Hi there, love the project!
Are there plans to support timestep-level caching techniques such as TeaCache, MagCache, TaoCache, EasyCache, Chipmunk, and more? These methods allow for skipping entire diffusion timesteps, which would complement the existing block-level caching beautifully.
There is debate about supporting timestep-level caching directly in sglang sgl-project/sglang#18498.
Thanks!