What's Changed
- [TurboQuant] Keep the K block scale a normal float16 for constant blocks by @abhijithneilabraham in #891
- [Perf] Cap automatic TurboQuant prefill workspace reservations by @suntp in #895
- [Spec Decode] Qualify the native Qwen3 DFlash block forward by @StevenWang-CY in #900
- [Kernel] Drop the redundant neutral-state guards left by #884 by @begcdn in #896
Full Changelog: v0.30.0.dev20260929203507...v0.30.0.dev20260929204817