Backend API¶

`get_device(log_info=False, *, cuda_precision='bf16-mixed')` ¶

Detect the accelerator and the precision a trainer runs at on it.

Parameters:

Name	Type	Description	Default
`log_info`	`bool`	If True, log the Python, torch and GPU backend versions.	`False`
`cuda_precision`	`str`	The precision on CUDA. `train` passes `training.trainer.precision` (spec 4.2); the default is the shipped `bf16-mixed`.	`'bf16-mixed'`

Returns:

Type	Description
`str`	`(device, precision, num_gpus)`. The device is `cuda`, `mps` or `cpu`. The
`str`	precision is `cuda_precision` on CUDA and `32-true` everywhere else. `num_gpus` is
`int`	1 on MPS and 0 on CPU.