Deployment¶

Bare metal / VM¶

pip install "laya-serve[inference]"
LAYA_SERVE_BACKEND=laya LAYA_SERVE_PRELOAD=true laya-serve
# -> http://0.0.0.0:8000

PRELOAD=true is recommended for servers: it loads all checkpoints at startup instead of on first request. Tune residency with LAYA_SERVE_MAX_LOADED=1 and placement with LAYA_SERVE_DEVICE=cuda (or cpu).

Host/port use uvicorn env/config or defaults (0.0.0.0:8000). Put TLS termination and rate limiting in front (reverse proxy / gateway); 429 handling in-app is future work.

Docker¶

docker build -t laya-serve .
docker run -p 8000:8000 \
  -e LAYA_SERVE_API_KEY="$LAYA_SERVE_API_KEY" \
  laya-serve

The shipped Dockerfile sets LAYA_SERVE_BACKEND=laya and LAYA_SERVE_PRELOAD=true and installs the inference extra.

Health checks¶

• Liveness: GET /healthz → {"status": "ok"}. Use it for Kubernetes liveness/readiness probes and load-balancer checks.
• GET /v1/models verifies the serving model and alias set after a deploy.

Configuration checklist¶

• [ ] LAYA_SERVE_SERVING_MODEL matches the deployed checkpoint id.
• [ ] LAYA_SERVE_EXTRA_MODELS covers any legacy client model strings.
• [ ] LAYA_SERVE_API_KEY set; clients send Authorization: Bearer <key>.
• [ ] Logs: 500s log server-side with tracebacks; 529s log a warning and return Retry-After: 1 so clients back off.