Add LCM scheduler

[wtedw]

This is a simple version of LCMScheduler that supports txt2img.
Here's a livebook that makes use of the scheduler:

• https://gist.github.com/wtedw/8af037108bf2371af6dcd24080e8e66b

To support LCM-LoRA natively in Axon + Bumblebee, we'd need these PRs

• Load lora weights from Bumblebee
• Add LoRA support to Axon? or keep separate. Right now it's using my own library to load LoRA .safetensors

also, seed + strength should be forwarded to the scheduler_init function, but I'll create a separate PR for this.

[josevalim]

I believe @seanmor5 wants LoRA support in Axon. :)

[jonatanklosko]

@wtedw thanks a lot for the PR! I'm doing minor some adjustments and docs, will push and merge tomorrow :)

seed + strength should be forwarded to the scheduler_init function

Good call, we should pass the prng key when calling any of the schedulers, I will change it afterwards. The strength is only relevant for img2img, so we can wait until we support it before we expose it.

[wtedw]

@jonatanklosko , Awesome thanks!

[jonatanklosko]

Btw. we will need a separate serving for LCM Diffusion, corresponding to LatentConsistencyModelPipeline in hf/diffusers. There are some differences from Stable Diffusion, in particular LCM doesn't really support negative prompt and the guidance scale is incorporated into the U-Net forward pass as additional timestamp embeddings (which is likely why using different guiding scales didn't work very well for you). But we will get to this once we support loading LoRA!

[wtedw]

@jonatanklosko Ah that's some good info. Thanks for looking into it!