FNSpd

I'm a few commits behind master. I'll test with PyTorch 2 to see if it'll work on mine version. Maybe, something breaks it in recent updates

> Looks good, except fully disabling autocast shouldn't be necessary. If there is some problematic section of code it can be disabled for that specifically, but otherwise fully disabling autocast...

So, looks like I found fix for PyTorch 2 issue but speed is the same as without upcast. I'll include it just in case but I would suggest to use...

> > I changed it to not confuse people and make one argument dependant on the other but yeah, --precision full works too. Without it GTX 1650 used to emulate...

> I tested the current code with 1.13 instead of 2.0 and basic generation works but LoRAs do crash the pipeline with them providing half but float is expected (I'll...

Do xformers do anything here? I think that only first optimization is applied

V1 uses less memory, iirc

Do xformers work with PyTorch 2.1?

Helps to keep shape of object the same at higher denoising values