RFD 1129 details: the two stages, and what an export has to carry
Sparse submanifold convolution
o_voxel and flex_gemm convolve over occupied voxels only, and the saving is the whole point: a dense 1024 grid is 2^30 cells while the occupied set is a thin shell.
ONNX has Conv and no submanifold sparse convolution. An export therefore either densifies, which destroys the saving and the memory budget with it, or emits a gather, matmul and scatter pattern whose support is exactly the open question.
Neighborhood attention, reached through a model nobody declares
Pixal3D’s image conditioning loads valeoai/NAF over torch.hub, and NAF’s attention layers call natten:
naf.py:115 -> attentions.py:72 -> natten.functional.na2d
-> neighborhood_attention_generic -> cutlass_fna_genericNothing in Pixal3D imports natten, which is why grepping for it finds only a README line and concludes wrongly that it is unused.
It is not optional. IMAGE_COND_CONFIGS sets use_naf_upsample: True for three of the four conditioning models, and the projection width depends on the flag – proj_channels = embed_dim * 2 if use_naf_upsample – so the published weights were trained with it in place. Turning it off mismatches the checkpoint rather than skipping a step.
What the gate already knows
scripts/gate_onnx_device.py exports the device half and scripts/gate_dfc_parse.py runs the compiler against it. The two run as a pair, and their disagreement is the result:
| macOS gate | DFC | meaning |
|---|---|---|
| PASS | parses | the allowlist held |
| PASS | rejects X | DEVICE_OPS is too generous |
| FAIL on X | parses | DEVICE_OPS is too strict |
weftspun-hailo-dfc:5.3.0 is built and hailo_sdk_client imports. The wheel is Linux-only, which is why that image exists at all.
The order that saves time
Export the attention layer before the sparse convolution. It is smaller, it is the dependency nobody expected, and a rejection there blocks the conditioning path – which makes the convolution question moot until it is answered.