Before this change, the model is loaded into GPU regardless of the value of "device" argument in CLI. (e.g. whisper "test.wav" --device cpu loads into GPU anyway)