run.py

import argparse
import getpass
import os
import os.path as osp
from datetime import datetime

from mmengine.config import Config, DictAction

from opencompass.partitioners import (MultimodalNaivePartitioner,
                                      NaivePartitioner, SizePartitioner)
from opencompass.registry import PARTITIONERS, RUNNERS
from opencompass.runners import SlurmRunner
from opencompass.utils import LarkReporter, Summarizer, get_logger
from opencompass.utils.run import (exec_eval_runner, exec_infer_runner,
                                   exec_mm_infer_runner, get_config_from_arg)


def parse_args():
    parser = argparse.ArgumentParser(description='Run an evaluation task')
    parser.add_argument('config', nargs='?', help='Train config file path')

    # add mutually exclusive args `--slurm` `--dlc`, defaults to local runner
    # if "infer" or "eval" not specified
    launch_method = parser.add_mutually_exclusive_group()
    launch_method.add_argument('--slurm',
                               action='store_true',
                               default=False,
                               help='Whether to force tasks to run with srun. '
                               'If True, `--partition(-p)` must be set. '
                               'Defaults to False')
    launch_method.add_argument('--dlc',
                               action='store_true',
                               default=False,
                               help='Whether to force tasks to run on dlc. If '
                               'True, `--aliyun-cfg` must be set. Defaults'
                               ' to False')
    # multi-modal support
    parser.add_argument('--mm-eval',
                        help='Whether or not enable multimodal evaluation',
                        action='store_true',
                        default=False)
    # Add shortcut parameters (models and datasets)
    parser.add_argument('--models', nargs='+', help='', default=None)
    parser.add_argument('--datasets', nargs='+', help='', default=None)
    # 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('--dry-run',
                        help='Dry run mode, in which the scheduler will not '
                        'actually run the tasks, but only print the commands '
                        'to run',
                        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 '
                        './outputs/default.',
                        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 an infer task. Only '
                        'effective when "infer" is missing from the config.',
                        type=int,
                        default=2000),
    parser.add_argument(
        '--gen-task-coef',
        help='The dataset cost measurement coefficient for generation tasks, '
        'Only effective when "infer" is missing from the config.',
        type=int,
        default=20)
    parser.add_argument('--max-num-workers',
                        help='Max number of workers to run in parallel. '
                        'Will be overrideen by the "max_num_workers" argument '
                        'in the config.',
                        type=int,
                        default=32)
    parser.add_argument('--max-workers-per-gpu',
                        help='Max task to run in parallel on one GPU. '
                        'It will only be used in the local runner.',
                        type=int,
                        default=1)
    parser.add_argument(
        '--retry',
        help='Number of retries if the job failed when using slurm or dlc. '
        'Will be overrideen by the "retry" argument in the config.',
        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)
    # set hf args
    hf_parser = parser.add_argument_group('hf_args')
    parse_hf_args(hf_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 launching 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=None,
                              type=str)
    slurm_parser.add_argument('--qos',
                              help='Slurm quality of service',
                              default=None,
                              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 parse_hf_args(hf_parser):
    """These args are all for the quick construction of HuggingFace models."""
    hf_parser.add_argument('--hf-path', type=str)
    hf_parser.add_argument('--peft-path', type=str)
    hf_parser.add_argument('--tokenizer-path', type=str)
    hf_parser.add_argument('--model-kwargs', nargs='+', action=DictAction)
    hf_parser.add_argument('--tokenizer-kwargs', nargs='+', action=DictAction)
    hf_parser.add_argument('--max-out-len', type=int)
    hf_parser.add_argument('--max-seq-len', type=int)
    hf_parser.add_argument('--no-batch-padding',
                           action='store_true',
                           default=False)
    hf_parser.add_argument('--batch-size', type=int)
    hf_parser.add_argument('--num-gpus', type=int)


def main():
    args = parse_args()
    if args.dry_run:
        args.debug = True
    # initialize logger
    logger = get_logger(log_level='DEBUG' if args.debug else 'INFO')

    cfg = get_config_from_arg(args)
    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':
            if not os.path.exists(cfg.work_dir) or not os.listdir(
                    cfg.work_dir):
                logger.warning('No previous results to reuse!')
            else:
                dirs = os.listdir(cfg.work_dir)
                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, format_python_code=False)

    # 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']:
        # When user have specified --slurm or --dlc, or have not set
        # "infer" in config, we will provide a default configuration
        # for infer
        if (args.dlc or args.slurm) and cfg.get('infer', None):
            logger.warning('You have set "infer" in the config, but '
                           'also specified --slurm or --dlc. '
                           'The "infer" configuration will be overridden by '
                           'your runtime arguments.')
        # Check whether run multimodal evaluation
        if args.mm_eval:
            partitioner = MultimodalNaivePartitioner(
                osp.join(cfg['work_dir'], 'predictions/'))
            tasks = partitioner(cfg)
            exec_mm_infer_runner(tasks, args, cfg)
            return
        elif args.dlc or args.slurm or cfg.get('infer', None) is None:
            # 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)
            if args.dry_run:
                return
            # execute the infer subtasks
            exec_infer_runner(tasks, args, cfg)
        # If they have specified "infer" in config and haven't used --slurm
        # or --dlc, just follow the config
        else:
            if args.partition is not None:
                if RUNNERS.get(cfg.infer.runner.type) == SlurmRunner:
                    cfg.infer.runner.partition = args.partition
                    cfg.infer.runner.quotatype = args.quotatype
            else:
                logger.warning('SlurmRunner is not used, so the partition '
                               'argument is ignored.')
            if args.debug:
                cfg.infer.runner.debug = True
            if args.lark:
                cfg.infer.runner.lark_bot_url = cfg['lark_bot_url']
            cfg.infer.partitioner['out_dir'] = osp.join(
                cfg['work_dir'], 'predictions/')
            partitioner = PARTITIONERS.build(cfg.infer.partitioner)
            tasks = partitioner(cfg)
            if args.dry_run:
                return
            runner = RUNNERS.build(cfg.infer.runner)
            runner(tasks)

    # evaluate
    if args.mode in ['all', 'eval']:
        # When user have specified --slurm or --dlc, or have not set
        # "eval" in config, we will provide a default configuration
        # for eval
        if (args.dlc or args.slurm) and cfg.get('eval', None):
            logger.warning('You have set "eval" in the config, but '
                           'also specified --slurm or --dlc. '
                           'The "eval" configuration will be overridden by '
                           'your runtime arguments.')
        if args.dlc or args.slurm or cfg.get('eval', None) is None:
            # Use NaivePartitioner，not split
            partitioner = NaivePartitioner(
                osp.join(cfg['work_dir'], 'results/'))
            tasks = partitioner(cfg)
            if args.dry_run:
                return
            # execute the eval tasks
            exec_eval_runner(tasks, args, cfg)
        # If they have specified "eval" in config and haven't used --slurm
        # or --dlc, just follow the config
        else:
            if args.partition is not None:
                if RUNNERS.get(cfg.infer.runner.type) == SlurmRunner:
                    cfg.eval.runner.partition = args.partition
                    cfg.eval.runner.quotatype = args.quotatype
                else:
                    logger.warning('SlurmRunner is not used, so the partition '
                                   'argument is ignored.')
            if args.debug:
                cfg.eval.runner.debug = True
            if args.lark:
                cfg.eval.runner.lark_bot_url = cfg['lark_bot_url']
            cfg.eval.partitioner['out_dir'] = osp.join(cfg['work_dir'],
                                                       'results/')
            partitioner = PARTITIONERS.build(cfg.eval.partitioner)
            tasks = partitioner(cfg)
            if args.dry_run:
                return
            runner = RUNNERS.build(cfg.eval.runner)
            runner(tasks)

    # visualize
    if args.mode in ['all', 'eval', 'viz']:
        summarizer = Summarizer(cfg)
        summarizer.summarize(time_str=cfg_time_str)


if __name__ == '__main__':
    main()