Closing the Train-Inference Divide for Asynchronous RL at Scale | Lila Sciences | Ray Summit 2026 @anyscale
Closing the Train-Inference Divide for Asynchronous RL at Scale | Lila Sciences | Ray Summit 2026  @anyscale
Uploaded September 2026 | Updated November 2026, 1 day ago
In asynchronous RL with disaggregated training and inference, even slight differences in model behavior between the training and inference stack can cause significant training instabilities that are painful to debug.

At Ray Summit 2026, Dominic Yurk, AI Researcher at Lila Sciences, walks through two contributions upstreamed to SkyRL that close that gap: faithful replay of expert-router decisions and of per-token sample support, both captured at rollout and reconstructed exactly on the training side. He digs into the engineering: a consistent abstraction for attaching token-level metadata, efficient encoding and transport of large-scale router replay data, and a sparse representation that makes sample-support computation fast at scale.

You'll leave with empirical results showing the effect on training stability and throughput.

Liked this video? Check out other Ray Summit breakout session recordings: youtube.com/playlist?list=PLZNaJyYZRPdo

Subscribe to our YouTube channel to stay up-to-date on the future of AI: youtube.com/c/anyscale

đź”— Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com
Closing the Train-Inference Divide for Asynchronous RL at Scale | Lila Sciences | Ray Summit 2026SkyRL tx: A unified training and inference engine | Ray Summit 2025Scaling Production LLM Inference Using EKS Auto Mode & Ray Serve | Ray Summit 2025From Self-Managed Ray to Anyscale Services: Migrating ML Serving | Coinbase | Ray Summit 2026Where’s My Training Job? Ray & GPU Observability at Datadog | Datadog | Ray Summit 2026Serving a Frontier Open Model with Ray & vLLM | Nscale | Ray Summit 2026Multi-Region, Environment Isolated Ray Environments in the Notion Way | Notion | Ray Summit 2026Pinterest’s Approach to Real-Time ML Experimentation Using Ray | Ray Summit 2025From OOMs to 2x Throughput: Scaling Multimodal Data Curation on Ray | Zoox | Ray Summit 2026Custom Pipeline Parallelism for Frontier-Scale Model Serving | Apple | Ray Summit 2026Scaling LinkedIns Online Training Solution with Ray | Ray Summit 2025AI workloads introduce a system design shift ...
Anyscale |

Closing the Train-Inference Divide for Asynchronous RL at Scale | Lila Sciences | Ray Summit 2026

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER