Now that we have quite a few model classes (and transformer model classes), I wonder if it makes sense to have an AutoModel class. Oftentimes, users initialize the model separately and pass it to a DiffusionPipeline:

transformer = "..."
pipeline = DiffusionPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", transformer=transformer)

It's common to see this happening, especially in the realm of quantization. And in these cases, an AutoModel class could be quite nice with the following developer experience:

transformer = AutoModel.from_pretrained("...", subfolder="...")