AWS released SkyRL on Amazon SageMaker HyperPod, enabling reinforcement learning for vision-language models. It is the company's first major update to its machine learning infrastructure since the launch of SageMaker HyperPod in 2023.
AWS reported a 95% maze solve rate for Qwen3-VL-8B models trained with SkyRL, up from 43.75% using a supervised fine-tuning checkpoint. That compares with the earlier 43.75% baseline from the VisGym SFT checkpoint.
SkyRL is built on Amazon SageMaker HyperPod and targets multi-modal reinforcement learning tasks. Availability begins with the release of the SageMaker HyperPod cluster, initially for developers and researchers.
"Starting from the VisGym SFT checkpoint, GRPO post-training on HyperPod improves the maze solve rate from 43.75% to more than 95% on a fixed 64-maze evaluation set," said the AWS blog. The framework uses Group Relative Policy Optimization to grade runs against one another.
The announcement follows AWS's focus on scalable machine learning infrastructure. AWS said the update supports long multi-node RL runs by enabling checkpointing and fault tolerance.
AWS did not specify the exact number of GPU instances needed for training, and the source raises the question of how the framework will scale to more complex environments. The training job uses Amazon FSx for Lustre shared storage for checkpoint synchronization.
Source: awsml
