Is it possible to use scan option to run diffusion (or other text-to-image) models with dummy inputs and gather shapes of all modules?
1 Like
Note sure!
The prerequisite for this would be to be able to load the modal on the meta device (which is work in progress).