GRPO for Automated Program Repair
CUDA_VISIBLE_DEVICES=0 trl vllm-serve --model /root/autodl-tmp/apr-rl/Qwen2.5-Coder-3B-Instruct --enforce-eager true
ACCELERATE_LOG_LEVEL=info CUDA_VISIBLE_DEVICES=1 accelerate launch grpo.py --config config_full.yamlACCELERATE_LOG_LEVEL=info CUDA_VISIBLE_DEVICES=0,1 accelerate launch sft.py --config sft_config.yaml@article{hu2025repair,
title={Repair-r1: Better test before repair},
author={Hu, Haichuan and Xie, Xiaochen and Zhang, Quanjun},
journal={arXiv preprint arXiv:2507.22853},
year={2025}
}