You can probably run the model on 4 (or maybe 8?) consumer GPUs, no?

→ View original post on X — @jeffdean