CREST: Effectively Compacting a Datastore For Retrieval-Based Speculative Decoding

TL;DR

CREST is a redesigned datastore for speculative decoding that stores a selective subset of n-grams, reducing storage needs while maintaining or improving decoding performance on benchmark tasks.

Contribution

CREST introduces a method to store only the most common and smallest n-grams, making the datastore more compact and efficient without sacrificing accuracy.

Findings

CREST reduces storage space by 10.6-13.5x compared to REST.

CREST achieves 16.5-17.1% higher acceptance length than REST.

CREST maintains performance while significantly decreasing storage requirements.

Abstract

We present CREST (Compact Retrieval-Based Speculative Decoding), a redesign of REST that allows it to be effectively "compacted". REST is a drafting technique for speculative decoding based on retrieving exact n-gram matches of the most recent n tokens generated by the target LLM from a datastore. The key idea of CREST is to only store a subset of the smallest and most common n-grams in the datastore with the hope of achieving comparable performance with less storage space. We found that storing a subset of n-grams both reduces storage space and improves performance. CREST matches REST's accepted token length with 10.6-13.5x less storage space and achieves a 16.5-17.1% higher acceptance length than REST using the same storage space on the HumanEval and MT Bench benchmarks.