BESA is a differentiable weight pruning technique for large language models. - View it on GitHub