A host-centric framework for scaling KV cache across host memory, filesystems, object stores, and remote peers — reducing recomputation and increasing serving capacity.