Describe the bug

The following code assumes that the only devices supported in PyTorch are CPU and CUDA. macOS devices now come with a Torch-supported built-in GPU known as "mps". Specifying "mps" as the device type will result in this error: AttributeError: 'ASRBrain' object has no attribute 'device_type'

        if self.device == "cpu":
            self.device_type = "cpu"
        elif "cuda" in self.device:
            self.device_type = "cuda"

            # Set cuda device based on device string
            try:
                _, device_index = self.device.split(":")
                torch.cuda.set_device(int(device_index))
            except ValueError:
                torch.cuda.set_device(0)

The code below will then check device_type and crash:

        if self.device_type == "cpu" and (
            self.precision == "fp16" or self.eval_precision == "fp16"
        ):
            raise ValueError(
                "The option `--precision` or `--eval_precision` is set to fp16. "
                "This option is not yet supported on CPU. "
                "Please use `--precision=bf16` or `--eval_precision=bf16` instead "
                "to enable mixed precision on CPU."
            )

While it is understandable that full compatibility with non-CUDA GPUs cannot be guaranteed, I have verified that basic training/inference is possible with MPS, and I would expect this to also affect AMD GPUs and others.

Expected behaviour

Recipes should be able to load on non-CUDA GPUs as long as they are not using any CUDA-specific functionality or Torch functionality unsupported by the device itself.

To Reproduce

No response

Environment Details

No response

Relevant Log Output

Additional Context

No response