SkyRL tx: A unified training and inference engine | Ray Summit 2025 скачать в хорошем качестве

SkyRL tx: A unified training and inference engine | Ray Summit 2025

At Ray Summit 2025, Philipp Moritz from Anyscale and Tyler Griggs from UC Berkeley share how SkyRL tx brings a groundbreaking unification of transformer training and inference through a single, REST-based interface inspired by Thinking Machines’ Tinker API. They begin by introducing SkyRL tx as an open-source system that treats post-training not as a separate pipeline, but as an extension of the inference engine itself—supporting both forward and backward passes through the same unified interface. This design eliminates the complexity of maintaining separate training and inference stacks and dramatically simplifies how models are developed, fine-tuned, and deployed. The speakers then explain how SkyRL tx uses LoRA-based adapters to enable cost-efficient multi-tenancy, allowing many users to share a common base model while customizing behavior through lightweight, isolated adapters. This approach makes it easier for organizations and research teams to support multiple users, tasks, and experiments without duplicating full models. Philipp and Tyler walk through: The architecture and core implementation of SkyRL tx Key design decisions behind unifying training and inference How SkyRL tx achieves compatibility with the Tinker API The project’s roadmap, including upcoming capabilities and extensibility points Opportunities for the community to contribute and adopt the stack SkyRL tx is designed for researchers and developers who want to experiment with unified model execution patterns, as well as organizations interested in running their own Tinker-compatible backend on self-managed hardware. Attendees will learn how SkyRL tx simplifies model operations, reduces infrastructure overhead, and opens new avenues for scalable, flexible transformer development. Liked this video? Check out other Ray Summit breakout session recordings    • Ray Summit 2025 - Breakout Sessions   Subscribe to our YouTube channel to stay up-to-date on the future of AI!    / anyscale   🔗 Connect with us: LinkedIn:   / joinanyscale   X: https://x.com/anyscalecompute Website: https://www.anyscale.com/