GPU-box hosting

Artifacts that run on a GPU instance (Vast.ai now, RunPod later). They are not part of the asfquart process.

Design (why): docs/vllm-fleet-design.md. servers.yaml schema is defined there (§5); this tree does not duplicate it.

Control plane: asfquart serves GET /vllm/config/{set_id} with the fleet key. That endpoint is application code (api.py / llmao/), not here.

Layout

Path	Role
`fetch_config.py`	`FLEET_KEY` + `VLLM_SET` + `ASFQUART_URL` → write `servers.yaml`
`launcher.py`	parse YAML, spawn/monitor `vllm serve`, SIGTERM
`onstart.sh`	provider on-start: fetch then exec launcher
`vast/README.md`	Vast.ai runbook (stock vLLM template + env)

Same scripts for every provider. Only the runbook (how to set env / on-start) differs.

Box environment

Required:

FLEET_KEY — shared secret; Authorization: Bearer
VLLM_SET — set id (semantic or box-a; control plane owns names)
ASFQUART_URL — origin only, e.g. https://llmao.apache.org (no trailing path). Required. Not hard-coded; dev and prod hosts differ.

Optional:

SERVERS_YAML — default /workspace/servers.yaml
LAUNCHER_MAX_RESTARTS — default 5 (per server); then stay up for SSH

Boot

onstart.sh → fetch_config.py → launcher.py

v1 uses the stock Vast vLLM image plus this on-start. A derived image is not justified until the layer is more than these three files (see Vast runbook).

Pin the git tag or commit in the provider on-start snippet. Do not curl floating main.