* ci: declare minimal GITHUB_TOKEN permissions (contents: read)
Clears the CodeQL actions/missing-workflow-permissions alert. These CI jobs only check out code and run tests, so contents: read is sufficient.
* chore: refresh packaging and retain node integration smoke
* test: retain budget integration in node cleanup
* style: clean budget integration test EOF
---------
Co-authored-by: Muse Code <muse@example.com>
pipeline.models[key] is None at patch time for GGUF Trellis2 pipelines
(Aero-Ex/ComfyUI-Trellis2-GGUF) that lazy-load the flow DiT inside the
sampler. nn.Module.__setattr__ only routes real Modules into _modules, so
self.inner = None became a plain __dict__ attribute that nn.Module.__getattr__
never searches, and the fallback raised 'object has no attribute inner'.
- store inner via _set_inner: real Modules stay registered submodules, None/
lazy placeholders are kept as a plain attribute; __getattr__ resolves both
- clear errors when a compute step / attribute access hits an unloaded model
- bind_inner(model) to attach the real DiT once it materializes
- _LazyPatchDict binds the model transparently when the sampler assigns it
into pipeline.models[key] (TRELLIS repos), so the cache survives lazy load
End-to-end validation on RTX 5090 (ComfyUI 0.24.0, Hunyuan3DWrapper master,
Hunyuan3D-2mini fp16, 30-step image-to-mesh workflow via the ComfyUI API):
- patch engages exactly on schedule (hermite i3: 11 computed + 19 skipped;
dmd i5: 7 + 23), sampling 16.1 -> 43.1 it/s (2.68x) for hermite i3
- quality A/B same seed vs unaccelerated mesh: hermite i3 F1@0.05 0.825
(different-seed floor 0.751); dmd i5 0.719, at the floor on mini, so the
upstream DMD lossless claim is not repeated for this node
- repeat runs in one session reset per-run state correctly
Two bugs found by the validation, both fixed with regression tests:
1. ComfyUI cache aliasing: the node mutated and returned the same pipeline
object, so a cached node output could silently run a different config
after an intervening run re-patched the shared pipeline (observed live:
a cached hermite-i3 output ran dmd-i5). apply/remove now return a
shallow copy that owns its patch; the input pipeline is never mutated.
2. Run-boundary reset missed back-to-back single-step runs (t=0 then t=0);
the boundary check is now non-increasing instead of strictly decreasing.
Defaults switch to the best-measured config (hermite, interval 3).
Tests 35 -> 38. README validation section rewritten from measurements;
REGISTRY-SUBMIT.md documents the registry.comfy.org publish flow.
One node, HiCacheAccelerate: patches the HY3DMODEL shape pipeline's DiT
(kijai/ComfyUI-Hunyuan3DWrapper) to skip DiT forwards and forecast the
flow-matching velocity via hicache-pp (HiCache Hermite / HiCache++ DMD /
auto holdout selection).
- zero comfy-internal imports: the patch targets the pipeline object the
wrapper passes between nodes, so it covers hy3dgen (2.0), hy3dshape (2.1)
and mini/turbo checkpoints alike
- 35 CPU unit tests: standalone node-pack load contract, schedule adherence
on a mock DiT with the wrapper's exact loop shape, bit-identity with
driving hicache-pp directly, DMD exactness on exponential trajectories,
run-boundary reset, unpatch, parameter validation
- registry.comfy.org pyproject metadata (publish deferred until GPU-in-
ComfyUI validation; repo ships as beta)
Krishi Attri