Algorithm and Hardness for Dynamic Attention Maintenance in Large Language Models

Jan Van Den Brand, Zhao Song, Tianyi Zhou
Proceedings of the 41st International Conference on Machine Learning, PMLR 235:49008-49028, 2024.

Abstract

The attention scheme is one of the key components over all the LLMs, such as BERT, GPT-1, Transformers, GPT-2, 3, 3.5 and 4. Inspired by previous theoretical study of static version of the attention multiplication problem [Zandieh, Han, Daliri, and Karbasi ICML 2023, Alman and Song NeurIPS 2023], we formally define a dynamic version of attention matrix multiplication problem. In each iteration we update one entry in key matrix K∈Rn×d or value matrix V∈Rn×d. In the query stage, we receive (i,j)∈[n]×[d] as input, and want to answer (D−1AV)i,j, where A:=exp(QK⊤)∈Rn×n is a square matrix and D:=diag(A1n)∈Rn×n is a diagonal matrix and 1n denotes a length-n vector that all the entries are ones. We provide two results: an algorithm and a conditional lower bound. Inspired by the lazy update idea from [Demetrescu and Italiano FOCS 2000, Sankowski FOCS 2004, Cohen, Lee and Song STOC 2019, Brand SODA 2020], we provide a data-structure that uses O(nω(1,1,τ)−τ) amortized update time, and O(n1+τ) worst-case query time, where nω(1,1,τ) denotes (n,n,nτ) with matrix multiplication exponent ω and τ denotes a constant in (0,1]. We also show that unless the hinted matrix vector multiplication conjecture [Brand, Nanongkai and Saranurak FOCS 2019] is false, there is no algorithm that can use both O(nω(1,1,τ)−τ−Ω(1)) amortized update time, and O(n1+τ−Ω(1)) worst query time.

Cite this Paper

BibTeX
@InProceedings{pmlr-v235-van-den-brand24a, title = {Algorithm and Hardness for Dynamic Attention Maintenance in Large Language Models}, author = {Van Den Brand, Jan and Song, Zhao and Zhou, Tianyi}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {49008--49028}, year = {2024}, editor = {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix}, volume = {235}, series = {Proceedings of Machine Learning Research}, month = {21--27 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v235/main/assets/van-den-brand24a/van-den-brand24a.pdf}, url = {https://proceedings.mlr.press/v235/van-den-brand24a.html}, abstract = {The attention scheme is one of the key components over all the LLMs, such as BERT, GPT-1, Transformers, GPT-2, 3, 3.5 and 4. Inspired by previous theoretical study of static version of the attention multiplication problem [Zandieh, Han, Daliri, and Karbasi ICML 2023, Alman and Song NeurIPS 2023], we formally define a dynamic version of attention matrix multiplication problem. In each iteration we update one entry in key matrix $K \in \mathbb{R}^{n \times d}$ or value matrix $V \in \mathbb{R}^{n \times d}$. In the query stage, we receive $(i,j) \in [n] \times [d]$ as input, and want to answer $(D^{-1} A V)_{i,j}$, where $A:=\exp(QK^\top) \in \mathbb{R}^{n \times n}$ is a square matrix and $D := \mathrm{diag}(A {\bf 1}_n) \in \mathbb{R}^{n \times n}$ is a diagonal matrix and ${\bf 1}_n$ denotes a length-$n$ vector that all the entries are ones. We provide two results: an algorithm and a conditional lower bound. Inspired by the lazy update idea from [Demetrescu and Italiano FOCS 2000, Sankowski FOCS 2004, Cohen, Lee and Song STOC 2019, Brand SODA 2020], we provide a data-structure that uses $O(n^{\omega(1,1,\tau)-\tau})$ amortized update time, and $O(n^{1+\tau})$ worst-case query time, where $n^{\omega(1,1,\tau)}$ denotes $\mathrm(n,n,n^\tau)$ with matrix multiplication exponent $\omega$ and $\tau$ denotes a constant in $(0,1]$. We also show that unless the hinted matrix vector multiplication conjecture [Brand, Nanongkai and Saranurak FOCS 2019] is false, there is no algorithm that can use both $O(n^{\omega(1,1,\tau) - \tau- \Omega(1)})$ amortized update time, and $O(n^{1+\tau-\Omega(1)})$ worst query time.} }
Endnote
%0 Conference Paper %T Algorithm and Hardness for Dynamic Attention Maintenance in Large Language Models %A Jan Van Den Brand %A Zhao Song %A Tianyi Zhou %B Proceedings of the 41st International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2024 %E Ruslan Salakhutdinov %E Zico Kolter %E Katherine Heller %E Adrian Weller %E Nuria Oliver %E Jonathan Scarlett %E Felix Berkenkamp %F pmlr-v235-van-den-brand24a %I PMLR %P 49008--49028 %U https://proceedings.mlr.press/v235/van-den-brand24a.html %V 235 %X The attention scheme is one of the key components over all the LLMs, such as BERT, GPT-1, Transformers, GPT-2, 3, 3.5 and 4. Inspired by previous theoretical study of static version of the attention multiplication problem [Zandieh, Han, Daliri, and Karbasi ICML 2023, Alman and Song NeurIPS 2023], we formally define a dynamic version of attention matrix multiplication problem. In each iteration we update one entry in key matrix $K \in \mathbb{R}^{n \times d}$ or value matrix $V \in \mathbb{R}^{n \times d}$. In the query stage, we receive $(i,j) \in [n] \times [d]$ as input, and want to answer $(D^{-1} A V)_{i,j}$, where $A:=\exp(QK^\top) \in \mathbb{R}^{n \times n}$ is a square matrix and $D := \mathrm{diag}(A {\bf 1}_n) \in \mathbb{R}^{n \times n}$ is a diagonal matrix and ${\bf 1}_n$ denotes a length-$n$ vector that all the entries are ones. We provide two results: an algorithm and a conditional lower bound. Inspired by the lazy update idea from [Demetrescu and Italiano FOCS 2000, Sankowski FOCS 2004, Cohen, Lee and Song STOC 2019, Brand SODA 2020], we provide a data-structure that uses $O(n^{\omega(1,1,\tau)-\tau})$ amortized update time, and $O(n^{1+\tau})$ worst-case query time, where $n^{\omega(1,1,\tau)}$ denotes $\mathrm(n,n,n^\tau)$ with matrix multiplication exponent $\omega$ and $\tau$ denotes a constant in $(0,1]$. We also show that unless the hinted matrix vector multiplication conjecture [Brand, Nanongkai and Saranurak FOCS 2019] is false, there is no algorithm that can use both $O(n^{\omega(1,1,\tau) - \tau- \Omega(1)})$ amortized update time, and $O(n^{1+\tau-\Omega(1)})$ worst query time.
APA
Van Den Brand, J., Song, Z. & Zhou, T.. (2024). Algorithm and Hardness for Dynamic Attention Maintenance in Large Language Models. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:49008-49028 Available from https://proceedings.mlr.press/v235/van-den-brand24a.html.

Related Material