[enhancement]: Lossless speed enhancement for Flux + SD3x

[iwr-redmond]

Is there an existing issue for this?

• [x] I have searched the existing issues

Contact Details

No response

What should this feature add?

This FR follows up on #7592, in which I suggested that ParaAttention was incompatible with Invoke due to its restrictive license.

Two novel algorithms for DiT based models have recently become available for Diffusers, and are well worth considering for inclusion in Invoke:

• TeaCache for Flux (pipeline)
• Microsoft RAS for SD3 (quickstart)

MSRAS is MIT-licensed, and TeaCache has an Apache 2.0 license that should be compatible with Invoke. MSRAS promises a speed-up of 1.6x with comparable generation quality (demo), while TeaCache supports a lossless speed-up of 1.4x (samples).

Both of the linked repositories are based on Diffusers and should be comparatively easy to integrate.

Alternatives

No response

Additional Content

No response

[hipsterusername]

We've reviewed both of these, and agree that they're compatible -- Our Flux implementation is not based on diffusers but the original BFL code.

We're open to contributions on both of these items