Dynamic Speculation Lookahead Accelerates Speculative Decoding of Large Language Models

TL;DR

This paper introduces DISCO, a dynamic method for selecting speculation lookahead in large language models, significantly improving inference speed without sacrificing output quality.

Contribution

DISCO is the first approach to dynamically optimize speculation lookahead, outperforming static methods in large language model decoding.

Findings

Achieves 10% average speedup over static lookahead methods

Maintains identical text output to static methods

Demonstrates effectiveness across four datasets

Abstract

Speculative decoding is commonly used for reducing the inference latency of large language models. Its effectiveness depends highly on the speculation lookahead (SL)-the number of tokens generated by the draft model at each iteration. In this work we show that the common practice of using the same SL for all iterations (static SL) is suboptimal. We introduce DISCO (DynamIc SpeCulation lookahead Optimization), a novel method for dynamically selecting the SL. Our experiments with four datasets show that DISCO reaches an average speedup of 10% compared to the best static SL baseline, while generating the exact same text.