fix(vllm): pin the model so the app id cannot lie

[rickstaa]

runners.json advertises vllm/qwen2.5-0.5b-instruct by name, but compose.yml read ${VLLM_MODEL:-Qwen/Qwen2.5-0.5B-Instruct} and .env.example presented it as an ordinary knob. Turn it and the app id becomes false: discovery republishes it unchanged at whatever price the operator set, and a caller filtering for Qwen gets Llama with no way to tell.

The fix pins the model as a literal in compose.yml and drops VLLM_MODEL from .env.example. Serving a different model now means editing the compose command and runners.json together.

Deriving the id from the loaded model (what realtime-transcription does) is not available here: vllm is a static runner whose container is the stock vllm/vllm-openai image with zero Livepeer code, which is what the example exists to show. So the README states the asymmetry instead. A dynamic app advertises what it loaded and cannot drift; a static one stays true only if the operator keeps it so, and the orchestrator health-polls the runner without ever checking it serves what runners.json claims.

Verified: stack up with the literal, vLLM served, runner registered as vllm/qwen2.5-0.5b-instruct.

[Copilot]

Pull request overview

This PR removes configuration drift in the vllm example by pinning the served model in compose.yml so the statically advertised app id in runners.json can’t become inaccurate via an environment override. It also updates vllm/README.md to explain the static-runner “contract” (config must remain truthful) and points readers to realtime-transcription for the dynamic alternative.

Changes:

• Pin the vLLM model in vllm/compose.yml instead of allowing an override via VLLM_MODEL.
• Remove VLLM_MODEL from vllm/.env.example to avoid implying the model is an operator “knob”.
• Update vllm/README.md to document why the model is pinned for static registration and how to change it safely.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
vllm/README.md	Documents static vs dynamic runner registration and clarifies that the vLLM model is intentionally pinned to keep the advertised app id truthful.
vllm/compose.yml	Pins --model to Qwen/Qwen2.5-0.5B-Instruct and updates comments explaining why this must not be an env override.
vllm/.env.example	Removes VLLM_MODEL to prevent accidental app-id/model drift through environment configuration.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.

Suppressed comments (1)

vllm/compose.yml:33

• The comment says runners.json “advertises this model by name”, but vllm/runners.json only contains the app id (vllm/qwen2.5-0.5b-instruct), not the Hugging Face model string. Rewording avoids suggesting the exact model name lives in runners.json.

    # Pinned, not a knob: runners.json advertises this model by name, so change both
    # together. --gpu-memory-utilization caps VRAM (0.9 default can exceed free).