LightRFT
A reinforcement fine-tuning framework for large language and vision-language models, covering reinforcement learning with verifiable rewards, RLHF, and reward-model-driven policy optimization on top of PyTorch distributed training and rollout backends.
LightRFT is a reinforcement fine-tuning framework for large language models and vision-language or multimodal generative models. It provides a structured workflow for reinforcement learning with verifiable rewards, reinforcement learning from human feedback, and model-reward-driven policy optimization, covering policy sampling, reward computation, advantage estimation, and policy updates. It also includes reward-model training and on-policy distillation workflows, with example code paths for text, image, video, and audio tasks.
It runs on PyTorch distributed communication, with a unified strategy interface connecting FSDP v2 and DeepSpeed ZeRO training backends to SGLang and vLLM rollout backends. Training and rollout share the same GPU process set by phase; the rollout engine can sleep during policy updates and reload refreshed weights before the next rollout.
It is written in Python and released under the Apache-2.0 license.