What's Changed
- [Bugfix] Synchronize before reading the MLX cache in profile_run by @abhijithneilabraham in #836
- [Spec Decode] Extract shared target feature capture by @StevenWang-CY in #834
- [Perf] Materialize MLA prefill K/V for chunks with cached context by @begcdn in #833
- [Attention] Bidirectional attention inside Gemma 4 image blocks by @Lavmee in #830
- [Multimodal] Keep only the most recent backbone mode decisions by @Lavmee in #829
Full Changelog: v0.30.0.dev20260926140952...v0.30.0.dev20260926202403