FEATURE 5 — macOS support for all models

As a macOS (Apple Silicon) user, I want the generation models to run on my Mac, so I can generate 3D meshes without a CUDA GPU.

The backend currently assumes CUDA (torch.cuda.is_available() in api/services/generators/base.py), so the four models only run on Windows. macOS needs the Metal (MPS) backend, with a CPU fallback. This ticket adds device detection and makes each model run on macOS.

Models

• TripoSG
• Trellis2
• Hunyuan3D
• Pixal3D

Scope

• Auto-detect device: CUDA → MPS → CPU based on the platform.
• Each model runs on MPS and produces a valid GLB on Apple Silicon.
• Handle MPS-specific issues (dtype, unsupported ops) with a CPU fallback where needed.
• Weights download and load correctly on macOS.

Implementation

• api/services/generators/base.py: replace the CUDA-only assumption with a shared device-selection helper (cuda → mps → cpu).
• Each model extension (generator.py): use the selected device instead of hardcoded .cuda(); guard ops that MPS doesn't support.
• Reuse the same helper across all four models so device logic lives in one place, not duplicated per extension.
• No frontend, store, or API surface changes.

Acceptance criteria

• Each of the four models generates a valid GLB on an Apple Silicon Mac.
• No device-related errors on macOS (falls back to CPU when MPS is missing an op).
• Windows/CUDA behavior is unchanged.

Estimate: M