TRELLIS.2 3D MODEL NOW RUNS ON MAC SILICON

The original TRELLIS.2 model, a 4-billion-parameter image-to-3D converter, relied on CUDA-exclusive components including flash_attn, nvdiffrast, and custom sparse convolution kernels—none compatible with macOS. The port implements pure-PyTorch replacements across nine files: a gather-scatter sparse 3D convolution, SDPA attention for sparse transformers, and Python-based mesh extraction that replaces CUDA hashmap operations. Total modifications span a few hundred lines. The changes enable the model to run via PyTorch's Metal Performance Shaders (MPS) backend on Apple Silicon chips. This makes advanced 3D generation accessible to Mac users without requiring external GPU hardware or cloud services. The project reduces barriers to running state-of-the-art generative models locally on consumer hardware.