Fix DINOv3 layer access for transformers >= 5.0

[Eyalm321]

Summary

• In transformers >= 5.0, DINOv3ViTModel wraps its encoder under .model, so the layer ModuleList lives at self.model.model.layer (the inner model is DINOv3ViTEncoder). Older transformers exposed .layer directly on the top-level model.
• extract_features in trellis2/modules/image_feature_extractor.py only handled the old layout, breaking inference on a fresh install with a current transformers release.
• Pick whichever exists via getattr(self.model, "model", self.model) so the path works on both old and new transformers.

Repro

On a fresh setup.sh --basic --flash-attn --nvdiffrast --nvdiffrec --cumesh --o-voxel install (transformers 5.6.2 from PyPI), running Trellis2ImageTo3DPipeline.run(image) raises:

File "trellis2/modules/image_feature_extractor.py", line 86, in extract_features
    for i, layer_module in enumerate(self.model.layer):
AttributeError: 'DINOv3ViTModel' object has no attribute 'layer'

The fix unblocks inference on the latest pinned transformers without affecting older versions.

Verified

Inference end-to-end on RTX A6000 + transformers 5.6.2 + torch 2.5.1+cu124 produced a valid GLB (504K verts / 945K faces) from a single 1024×1024 PNG input.

Test plan

• Inference completes and exports GLB on transformers 5.6.2
• No-op on transformers <5 where self.model.layer already exists (getattr returns the same model object, behavior unchanged)

🤖 Generated with Claude Code