Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning

Chaofan Lin, Jiaming Tang, Shuo Yang 0011, Hanshuo Wang, Tian Tang 0001, Boyu Tian, Ion Stoica, Song Han 0003, Mingyu Gao 0001. Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning. In Danielle Belgrave, Cheng Zhang 0005, Laura N. Montoya, Hsuan-Tien Lin, Razvan Pascanu, Piotr Koniusz, Marzyeh Ghassemi, Nancy Chen, Iván Vladimir Meza Ruíz, Arturo Loaiza-Bonilla, editors, Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2025, NeurIPS 2025, San Diago, CA, USA, December 2-7, 2025 / Mexico City, Mexico, November 30 - December 5, 2025. 2025. [doi]

Authors

Chaofan Lin

This author has not been identified. Look up 'Chaofan Lin' in Google

Jiaming Tang

This author has not been identified. Look up 'Jiaming Tang' in Google

Shuo Yang 0011

This author has not been identified. Look up 'Shuo Yang 0011' in Google

Hanshuo Wang

This author has not been identified. Look up 'Hanshuo Wang' in Google

Tian Tang 0001

This author has not been identified. Look up 'Tian Tang 0001' in Google

Boyu Tian

This author has not been identified. Look up 'Boyu Tian' in Google

Ion Stoica

This author has not been identified. Look up 'Ion Stoica' in Google

Song Han 0003

This author has not been identified. Look up 'Song Han 0003' in Google

Mingyu Gao 0001

This author has not been identified. Look up 'Mingyu Gao 0001' in Google