What's Changed
- [Perf] Allocate the pooled KV cache without committing every page by @pythongiant in #898
- [Bugfix] Keep no frames of a failed sample_tokens step by @Lavmee in #907
- [Config] Validate the vllm-metal environment in one startup pass by @Lavmee in #873
- [Perf] Batch short-context absorbed MLA decode rows by @begcdn in #908
- [Test] Pin the window-mode step-3 sentinel guard with an inf V slot by @begcdn in #903
Full Changelog: v0.30.0.dev20260930084459...v0.30.0.dev20260930094010