https://github.com/IceFog72/llama.cpp I added an experimental sampler to llama.cpp called scatter. The short version: it slightly smooths the model’s next-token probability distribution inside the already-selected top candidates. It is meant to make generation less rigid without doing the usual “rai
