RFD 1132 details: the conversion order and what it rests on

Every model in one table, from two sources that did not agree.

RFD 1016 inventories fifteen, drawn from src/library/aiModelsCatalog.js – the studio’s catalog. default.xml records what is checked out, per goal. Neither is a subset of the other, and an order built from one alone orders work that is partly somewhere else.

in RFD 1016 and on this goal's manifest          4
in RFD 1016 but pinned to the mesh-latents goal  5
in RFD 1016 and in no manifest here             6
on a manifest and absent from RFD 1016          8
in use with no repo and no catalog entry        2

Ranked measured

goal is what default.xml says, which is what decides placement: main for the keypoint goal, the mesh-latents tag for the other. A dash means the catalog names it and no manifest here checks it out, so nobody can census it.

Qwen3.5-Defiant is not in the table, and Gemma 4 is chosen over it. Both are RunPod serving interactors rather than accelerator candidates today, so the choice is which one to carry forward, and it turns on audio. Gemma 4’s card supports text, image and video on every size and audio on E2B, E4B and 12B – the served build is the 12B. Qwen3.5-9B has vision, needing a separate mmproj, and no audio at all, so it is not a smaller version of the same thing.

Architecture agrees with capability here, which is unusual enough to say. Gemma 4 12B Unified is encoder-free and dense; Qwen3.5 is Gated Delta Networks over a sparse MoE, and MoE routing is data-dependent – the one Refused category no rewrite reaches. The 9B is 0.87 GB smaller at Q4_K_M and that is the only axis it wins.

Unverified, and it is about the fine-tune rather than the architecture. The served build calls itself specialized for coding and agentic work, and says its multimodal examples are inherited from base. Whether it keeps working audio and vision towers, and whether the GGUF ships their projectors, is settled by loading it and feeding it a clip. Nothing here has done that.

TRELLIS2 is superseded by Pixal3D and is not in the table. Both draw their views from sphere_hammersley_sequence – render_view.py names them together – so Pixal3D inherits the structured-latent lineage, and the mesh-painting stage TRELLIS2 provided is derivable from it rather than lost. The three trellis2-* repositories are still checked out on the mesh-latents tag and no RFD records the supersession; this is the first place it is written down, which means RFD 1038 and RFD 1039 package models nothing now plans to convert.

Two models were in use with no repo of their own, and now have one. OmniGen2 and EditScore are pip dependencies inside 6-datasource/anny-render-corpus, declared as the omnigen2 and editscore pixi features. When this was written neither appeared in RFD 1016 nor as a <project> anywhere. They were not missing by oversight – they were used by the corpus pipeline rather than served from the catalog, so nothing that enumerates model IMAGES would ever find them. That is the third way a model can be absent from a list of models, after “on the other goal” and “in no manifest at all”.

Both are now forked from VectorSpaceLab at Apache-2.0 and placed on 3-interactor in weftspun-keypoint. The observation stands even though its examples no longer do: being a dependency rather than an image is still a way to be invisible to a catalog, and the fix was to place them, not to argue they were findable.

model id goal status
rfdetr keypoint keypoint rank 1, measured
omnigen2 keypoint placed, uncensused, absent 1010
editscore keypoint placed, uncensused, absent 1010
gemma4_composer mesh-latents uncensused, absent from 1010
seethrough_layerdiff keypoint placed, uncensused
seethrough_marigold_depth keypoint placed, uncensused
seethrough_vae keypoint placed, uncensused
seethrough_partseg keypoint placed, uncensused
skintokens_auto_rig keypoint uncensused
cyclegan_style_transfer keypoint uncensused, absent from 1010
pose_consensus keypoint uncensused, absent from 1010
anny keypoint placed, uncensused, absent 1010
soma-x keypoint placed, uncensused, absent 1010
mujoco keypoint placed, uncensused, absent 1010
pixal3d_image_to_textured_mesh mesh-latents uncensused
pixal3d_image_mesh_painting mesh-latents placed, uncensused
voxhammer_text_mesh_editing mesh-latents uncensused
voxhammer_image_mesh_editing mesh-latents uncensused
kimodo_text_to_motion keypoint uncensused
tropes_removal_model keypoint placed 3-interactor, absent 1010

The device half at 576 with num_windows=1 is 825 nodes over 22 distinct operators, and every operator is inside gate_onnx_device.py’s allowlist. The numeric check against PyTorch holds at 3.3e-6. Reproduce with scripts/gate_onnx_device.py --num-windows 1.

It does not yet compile, and that is why it stays first rather than a reason to move it. Parsing takes 36 s and optimization 1001 s, and then the allocator refuses:

slice1 failed on kernel validation: slice1 has 2 APUs but max allowed is 1
BackendAllocatorException: No successful assignments: ne_activation_ew_sub_softmax4

That is a resource-placement failure on one layer, not a memory limit. The weights are about 25 MB at int8 against a module carrying 4-8 GB of LPDDR4, so capacity is not what stops it. Every allocator error carried NO_CONTEXT_ENABLED, so multi-context compilation is the untried lever.

A second model would inherit this obstacle without adding information about it.

The two with no repo

OmniGen2 generates the posed corpus in gen_posed_from_reference.py, and BLOCKLIST.md names it as the replacement for Qwen-Image-Edit: 7.8B, Apache-2.0, clean output on the same input that the blocked model corrupted. EditScore grades those edits in score_edits.py – a reward model rather than a generator, which is why the measured refusal count there was zero of twelve and the search for an uncensored base was unnecessary.

Neither is a conversion candidate today. Both run in the corpus pipeline on the desk, not on the accelerator, and converting a 7.8B generator to a 24 GB-class edge device is a different question from converting a detector. They are in the table because a list of models that omits the two doing daily work is not a list of models.

The ten that are excluded

RFD 1133 carries them, and this document does not restate the rows. It was split out of this one because the ten did not share a reason: three are blocklisted by the agreements, four are abandoned with the world-building scope, and three name nothing found in this repository. One table calling all ten blocklisted said the reversible and the closed were the same thing.

What ordering the rest requires

One number per model, and the tooling to produce it already exists:

  1. Export to ONNX and census the operators, as gate_onnx_device.py does.
  2. Run check_device_ops.py, which sorts every operator into observed and documented and reports which fall outside both.
  3. Count how many operators need a secondary-path rewrite. 139 of the 178 operators at opset 17 reduce to the 58 the device executes, so a model drawing only on those needs no new work; one reaching into the 39 refused operators may not be convertible at all.

The rank then follows from a measurement rather than a preference: fewest operators outside the executable set goes first.

The refused set is where a model becomes unconvertible rather than inconvenient. Four categories are permanent – data-dependent output shape, control flow, non-tensor types, and nondeterminism – and no rewrite reaches them. A model needing NonZero, or TopK with a data-dependent count, or If does not get a lower rank; it gets a different deployment target.

What this leaves for other RFDs

RFD 1016 is missing eight models that are checked out, three on this goal and five on mesh-latents. Either its catalog is narrower than the workspace on purpose, or it has drifted; this records the difference rather than deciding which.

Four catalog models have no checkout in either manifest here. They cannot be censused until somebody places them, and CLAUDE.md’s Sides rule says an unplaced project is the drift the six words exist to stop.

It does not decide whether a model that fails the census should be converted partially, split across the accelerator and the backup runtime, or left on the backup entirely. nx_shuttle already reports operators that run on one target and not the other, so the split is measurable; whether to take it is a decision nobody has made.