new API always uses GPU

In inference.py line 193, the model class (new api) calls the older predict method but skips passing in device. Hence cuda device is always used by default, though model.device may be cpu.