1 utils.precision

utils.precision

Precision-context helpers for float32 geometric computations.

1.1 Functions

Name	Description
float32_geometry_context	Disable autocast for CPU/CUDA geometry and use a no-op context on MPS.

1.1.1 float32_geometry_context

utils.precision.float32_geometry_context(device)

Disable autocast for CPU/CUDA geometry and use a no-op context on MPS.

Callers explicitly cast geometric inputs to float32 inside this context. Torch 2.4 cannot construct torch.autocast(device_type="mps", enabled=False); MPS therefore uses :func:contextlib.nullcontext while retaining those explicit float32 casts. This does not claim all-native MPS AMP support.

# utils.precision { #aria_nbv.utils.precision }

`utils.precision`

Precision-context helpers for float32 geometric computations.

## Functions

| Name | Description |
| --- | --- |
| [float32_geometry_context](#aria_nbv.utils.precision.float32_geometry_context) | Disable autocast for CPU/CUDA geometry and use a no-op context on MPS. |

### float32_geometry_context { #aria_nbv.utils.precision.float32_geometry_context }

```python
utils.precision.float32_geometry_context(device)
```

Disable autocast for CPU/CUDA geometry and use a no-op context on MPS.

Callers explicitly cast geometric inputs to float32 inside this context. Torch
2.4 cannot construct ``torch.autocast(device_type="mps", enabled=False)``;
MPS therefore uses :func:`contextlib.nullcontext` while retaining those
explicit float32 casts. This does not claim all-native MPS AMP support.