Hi,
I have been use examples\controlnet\train_controlnet_sd3.py for controlnet training for a while, and I have some confusion and would like your advice

1. In the line 1097:
   vae.to(accelerator.device, dtype=torch.float32)
   It seems we should use fp32 for VAE, but as far as I know, SD3 currently has no fp32 checkpoints, so does it really work if we populate fp16 into fp32?

2. Before running the train script, accelerate config can specify whether to use mixed precision or not, since SD3 only has fp16 checkpoint at present, I don't know how to choose this option, whether to choose 'fp16' or 'no'.

Really appreciate your advice!
@sayakpaul @DavyMorgan