RFD 1128: Does the cascade survive four bits
Problem
The ASUS UGen300 is a Hailo-10H with 8 GB, specified for 4-bit weights. Pixal3D’s checkpoints are 24.045 GB in bf16, three times the device. At four bits they are about 6 GB, which fits.
So the arithmetic permits it and nothing says the model survives it. A 1.3B diffusion transformer at four bits may produce a mesh nobody can tell from the original, or a worse one, or a broken one. RFD 1043 records the same open question for another model.
Decision
Measure tolerance first, because a failure here makes RFD 1129’s operator work pointless.
The 3090 is the instrument. Its 24 GB holds bf16 and four-bit forms of one stage at once, so the same input runs both ways and the difference is the quantity. The device cannot do this: 8 GB holds one form.
Three rules make it mean something: the input is identical rather than similar, since render_view.py is bit-reproducible; the baseline is bf16 on this card rather than a paper; and the error is millimetres of surface deviation rather than an impression.
Quantise with the Dataflow Compiler: bitsandbytes measures a four-bit, and the DFC’s host emulator runs the device’s.
See DETAILS.md for the apparatus, and SKILL.md for the order.