>>105928363
anon was curious. IQ4_XS is the minimum to get at least something but even then it's horrid shit. still trying to come up with a game plan for llama and sdcpp to use the same ggml backend and the best way to keep them synced. should I just make a fork with both as submodules and tie them together on the same backend? feels like llama would always be a bit behind so sdcpp can keep working