Feature Request: Auto-sleep / idle unload for flm serve #562

Please, add an option to flm serve that automatically unloads the model from memory after a configurable period of inactivity, and transparently reloads it on the next incoming request.

Same idea and semantics as the recently merged llama.cpp feature: ggml-org/llama.cpp#18228 (--sleep-idle-seconds).