chengzeyi / stable-fast

Best inference performance optimization framework for HuggingFace Diffusers on NVIDIA GPUs.
MIT License
1.19k stars 73 forks source link

[Bug] MemoryEfficientAttentionTritonFwdFlashBwOp no longer availale #140

Open imD-5 opened 7 months ago

imD-5 commented 7 months ago

Hi! thanks for this great repo it's helping me a lot. So i've upgraded to xformers == 0.0.25, stable-fast==1.0.4, and i get the following error.

AttributeError: module 'xformers.ops' has no attribute 'MemoryEfficientAttentionTritonFwdFlashBwOp'. Did you mean: 'MemoryEfficientAttentionCutlassFwdFlashBwOp'?

I did some research and apparently xformers deleted the function MemoryEfficientAttentionTritonFwdFlashBwOp because it wasn't as efficient and also hgave some errors under some circumstances.

reference: https://github.com/facebookresearch/xformers/blob/main/CHANGELOG.md

can you fix this?

jryebread commented 7 months ago

same issue :(