[dofm]

To briefly follow up, as of yesterday llama.cpp can do Gemma 4's MTP, so I have this working at least initially — details here:

https://news.ycombinator.com/item?id=48441450