TRELLIS.2

IMO TRELLIS.2 is slightly different case from the HF models scenario. It depends on five compiled CUDA-only extensions -- flex_gemm for sparse convolution, flash_attn, o_voxel for CUDA hashmap ops, cumesh for mesh processing, and nvdiffrast for differentiable rasterization. These aren't PyTorch ops that fall back to MPS -- they're custom C++/CUDA kernels. The upstream setup.sh literally exits with "No supported GPU found" if nvidia-smi isn't present. The only reason I picked this up because I thought it was cool and no one was working on this open issue for Silicon back then (github.com/microsoft/TRELLIS.2/issues/74) requesting non-CUDA support.

They're still going. I just bought a 5090 for myself this Christmas to do more interesting things. I mostly use them for game assets. Trellis2 is very cool. Ive managed to put together a sdxl -> trellis -> unirig pipeline to generate 3d characters with mixamo skeletons that's working pretty well. On the llm front, deepseek and qwen are still cranking away. Qwen3 a22b instruct, imho does a better job than gemini in some cases with ocr and translation of handwritten documents. The problem with these frontier open weight models is that running them locally is not exactly tenable. You either have to get a cloud GPU instance, or go through a provider. - https://github.com/microsoft/TRELLIS.2 - https://github.com/VAST-AI-Research/UniRig