AI model approvals (GPU approval queue)

What it's for

A guardrail so GPU-backed models (gpu-t4, gpu-a100) don't silently burn expensive quota — Owner/Admin review and approve or reject each GPU deploy request before it goes live.

How to review pending approvals

1. Click your avatar (top-right) → Approvals. This is a global, cross-project page — it's not in any per-project nav, only in the account menu, and only appears there if you're Owner/Admin on at least one project.
2. Each row shows: Project, Resource (kind + name, e.g. AIModel / iris-classifier), Action (e.g. CreateAIModel), Requested by, Age, and a short Summary (model type, profile, and source for model-deploy requests).
3. Click Approve to let it proceed — takes effect immediately, no confirmation step.
4. Click Reject to open a modal — you must type a Reason (the submit button stays disabled until you do) before it will let you reject.
5. Use the search box to filter by project/resource/action/requester, and Refresh to pull the latest queue.

Gotchas

• This page only exists for people who are Owner/Admin on at least one project — everyone else who somehow reaches /admin/approvals gets an access-denied message, not a 404.
• A rejection reason is mandatory — there's no "reject with no explanation" path, by design.
• Approve has no confirmation dialog — double-check the row (project + resource + summary) before clicking; there's no "undo" button once you approve (you'd have to have the requester delete/redeploy the model instead).
• Today this queue is effectively GPU-model-request-only in practice: the Summary column only knows how to describe CreateAIModel payloads — other resource kinds would show a blank summary if the underlying approval system is later extended to gate anything else.

Not yet supported

• Notifications when a request needs approval or gets a decision — you have to visit this page to know something's waiting.
• Bulk approve/reject.
• Any visibility into why a request required approval beyond "GPU profile + project GPU quota was 0 at request time" (see AI Models).