Note

Medium Risk
Changes the timing of heartbeat thread startup in the vLLM multiprocess adapter, which can affect health detection and degraded-mode behavior during startup and early requests. Concurrency/thread-lifecycle adjustments may introduce subtle race or ordering issues.

Overview
Defers starting the LMCache HeartbeatThread in LMCacheMPWorkerAdapter until the first submit_store_request or submit_retrieve_request, instead of starting it immediately after register_kv_caches completes.

Replaces the eager _start_heartbeat call with an idempotent _ensure_heartbeat_started helper and updates comments to clarify the new “start on first use” behavior once vLLM initialization (model load, KV allocation, warmup/CUDA graph capture) is complete.

^{Written by Cursor Bugbot for commit 2775a48. This will update automatically on new commits. Configure here.}