I see Diffusers#772 was included with today’s diffusers release, which means I should be able to pass some kind of
device_map when I construct the pipeline and direct which device each submodel is loaded on, right?
But I’ve got
device_map=dict(unet='cuda') and am running in to errors that indicate it’s trying to run the unet on CPU.
It may be that
unet is not the right key? I’ve seen some indication that the text encoder might be
text_model, and not
text_encoder as I expected from the way the submodels are named in model_index.json.
How do I see what names it knows things by?