v0.4.0: park race (409 on /v1/unload) latches the LLM, ComfyUI emptied on short LLM calls, stop on LATCHED

Aperta
  • Qualcosa non va?
#275 1 commento 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
4/5
Tempo stimato
3-5 giorni
Idoneità per principianti
22/100
Tipo di issue
Bug
Chiarezza
Abbastanza chiara
Stato di attività
Attiva
Stack tecnologico
go
Ambito
backend

Direzione di ricerca

Start from the park path that calls POST /v1/unload and gets a 409 when a concurrent ComfyUI lease exists during the LLM resume, and from the min_turn lease-turn logic that is meant to prevent it. Reproduce by sending a ComfyUI POST /prompt while the LLM is still resuming, then check whether the engine ends up LATCHED with the request unanswered. Done means the second lease waits or the 409 is retried, and the pending LLM request gets an explicit response. Points 2 and 3 are design questions and should be settled with the maintainer first.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

Version: v0.4.0, Windows (berth daemon + berth gateway), one card, two engines: a llama.cpp-style LLM server and ComfyUI. min_turn = 45 s, keep_comfyui_up = true, floor_gib = 15. An upstream scheduler sends ComfyUI POST /prompt and OpenAI-style LLM requests through the gateway (LLM on one listen port, ComfyUI on another).

Seen in real use over one day. These are three separate points; I can split them into separate issues if you prefer.

1. Race: a concurrent lease during the LLM resume makes the park fail, and the LLM ends up latched with its request left hanging
  • A request for the LLM arrives, so ComfyUI is parked and the LLM is resumed.
  • While the LLM is still resuming, a new ComfyUI POST /prompt arrives and creates a concurrent lease.
  • The park call fails with POST /v1/unload → 409 Conflict. Berth then stops the LLM engine, which stays LATCHED. The pending LLM request never gets a response; the client times out.
  • min_turn = 45 s narrows the window but does not close it.
  • Expected: the second lease waits for the turn to settle (or the 409 is retried), and the pending request is answered or failed explicitly.
2. A short LLM call empties ComfyUI even with keep_comfyui_up

Each switch to the LLM unloads ComfyUI's models. The next render then reloads them from cold: about 20–30 s before ComfyUI is ready, plus about 100 s of model loading, so roughly 2 min of near-idle GPU per short LLM call.
Request: an option to keep ComfyUI's memory as is for a short LLM lease, when VRAM and RAM allow it (the LLM uses about 10.5 GB of VRAM, ComfyUI about 8 GB idle, on a 12 GB card: it does not always fit), or at least to park ComfyUI without making it drop its model cache.

3. berth stop on a LATCHED engine

What should berth stop (or POST /v1/berth/stop) do on an engine that is LATCHED after point 1? Clear the latch and stop it, or refuse? The docs do not say, and we would like to use it as a recovery step.

Logs (berthd and gateway) are available if useful; I would just strip local paths from them first.

Lingua principale
Go
Stelle
226
Fork
20
Merge medio
1h 25m
PR unite (30g)
238

Preparare l'ambiente

Questo progetto non fornisce container di sviluppo, Dockerfile né guida per i contributori, quindi l'ambiente è a tuo carico: parti dal suo README e consulta la nostra guida al primo contributo per i passaggi generali.

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di T-Crypt/berth

Tutte le issue di T-Crypt/berth

Issue simili

Altre issue su Go

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.