Another cool feature! Rebased it on my testing-branch+master to test it out!

Minimal test as a global:

; llama-server --port 8082 --models-max 1 --models-preset backend.ini --webui-config-file frontend.json[*]fit = off  ; Disable automatic memory fittingngl = 999  ; Full GPU offloadctk = q8_0 ; KV cache key quantizationctv = q8_0 ; KV cache value quantizationfa = on    ; Enable flash attentionmlock = on ; Lock model in RAMnp = 4     ; Parallel request batchingkvu = on   ; Unified KV cache buffersleep_idle_seconds = 60 ; Testing[Dense-Devstral-Small-2-24B-Instruct-2512]m = unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF/Devstral-Small-2-24B-Instruct-2512-UD-Q6_K_XL.ggufc = 131072load-on-startup = 1[my-other-models]...

Seems to have no effect: the model stays loaded continuously. Perhaps 'load-on-startup = 1' combined with 'sleep_idle_seconds' doesn't work yet? Same for any models I load manualy.

This feature will be useful for my real use case: unloading large MoE models that spill over from VRAM into system RAM

[MoE-Uncensored-GLM-4.5-Air-Derestricted-106B]m = bartowski/ArliAI_GLM-4.5-Air-Derestricted-GGUF/ArliAI_GLM-4.5-Air-Derestricted-Q4_K_M-00001-of-00002.ggufn-cpu-moe = 30c = 32768sleep_idle_seconds = 60

I try this also.

Seems to have no effect: the model stays loaded continuously. Perhaps 'load-on-startup = 1' combined with 'sleep_idle_seconds' doesn't work yet? Same for any models I load manualy.

this feature does not unload the model instance, it is independent from router mode

instead, monitor your log and you will see log lines like this:

que    start_loop: entering sleeping statesrv  handle_sleep: server is entering sleeping state

we don't unload the whole process because#18189 (comment)

Seems to have no effect: the model stays loaded continuously. Perhaps 'load-on-startup = 1' combined with 'sleep_idle_seconds' doesn't work yet? Same for any models I load manualy.

this feature does not unload the model instance, it is independent from router mode

instead, monitor your log and you will see log lines like this:

que    start_loop: entering sleeping statesrv  handle_sleep: server is entering sleeping state

we don't unload the whole process because#18189 (comment)

Got it! I was expecting the child process to be killed, but it's an internal model unload within the process itself.
I'll monitor RSS and look for the "entering/exiting sleeping state" log lines.
Testing now with standalone mode first to validate the feature, then I'll integrate it with router mode. Thanks for the clarification!

The internal sleep approach (keeping process alive) is much cleaner than kill/respawn.
Looking forward to the future "sleep levels" feature to fine-tune which components get unloaded!