ROLL is an open-source library that enables efficient large-scale reinforcement learning training for LLMs on GPU clusters. It implements multi-role distributed architecture powered by Ray for flexible resource scheduling, and integrates modern inference engines like vLLM and SGLang alongside Megatron-Core for model training. The library focuses on improving LLM capabilities in human preference alignment (RLHF), multi-turn agentic interactions, and complex reasoning tasks.