Low step inference for speedups?

#1
by prottropprot - opened

Is it possible to achieve consistent low step (4 or 8 step) technique in Lumina family of models? Current requirement of 30+ steps is a steep barrier of entry. Do approaches like DMD2 translate tho this architecture?

Hi currently, there is not any effective solution to reduce step for lumina Image v2. I am still trying some approaches to handle this. So, when i complete i will publish and have a note for this.

This comment has been hidden (marked as Resolved)

hi @Y4iges . This problem related to the sage attention. If you use it please turn it off.

Sign up or log in to comment