It shouldn't require retraining, nope. I believe for INT4 there is a small adapter layer added that needs to be trained, but it is small and wouldn't require mu

It shouldn't require retraining, nope. I believe for INT4 there is a small adapter layer added that needs to be trained, but it is small and wouldn't require much data or computation to do so. Will know more once we actually start implementing :)