Legacy (In-Process Mode)

Warning

These pages document LMCache’s original in-process mode, where LMCache ran inside the inference engine process (e.g. via LMCacheConnectorV1 on vLLM). In-process mode is deprecated; new deployments should use Multiprocess (MP) mode.

Background

LMCache began as an in-process library embedded directly in the serving engine. The multiprocess refactor moved LMCache into a standalone lmcache server and made an asynchronous prefetching architecture the default – a LOOKUP followed by background L2→L1 loads – alongside process isolation, shared caching across engine instances, and multi-tier (L1/L2) storage.

MP is now the recommended mode and is on track to support essentially everything in-process mode did. The pages below are kept for users still on in-process mode and as a historical reference while MP closes any remaining gaps; where a feature already has an MP equivalent, prefer the MP docs.