Llama.cpp speculative sampling: 2x faster inference for large models (github.com)

Small caveat: this is only true for generating text with simple grammar, like code. For human text this doesn't work so well.