An attention memory-management algorithm that organizes the key-value cache in blocks for efficient model serving.
Sign in to savePagedAttention applies ideas from virtual memory and paging to the key-value cache used during language-model generation. It reduces fragmentation and supports cache sharing, helping a serving system use memory more effectively. The original paper introduces vLLM as a serving system built around this algorithm.
Research paper first submitted to arXiv on 2023-09-12.
Relevant to inference infrastructure and serving throughput. Distinguish the memory-management technique from model architecture changes or claims about answer quality. Catalog note: this is a source-verified reference entry, not a measured current-trend signal. No search-volume or growth claim is made.
Pro opens the same-day prioritize / watch / skip table, including names that are still too new to publish here.
This name has not appeared in an imported daily report yet.