Accelerate multimodal RL training with SkyRL on Amazon SageMaker HyperPod

Amazon Web Services · 2 weeks ago · 43 ● 2 sources

Amazon SageMaker HyperPod runs SkyRL on a Ray cluster to post-train a Qwen3-VL-8B vision-language model with GRPO for maze navigation while using checkpointing and automatic node replacement to survive hardware failures. The maze solve rate increases from 43.75% to more than 95% on a fixed 64-maze evaluation set. The setup changes to colocate vLLM rollout generation and FSDP policy training on the same GPUs with LoRA adapter syncing through /shared storage on HyperPod.