How to use diffusers format when training SD3

[David-Hari]

I am using the sd3_train.py script from the sd3 branch, and I was under the impression that it could accept a model as either a safetensors file or diffusers folder. The pretrained_model_name_or_path argument says so.
Yet when I give it "stabilityai/stable-diffusion-3.5-medium", the code still seems to try loading it as safetensors, calling the load_safetensors function in utils.py. That then fails with FileNotFoundError.

[David-Hari]

Additionally, even when I do use safetensors file, it fails to find clip_l, clip_g and t5xxl in the state_dict. I downloaded it from here but that does not appear to include the text models.