Post-training attention sparsification reduces the quadratic cumulative attention cost of pretrained Transformers by selecting a small set of context units (…
机构:腾讯
来源:arXiv 2609.13141 | AI4Papers 论文推荐平台