RFD 1163: The flow rents a card, the accelerator stays on the desk
Problem
The four loops run on the desk card, and RFD 1140 recorded what that costs: one RTX 3090 at 24 GiB, where OmniGen2 pages into shared memory rather than failing, so the symptom is a run that stops being fast.
The UGen300 has no place in any loop. The compile that would give it one was killed at 30.26 GiB, Docker’s ceiling and also WSL’s under an absent .wslconfig. That file now grants 48 GB and WSL reports 47, so the ceiling is lifted and the compile is untried rather than blocked.
Renting answers the first and cannot answer the second: a pod has no USB, so no rented card runs a HEF. The compiler needs no device at all – gate_dfc_parse.py says so in its own docstring – and that splits the work rather than blocking it.
Decision
The flow runs on RunPod and the accelerator runs on the desk, and the HEF is the only thing that crosses. RunPod takes OmniGen2 with anny-camera-lora, EditScore, and the Dataflow Compiler; the desk keeps hailortcli and usb/004:013. RFD 1140’s order governs the renting: push before, tear down after, check the tear down.
Loop 1’s detect_keypoints goes first, because RF-DETR is the one graph already through translate: 825 nodes, 22 operators, and the compiler agreeing with the allowlist. EditScore is second and bounded; DETAILS.md has the measurement that bounds it.