Sinkhorn: Make LLMs even smaller through quantisation while maintaining accuracy (github.com)