Multi-Draft Speculative Sampling: Canonical Decomposition and Theoretical Limits

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Khisti, Ashish, Ebrahimi, M. Reza, Dbouk, Hassan, Behboodi, Arash, Memisevic, Roland, Louizos, Christos
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913827748052992
author Khisti, Ashish
Ebrahimi, M. Reza
Dbouk, Hassan
Behboodi, Arash
Memisevic, Roland
Louizos, Christos
author_facet Khisti, Ashish
Ebrahimi, M. Reza
Dbouk, Hassan
Behboodi, Arash
Memisevic, Roland
Louizos, Christos
contents We consider multi-draft speculative sampling, where the proposal sequences are sampled independently from different draft models. At each step, a token-level draft selection scheme takes a list of valid tokens as input and produces an output token whose distribution matches that of the target model. Previous works have demonstrated that the optimal scheme (which maximizes the probability of accepting one of the input tokens) can be cast as a solution to a linear program. In this work we show that the optimal scheme can be decomposed into a two-step solution: in the first step an importance sampling (IS) type scheme is used to select one intermediate token; in the second step (single-draft) speculative sampling is applied to generate the output token. For the case of two identical draft models we further 1) establish a necessary and sufficient condition on the distributions of the target and draft models for the acceptance probability to equal one and 2) provide an explicit expression for the optimal acceptance probability. Our theoretical analysis also motives a new class of token-level selection schemes based on weighted importance sampling. Our experimental results demonstrate consistent improvements in the achievable block efficiency and token rates over baseline schemes in a number of scenarios.
format Preprint
id arxiv_https___arxiv_org_abs_2410_18234
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Multi-Draft Speculative Sampling: Canonical Decomposition and Theoretical Limits
Khisti, Ashish
Ebrahimi, M. Reza
Dbouk, Hassan
Behboodi, Arash
Memisevic, Roland
Louizos, Christos
Computation and Language
Distributed, Parallel, and Cluster Computing
Information Theory
Machine Learning
We consider multi-draft speculative sampling, where the proposal sequences are sampled independently from different draft models. At each step, a token-level draft selection scheme takes a list of valid tokens as input and produces an output token whose distribution matches that of the target model. Previous works have demonstrated that the optimal scheme (which maximizes the probability of accepting one of the input tokens) can be cast as a solution to a linear program. In this work we show that the optimal scheme can be decomposed into a two-step solution: in the first step an importance sampling (IS) type scheme is used to select one intermediate token; in the second step (single-draft) speculative sampling is applied to generate the output token. For the case of two identical draft models we further 1) establish a necessary and sufficient condition on the distributions of the target and draft models for the acceptance probability to equal one and 2) provide an explicit expression for the optimal acceptance probability. Our theoretical analysis also motives a new class of token-level selection schemes based on weighted importance sampling. Our experimental results demonstrate consistent improvements in the achievable block efficiency and token rates over baseline schemes in a number of scenarios.
title Multi-Draft Speculative Sampling: Canonical Decomposition and Theoretical Limits
topic Computation and Language
Distributed, Parallel, and Cluster Computing
Information Theory
Machine Learning
url https://arxiv.org/abs/2410.18234