Has anyone managed to get --upcast-sampling working on CUDA?

[FNSpd]

Doesn't seem to give any performance effect for GTX 1650 (which should work twice as fast using FP32)

[fractal-fumbler]

i am getting different results in image generation without --upcast-sampling and with it enabled
also getting different result with --no-half --precision full

atm what i can't able to use is Upcast cross attention layer to float32 cause getting error

ValueError: Query/Key/Value should all have the same dtype
  query.dtype: torch.float32
  key.dtype  : torch.float32
  value.dtype: torch.float16

gpu is 3060

[fractal-fumbler]

--xformers --xformers-flash-attention --listen --no-half-vae
p.s. don't use --no-half

1. without --upcast-sampling


2. with --upcast-sampling


3. with --upcast-sampling and Upcast cross attention layer to float32


[FNSpd]

Maybe it's due to compression but they are slightly different. Overlayed them in GIMP with "Difference" blend mode and there still was some noise left

[fractal-fumbler]

yes, there are some difference in images.
huge difference between fp16 and fp32 i noticed on high resolution images and with highres fix.
gonna try it out

[fractal-fumbler]

tried different checkpoint (illuminati v1.1)

1. with --upcast-sampling


2. with --no-half --precision full


3. without --upcast-sampling or --no-half --precision full, but with ATTN_PRECISION=fp16


[DimaSpell]

Hi! At the last step my pic become like I attaches, only grey. How to fix it? Help me, please