kvcached adds vLLM 0.28 support, including Model Runner V2 and packed K/V storage
GitHub·low signal
Merges on 24 and 25 Sep (#493, #494, #500) add an attention-only adapter for vLLM 0.28's Model Runner V2. V2 had bypassed kvcached's hooks by calling a module-level init_kv_cache(). The merges also add packed four-dimensional K/V storage and hybrid Mamba page binding for V1. V2 still rejects hybrid/Mamba and cross-layer sharing, and ROCm stays on split K/V. The repo has 1,504 stars and no tagged release since v0.1.5 in April, so these changes are only on main.