[bug]: SD 2.x models return only black in float16

[keturn]

• OS: Linux (but also reported on Windows)
• GPU: cuda

What happened?

When using a 768px Stable Diffusion 2.x model, the results are all black (zeros).

• full-precision works fine, only an issue when running at float16
• xformers works fine, only an issue when xformers is disabled (or not installed)
• the stable-diffusion-2.1-base (512 px) model works fine, only an issue on the full 768px v-prediction model.

Additional context

Likely related to huggingface/diffusers#1614

[keturn]

With #2349 applied, SD 2.1 also works with diffusers enable_attention_slicing, so I think the problem must be in the attention code that InvokeAI monkey-patches in.

[keturn]

Since merging #2385, I can now use stable-diffusion-2.1 with --no-xformers. Can anyone else confirm?

[keturn]

Fixed by #2385