Llama Surgery: Continuous Sparsification of Pre-Trained Language Models via Differentiable Ultrametric Topology Injection(reddit.com)