| from dataclasses import dataclass |
| import gc |
| import os |
| from pathlib import Path |
| import subprocess |
|
|
| from omegaconf import OmegaConf |
| from rich import print |
| import torch |
|
|
| from refined_v4.train import Config as TrainConfig, run |
| from refined_v4.study import model_inventory |
| from .data import ROOT, TASKS |
|
|
|
|
| @dataclass |
| class Config: |
| root: str = str(ROOT / 'smoke') |
| output: str = str(ROOT / 'smoke-flow') |
| task: str = '' |
| device: str = 'auto' |
|
|
|
|
| def main(config: Config): |
| base = next(row for row in model_inventory() if row['name'] == '8m') |
| for task in TASKS: |
| if config.task and config.task != task: |
| continue |
| settings = TrainConfig(task=task, model='8m', root=config.root, |
| checkpoint=base['checkpoint'], output=str(Path(config.output) / task), |
| sft_tokens=4096, rl_rollouts=128, samples=32, groups=4, |
| eval_queries=16, eval_samples=2, interval=128, recovery_interval=128, |
| workers=2, threads=2, wandb_mode='disabled', smoke=True, device=config.device) |
| if not run(settings): |
| if os.environ.get('SLURM_JOB_ID'): |
| subprocess.run(['scontrol', 'requeue', os.environ['SLURM_JOB_ID']], check=True) |
| return |
| raise RuntimeError('Smoke training interrupted') |
| print(f'[green]{task}: real SFT + RL smoke complete[/green]') |
| gc.collect() |
| if torch.cuda.is_available(): |
| torch.cuda.empty_cache() |
|
|
|
|
| if __name__ == '__main__': |
| main(OmegaConf.to_object(OmegaConf.merge(OmegaConf.structured(Config), OmegaConf.from_cli()))) |
|
|
