FR-Spec: Accelerating Large-Vocabulary Language Models via Frequency-Ranked Speculative Sampling

TL;DR

FR-Spec introduces a frequency-ranked speculative sampling method that significantly reduces computational overhead and accelerates large-vocabulary language model generation by prioritizing frequent tokens.

Contribution

It proposes a novel vocabulary compression technique for speculative sampling, improving efficiency for large-vocabulary LLMs without sacrificing output quality.

Findings

75% reduction in LM Head computation

Average 1.12× speedup over EAGLE-2

Maintains output distribution equivalence

Abstract

Speculative sampling has emerged as an important technique for accelerating the auto-regressive generation process of large language models (LLMs) by utilizing a draft-then-verify mechanism to produce multiple tokens per forward pass. While state-of-the-art speculative sampling methods use only a single layer and a language modeling (LM) head as the draft model to achieve impressive layer compression, their efficiency gains are substantially reduced for large-vocabulary LLMs, such as Llama-3-8B with a vocabulary of 128k tokens. To address this, we present FR-Spec, a frequency-ranked speculative sampling framework that optimizes draft candidate selection through vocabulary space compression. By constraining the draft search to a frequency-prioritized token subset, our method reduces LM Head computation overhead by 75% while ensuring the equivalence of the final output distribution. Experiments across multiple datasets demonstrate an average of 1.12$\times$ speedup over the state-of-the-art speculative sampling method EAGLE-2. Code available at https://github.com/thunlp/FR-Spec.