What's Changed
- [Cleanup] Move the MLA block-size set next to KERNEL_BLOCK_SIZES by @begcdn in #926
- Build the GGUF skeleton lazily by @ricky-chaoju in #929
- [Spec Decode] Add experimental DFlash serving on scheduler-owned KV by @StevenWang-CY in #928
- [Perf] Sort batched decode rows by context and key the index cache by @begcdn in #927
Full Changelog: v0.30.0.dev20260930202258...v0.30.0.dev20261001081948