Speculative Sampling via Exponential Races

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Kobus, Szymon, Gündüz, Deniz
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866909588222115840
author Kobus, Szymon
Gündüz, Deniz
author_facet Kobus, Szymon
Gündüz, Deniz
contents Speculative decoding accelerates large language model inference using a smaller draft model. In this paper, we establish a surprising connection between speculative decoding and channel simulation, which aims at simulating a noisy channel using as few bits as possible. This connection allows us to provide an information-theoretic analysis of the speed up that can be achieved by speculative decoding. Leveraging this link, we derive an explicit relation between generation speed-up and the number of tokens $k$ generated by the draft model for large $k$, which serves as an upper bound for all $k$. We also propose a novel speculative decoding method via exponential race ERSD that matches state-of-the-art performance.
format Preprint
id arxiv_https___arxiv_org_abs_2504_15475
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Speculative Sampling via Exponential Races
Kobus, Szymon
Gündüz, Deniz
Computation and Language
Information Theory
Speculative decoding accelerates large language model inference using a smaller draft model. In this paper, we establish a surprising connection between speculative decoding and channel simulation, which aims at simulating a noisy channel using as few bits as possible. This connection allows us to provide an information-theoretic analysis of the speed up that can be achieved by speculative decoding. Leveraging this link, we derive an explicit relation between generation speed-up and the number of tokens $k$ generated by the draft model for large $k$, which serves as an upper bound for all $k$. We also propose a novel speculative decoding method via exponential race ERSD that matches state-of-the-art performance.
title Speculative Sampling via Exponential Races
topic Computation and Language
Information Theory
url https://arxiv.org/abs/2504.15475