It seems like it quantizes FP16 into INT8. That's a waste, because many models can work perfectly fine with binary/ternary weights, and some can even tolerate l

It seems like it quantizes FP16 into INT8. That's a waste, because many models can work perfectly fine with binary/ternary weights, and some can even tolerate low precision activations.