RFD 1167 details: one table per rung, and the evidence for each place
Ranks in brackets are RFD 1166’s, carried so the two can be read together without being confused for each other.
The rungs run low to high, which is also early to late: a model climbs, so the document is read in the order the work happens.
Rung 0 – assessed, never exported
model [1166] what the score rests on
cyclegan_style_transfer [2] read only
OmniGen2 [3] measured at runtime, not exported:
131 s bf16, 133 s NF4, 14.75 GiB
EditScore, Qwen3-VL-8B [4] measured at runtime: 6.75 GiB NF4,
28-36 s a score
Kimodo [5] RFD 1026 estimate
MoGe [6] read only; a Pixal3D dependency
SkinTokens [7] RFD 1026 estimate
MuJoCo MJX [8] read only
Mitsuba 3 shading [9] measured at runtime: 14.1 ns a
pixel shading, 48.1 G-bufferNine of eleven. Four carry runtime measurements, which is not nothing, but a model that runs on a 3090 has said nothing about whether it compiles for a dataflow part.
Rung 1 – exports to ONNX
model [1166] evidence
VoxHammer, DINOv2 stage [12] 1343 nodes, 22 operators, at
[1,3,518,518] fixed. onnxruntime
against torch: max|diff| 6.027e-04
over max|ref| 120.355, 5.008e-06
relative.It went straight to rung 2 in the same run, so it is listed there rather than resting here. It is the first model other than rf-detr to earn a rung instead of being estimated onto one.
THE EARLIER WORDING OF THIS SECTION FLATTERED US, AND THE EARLIER WORDING FLATTERED US. It said everything reaching rung 1 went further in the same sitting, which reads as though the rung were merely a place nobody lingers. Enumerating the checkouts says otherwise: across the whole workspace, torch.onnx.export appears in rf-detr-cpp and in two projects that are not candidates at all. Ten of the eleven others have no export site of any kind.
Rung 1 is empty because nobody has written the export, not because everybody sprinted past it. The rung is still the cheap one – no weights, no device, no rented card – and that is now an argument about work not yet started rather than work that left no trace.
Rung 2 – the operator set is known
MoGe’s rung-2 entry is about v1, and RFD 1171 has since taken v3. The export targets dinov2_vitb14, which is v1’s encoder; v2 and v3 are ViT-L, so the census below has to be redone rather than carried over. The measurement stands as made.
model [1166] evidence
VoxHammer, DINOv2 stage [12] 22 operators, every one inside
`DEVICE_OPS` -- the same allowlist
rf-detr's device half carried
through DFC 5.3.0 at 825 nodes.
Nothing refused, so rung 3 is the
next thing to run rather than the
next thing to argue about.
TRELLIS.2 / Pixal3D [11] the SS stage exports at 544 nodes
with RoPE rewritten in reals,
bit-identical at max|diff| 0.000e+00.
28 operators, of which `Cos`, `Sin`
and `ReduceL2` sit outside
DEVICE_OPS. Rung 3 refused it in
`_add_input_layers`, four times, on
input rank rather than arithmetic.No script in this workspace reproduces that row. pixal3d-upstream has no export site, and the RoPE rewrite that produced 544 nodes lived in a scratch file that is gone. The measurement is recorded and the apparatus is not, which is the half CLAUDE.md asks for by name. Rung 2 therefore rests on a number nobody here can re-derive, and re-earning it means writing the export that should have been committed the first time.
Rung 3 – the Dataflow Compiler accepts the graph
model [1166] evidence
rf-detr keypoint [1] 825 nodes, 22 operators, parse OK on
arch hailo10h, operators outside the
allowlist: none. 61 s to translate.One model. It is also the only row in RFD 1166 whose scores rest on something run rather than read.
Rung 4 – quantises
model evidenceEmpty, and rf-detr has attempted it twice. Statistics Collector took SIGKILL at Docker’s 30.26 GiB, cleared by a 48 GB ceiling; then Quantization-Aware Fine-Tuning exhausted the 3090’s 24 GiB at the default batch and again at batch_size=1, epochs=1 over 64 frames. RFD 1165 carries that retraction.
Rung 5 – runs on the device
model evidenceEmpty. No HEF has executed on usb/004:013 from this workspace. The device itself is measured – HAILO10H, firmware 5.3.2, 2.27 ms hardware latency on a zoo classifier – but nothing of ours has reached it.
torch.onnx.export was the convenient proxy, and it understated
The audit above counted export CALL SITES, found them in rf-detr-cpp and nowhere else among the candidates, and concluded the other ten would each need an export written from scratch. The count is right and the conclusion it invited is too pessimistic.
MoGe v2 has no export call and is nevertheless the readiest of them. DINOv2Encoder carries an onnx_compatible_mode setter that upstream wrote for this purpose, and it switches off precisely the two things that break a fixed-shape export: the positional-embedding fast path that branches on npatch == N and w == h, and antialias=True in the F.interpolate that resizes to a patch multiple.
That second one is the same obstacle rf-detr’s gate meets and answers differently. gate_onnx_device.py constant-folds the antialiased resize, which is exact and asserts the fold fired; MoGe turns antialias off, which changes the numbers. Two models, one obstacle, and the choice between them is exactness against convenience.
So the readiness signal is not whether a repository calls the exporter. It is whether the graph has a fixed-shape mode at all, and asking the easy question gave the gloomier answer. Rule 1, on our own audit.
What rung 0 does not distinguish, and should
Rung 0 reads blocklist clear, checkout present, and eleven rows satisfy it. They do not satisfy it in the same way, and the single label hides two kinds of row that can never climb as written.
Three rows have no model on disk. Kimodo and SkinTokens are server.py and nothing else: every path outside WEFTSPUN_STUB=1 raises NotImplementedError, so what is checked out is an HTTP contract, not a network. cyclegan_style_transfer is the same shape with a checkpoint path in it – one file, no nn.Module anywhere. For these three, rung 1 is not the next step; acquiring the model is.
Two rows are not neural networks at all. MuJoCo MJX is a physics engine and Mitsuba 3 is a renderer. Neither has weights, neither has a graph to export, and the ladder from ONNX through the Dataflow Compiler is the wrong instrument for both. They sit at rung 0 the way a boat sits at the bottom of a staircase. Accelerating them is a real question and it is a different one, so the honest entry is not on this ladder rather than a rung.
That leaves eight rows the ladder actually measures: rf-detr, OmniGen2, EditScore, MoGe, Pixal3D/TRELLIS.2, VoxHammer and – once it has a model – CycleGAN. Six of the seven hold real module and checkpoint code, so for them rung 1 is genuinely the next thing rather than a placeholder.
See-Through left the field entirely. RFD 1166 dropped it: every checkpoint its inference scripts load states no licence, and the depth one is a fine-tune of an OpenRAIL++-M model whose restrictions travel into derivatives. It is kept as a reference for layer taxonomy, which needs no weights and so has no rung.
What the shape of this says
The ranking’s top three have never been exported and its eleventh has been further up the ladder than any of them. That is not a fault in the ranking – RFD 1166 asks what is worth doing – but it is why the two tables are separate documents.
That was written with rung 1 empty, and the cheapest thing that would change the picture was said to be an export of anything – VoxHammer’s DINOv2 stage being the shortest, a stock checkpoint at a fixed resolution in a family already cleared at rung 3.
It was, and it cost one script. 1343 nodes, no refused operator, and two rungs in a single run. The prediction is kept because it was made before the measurement rather than after it, which is the only condition under which a prediction is worth anything.
Six of the eight measurable rows have still never been exported, and the same argument applies to each of them unchanged.