MolecularDiffusion.device¶

Device selection shared by training, generation and analysis.

Order is CUDA -> Apple MPS -> CPU, so a CUDA machine picks exactly what the old "cuda" if torch.cuda.is_available() else "cpu" idiom picked; MPS only enters the picture on a Mac, where CUDA can never be available.

MOLCRAFT_DEVICE (e.g. cpu, mps, cuda:1) overrides the choice.

Functions¶

`get_device`(→ torch.device)	Return the preferred torch device: $MOLCRAFT_DEVICE, else cuda, mps, cpu.
`mps_available`(→ bool)
`to_device`(obj, device, args, *kwargs)	Move any nested container of tensors to `device` (device-generic `utils.cuda`).

Module Contents¶

MolecularDiffusion.device.get_device() → torch.device¶: Return the preferred torch device: $MOLCRAFT_DEVICE, else cuda, mps, cpu.

MolecularDiffusion.device.mps_available() → bool¶

MolecularDiffusion.device.to_device(obj, device, *args, **kwargs)¶: Move any nested container of tensors to device (device-generic utils.cuda).