huggingface
Magcache Support.
Id like to see magcache acceleration supported if possible. My own testing shows it to cut video generation times in half with almost no quality loss at all. It seems to be considerably better quality than something like teacache or more efficient attention mechanisms.
https://github.com/Zehong-Ma/MagCache
Kandinsky5 I2V Normal inference: 1168 seconds
https://github.com/user-attachments/assets/3404e1bc-8dfc-4303-b189-c49f549eafab
Kandinsky5 I2V Magcache Inference: 592 seconds
https://github.com/user-attachments/assets/4106a91c-bc4a-44a4-b1b7-4c4a34840551
@leffff @Zehong-Ma
Hi! We have a MagCache version in our main repo.
@sayakpaul hiw can we integrate MagCache into Diffusers pipeline?
We support a bunch of caching techniques in Diffusers: https://huggingface.co/docs/diffusers/main/en/api/cache
I think we could ship MagCache in a similar manner. Good reference would be FirstBlock cache: https://huggingface.co/docs/diffusers/main/en/api/cache#diffusers.FirstBlockCacheConfig
