RFD 1140: Rented GPUs on RunPod
Problem
The desk card is one RTX 3090 with 24 GiB. An OmniGen2 LoRA trains here at 256 square in 22 minutes for 200 steps, and does not finish one step in twelve minutes at 512 square. Neither run reports an error. Windows pages into shared memory instead of failing, so the only symptom is a run that stops being fast. Pixal3D dies one stage short of geometry, because NATTEN has no sm_86 kernel and every stage after the first reaches it.
Renting fixes both. It also loses whatever is not in a git repository when the pod goes away, and it bills while it waits.
Decision
Rented GPU work runs on RunPod. Push before renting, tear down after use, and check the tear down.
The API is REST at rest.runpod.io/v1 with a bearer token. GraphQL is retired and answers 403. A key that answers 401 is stale rather than absent. The CLI reads a [default] profile, not a bare top-level apikey.
Prefer batch over a pod when the work is a queue of independent items. A camera sweep is that shape. A pod bills while it waits for an instruction; a batch does not.
References
See SKILL.md for the order, and DETAILS.md for the measurements.