Hi everyone,
๐ LWS v1.0.0 is released!ย ๐
Release notes ยท
Installation and upgradesDistributed inference APIs for Kubernetes, with LeaderWorkerSet and DisaggregatedSet.
We're happy to share this milestone with everyone who has helped LWS grow.
๐ Here's what's new:
โข ๐ Fractional rollout window: DisaggregatedSet roles advance proportionally, making better use of rollout budgets.
โข ๐ Scaling during rollouts: DisaggregatedSet scale-up continues during updates, with opt-in scale-down.
โข ๐งฉ Independently scalable sub-roles: Separate scaling and routing pools within a DisaggregatedSet role.
โข ๐ค Native gang scheduling reliability: Fixes for group recreation and resizing; still alpha and opt-in.
โข ๐ก๏ธ Bounded group recovery: Restart limits and backoff bound repeated group recreations.
โข โก Parallel Pod termination: Opt in to terminating leaders and workers together; sequential remains the default.
โข โ
Stable configuration API: Controller configuration graduates to v1; v1alpha1 remains supported but deprecated.
๐ Upgrading? Read the Urgent Upgrade Notes for default changes and migration requirements, and follow the upgrade guide.
๐ A heartfelt thank you to our contributors and users for helping LWS reach 1.0!
--
Best regards,
Kay Yan
The LWS Team