Look at integrating crumb's RLHF prompter

[philpax]

@aicrumb built a really cool RLHF-trained Stable Diffusion prompter on BLOOM: https://huggingface.co/crumb/bloom-560m-RLHF-SD2-prompter

I believe it's possible to convert it to ONNX and then run it under Tract or similar. CPU inference should be fine for this; it obviously won't be as fast as GPU inference, but we only need a few tokens, anyway

[philpax]

https://huggingface.co/blog/convert-transformers-to-onnx