Explore non-CUDA backends (MPS / AMD / Intel) for Trellis2

Describe the problem
Trellis2 is a strong project, but its current CUDA-only dependency limits who can use and contribute to it.

As the AI hardware ecosystem evolves, more developers are working on platforms outside the NVIDIA stack, and Trellis2 is currently inaccessible to them.

Why this matters
Recent advances have made alternative platforms increasingly compelling:

• Apple Silicon (MPS / Metal) offers excellent local inference performance and efficiency
• AMD GPUs are broadly available and improving quickly (ROCm / Vulkan)
• Intel ARC has become a realistic option for AI workloads (Level Zero / OpenCL)

Supporting additional backends would significantly broaden Trellis2’s reach.

Impact

• Greater accessibility across platforms
• Increased contributor and user base
• Reduced reliance on a single vendor ecosystem

Requested improvement
If possible, please consider:

• A roadmap or discussion around MPS support
• Exploration of AMD-compatible backends
• Experimental investigation into Intel ARC acceleration

Even early-stage or partial support would be valuable to many users.

Additional context
Projects that adopt multi-backend acceleration tend to attract more contributors and remain adaptable as hardware continues to change. I believe Trellis2 would benefit long-term from this direction.

Happy to help test or contribute where useful.

Start by reviewing Trellis2's CUDA-only dependency and the discussion around MPS, AMD, and Intel ARC support. The issue names no files, entry points, or tests; done would require a defined backend direction or scoped implementation plan before coding can begin.