Update on the DwarfStar laguna-s2.1 branch, the GGUF files got updated, now the support for the file with Q8 projections and smaller routed experts (68GB total IIRC) is implemented as well. Speed improved significantly to 60 t/s generation, 550 t/s prefill.