Massive KV caches can cause severe memory-bandwidth bottlenecks during long-context decoding. Sparse attention methods mitigate this via selective loading, b…
机构:Google
来源:arXiv 2609.20888 | AI4Papers 论文推荐平台