RFD 1154: Pixal3D fits only at four bits, as a whole
Problem
Pixal3D is 12.02 B parameters: 24.05 GB at bf16 and 12.02 GB at eight bits against 8 GB of device memory, fitting only at four bits and 6.61 GB. One precision reaches it, so RFD 1128 was a precondition.
THAT IS THE WRONG UNIT, AND THE ABANDONMENT WAS DECIDED ON IT. Pixal3D is four stages. The sparse structure stage is ss_flow_img_dit_1_3B_* at 1.3 B: 2.6 GB at bf16, 0.72 GB at four, fitting at every precision and never meeting RFD 1128. It is also the only stage that runs here, its two configs being the only ones without use_naf_upsample, so they never reach NATTEN.
Decision
THE PREDICTION THAT REPLACED IT WAS ALSO WRONG. This RFD said sparse convolution would supply RFD 1131’s gather and scatter. sparse_structure_flow.py imports nothing from modules.sparse: it is a dense DiT, named for what it emits, not how it computes.
Abandoned at 10 of 25 and it stays abandoned, and the reason is now measured. Complex RoPE blocked the export, an encoding rather than arithmetic: rewritten in reals it is bit-identical and the graph exports at 544 nodes, carrying none of the refused operators.
The compiler refuses it anyway, and not for arithmetic. Every attempt dies in _add_input_layers: the parser wants each input image-shaped, and this stage takes a voxel grid, a timestep and two conditioning tensors. Cos, Sin and ReduceL2 stay unjudged; DETAILS.md has the ladder.