Diffusion-model (LLaDA) support for Llama.cpp?

[kannony]

Mercury Coder is a diffusion-model based LLM which its creators claim offers huge efficiency improvements over standard LLMs. Is there any possibility that llama.cpp will add support for it?

https://huggingface.co/spaces/multimodalart/LLaDA
https://ml-gsai.github.io/LLaDA-demo/
https://github.com/ML-GSAI/LLaDA

[jhowilbur]

Hey community
A good starting point could be a mix with this StableDiffusionGGML

Here we have a good explanation to help you run the python version locally
https://www.youtube.com/watch?v=TIyUD0uJcsM

[andrewkchan]

This is newly relevant with the release of Gemini Diffusion yesterday: https://simonwillison.net/2025/May/21/gemini-diffusion/

I might be interested in exploring this. Wrapping up a quantization PR for Iwan's fork (ikawrakow/ik_llama.cpp#441) and am interested in what it would take to add support for a whole new architecture to llama.cpp (although I can't guarantee immediate bandwidth).