1
SpotAttention: Plug-In Block-Sparse Routing for Pretrained Long-Context Transformers
SpotAttention以即插即用的块稀疏路由，高效解决长上下文Transformer的推理成本难题，值得AI研究者关注。
arXiv:2606.22874v1 Announce Type: cross Abstract: Long contexts have become standard in pretrained LLMs, yet they remain expensive to run: prefill com…