The base stays shared
Deploying a specialist is loading an adapter onto a base that already serves other traffic. You do not ship a second copy of Qwen2.5-1.5B-Instruct for every job. vLLM's LoRA serving (and peers on Modal or RunPod) keeps one base in GPU memory and swaps adapters per request.
That is the cost story. A 1.5B 4-bit base plus a few-megabyte adapter is a different bill than a fine-tuned 70B you cannot evict. Latency is the base decode plus a small adapter tax — not a new model cold start per specialist.
What you must not claim in deploy-notes.md: that the base "got smarter." The base did not change. The request that names ci-triage-labeler loads that adapter. A request that names nothing gets the stock instruct model, which still drifts on the three labels.
If you cannot run vLLM this week, write the notes as if you could: image, base id, adapter path, the request field that selects the LoRA, and the rollback (unload adapter, keep serving the base). Simulated deploy is allowed. Inventing a generally smarter checkpoint is not.