What it is
A Python framework for RL post-training of LLMs that connects Megatron for training with SGLang for rollout generation through a shared data buffer. Users can plug in their own generation logic, reward scoring, verifiers, and environments through documented interfaces, which makes multi-turn and agentic RL possible. ML research teams training models such as GLM, Qwen, DeepSeek, and Llama use it to run large-scale RL with checkpointing, fault tolerance, and asynchronous rollout.