AWS Demonstrates SkyRL Post-Training for Qwen3-VL-8B on SageMaker HyperPod

AWS published a technical walkthrough on September 25, 2026, showing how to run the open-source reinforcement-learning framework SkyRL on Amazon SageMaker HyperPod to post-train the Qwen3-VL-8B vision-language model. The workflow starts from a VisGym supervised fine-tuning checkpoint and uses Group Relative Policy Optimization (GRPO), updating the model based on outcomes from repeated maze runs.

In AWS’s fixed 64-maze evaluation, the starting solve rate was 43.75%. It reached 75% around training step 100 and peaked at 96.875%—62 of 64 mazes—at step 160. These figures describe the tutorial’s experiment, not a performance comparison across different models or workloads.

The example uses Ray on HyperPod, with a CPU head node managing three GPU worker nodes equipped with six GPUs in total. SkyRL runs vLLM inference engines to generate rollouts and uses FSDP for policy training; Amazon FSx for Lustre provides shared storage for LoRA adapter synchronization. The walkthrough also covers creating the Ray cluster from SageMaker Studio, submitting and monitoring jobs, and hosting the trained adapter for inference.

AWS says HyperPod monitors node health and automatically replaces faulty nodes; with checkpointing, a job can resume from its last saved step. The tutorial uses Ray Dashboard and Amazon Managed Grafana for monitoring. In a separate AWS post on MoE reinforcement learning, the company reported 40% higher aggregate rollout throughput with EKS, EFA, and DeepEP across 48 P5en instances. That result comes from a different architecture and benchmark and is not directly comparable with the SkyRL maze experiment.