Paged Attention

A more memory-efficient attention algorithm that partitions the KV cache into fixed-size blocks stored non-contiguously in GPU memory, thereby supporting larger batches and higher throughput.