AWS details workflow for running open-source SkyRL on SageMaker HyperPod to post-train Qwen3-VL-8B

A new AWS Machine Learning Blog post describes a workflow for running SkyRL, an open-source reinforcement learning framework, on Amazon SageMaker HyperPod to post-train the Qwen3-VL-8B vision-language model using GRPO. The post-training process involves generating trajectories, receiving rewards, and updating the policy based on outcomes, which is standard for building capable language model agents.

The workflow outlined in the post includes building a container image for SkyRL, launching a Ray cluster directly from SageMaker Studio, submitting the training job remotely, and monitoring progress through SageMaker’s interfaces. After training completes, the resulting LoRA adapter can be hosted for inference.

Amazon SageMaker HyperPod provides infrastructure designed for large-scale ML workloads running on Amazon Elastic Kubernetes Service (Amazon EKS). Its cluster resiliency features continuously monitor node health and automatically replace faulty nodes, preventing a single hardware failure from terminating a long-running training job. This is particularly valuable for multi-node reinforcement learning runs that can require hundreds of GPU-hours of rollouts per training run.

The post highlights that checkpointing in SageMaker HyperPod allows a training job to resume from the last saved step rather than restarting from scratch after a failure, which reduces wasted compute and accelerates iteration cycles for RL workloads.