>>105672352
>>105672416
from a quick test, there is zero or negligible difference from using TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 and/or main.py --use-pytorch-cross-attention

this kind of thing is where AMD really needs to get their shit together. instead the retards in their AI department are busy releasing "optimized" SDXL and SD35 versions that nobody will fucking use. they should instead be contributing to comfyui and underlying libraries to make sure consumer AI runs fast.