post-training-pruning

PALS adjusts per-layer sparsity ratios for LLM pruning based on the 99th percentile of activation magnitudes, achieving significant perplexity improvements on LLaMA-2-7B compared to uniform sparsity, with negligible added cost.