performance-mps-device-support-trainingarguments-does-not-detect-mps-gpu-on-macos-0398bf15

THEN Subclass TrainingArguments and override the `device` property to check for MPS availability using `torch.backends.mps.is_available()`. Additionally, set the environment variable `PYTORCH_ENABLE_MPS_FALLBACK=1` to handle operations not yet implemented in MPS, falling back to CPU with a warning.

import torch
from transformers import TrainingArguments

class TrainingArgumentsWithMPS(TrainingArguments):
    @property
    def device(self) -> torch.device:
        if torch.cuda.is_available():
            return torch.device("cuda")
        elif torch.backends.mps.is_available():
            return torch.device("mps")
        else:
            return torch.device("cpu")