Deploy adapters — step 3 of 3

What deploy-notes.md has to say

Write the notes so an on-call engineer can load the adapter without rereading this chapter. A sufficient page, even if the deploy is simulated:

base: Qwen/Qwen2.5-1.5B-Instruct
adapter: specialist-adapter-receipt/adapter/   # or Colab export path
serve: vLLM with LoRA enabled
select: request field lora_name=ci-triage-labeler
rollback: unload that lora_name; base keeps serving
cost: one 1.5B (4-bit) resident; adapter is megabytes
out of scope: claiming the base improved on MMLU / "general chat"
escape hatches: Modal or RunPod if you need a GPU and Colab quota is gone

vLLM's LoRA feature is the named default because it matches "one base, many adapters." If your team uses another server that does the same thing, say so and keep the fields. If your team wants to merge adapters into a single checkpoint for an air-gapped box, say that too — and keep the unmerged adapter so you can still roll back.

Paid infra is optional. The notes are not. A receipt folder with a blank deploy-notes.md is an unfinished artifact even when the eval table says ship.