not quite there yet, but on possibility is train multiple models and switch between them on the fly

another is add more params to the models to take the knob states into the account when doing predictions (which can impact the performance)