Relax

Why we included this project

Relax targets a specific pain point: reinforcement-learning post-training on large multimodal models, where training and rollout inference tend to compete for the same GPUs. The framework splits those jobs apart. Built on Ray Serve with a service-oriented design, it runs each stage as its own deployment, so Megatron-LM handles training, SGLang handles inference, and either can scale or recover without taking the whole pipeline down. A data-transfer layer called TransferQueue streams data between them, which lets rollout, actor, reference, and advantage computation each run on separate GPU clusters instead of blocking one another. That decoupling is what makes it one of the few open frameworks that can do end-to-end RL post-training across text, vision, and audio, including omni models like Qwen3-Omni, with PPO, GRPO, and other algorithms included. It is a real distributed system.