[I found a real fix to work in half precision] I identified when the GeForce 16xx crash in half-precision

[KarhellTwitch]

So i made some test with my 1660 Super and I found what triggers the problem with the half-precision.
To do my tests I started with only --xformers as parameter

And everything works fine.... Until....

The problem seems to be related to the method which split the prompt when it exceeds 75 to put them in queue (also for the negative prompt)

Here a test without exceeding 75 in the 2 prompts (works 100% of the time)

Here a test with both at 75/75

Here a test when one of them exceed 75

So we can say that the problem comes from how Unet splits the prompts to put them in queue

[KarhellTwitch]

Ok i found a fix ! I'll test it longer but i can already generate even with a prompt over 75 now

I activated this option in optimizations (settings)

And now it works

[KarhellTwitch]

It works at any lenght now with only --xformers in launch parameters

I hope it can help other GeForce 16xx users to get benefit of half precision with their already low VRAM GPU