RFD 1156: OmniGen2 fits as stages, not as a pipeline

Problem

OmniGen2 is three components on disk, measured rather than estimated: a 15.87 GB transformer, a 15.02 GB mllm and a 0.34 GB vae, stored at float32. That is 7.81 B parameters.

Resident together at eight bits they are 7.81 GB against 8 GB. The arithmetic fits and the device does not, because nothing is left for activations. At four bits they are 4.30 GB and the pressure lifts.

The components run in sequence. If they need not be resident at once, the binding number is the largest single stage at 3.97 B, which is comfortable at eight bits. Whether this part can stage residency is unknown, and on a USB accelerator swapping a 4 GB stage per invocation is the streaming case RFD 1130 exists to measure.

Decision

Abandoned on 2026-08-28. The accelerator work is scoped to rf-detr keypoint and RFD 1157, and this scored 13 of 25 against RFD 1157’s 18.

Treat OmniGen2 as three placement questions, not one. Ask whether the device can hold one stage and swap, and measure the transfer that costs before choosing a precision.

Take the vae first. At 0.085 B it is the cheapest graph here, it is a plain encoder-decoder, and it answers the operator question for this family without committing to the 4 B transformer.

Related

RFD 1130 separates resident from streaming. RFD 1026 gives the catalog memory. RFD 1128 decides four bits.