Published SGLang image refuses SGLang embedders, so an embedder and a guard can't share one SIE server

Aperta
#969 0 commenti 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
42/100
Tipo di issue
Bug
Chiarezza
Abbastanza chiara
Stato di attività
Attiva
Stack tecnologico
docker, python

Direzione di ricerca

Start with the baked bundle compatibility check and compare packages/sie_server/bundles/sglang.yaml with packages/sie_server/bundles/sglang-embedding.yaml; the issue says their dependencies match and the split is for worker pool isolation. Check the image build configuration and docs/compatibility.md to see how bundles are selected and which images are published. Done means the reported embedder and guard can both run from a published image, or the compatibility docs clearly explain the supported setup.

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

Descrizione

What we were testing

We wanted to check that a reader can reproduce a simple setup with a published SIE image: an SGLang embedding model and an SGLang guard model sharing one 24 GB L4, served by one SIE server. The pair is Qwen/Qwen3-Embedding-4B and Qwen/Qwen3Guard-Gen-0.6B. It's the case SIE's SGLang memory sizing (#763) was built for: plain SGLang at its defaults runs out of memory on the second server, and SIE sizes each one from the model files so both fit.

The test is the cookbook recipe why_sie/share_one_gpu_between_two_sglang_models (superlinked/cookbook#39). It runs three setups on one GPU, all from ghcr.io/superlinked/sie-server:v0.11.1-cuda12-sglang:

  1. Plain SGLang at its defaults (the guard runs out of memory, as expected).
  2. Plain SGLang with --mem-fraction-static set by hand (both start).
  3. One SIE server with both models and KV budgets of 8,192 and 16,384 tokens.

Setups 1 and 2 behaved as expected. Setup 3 didn't start.

What happened

Inside the published image, on one NVIDIA L4:

python -m sie_server.cli serve --host 127.0.0.1 --port 8080 \
  --models-dir <dir with the two v0.11.1 model configs, kv_budget_tokens set> \
  --models Qwen/Qwen3-Embedding-4B
Error: Model(s) Qwen/Qwen3-Embedding-4B are incompatible with baked bundle 'sglang'; use an image built for bundle 'sglang-embedding'

The guard is in the sglang bundle and the embedder in sglang-embedding. The image is baked for sglang, so it refuses the embedder, and there's no published image for sglang-embedding (docs/compatibility.md lists all five SGLang embedders as "no published image serves the sglang-embedding bundle", on v0.11.1 and on main). An image built with --build-arg BUNDLE=sglang-embedding would refuse the guard instead.

So with the released images, one SIE server can't serve an SGLang embedder next to an SGLang generation model, and none of the SGLang embedders can be served at all.

The same two models in one server, from the same source in an environment that isn't baked for a bundle, start and serve fine: SIE reserved 12.06 GiB for the embedder and 6.27 GiB for the guard and both answered with no errors.

Why it looks avoidable

packages/sie_server/bundles/sglang-embedding.yaml says its dependencies are kept in lockstep with sglang.yaml and that "the bundle split exists only for worker pool isolation". The packages are the same; only the baked-bundle check keeps the two apart.

What would fix it

Any of these would let the setup above run from a published image:

  • Let an image baked for sglang also serve sglang-embedding models (same dependencies), or mark the two bundles compatible.
  • Publish a cuda12-sglang-embedding image, and document how to run an embedder and a guard together (two servers, or one image that serves both).

Until then, docs/compatibility.md could say plainly that SGLang embedders need a custom image build.

Lingua principale
Python
Stelle
3.5k
Fork
328
Merge medio
16h 25m
PR unite (30g)
121

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 superlinked/sie

Tutte le issue di superlinked/sie

Issue simili

Altre issue su Python

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.