mirror of
https://github.com/open-compass/opencompass.git
synced 2025-05-30 16:03:24 +08:00
8ea2c404d7
* enable HuggingFacewithChatTemplate with --accelerator via cli * rm vllm_internlm2_chat_7b
344 lines
16 KiB
Python
344 lines
16 KiB
Python
# flake8: noqa
|
|
# yapf: disable
|
|
import os
|
|
from typing import List, Tuple, Union
|
|
|
|
import tabulate
|
|
from mmengine.config import Config
|
|
|
|
from opencompass.datasets.custom import make_custom_dataset_config
|
|
from opencompass.models import (VLLM, HuggingFace, HuggingFaceBaseModel,
|
|
HuggingFaceCausalLM, HuggingFaceChatGLM3,
|
|
HuggingFacewithChatTemplate, TurboMindModel,
|
|
TurboMindModelwithChatTemplate,
|
|
VLLMwithChatTemplate)
|
|
from opencompass.partitioners import NaivePartitioner, NumWorkerPartitioner
|
|
from opencompass.runners import DLCRunner, LocalRunner, SlurmRunner
|
|
from opencompass.tasks import OpenICLEvalTask, OpenICLInferTask
|
|
from opencompass.utils import get_logger, match_files
|
|
|
|
|
|
def match_cfg_file(workdir: str,
|
|
pattern: Union[str, List[str]]) -> List[Tuple[str, str]]:
|
|
"""Match the config file in workdir recursively given the pattern.
|
|
|
|
Additionally, if the pattern itself points to an existing file, it will be
|
|
directly returned.
|
|
"""
|
|
if isinstance(pattern, str):
|
|
pattern = [pattern]
|
|
pattern = [p + '.py' if not p.endswith('.py') else p for p in pattern]
|
|
files = match_files(workdir, pattern, fuzzy=False)
|
|
if len(files) != len(pattern):
|
|
nomatched = []
|
|
ambiguous = []
|
|
err_msg = ('The provided pattern matches 0 or more than one '
|
|
'config. Please verify your pattern and try again. '
|
|
'You may use tools/list_configs.py to list or '
|
|
'locate the configurations.\n')
|
|
for p in pattern:
|
|
files = match_files(workdir, p, fuzzy=False)
|
|
if len(files) == 0:
|
|
nomatched.append([p[:-3]])
|
|
elif len(files) > 1:
|
|
ambiguous.append([p[:-3], '\n'.join(f[1] for f in files)])
|
|
if nomatched:
|
|
table = [['Not matched patterns'], *nomatched]
|
|
err_msg += tabulate.tabulate(table,
|
|
headers='firstrow',
|
|
tablefmt='psql')
|
|
if ambiguous:
|
|
table = [['Ambiguous patterns', 'Matched files'], *ambiguous]
|
|
err_msg += tabulate.tabulate(table,
|
|
headers='firstrow',
|
|
tablefmt='psql')
|
|
raise ValueError(err_msg)
|
|
return files
|
|
|
|
|
|
def try_fill_in_custom_cfgs(config):
|
|
for i, dataset in enumerate(config['datasets']):
|
|
if 'type' not in dataset:
|
|
config['datasets'][i] = make_custom_dataset_config(dataset)
|
|
if 'model_dataset_combinations' not in config:
|
|
return config
|
|
for mdc in config['model_dataset_combinations']:
|
|
for i, dataset in enumerate(mdc['datasets']):
|
|
if 'type' not in dataset:
|
|
mdc['datasets'][i] = make_custom_dataset_config(dataset)
|
|
return config
|
|
|
|
|
|
def get_config_from_arg(args) -> Config:
|
|
"""Get the config object given args.
|
|
|
|
Only a few argument combinations are accepted (priority from high to low)
|
|
1. args.config
|
|
2. args.models and args.datasets
|
|
3. Huggingface parameter groups and args.datasets
|
|
"""
|
|
logger = get_logger()
|
|
if args.config:
|
|
config = Config.fromfile(args.config, format_python_code=False)
|
|
config = try_fill_in_custom_cfgs(config)
|
|
# set infer accelerator if needed
|
|
if args.accelerator in ['vllm', 'lmdeploy']:
|
|
config['models'] = change_accelerator(config['models'], args.accelerator)
|
|
if 'eval' in config and 'partitioner' in config['eval']:
|
|
config['eval']['partitioner']['models'] = change_accelerator(config['eval']['partitioner']['models'], args.accelerator)
|
|
if config.get('eval', {}).get('partitioner', {}).get('judge_models') is not None:
|
|
config['eval']['partitioner']['judge_models'] = change_accelerator(config['eval']['partitioner']['judge_models'], args.accelerator)
|
|
return config
|
|
|
|
# parse dataset args
|
|
if not args.datasets and not args.custom_dataset_path:
|
|
raise ValueError('You must specify "--datasets" or "--custom-dataset-path" if you do not specify a config file path.')
|
|
datasets = []
|
|
if args.datasets:
|
|
datasets_dir = os.path.join(args.config_dir, 'datasets')
|
|
for dataset_arg in args.datasets:
|
|
if '/' in dataset_arg:
|
|
dataset_name, dataset_suffix = dataset_arg.split('/', 1)
|
|
dataset_key_suffix = dataset_suffix
|
|
else:
|
|
dataset_name = dataset_arg
|
|
dataset_key_suffix = '_datasets'
|
|
|
|
for dataset in match_cfg_file(datasets_dir, [dataset_name]):
|
|
logger.info(f'Loading {dataset[0]}: {dataset[1]}')
|
|
cfg = Config.fromfile(dataset[1])
|
|
for k in cfg.keys():
|
|
if k.endswith(dataset_key_suffix):
|
|
datasets += cfg[k]
|
|
else:
|
|
dataset = {'path': args.custom_dataset_path}
|
|
if args.custom_dataset_infer_method is not None:
|
|
dataset['infer_method'] = args.custom_dataset_infer_method
|
|
if args.custom_dataset_data_type is not None:
|
|
dataset['data_type'] = args.custom_dataset_data_type
|
|
if args.custom_dataset_meta_path is not None:
|
|
dataset['meta_path'] = args.custom_dataset_meta_path
|
|
dataset = make_custom_dataset_config(dataset)
|
|
datasets.append(dataset)
|
|
|
|
# parse model args
|
|
if not args.models and not args.hf_path:
|
|
raise ValueError('You must specify a config file path, or specify --models and --datasets, or specify HuggingFace model parameters and --datasets.')
|
|
models = []
|
|
if args.models:
|
|
model_dir = os.path.join(args.config_dir, 'models')
|
|
for model in match_cfg_file(model_dir, args.models):
|
|
logger.info(f'Loading {model[0]}: {model[1]}')
|
|
cfg = Config.fromfile(model[1])
|
|
if 'models' not in cfg:
|
|
raise ValueError(f'Config file {model[1]} does not contain "models" field')
|
|
models += cfg['models']
|
|
else:
|
|
if args.hf_type == 'chat':
|
|
mod = HuggingFacewithChatTemplate
|
|
else:
|
|
mod = HuggingFaceBaseModel
|
|
model = dict(type=f'{mod.__module__}.{mod.__name__}',
|
|
abbr=args.hf_path.split('/')[-1] + '_hf',
|
|
path=args.hf_path,
|
|
model_kwargs=args.model_kwargs,
|
|
tokenizer_path=args.tokenizer_path,
|
|
tokenizer_kwargs=args.tokenizer_kwargs,
|
|
peft_path=args.peft_path,
|
|
peft_kwargs=args.peft_kwargs,
|
|
max_seq_len=args.max_seq_len,
|
|
max_out_len=args.max_out_len,
|
|
batch_size=args.batch_size,
|
|
pad_token_id=args.pad_token_id,
|
|
stop_words=args.stop_words,
|
|
run_cfg=dict(num_gpus=args.num_gpus))
|
|
logger.debug(f'Using model: {model}')
|
|
models.append(model)
|
|
# set infer accelerator if needed
|
|
if args.accelerator in ['vllm', 'lmdeploy']:
|
|
models = change_accelerator(models, args.accelerator)
|
|
# parse summarizer args
|
|
summarizer_arg = args.summarizer if args.summarizer is not None else 'example'
|
|
summarizers_dir = os.path.join(args.config_dir, 'summarizers')
|
|
|
|
# Check if summarizer_arg contains '/'
|
|
if '/' in summarizer_arg:
|
|
# If it contains '/', split the string by '/'
|
|
# and use the second part as the configuration key
|
|
summarizer_file, summarizer_key = summarizer_arg.split('/', 1)
|
|
else:
|
|
# If it does not contain '/', keep the original logic unchanged
|
|
summarizer_key = 'summarizer'
|
|
summarizer_file = summarizer_arg
|
|
|
|
s = match_cfg_file(summarizers_dir, [summarizer_file])[0]
|
|
logger.info(f'Loading {s[0]}: {s[1]}')
|
|
cfg = Config.fromfile(s[1])
|
|
# Use summarizer_key to retrieve the summarizer definition
|
|
# from the configuration file
|
|
summarizer = cfg[summarizer_key]
|
|
|
|
return Config(dict(models=models, datasets=datasets, summarizer=summarizer), format_python_code=False)
|
|
|
|
|
|
def change_accelerator(models, accelerator):
|
|
models = models.copy()
|
|
logger = get_logger()
|
|
model_accels = []
|
|
for model in models:
|
|
logger.info(f'Transforming {model["abbr"]} to {accelerator}')
|
|
# change HuggingFace model to VLLM or TurboMindModel
|
|
if model['type'] in [HuggingFace, HuggingFaceCausalLM, HuggingFaceChatGLM3]:
|
|
gen_args = dict()
|
|
if model.get('generation_kwargs') is not None:
|
|
generation_kwargs = model['generation_kwargs'].copy()
|
|
gen_args['temperature'] = generation_kwargs.get('temperature', 0.001)
|
|
gen_args['top_k'] = generation_kwargs.get('top_k', 1)
|
|
gen_args['top_p'] = generation_kwargs.get('top_p', 0.9)
|
|
gen_args['stop_token_ids'] = generation_kwargs.get('eos_token_id', None)
|
|
generation_kwargs['stop_token_ids'] = generation_kwargs.get('eos_token_id', None)
|
|
generation_kwargs.pop('eos_token_id')
|
|
else:
|
|
# if generation_kwargs is not provided, set default values
|
|
generation_kwargs = dict()
|
|
gen_args['temperature'] = 0.0
|
|
gen_args['top_k'] = 1
|
|
gen_args['top_p'] = 0.9
|
|
gen_args['stop_token_ids'] = None
|
|
|
|
if accelerator == 'lmdeploy':
|
|
logger.info(f'Transforming {model["abbr"]} to {accelerator}')
|
|
mod = TurboMindModel
|
|
acc_model = dict(
|
|
type=f'{mod.__module__}.{mod.__name__}',
|
|
abbr=model['abbr'].replace('hf', 'lmdeploy') if '-hf' in model['abbr'] else model['abbr'] + '-lmdeploy',
|
|
path=model['path'],
|
|
engine_config=dict(session_len=model['max_seq_len'],
|
|
max_batch_size=model['batch_size'],
|
|
tp=model['run_cfg']['num_gpus']),
|
|
gen_config=dict(top_k=gen_args['top_k'],
|
|
temperature=gen_args['temperature'],
|
|
top_p=gen_args['top_p'],
|
|
max_new_tokens=model['max_out_len'],
|
|
stop_words=gen_args['stop_token_ids']),
|
|
max_out_len=model['max_out_len'],
|
|
max_seq_len=model['max_seq_len'],
|
|
batch_size=model['batch_size'],
|
|
concurrency=model['batch_size'],
|
|
run_cfg=model['run_cfg'],
|
|
)
|
|
for item in ['meta_template']:
|
|
if model.get(item) is not None:
|
|
acc_model[item] = model[item]
|
|
elif accelerator == 'vllm':
|
|
logger.info(f'Transforming {model["abbr"]} to {accelerator}')
|
|
|
|
acc_model = dict(
|
|
type=f'{VLLM.__module__}.{VLLM.__name__}',
|
|
abbr=model['abbr'].replace('hf', 'vllm') if '-hf' in model['abbr'] else model['abbr'] + '-vllm',
|
|
path=model['path'],
|
|
model_kwargs=dict(tensor_parallel_size=model['run_cfg']['num_gpus']),
|
|
max_out_len=model['max_out_len'],
|
|
max_seq_len=model['max_seq_len'],
|
|
batch_size=model['batch_size'],
|
|
generation_kwargs=generation_kwargs,
|
|
run_cfg=model['run_cfg'],
|
|
)
|
|
for item in ['meta_template', 'end_str']:
|
|
if model.get(item) is not None:
|
|
acc_model[item] = model[item]
|
|
else:
|
|
raise ValueError(f'Unsupported accelerator {accelerator} for model type {model["type"]}')
|
|
elif model['type'] in [HuggingFacewithChatTemplate]:
|
|
if accelerator == 'vllm':
|
|
mod = VLLMwithChatTemplate
|
|
acc_model = dict(
|
|
type=f'{mod.__module__}.{mod.__name__}',
|
|
abbr='-hf'.join(model['abbr'].split('-hf')[:-1]) + '-vllm',
|
|
path=model['path'],
|
|
model_kwargs=dict(tensor_parallel_size=model['run_cfg']['num_gpus']),
|
|
max_out_len=model['max_out_len'],
|
|
batch_size=32768,
|
|
run_cfg=model['run_cfg'],
|
|
stop_words=model.get('stop_words', []),
|
|
)
|
|
elif accelerator == 'lmdeploy':
|
|
mod = TurboMindModelwithChatTemplate
|
|
acc_model = dict(
|
|
type=f'{mod.__module__}.{mod.__name__}',
|
|
abbr='-hf'.join(model['abbr'].split('-hf')[:-1]) + '-turbomind',
|
|
path=model['path'],
|
|
engine_config=dict(max_batch_size=model.get('batch_size', 16), tp=model['run_cfg']['num_gpus']),
|
|
gen_config=dict(top_k=1, temperature=1e-6, top_p=0.9),
|
|
max_seq_len=model.get('max_seq_len', 2048),
|
|
max_out_len=model['max_out_len'],
|
|
batch_size=32768,
|
|
run_cfg=model['run_cfg'],
|
|
stop_words=model.get('stop_words', []),
|
|
)
|
|
else:
|
|
raise ValueError(f'Unsupported accelerator {accelerator} for model type {model["type"]}')
|
|
else:
|
|
raise ValueError(f'Unsupported model type {model["type"]}')
|
|
model_accels.append(acc_model)
|
|
return model_accels
|
|
|
|
|
|
def get_config_type(obj) -> str:
|
|
return f'{obj.__module__}.{obj.__name__}'
|
|
|
|
|
|
def fill_infer_cfg(cfg, args):
|
|
new_cfg = dict(infer=dict(
|
|
partitioner=dict(type=get_config_type(NumWorkerPartitioner),
|
|
num_worker=args.max_num_workers),
|
|
runner=dict(
|
|
max_num_workers=args.max_num_workers,
|
|
debug=args.debug,
|
|
task=dict(type=get_config_type(OpenICLInferTask)),
|
|
lark_bot_url=cfg['lark_bot_url'],
|
|
)), )
|
|
if args.slurm:
|
|
new_cfg['infer']['runner']['type'] = get_config_type(SlurmRunner)
|
|
new_cfg['infer']['runner']['partition'] = args.partition
|
|
new_cfg['infer']['runner']['quotatype'] = args.quotatype
|
|
new_cfg['infer']['runner']['qos'] = args.qos
|
|
new_cfg['infer']['runner']['retry'] = args.retry
|
|
elif args.dlc:
|
|
new_cfg['infer']['runner']['type'] = get_config_type(DLCRunner)
|
|
new_cfg['infer']['runner']['aliyun_cfg'] = Config.fromfile(
|
|
args.aliyun_cfg)
|
|
new_cfg['infer']['runner']['retry'] = args.retry
|
|
else:
|
|
new_cfg['infer']['runner']['type'] = get_config_type(LocalRunner)
|
|
new_cfg['infer']['runner'][
|
|
'max_workers_per_gpu'] = args.max_workers_per_gpu
|
|
cfg.merge_from_dict(new_cfg)
|
|
|
|
|
|
def fill_eval_cfg(cfg, args):
|
|
new_cfg = dict(
|
|
eval=dict(partitioner=dict(type=get_config_type(NaivePartitioner)),
|
|
runner=dict(
|
|
max_num_workers=args.max_num_workers,
|
|
debug=args.debug,
|
|
task=dict(type=get_config_type(OpenICLEvalTask)),
|
|
lark_bot_url=cfg['lark_bot_url'],
|
|
)))
|
|
if args.slurm:
|
|
new_cfg['eval']['runner']['type'] = get_config_type(SlurmRunner)
|
|
new_cfg['eval']['runner']['partition'] = args.partition
|
|
new_cfg['eval']['runner']['quotatype'] = args.quotatype
|
|
new_cfg['eval']['runner']['qos'] = args.qos
|
|
new_cfg['eval']['runner']['retry'] = args.retry
|
|
elif args.dlc:
|
|
new_cfg['eval']['runner']['type'] = get_config_type(DLCRunner)
|
|
new_cfg['eval']['runner']['aliyun_cfg'] = Config.fromfile(
|
|
args.aliyun_cfg)
|
|
new_cfg['eval']['runner']['retry'] = args.retry
|
|
else:
|
|
new_cfg['eval']['runner']['type'] = get_config_type(LocalRunner)
|
|
new_cfg['eval']['runner'][
|
|
'max_workers_per_gpu'] = args.max_workers_per_gpu
|
|
cfg.merge_from_dict(new_cfg)
|