Comment by beacon294

Try the llama.cpp fork by thetom. It's called turboquant after the technique