TRELLIS.2 runs on AMD ROCm

This is a nice reminder that AMD compatibility for frontier AI often arrives through community patches, not vendor blessing. The practical result matters: TRELLIS.2’s image-to-3D pipeline is now usable end to end on Radeon, which broadens the local 3D generation crowd.

• –Official TRELLIS.2 docs still assume Linux plus NVIDIA CUDA, so this fills a real gap for Radeon users.
• –The chunked linear workaround suggests ROCm can stumble on large tensor ops in 3D pipelines, not just LLM inference.
• –Swapping `cudaMemcpy2D` for `cudaMemcpy` shows how HIPified C++ can quietly corrupt geometry rather than crash.
• –End-to-end preview and GLB export are the real milestones here, even if the preview still drops normals.
• –A ~280-second run at 1024 resolution and 20 steps is acceptable for a demo, but still far from a casual creator workflow.