Enregistré dans:
Détails bibliographiques
Auteurs principaux: Key, Oscar, Ribar, Luka, Cattaneo, Alberto, Hudlass-Galley, Luke, Orr, Douglas
Format: Preprint
Publié: 2024
Sujets:
Accès en ligne:https://arxiv.org/abs/2412.04358
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866912145781817344
author Key, Oscar
Ribar, Luka
Cattaneo, Alberto
Hudlass-Galley, Luke
Orr, Douglas
author_facet Key, Oscar
Ribar, Luka
Cattaneo, Alberto
Hudlass-Galley, Luke
Orr, Douglas
contents We present an evaluation of bucketed approximate top-$k$ algorithms. Computing top-$k$ exactly suffers from limited parallelism, because the $k$ largest values must be aggregated along the vector, thus is not well suited to computation on highly-parallel machine learning accelerators. By relaxing the requirement that the top-$k$ is exact, bucketed algorithms can dramatically increase the parallelism available by independently computing many smaller top-$k$ operations. We explore the design choices of this class of algorithms using both theoretical analysis and empirical evaluation on downstream tasks. Our motivating examples are sparsity algorithms for language models, which often use top-$k$ to select the most important parameters or activations. We also release a fast bucketed top-$k$ implementation for PyTorch.
format Preprint
id arxiv_https___arxiv_org_abs_2412_04358
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Approximate Top-$k$ for Increased Parallelism
Key, Oscar
Ribar, Luka
Cattaneo, Alberto
Hudlass-Galley, Luke
Orr, Douglas
Machine Learning
We present an evaluation of bucketed approximate top-$k$ algorithms. Computing top-$k$ exactly suffers from limited parallelism, because the $k$ largest values must be aggregated along the vector, thus is not well suited to computation on highly-parallel machine learning accelerators. By relaxing the requirement that the top-$k$ is exact, bucketed algorithms can dramatically increase the parallelism available by independently computing many smaller top-$k$ operations. We explore the design choices of this class of algorithms using both theoretical analysis and empirical evaluation on downstream tasks. Our motivating examples are sparsity algorithms for language models, which often use top-$k$ to select the most important parameters or activations. We also release a fast bucketed top-$k$ implementation for PyTorch.
title Approximate Top-$k$ for Increased Parallelism
topic Machine Learning
url https://arxiv.org/abs/2412.04358