.filter_api)
-    config_path = os.path.realpath(args.config_path) if args.config_path else None
-    result_csv_path = args.result_csv_path or os.path.join(out_path, f"accuracy_checking_result_{time.strftime('%Y%m%d%H%M%S')}.csv")
+    config_path = args.config_path if args.config_path else None
+    if config_path:
+        config_path_checker = FileChecker(config_path, FileCheckConst.FILE, 
+                                          FileCheckConst.READ_ABLE, FileCheckConst.JSON_SUFFIX)
+        config_path = config_path_checker.common_check()
+    result_csv_path = args.result_csv_path or os.path.join(
+                      out_path, f"accuracy_checking_result_{time.strftime('%Y%m%d%H%M%S')}.csv")
     if not args.result_csv_path:
         details_csv_path = os.path.join(out_path, f"accuracy_checking_details_{time.strftime('%Y%m%d%H%M%S')}.csv")
         comparator = Comparator(result_csv_path, details_csv_path, False)
@@ -190,7 +215,8 @@ def prepare_config(args):
 def main():
     parser = argparse.ArgumentParser(description='Run UT in parallel')
     _run_ut_parser(parser)
-    parser.add_argument('-n', '--num_splits', type=int, choices=range(1, 65), default=8, help='Number of splits for parallel processing. Range: 1-64')
+    parser.add_argument('-n', '--num_splits', type=int, choices=range(1, 65), default=8, 
+                        help='Number of splits for parallel processing. Range: 1-64')
     args = parser.parse_args()
     config = prepare_config(args)
     run_parallel_ut(config)

msprobe/pytorch/api_accuracy_checker/run_ut/run_overflow_check.py

@@ -1,3 +1,20 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# 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 argparse
 import os
 import sys
@@ -11,11 +28,12 @@ else:
 import torch
 from tqdm import tqdm
 from msprobe.pytorch.api_accuracy_checker.run_ut.run_ut import generate_device_params, get_api_info
-from msprobe.pytorch.api_accuracy_checker.run_ut.run_ut_utils import exec_api
-from msprobe.core.common.file_utils import check_link
+from msprobe.pytorch.api_accuracy_checker.run_ut.run_ut_utils import exec_api, is_unsupported_api
+from msprobe.core.common.file_utils import check_link, FileChecker
+from msprobe.pytorch.api_accuracy_checker.common.utils import extract_basic_api_segments
+from msprobe.core.common.const import FileCheckConst, Const
 from msprobe.pytorch.common.log import logger
 from msprobe.pytorch.common.parse_json import parse_json_info_forward_backward
-from msprobe.core.common.const import Const
 
 
 def check_tensor_overflow(x):
@@ -24,8 +42,8 @@ def check_tensor_overflow(x):
             tensor_max = x.cpu().detach().float().numpy().tolist()
             tensor_min = tensor_max
         else:
-            tensor_max = torch._C._VariableFunctionsClass.max(x).cpu().detach().float().numpy().tolist()
-            tensor_min = torch._C._VariableFunctionsClass.min(x).cpu().detach().float().numpy().tolist()
+            tensor_max = torch.max(x).cpu().detach().float().numpy().tolist()
+            tensor_min = torch.min(x).cpu().detach().float().numpy().tolist()
         # inf
         if tensor_max == float('inf') or tensor_min == float('-inf'):
             return True
@@ -57,23 +75,25 @@ def run_overflow_check(forward_file):
     logger.info("start UT test")
     forward_content, _, real_data_path = parse_json_info_forward_backward(forward_file)
     for api_full_name, api_info_dict in tqdm(forward_content.items()):
+        if is_unsupported_api(api_full_name, is_overflow_check=True):
+            continue
         try:
             run_torch_api(api_full_name, api_info_dict, real_data_path)
         except Exception as err:
             _, api_name, _ = api_full_name.split(Const.SEP)
             if "not implemented for 'Half'" in str(err):
-                logger.warning(f"API {api_name} not support half tensor in CPU, please add {api_name} to CONVERT_API "
-                               f"'fp16_to_fp32' list in accuracy_tools/api_accuracy_check/common/utils.py file.")
+                logger.warning(f"API {api_name} not support half tensor in CPU. This API does not support overflow "
+                               "check, so it will be skipped.")
             elif "expected scalar type Long" in str(err):
                 logger.warning(f"API {api_name} not support int32 tensor in CPU, please add {api_name} to CONVERT_API "
-                               f"'int32_to_int64' list in accuracy_tools/api_accuracy_check/common/utils.py file.")
+                               "'int32_to_int64' list in accuracy_tools/msprobe/core/common/const.py file.")
             else:
                 logger.error(f"Run {api_full_name} UT Error: %s" % str(err))
 
 
 def run_torch_api(api_full_name, api_info_dict, real_data_path):
     torch.npu.clear_npu_overflow_flag()
-    api_type, api_name, _ = api_full_name.split(Const.SEP)
+    api_type, api_name = extract_basic_api_segments(api_full_name)
     args, kwargs, need_grad = get_api_info(api_info_dict, api_name, real_data_path)
     if not need_grad:
         logger.warning("%s function with out=... arguments don't support automatic differentiation, skip backward." 
@@ -118,8 +138,9 @@ def _run_overflow_check(parser=None):
 def _run_overflow_check_command(args):
     torch.npu.set_compile_mode(jit_compile=args.jit_compile)
     npu_device = "npu:" + str(args.device_id)
-    check_link(args.api_info_file)
-    api_info = os.path.realpath(args.api_info_file)
+    api_info_file_checker = FileChecker(file_path=args.api_info_file, path_type=FileCheckConst.FILE, 
+                                            ability=FileCheckConst.READ_ABLE, file_type=FileCheckConst.JSON_SUFFIX)
+    api_info = api_info_file_checker.common_check()
     try:
         torch.npu.set_device(npu_device)
     except Exception as error:

msprobe/pytorch/api_accuracy_checker/run_ut/run_ut.py

@@ -1,6 +1,23 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+# 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 argparse
 import os
-import csv
+import re
 import sys
 import time
 import gc
@@ -17,43 +34,34 @@ else:
 import torch
 from tqdm import tqdm
 
-from msprobe.pytorch.api_accuracy_checker.run_ut.run_ut_utils import Backward_Message, hf_32_standard_api, UtDataInfo, \
-    get_validated_result_csv_path, get_validated_details_csv_path, exec_api
+from msprobe.pytorch.api_accuracy_checker.run_ut.run_ut_utils import BackwardMessage, UtDataInfo, \
+    get_validated_result_csv_path, get_validated_details_csv_path, exec_api, record_skip_info, is_unsupported_api
 from msprobe.pytorch.api_accuracy_checker.run_ut.data_generate import gen_api_params, gen_args
 from msprobe.pytorch.api_accuracy_checker.common.utils import api_info_preprocess, \
     initialize_save_path, UtDataProcessor, extract_basic_api_segments, ApiData
 from msprobe.pytorch.api_accuracy_checker.compare.compare import Comparator
 from msprobe.pytorch.api_accuracy_checker.compare.compare_column import CompareColumn
-from msprobe.pytorch.api_accuracy_checker.common.config import msCheckerConfig
+from msprobe.pytorch.api_accuracy_checker.common.config import CheckerConfig
 from msprobe.pytorch.common.parse_json import parse_json_info_forward_backward
-from msprobe.core.common.file_utils import FileOpen, FileChecker, \
-    change_mode, check_path_before_create, create_directory, get_json_contents
+from msprobe.core.common.file_utils import FileChecker, change_mode, \
+    create_directory, get_json_contents, read_csv, check_file_or_directory_path, check_crt_valid
 from msprobe.pytorch.common.log import logger
 from msprobe.pytorch.pt_config import parse_json_config
 from msprobe.core.common.const import Const, FileCheckConst, CompareConst
+from msprobe.core.common.utils import safe_get_value
 from msprobe.pytorch.api_accuracy_checker.tensor_transport_layer.attl import ATTL, ATTLConfig, move2device_exec
 from msprobe.pytorch.api_accuracy_checker.tensor_transport_layer.device_dispatch import ConsumerDispatcher
+from msprobe.pytorch.api_accuracy_checker.run_ut.run_ut_utils import generate_cpu_params, generate_device_params
 
 
 current_time = time.strftime("%Y%m%d%H%M%S")
 UT_ERROR_DATA_DIR = 'ut_error_data' + current_time
 RESULT_FILE_NAME = "accuracy_checking_result_" + current_time + ".csv"
 DETAILS_FILE_NAME = "accuracy_checking_details_" + current_time + ".csv"
-RunUTConfig = namedtuple('RunUTConfig', ['forward_content', 'backward_content', 'result_csv_path', 'details_csv_path',
-                                         'save_error_data', 'is_continue_run_ut', 'real_data_path', 'white_list',
-                                         'black_list', 'error_data_path', 'online_config'])
 
-OnlineConfig = namedtuple('OnlineConfig', ['is_online', 'nfs_path', 'host', 'port', 'rank_list', 'tls_path'])
 
 not_backward_list = ['repeat_interleave']
-not_detach_set = {'resize_', 'resize_as_', 'set_', 'transpose_', 't_', 'squeeze_', 'unsqueeze_'}
-not_raise_dtype_set = {'type_as'}
 
-RAISE_PRECISION = {
-    torch.float16: torch.float32,
-    torch.bfloat16: torch.float32,
-    torch.float32: torch.float64
-}
 
 tqdm_params = {
     'smoothing': 0,  # 平滑进度条的预计剩余时间，取值范围0到1
@@ -71,98 +79,6 @@ tqdm_params = {
 }
 
 
-def deal_detach(arg, to_detach=True):
-    return arg.detach() if to_detach else arg
-
-
-def raise_bench_data_dtype(api_name, arg, raise_dtype=None):
-    '''
-    将标杆数据的dtype转换为raise_dtype
-    输入：
-        api_name：api名称
-        arg：标杆输入
-        raise_dtype：需要转换的dtype
-    输出： 
-        arg: 转换dtype的标杆输入
-    '''
-    if api_name in hf_32_standard_api and arg.dtype == torch.float32:
-        return arg
-    if raise_dtype is None or arg.dtype not in RAISE_PRECISION or raise_dtype == arg.dtype:
-        return arg
-    return arg.type(raise_dtype)
-
-
-def generate_device_params(input_args, input_kwargs, need_backward, api_name):
-    def recursive_arg_to_device(arg_in, to_detach):
-        if isinstance(arg_in, (list, tuple)):
-            return type(arg_in)(recursive_arg_to_device(arg, to_detach) for arg in arg_in)
-        elif isinstance(arg_in, torch.Tensor):
-            if need_backward and arg_in.requires_grad:
-                arg_in = deal_detach(arg_in.clone(), to_detach).to(current_device).requires_grad_()
-                temp_arg_in = arg_in * 1
-                arg_in = temp_arg_in.type_as(arg_in)
-                arg_in.retain_grad()
-                return arg_in
-            else:
-                return deal_detach(arg_in.clone(), to_detach).to(current_device)
-        else:
-            return arg_in
-
-    is_detach = api_name not in not_detach_set
-    device_args = recursive_arg_to_device(input_args, is_detach)
-    device_kwargs = \
-        {key: recursive_arg_to_device(value, key != "out" and is_detach) for key, value in input_kwargs.items()}
-    return device_args, device_kwargs
-
-
-def generate_cpu_params(input_args, input_kwargs, need_backward, api_name):
-    def recursive_arg_to_cpu(arg_in, to_detach, raise_dtype=None):
-        if isinstance(arg_in, (list, tuple)):
-            return type(arg_in)(recursive_arg_to_cpu(arg, to_detach, raise_dtype=raise_dtype) for arg in arg_in)
-        elif isinstance(arg_in, torch.Tensor):
-            if need_backward and arg_in.requires_grad:
-                arg_in = deal_detach(raise_bench_data_dtype(
-                                     api_name, arg_in.clone(), raise_dtype=raise_dtype), to_detach).requires_grad_()
-                temp_arg_in = arg_in * 1
-                arg_in = temp_arg_in.type_as(arg_in)
-                arg_in.retain_grad()
-                return arg_in
-            else:
-                return deal_detach(raise_bench_data_dtype(api_name, arg_in.clone(), raise_dtype=raise_dtype), to_detach)
-        else:
-            return arg_in
-
-    def is_tensor_with_raise_precision(arg_in, check_kwargs=False):
-        if arg_in.dtype in RAISE_PRECISION:
-            return True
-        if check_kwargs and arg_in.dtype in [torch.half, torch.bfloat16]:
-            return True
-        return False
-
-    def recursive_find_dtypes(arg_in, kwargs=None, check_kwargs=False):
-        if isinstance(arg_in, (list, tuple)):
-            return set().union(*tuple(recursive_find_dtypes(arg, kwargs, check_kwargs=check_kwargs) for arg in arg_in))
-        elif isinstance(arg_in, torch.Tensor) and is_tensor_with_raise_precision(arg_in, check_kwargs):
-            return set([arg_in.dtype])
-        elif isinstance(arg_in, dict) and check_kwargs:
-            return set().union(*tuple(recursive_find_dtypes(v, kwargs, check_kwargs=True) for v in arg_in.values()))
-        return set()
-
-    raise_dtype = None
-    need_raise_dtypes = recursive_find_dtypes(input_args)
-    need_raise_dtypes.update(recursive_find_dtypes(input_kwargs, check_kwargs=True))
-    if len(need_raise_dtypes) == 1:
-        raise_dtype = RAISE_PRECISION.get(need_raise_dtypes.pop(), torch.float32)
-    elif len(need_raise_dtypes) >= 2:
-        raise_dtype = torch.float32
-
-    raise_dtype = None if api_name in not_raise_dtype_set else raise_dtype
-    is_detach = api_name not in not_detach_set
-    cpu_args = recursive_arg_to_cpu(input_args, is_detach, raise_dtype=raise_dtype)
-    cpu_kwargs = {key: recursive_arg_to_cpu(value, key != "out" and is_detach, raise_dtype=raise_dtype) for key, value in input_kwargs.items()}
-    return cpu_args, cpu_kwargs
-
-
 def run_ut(config):
     logger.info("start UT test")
     if config.online_config.is_online:
@@ -179,10 +95,12 @@ def run_ut(config):
     if config.online_config.is_online:
         run_api_online(config, compare)
     else:
-        with FileOpen(config.result_csv_path, 'r') as file:
-            csv_reader = csv.reader(file)
-            next(csv_reader)
-            api_name_set = {row[0] for row in csv_reader}
+        csv_df = read_csv(config.result_csv_path)
+        try:
+            api_name_set = {row[0] for row in csv_df.itertuples(index=False, name=None)}
+        except IndexError:
+            logger.error(f"Read {config.result_csv_path} error, api_name_set is empty.")
+            api_name_set = set()
         run_api_offline(config, compare, api_name_set)
     for result_csv_path, details_csv_path in zip(compare.save_path_list, compare.detail_save_path_list):
         change_mode(result_csv_path, FileCheckConst.DATA_FILE_AUTHORITY)
@@ -198,17 +116,23 @@ def run_api_offline(config, compare, api_name_set):
         if api_full_name in api_name_set:
             continue
         if is_unsupported_api(api_full_name):
+            skip_message = f"API {api_full_name} not support for run ut. SKIP."
+            compare_alg_results = err_column.to_column_value(CompareConst.SKIP, skip_message)
+            record_skip_info(api_full_name, compare, compare_alg_results)
             continue
         _, api_name = extract_basic_api_segments(api_full_name)
         if not api_name:
             err_message = f"API {api_full_name} not support for run ut. SKIP."
             logger.error(err_message)
-            fwd_compare_alg_results = err_column.to_column_value(CompareConst.SKIP, err_message)
-            result_info = (api_full_name, CompareConst.SKIP, CompareConst.SKIP, [fwd_compare_alg_results], None, 0)
-            compare.record_results(result_info)
+            compare_alg_results = err_column.to_column_value(CompareConst.SKIP, err_message)
+            record_skip_info(api_full_name, compare, compare_alg_results)
             continue
         try:
             if blacklist_and_whitelist_filter(api_name, config.black_list, config.white_list):
+                skip_message = f"API {api_name} in black list or not in white list. SKIP."
+                logger.info(skip_message)
+                compare_alg_results = err_column.to_column_value(CompareConst.SKIP, skip_message)
+                record_skip_info(api_full_name, compare, compare_alg_results)
                 continue
             data_info = run_torch_api(api_full_name, config.real_data_path, config.backward_content, api_info_dict)
             is_fwd_success, is_bwd_success = compare.compare_output(api_full_name, data_info)
@@ -217,12 +141,11 @@ def run_api_offline(config, compare, api_name_set):
         except Exception as err:
             if "expected scalar type Long" in str(err):
                 logger.warning(f"API {api_name} not support int32 tensor in CPU, please add {api_name} to CONVERT_API "
-                               f"'int32_to_int64' list in accuracy_tools/api_accuracy_check/common/utils.py file.")
+                               "'int32_to_int64' list in accuracy_tools/msprobe/core/common/const.py file.")
             else:
                 logger.error(f"Run {api_full_name} UT Error: %s" % str(err))
-            fwd_compare_alg_results = err_column.to_column_value(CompareConst.SKIP, str(err))
-            result_info = (api_full_name, CompareConst.SKIP, CompareConst.SKIP, [fwd_compare_alg_results], None, 0)
-            compare.record_results(result_info)
+            compare_alg_results = err_column.to_column_value(CompareConst.SKIP, str(err))
+            record_skip_info(api_full_name, compare, compare_alg_results)
         finally:
             if is_gpu:
                 torch.cuda.empty_cache()
@@ -298,14 +221,6 @@ def blacklist_and_whitelist_filter(api_name, black_list, white_list):
     return False
 
 
-def is_unsupported_api(api_name):
-    split_name = api_name.split(Const.SEP)[0]
-    flag = split_name == Const.DISTRIBUTED
-    if flag:
-        logger.info(f"{split_name} api is not supported for run ut. SKIP.")
-    return flag
-
-
 def do_save_error_data(api_full_name, data_info, error_data_path, is_fwd_success, is_bwd_success):
     if not is_fwd_success or not is_bwd_success:
         processor = UtDataProcessor(error_data_path)
@@ -327,12 +242,12 @@ def run_torch_api(api_full_name, real_data_path, backward_content, api_info_dict
     in_fwd_data_list.append(kwargs)
     need_backward = api_full_name in backward_content
     if not need_grad:
-        logger.warning("%s %s" % (api_full_name, Backward_Message.UNSUPPORT_BACKWARD_MESSAGE))
-        backward_message += Backward_Message.UNSUPPORT_BACKWARD_MESSAGE
+        logger.warning("%s %s" % (api_full_name, BackwardMessage.UNSUPPORT_BACKWARD_MESSAGE))
+        backward_message += BackwardMessage.UNSUPPORT_BACKWARD_MESSAGE
     if api_name in not_backward_list:
         need_grad = False
-        logger.warning("%s %s" % (api_full_name, Backward_Message.NO_BACKWARD_RESULT_MESSAGE))
-        backward_message += Backward_Message.NO_BACKWARD_RESULT_MESSAGE
+        logger.info("%s %s" % (api_full_name, BackwardMessage.NO_BACKWARD_RESULT_MESSAGE))
+        backward_message += BackwardMessage.NO_BACKWARD_RESULT_MESSAGE
     need_backward = need_backward and need_grad
     if kwargs.get("device"):
         del kwargs["device"]
@@ -353,16 +268,20 @@ def run_torch_api(api_full_name, real_data_path, backward_content, api_info_dict
     if need_backward:
         if need_to_backward(grad_index, out):
             backward_args = backward_content[api_full_name].get("input")
-            grad = gen_args(backward_args, api_name, real_data_path=real_data_path)[0]
+            func_options = {
+                'real_data_path': real_data_path
+            }
+            grad = gen_args(backward_args, api_name, func_options)
+            grad = safe_get_value(grad, 0, "grad")
             bench_grad, _ = generate_cpu_params(grad, {}, False, api_name)
             bench_grad_out = run_backward(cpu_args, bench_grad, grad_index, out)
             device_grad = grad.clone().detach().to(current_device)
             device_grad_out = run_backward(device_args, device_grad, grad_index, device_out)
         else:
-            backward_message += Backward_Message.MULTIPLE_BACKWARD_MESSAGE
+            backward_message += BackwardMessage.MULTIPLE_BACKWARD_MESSAGE
     if api_name == "npu_fusion_attention":
-        out = out[0]
-        device_out = device_out[0]
+        out = safe_get_value(out, 0, "out")
+        device_out = safe_get_value(device_out, 0, "device_out")
 
     return UtDataInfo(bench_grad_out, device_grad_out, device_out, out, bench_grad, in_fwd_data_list, backward_message)
 
@@ -398,6 +317,9 @@ def need_to_backward(grad_index, out):
 
 def run_backward(args, grad, grad_index, out):
     if grad_index is not None:
+        if grad_index >= len(out):
+            logger.error(f"Run backward error when grad_index is {grad_index}")
+            raise IndexError(f"Run backward error when grad_index is {grad_index}")
         out[grad_index].backward(grad)
     else:
         out.backward(grad)
@@ -411,12 +333,11 @@ def run_backward(args, grad, grad_index, out):
 
 
 def initialize_save_error_data(error_data_path):
-    check_path_before_create(error_data_path)
     create_directory(error_data_path)
     error_data_path_checker = FileChecker(error_data_path, FileCheckConst.DIR,
                                           ability=FileCheckConst.WRITE_ABLE)
     error_data_path = error_data_path_checker.common_check()
-    error_data_path =initialize_save_path(error_data_path, UT_ERROR_DATA_DIR)
+    error_data_path = initialize_save_path(error_data_path, UT_ERROR_DATA_DIR)
     return error_data_path
 
 
@@ -477,7 +398,8 @@ def preprocess_forward_content(forward_content):
         if key not in arg_cache:
             filtered_new_args = [
                 {k: v for k, v in arg.items() if k not in ['Max', 'Min']}
-                for arg in value['input_args'] if isinstance(arg, dict)
+                for arg in value['input_args']
+                if isinstance(arg, dict)
             ]
             arg_cache[key] = (filtered_new_args, value['input_kwargs'])
 
@@ -512,7 +434,49 @@ def _run_ut(parser=None):
     run_ut_command(args)
 
 
+def checked_online_config(online_config):
+    if not online_config.is_online:
+        return
+    if not isinstance(online_config.is_online, bool):
+        raise ValueError("is_online must be bool type")
+    # rank_list
+    if not isinstance(online_config.rank_list, list):
+        raise ValueError("rank_list must be a list")
+    if online_config.rank_list and not all(isinstance(rank, int) for rank in online_config.rank_list):
+        raise ValueError("All elements in rank_list must be integers")
+
+    # nfs_path
+    if online_config.nfs_path:
+        check_file_or_directory_path(online_config.nfs_path, isdir=True)
+        return
+    # tls_path
+    if online_config.tls_path:
+        check_file_or_directory_path(online_config.tls_path, isdir=True)
+        check_file_or_directory_path(os.path.join(online_config.tls_path, "server.key"))
+        check_file_or_directory_path(os.path.join(online_config.tls_path, "server.crt"))
+        check_crt_valid(os.path.join(online_config.tls_path, "server.crt"))
+
+    # host and port
+    if not isinstance(online_config.host, str) or not re.match(Const.ipv4_pattern, online_config.host):
+        raise Exception(f"host: {online_config.host} is invalid.")
+    if not isinstance(online_config.port, int) or not (0 < online_config.port <= 65535):
+        raise Exception(f"port: {online_config.port} is invalid, port range 0-65535.")
+
+
 def run_ut_command(args):
+    if args.config_path:
+        config_path_checker = FileChecker(args.config_path, FileCheckConst.FILE, 
+                                          FileCheckConst.READ_ABLE, FileCheckConst.JSON_SUFFIX)
+        checked_config_path = config_path_checker.common_check()
+        _, task_config = parse_json_config(checked_config_path, Const.RUN_UT)
+        checker_config = CheckerConfig(task_config)
+    else:
+        checker_config = CheckerConfig()
+    
+    if not checker_config.is_online and not args.api_info_file:
+        logger.error("Please provide api_info_file for offline run ut.")
+        raise Exception("Please provide api_info_file for offline run ut.")
+
     if not is_gpu:
         torch.npu.set_compile_mode(jit_compile=args.jit_compile)
     used_device = current_device + ":" + str(args.device_id[0])
@@ -529,17 +493,16 @@ def run_ut_command(args):
     # 离线场景下，forward_content, backward_content, real_data_path从api_info_file中解析
     forward_content, backward_content, real_data_path = None, None, None
     if args.api_info_file:
-        api_info_file_checker = FileChecker(file_path = args.api_info_file, path_type = FileCheckConst.FILE, 
-                                            ability = FileCheckConst.READ_ABLE, file_type = FileCheckConst.JSON_SUFFIX)
+        api_info_file_checker = FileChecker(file_path=args.api_info_file, path_type=FileCheckConst.FILE, 
+                                            ability=FileCheckConst.READ_ABLE, file_type=FileCheckConst.JSON_SUFFIX)
         checked_api_info = api_info_file_checker.common_check()
         forward_content, backward_content, real_data_path = parse_json_info_forward_backward(checked_api_info)
         if args.filter_api:
-            logger.info("Start filtering the api in the forward_input_file.")
+            logger.info("Start filtering the api in the api_info_file.")
             forward_content = preprocess_forward_content(forward_content)
-            logger.info("Finish filtering the api in the forward_input_file.")
+            logger.info("Finish filtering the api in the api_info_file.")
 
-    out_path = os.path.realpath(args.out_path) if args.out_path else "./"
-    check_path_before_create(out_path)
+    out_path = args.out_path if args.out_path else Const.DEFAULT_PATH
     create_directory(out_path)
     out_path_checker = FileChecker(out_path, FileCheckConst.DIR, ability=FileCheckConst.WRITE_ABLE)
     out_path = out_path_checker.common_check()
@@ -550,40 +513,27 @@ def run_ut_command(args):
     if args.result_csv_path:
         result_csv_path = get_validated_result_csv_path(args.result_csv_path, 'result')
         details_csv_path = get_validated_details_csv_path(result_csv_path)
-    white_list = msCheckerConfig.white_list
-    black_list = msCheckerConfig.black_list
-    error_data_path = msCheckerConfig.error_data_path
-    is_online = msCheckerConfig.is_online
-    nfs_path = msCheckerConfig.nfs_path
-    host = msCheckerConfig.host
-    port = msCheckerConfig.port
-    rank_list = msCheckerConfig.rank_list
-    tls_path = msCheckerConfig.tls_path
-    if args.config_path:
-        config_path_checker = FileChecker(args.config_path, FileCheckConst.FILE, 
-                                          FileCheckConst.READ_ABLE, FileCheckConst.JSON_SUFFIX)
-        checked_config_path = config_path_checker.common_check()
-        _, task_config = parse_json_config(checked_config_path, Const.RUN_UT)
-        white_list = task_config.white_list
-        black_list = task_config.black_list
-        error_data_path = task_config.error_data_path
-        is_online = task_config.is_online
-        nfs_path = task_config.nfs_path
-        host = task_config.host
-        port = task_config.port
-        rank_list = task_config.rank_list
-        tls_path = task_config.tls_path
 
+    error_data_path = checker_config.error_data_path
     if save_error_data:
         if args.result_csv_path:
             time_info = result_csv_path.split('.')[0].split('_')[-1]
             global UT_ERROR_DATA_DIR
             UT_ERROR_DATA_DIR = 'ut_error_data' + time_info
         error_data_path = initialize_save_error_data(error_data_path)
-    online_config = OnlineConfig(is_online, nfs_path, host, port, rank_list, tls_path)
-    run_ut_config = RunUTConfig(forward_content, backward_content, result_csv_path, details_csv_path, save_error_data,
-                                args.result_csv_path, real_data_path, set(white_list), set(black_list), error_data_path,
-                                online_config)
+    online_config = checker_config.get_online_config()
+    checked_online_config(online_config)
+    config_params = {
+        'forward_content': forward_content,
+        'backward_content': backward_content,
+        'result_csv_path': result_csv_path,
+        'details_csv_path': details_csv_path,
+        'save_error_data': save_error_data,
+        'is_continue_run_ut': args.result_csv_path,
+        'real_data_path': real_data_path,
+        'error_data_path': error_data_path
+    }
+    run_ut_config = checker_config.get_run_ut_config(**config_params)
     run_ut(run_ut_config)