SkyRL on Amazon SageMaker HyperPod: Multimodal RL Training for Vision-Language Models

SkyRL is an open-source reinforcement learning framework used to post-train language and multimodal models. It can run on cloud infrastructure such as Amazon SageMaker HyperPod.

GRPO (Group Relative Policy Optimization) compares a group of outputs for the same input instead of relying on a separate value model, which simplifies training and lowers memory requirements. The article presents it within the Qwen3-VL-8B training example.

The framework is open source and SageMaker HyperPod is a managed cloud service, so it is accessible. Adoption still requires RL expertise, Ray cluster management, and GPU budget, so starting with a small model and limited data is advisable.

SageMaker HyperPod provides managed GPU cluster infrastructure, allowing distributed SkyRL jobs to run via Ray without manual server management.