The stated VRAM requirements on Github seem almost impossibly high for both versions of the model relative to the claimed parameter count
#1
by
DiffusionFanatic1
- opened
Is this perhaps related to the current hardcoded requirement of full precision T5-XXL, or something?
Yep, it's due to the T5. You can just free the T5 after this line --
Later this week, we'll make a PR to have the T5 free itself (optionally) in order to reduce overall VRAM.