LMCache stores KV cache beyond GPU memory in a tiered hierarchy and reloads previously computed KV cache across requests, sessions, and engine instances to reduce repeated prefill work.