João Marcelo Brito
Results
2
comments of
João Marcelo Brito
The answer from @rolson24 is correct in the sense that the weights are only converted before the linear operation and then dequantized just like the original paper describes(see image below)....
I was experiencing the same issue on 1.119 and iOS 26, migrated first for 1.136.0 then 1.143.1 and it worked after that.