Use train_grpo.py as detailed on https://huggingface.co/docs/trl/main/en/grpo_trainer to launch training. Rest of the repo is a copy of the trl library. Modified files are trainer/grpo_trainer.py, extras/multi_vllm_client.py, and trainer/grpo_config.py