DeepSeek-Math/evaluation/submit_eval_jobs.py

67 lines
2.1 KiB
Python
Raw Normal View History

2024-02-06 02:27:40 +00:00
import os
import argparse
configs = [
{
'output-dir': "outputs/DeepSeekMath-Base",
'model-path': "deepseek-ai/deepseek-math-7b-base",
'tokenizer-path': "deepseek-ai/deepseek-math-7b-base",
'model-size': "7b",
'overwrite': False,
'use-vllm': True,
'no-markup-question': True,
'test-conf': "configs/few_shot_test_configs.json",
'prompt_format': 'few_shot',
'expname': 'eval-deepseek-math-7b-base'
},
{
'output-dir': "outputs/DeepSeekMath-Instruct",
'model-path': "deepseek-ai/deepseek-math-7b-instruct",
'tokenizer-path': "deepseek-ai/deepseek-math-7b-instruct",
'model-size': "7b",
'overwrite': False,
'use-vllm': True,
'test-conf': "configs/zero_shot_test_configs.json",
'expname': 'eval-deepseek-math-7b-instruct'
},
{
'output-dir': "outputs/DeepSeekMath-RL",
'model-path': "deepseek-ai/deepseek-math-7b-rl",
'tokenizer-path': "deepseek-ai/deepseek-math-7b-rl",
'model-size': "7b",
'overwrite': False,
'use-vllm': True,
'test-conf': "configs/zero_shot_test_configs.json",
'expname': 'eval-deepseek-math-7b-rl'
}
]
base_conf, instruct_conf, rl_conf = configs
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--n-repeats", type=int ,default=1)
parser.add_argument("--temperature", type=float, default=0)
parser.add_argument("--n-gpus", type=int, default=8)
args = parser.parse_args()
conf = base_conf # TODO: your conf here
for rank in range(args.n_gpus):
cmd = "python run_subset_parallel.py"
for key, val in conf.items():
if key == 'expname':
continue
if isinstance(val, str):
cmd += f" --{key} {val}"
elif val:
cmd += f" --{key}"
cmd += f" --test-conf {conf['test-conf']}"
cmd += f" --n-repeats {args.n_repeats}"
cmd += f" --temperature {args.temperature}"
cmd += f" --rank {rank} &"
print(cmd, flush=True)
os.system(cmd)
if __name__ == '__main__':
main()