Tag
This article introduces an experimental sampler for llama.cpp called 'scatter' that applies a local diffusion over rank-sorted tokens to smooth the probability distribution among top candidates, enabling more creative text generation without increasing tail probabilities.