October 26, 2021
The quadratic computational and memory complexities of the Transformer’s at-tention mechanism have limited its scalability for modeling long sequences. Inthis paper, we propose Luna, a linear unified nested attention mechanism thatapproximates softmax attention withtwo nested linear attention functions, yieldingonly linear (as opposed to quadratic) time and space complexity. As compared toa more traditional attention mechanism, Luna introduces an additional sequencewith a fixed length as input and an additional corresponding output, which allowsLuna to perform attention operation linearly, while also storing adequate contextualinformation. We perform extensive evaluations on three benchmarks of sequencemodeling tasks: long-context sequence modeling, neural machine translation andmasked language modeling for large-scale pretraining. Competitive or even betterexperimental results demonstrate both the effectiveness and efficiency of Lunacompared to a variety of strong baseline methods including the full-rank attentionand other efficient sparse and dense attention methods. The implementation of ourmodel is available at https://github.com/XuezheMax/fairseq-apollo
Publisher
NeurIPS
Research Topics
May 14, 2025
Linnea Evanson, Christine Bulteau, Mathilde Chipaux, Georg DorfmĂĽller, Sarah Ferrand-Sorbets, Emmanuel Raffo, Sarah Rosenberg, Pierre Bourdillon, Jean Remi King
May 14, 2025
April 25, 2025
Rulin Shao, Qiao Rui, Varsha Kishore, Niklas Muennighoff, Victoria Lin, Daniela Rus, Bryan Kian Hsiang Low, Sewon Min, Scott Yih, Pang Wei Koh, Luke Zettlemoyer
April 25, 2025
April 17, 2025
Ansong Ni, Ruta Desai, Yang Li, Xinjie Lei, Dong Wang, Ramya Raghavendra, Gargi Ghosh, Daniel Li (FAIR), Asli Celikyilmaz
April 17, 2025
April 04, 2025
Olga Golovneva, Tianlu Wang, Jason Weston, Sainbayar Sukhbaatar
April 04, 2025
Our approach
Latest news
Foundational models