fix(rt-detr): use float32 on MPS in build_2d_sinusoidal_position_embedding

[Shylin26]

What does this PR do?

Fixes #46159

build_2d_sinusoidal_position_embedding in both RT-DETR and RT-DETRv2
hardcodes dtype=torch.float64 for internal sinusoidal arithmetic.
MPS (Apple Silicon) does not support float64, causing a TypeError when
running inference on Apple Silicon devices.

This PR uses float64 on CUDA/CPU for precision, and falls back to
float32 on MPS devices only.

Code Agent Policy

• I confirm that this is not a pure code agent PR.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a Github issue? Fixes [RT-DETRv2] MPS crash: build_2d_sinusoidal_position_embedding hardcodes torch.float64, breaking Apple Silicon / MPS inference #46159
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

@yonigozlan @molbap

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: rt_detr, rt_detr_v2

[Shylin26]

Submitted PR #46513 to fix this — uses float64 on CUDA/CPU for precision and falls back to float32 on MPS devices to avoid the TypeError on Apple Silicon.

[pidefrem]

Thanks for picking this up — we hit this exact MPS crash with RT-DETRv2 and have been patching it locally, so glad to see a fix in flight.

The device-conditional approach looks right — keeps float64 precision on CUDA/CPU where it's supported, and gracefully falls back to float32 on MPS where it's not.

Two suggestions:

1. Device check robustness: str(device) != "mps" will miss torch.device("mps:0") or other variants, and str(None) is "None" which happens to work but only by accident (the signature accepts device: torch.device | None = None). Consider:

   compute_dtype = torch.float32 if device is not None and torch.device(device).type == "mps" else torch.float64

2. Propagation via modular: The canonical definition lives in modular_vit_mae.py (line 44). modular_rt_detr.py and modular_aimv2.py import it from there, and the code-generation system inlines it into all 8 generated modeling_*.py files (vit_mae, rt_detr, rt_detr_v2, aimv2, deimv2, d_fine, pp_doclayout_v2, pp_doclayout_v3). The fix should go in modular_vit_mae.py and be propagated with make fix-repo — otherwise the other 6 files remain broken. PR fix(mps): build_2d_sinusoidal_position_embedding crashes on Apple Silicon due to float64 on MPS device #46174 already did this correctly.

On the CUDA perf concern raised on #46174: this is a one-time positional embedding computation (a few hundred positions), not a hot path. The device-conditional approach avoids the question entirely since CUDA keeps float64 on-device as before — no CPU↔device transfer, no perf regression.

Happy to help test on MPS if needed — we use RT-DETRv2 via docling for layout detection on Apple Silicon daily.

[Rocketknight1]

Duplicate of #46174