less-is-more-sparse-attention

【效率与协作】Improves inference efficiency through training-free sparse attention using global token selection patterns aggregated from local attention heads for unified cross-head ranking.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准