3 ms·
It's not quantising existing models, they're training new ones.
by IanCal 3y ago
It's not quantising existing models, they're training new ones.
- leroman 3y agoI understand this part but it seemed that the 16->8->4 etc is similar to compression of the "net" and seemed to lower quality below 8.