Thanks to @DN6 we support loading GGUF checkpoints and running on-the-fly dequants ([PR](https://github.com/huggingface/diffusers/pull/9964)). Currently, we don't support `save_pretrained()` on a `DiffusionPipeline` with GGUF, which, IMO, could be massively impactful, too.