This is the way. I managed to run (super) tiny models on CPU only with this approach.