why rebuild a vae in inference stage?

Thanks for ur effort for diffusion model.

I want to know why we need to rebuild a vae in inference stage. I think it will introduce extra GPU cost.
https://github.com/huggingface/diffusers/blob/a85b34e7fdc0a5fceb11aa0fa6199bd9afaca396/examples/text_to_image/train_text_to_image_sdxl.py#L1217C16-L1223C24

Start at examples/text_to_image/train_text_to_image_sdxl.py around lines 1217-1223 and trace how the VAE is constructed during inference. Compare that lifecycle with the training setup and document whether the rebuild is required and what the GPU cost is.

Vom Indexierungsmodell aus dem Issue-Text verfasst.