[Tracking] MPS (Apple Silicon) ops that fall back to CPU or error in FireANTs

[LalithShiyam]

Summary

This issue tracks PyTorch operators that are currently unsupported on MPS and either error or transparently fall back to CPU when running FireANTs on Apple Silicon.

Observed unsupported/fallback ops

• conv3d

  Error → RuntimeError: “Conv3D is not supported on MPS”
  Not yet implemented: [MPS] Add Conv3D support for MPS pytorch/pytorch#114183 (Unsupported)

• interpolate(..., mode="trilinear")

  aten::upsample_trilinear3d.out → Warning: “will fall back to CPU”

[LalithShiyam]

Update: the latest version of PyTorch supports conv3d.