Faster long-context AI by reusing attention computations across layers

Original: You Only Index Once: Cross-Layer Sparse Attention with Shared Routing

Source: arxiv.org ↗