forked from LLM-Dev-Open/opencompass_base
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrun.py
250 lines (225 loc) · 10.1 KB
/
run.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
import argparse
import getpass
import os
import os.path as osp
from datetime import datetime
from mmengine.config import Config
from opencompass.partitioners import NaivePartitioner, SizePartitioner
from opencompass.runners import DLCRunner, LocalRunner, SlurmRunner
from opencompass.utils import LarkReporter, Summarizer, get_logger
def parse_args():
parser = argparse.ArgumentParser(description='Run an evaluation task')
parser.add_argument('config', help='Train config file path')
# add mutually exclusive args `--slurm` `--dlc`, default to local runner
luach_method = parser.add_mutually_exclusive_group()
luach_method.add_argument('--slurm',
action='store_true',
default=False,
help='Whether to use srun to launch tasks, if '
'True, `--partition(-p)` must be set. Defaults'
' to False')
luach_method.add_argument('--dlc',
action='store_true',
default=False,
help='Whether to use dlc to launch tasks, if '
'True, `--aliyun-cfg` must be set. Defaults'
' to False')
# add general args
parser.add_argument('--debug',
help='Debug mode, in which scheduler will run tasks '
'in the single process, and output will not be '
'redirected to files',
action='store_true',
default=False)
parser.add_argument('-m',
'--mode',
help='Running mode. You can choose "infer" if you '
'only want the inference results, or "eval" if you '
'already have the results and want to evaluate them, '
'or "viz" if you want to visualize the results.',
choices=['all', 'infer', 'eval', 'viz'],
default='all',
type=str)
parser.add_argument('-r',
'--reuse',
nargs='?',
type=str,
const='latest',
help='Reuse previous outputs & results, and run any '
'missing jobs presented in the config. If its '
'argument is not specified, the latest results in '
'the work_dir will be reused. The argument should '
'also be a specific timestamp, e.g. 20230516_144254'),
parser.add_argument('-w',
'--work-dir',
help='Work path, all the outputs will be saved in '
'this path, including the slurm logs, the evaluation'
' results, the summary results, etc. If not specified,'
' the work_dir will be set to None',
default=None,
type=str)
parser.add_argument('-l',
'--lark',
help='Report the running status to lark bot',
action='store_true',
default=False)
parser.add_argument('--max-partition-size',
help='The maximum size of a task.',
type=int,
default=2000),
parser.add_argument(
'--gen-task-coef',
help='The dataset cost measurement coefficient for generation tasks',
type=int,
default=20)
parser.add_argument('--max-num-workers',
help='Max number of workers to run in parallel.',
type=int,
default=32)
parser.add_argument(
'--retry',
help='Number of retries if the job failed when using slurm or dlc.',
type=int,
default=2)
# set srun args
slurm_parser = parser.add_argument_group('slurm_args')
parse_slurm_args(slurm_parser)
# set dlc args
dlc_parser = parser.add_argument_group('dlc_args')
parse_dlc_args(dlc_parser)
args = parser.parse_args()
if args.slurm:
assert args.partition is not None, (
'--partition(-p) must be set if you want to use slurm')
if args.dlc:
assert os.path.exists(args.aliyun_cfg), (
'When luaching tasks using dlc, it needs to be configured'
'in "~/.aliyun.cfg", or use "--aliyun-cfg $ALiYun-CFG_Path"'
' to specify a new path.')
return args
def parse_slurm_args(slurm_parser):
"""these args are all for slurm launch."""
slurm_parser.add_argument('-p',
'--partition',
help='Slurm partition name',
default=None,
type=str)
slurm_parser.add_argument('-q',
'--quotatype',
help='Slurm quota type',
default='auto',
type=str)
def parse_dlc_args(dlc_parser):
"""these args are all for dlc launch."""
dlc_parser.add_argument('--aliyun-cfg',
help='The config path for aliyun config',
default='~/.aliyun.cfg',
type=str)
def main():
args = parse_args()
# initialize logger
logger = get_logger(log_level='DEBUG' if args.debug else 'INFO')
cfg = Config.fromfile(args.config)
if args.work_dir is not None:
cfg['work_dir'] = args.work_dir
else:
cfg.setdefault('work_dir', './outputs/default/')
# cfg_time_str defaults to the current time
cfg_time_str = dir_time_str = datetime.now().strftime('%Y%m%d_%H%M%S')
if args.reuse:
if args.reuse == 'latest':
dirs = os.listdir(cfg.work_dir)
assert len(dirs) > 0, 'No previous results to reuse!'
dir_time_str = sorted(dirs)[-1]
else:
dir_time_str = args.reuse
logger.info(f'Reusing experiements from {dir_time_str}')
elif args.mode in ['eval', 'viz']:
raise ValueError('You must specify -r or --reuse when running in eval '
'or viz mode!')
# update "actual" work_dir
cfg['work_dir'] = osp.join(cfg.work_dir, dir_time_str)
os.makedirs(osp.join(cfg.work_dir, 'configs'), exist_ok=True)
# dump config
output_config_path = osp.join(cfg.work_dir, 'configs',
f'{cfg_time_str}.py')
cfg.dump(output_config_path)
# Config is intentally reloaded here to avoid initialized
# types cannot be serialized
cfg = Config.fromfile(output_config_path)
# report to lark bot if specify --lark
if not args.lark:
cfg['lark_bot_url'] = None
elif cfg.get('lark_bot_url', None):
content = f'{getpass.getuser()}\'s task has been launched!'
LarkReporter(cfg['lark_bot_url']).post(content)
if args.mode in ['all', 'infer']:
# Use SizePartitioner to split into subtasks
partitioner = SizePartitioner(osp.join(cfg['work_dir'],
'predictions/'),
max_task_size=args.max_partition_size,
gen_task_coef=args.gen_task_coef)
tasks = partitioner(cfg)
# execute the infer subtasks
exec_infer_runner(tasks, args, cfg)
# evaluate
if args.mode in ['all', 'eval']:
# Use NaivePartitioner,not split
partitioner = NaivePartitioner(osp.join(cfg['work_dir'], 'results/'))
tasks = partitioner(cfg)
# execute the eval tasks
exec_eval_runner(tasks, args, cfg)
# visualize
if args.mode in ['all', 'eval', 'viz']:
summarizer = Summarizer(cfg)
summarizer.summarize(time_str=cfg_time_str)
def exec_infer_runner(tasks, args, cfg):
"""execute infer runner according to args."""
if args.slurm:
runner = SlurmRunner(dict(type='OpenICLInferTask'),
max_num_workers=args.max_num_workers,
partition=args.partition,
quotatype=args.quotatype,
retry=args.retry,
debug=args.debug,
lark_bot_url=cfg['lark_bot_url'])
elif args.dlc:
runner = DLCRunner(dict(type='OpenICLInferTask'),
max_num_workers=args.max_num_workers,
aliyun_cfg=Config.fromfile(args.aliyun_cfg),
retry=args.retry,
debug=args.debug,
lark_bot_url=cfg['lark_bot_url'])
else:
runner = LocalRunner(
task=dict(type='OpenICLInferTask'),
# max_num_workers = args.max_num_workers,
debug=args.debug,
lark_bot_url=cfg['lark_bot_url'])
runner(tasks)
def exec_eval_runner(tasks, args, cfg):
"""execute infer runner according to args."""
if args.slurm:
runner = SlurmRunner(dict(type='OpenICLEvalTask'),
max_num_workers=args.max_num_workers,
partition=args.partition,
quotatype=args.quotatype,
retry=args.retry,
debug=args.debug,
lark_bot_url=cfg['lark_bot_url'])
elif args.dlc:
runner = DLCRunner(dict(type='OpenICLEvalTask'),
max_num_workers=args.max_num_workers,
aliyun_cfg=Config.fromfile(args.aliyun_cfg),
retry=args.retry,
debug=args.debug,
lark_bot_url=cfg['lark_bot_url'])
else:
runner = LocalRunner(
task=dict(type='OpenICLEvalTask'),
# max_num_workers = args.max_num_workers,
debug=args.debug,
lark_bot_url=cfg['lark_bot_url'])
runner(tasks)
if __name__ == '__main__':
main()