Polybasic Speculative Decoding Under a Theoretical Perspective

Ruilin Wang, Huixia Li, Yuexiao Ma, Xiawu Zheng, Fei Chao, Xuefeng Xiao, Rongrong Ji

Keywords: Speculative Decoding; LLM Inference

Abstract:

Speculative decoding has emerged as a critical technique for accelerating inference in large language models, achieving significant speedups while ensuring consistency with the outputs of the original models.
However, there is currently a lack of theoretical guidance in speculative decoding.
As a result, most existing works are dualistic target-draft model paradigm, which significantly restricts the hinders potential application scenarios.
In this paper, we propose a polybasic speculative decoding framework supported by a solid theoretical foundation.
We first deduce a theorem to control the ideal inference time of speculative decoding systems which is then serve as a design criterion that effectively expands the original dualistic speculative decoding into a more efficient polybasic speculative decoding.
We further theoretically analyze the sampling process, identifying variables that can be optimized to enhance inference efficiency in multi-model systems.
We demonstrate, both theoretically and empirically, that this system accelerates inference for the target model, and that our approach is orthogonal to the majority of existing speculative methods, allowing for independent application or combination with other techniques.
Experimentally, we conducted comprehensive evaluations across a wide range of models, including those from the Vicuna, LLaMA2-Chat, and LLaMA3 families.
Our method achieved remarkable latency speedup ratios of $\textbf{3.31$\times$-4.01$\times$}$ for LLaMA2-Chat 7B, up to $\textbf{3.87$\times$}$ for LLaMA3-8B, and up to $\textbf{4.43$\times$}$ for Vicuna-7B, while maintaining the distribution of the generated text. Code is available in supplementary materials.

Primary Area: foundation or frontier models, including LLMs

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 322