Avoid PyTorch3D CUDA normals kernels

Avoid PyTorch3D CUDA normals kernels

Avoid PyTorch3D CUDA normals kernels

Broaden CUDA archs and fallback on nvdiffrast runtime errors

Rebuild nvdiffrast for ZeroGPU torch and add fallbacks

Avoid diffusers device_map for explicit CUDA placement

Use supported diffusers device map strategy

Use lazy CUDA nvdiffrast context for ZeroGPU

Deploy official local Gradio demo adapted for ZeroGPU