Kevin Turner

that's odd. It's visible here even if I'm not logged in to GitHub. not sure what I'd need to do differently for you to view it?

I just pulled main here, ran vite build, reloaded, reloaded harder, and this still reproduces without fail here on both Firefox 118 _and_ Chrome 117. Nothing suspicious-looking in the browser...

was not fixed by that.

Looks promising! I think separating out xformers and sliced attention like this helps readers as well as any code-tracers, as it no longer has those if-branches in the `forward` method....

I just learned that StructureDiffusion is no longer anonymous; see #878 for more on this cross-attention use case.

SD 1.x models also use a CLIPTextModel: https://huggingface.co/CompVis/stable-diffusion-v1-4/blob/main/text_encoder/config.json#L19 but I have no idea how you look at a config like that and figure out what is being "skipped." None of...

That explains a thing or two, thank you. But people are looking to do this during inference on various fine-tunings of SD 1.x. (How this makes any sense mathematically, I...

oh, I guess the other consideration here is that we're talking about the CLIPTextModel. So different diffusion "models" may be different diffusion UNets, but all referring to the same CLIPTextModel,...

Does this only cover the one weights file, or does it cover accompanying files like `config.json` or `vocab.json`? sha256 is slow for multi-gigabyte files. Consider https://github.com/escherba/python-metrohash or https://github.com/oconnor663/blake3-py (b3sum) or...

> Regarding sha256 being slow. I don't think this is fully related since we don't need to compute the sha256 - it's already computed when models are uploaded to the...