FutureFill: Fast Generation from Convolutional Sequence Models

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Agarwal, Naman, Chen, Xinyi, Dogariu, Evan, Shah, Devan, Strauss, Hubert, Feinberg, Vlad, Suo, Daniel, Bartlett, Peter, Hazan, Elad
Format: Preprint
Publié: 2024
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866913905771544576
author Agarwal, Naman
Chen, Xinyi
Dogariu, Evan
Shah, Devan
Strauss, Hubert
Feinberg, Vlad
Suo, Daniel
Bartlett, Peter
Hazan, Elad
author_facet Agarwal, Naman
Chen, Xinyi
Dogariu, Evan
Shah, Devan
Strauss, Hubert
Feinberg, Vlad
Suo, Daniel
Bartlett, Peter
Hazan, Elad
contents We address the challenge of efficient auto-regressive generation in sequence prediction models by introducing FutureFill, a general-purpose fast generation method for any sequence prediction algorithm based on convolutional operators. FutureFill reduces generation time from quadratic to quasilinear in the context length. Moreover, when generating from a prompt, it requires a prefill cache whose size grows only with the number of tokens to be generated, often much smaller than the caches required by standard convolutional or attention based models. We validate our theoretical claims with experiments on synthetic tasks and demonstrate substantial efficiency gains when generating from a deep convolutional sequence prediction model.
format Preprint
id arxiv_https___arxiv_org_abs_2410_03766
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle FutureFill: Fast Generation from Convolutional Sequence Models
Agarwal, Naman
Chen, Xinyi
Dogariu, Evan
Shah, Devan
Strauss, Hubert
Feinberg, Vlad
Suo, Daniel
Bartlett, Peter
Hazan, Elad
Machine Learning
Artificial Intelligence
Computation and Language
We address the challenge of efficient auto-regressive generation in sequence prediction models by introducing FutureFill, a general-purpose fast generation method for any sequence prediction algorithm based on convolutional operators. FutureFill reduces generation time from quadratic to quasilinear in the context length. Moreover, when generating from a prompt, it requires a prefill cache whose size grows only with the number of tokens to be generated, often much smaller than the caches required by standard convolutional or attention based models. We validate our theoretical claims with experiments on synthetic tasks and demonstrate substantial efficiency gains when generating from a deep convolutional sequence prediction model.
title FutureFill: Fast Generation from Convolutional Sequence Models
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2410.03766