SEED: Accelerating Reasoning Tree Construction via Scheduled Speculative Decoding

TL;DR

This paper presents SeeD, an innovative inference framework that accelerates reasoning tree construction in large language models by using scheduled speculative decoding, significantly reducing inference latency and memory usage.

Contribution

SeeD introduces a rounds-scheduled speculative execution strategy to optimize reasoning tree inference in LLMs, enhancing speed and efficiency.

Findings

Achieves significant speedup over baseline methods

Reduces GPU memory consumption during inference

Demonstrates effectiveness across multiple reasoning datasets

Abstract

Large Language Models (LLMs) demonstrate remarkable emergent abilities across various tasks, yet fall short of complex reasoning and planning tasks. The tree-search-based reasoning methods address this by surpassing the capabilities of chain-of-thought prompting, encouraging exploration of intermediate steps. However, such methods introduce significant inference latency due to the systematic exploration and evaluation of multiple thought paths. This paper introduces SeeD, a novel and efficient inference framework to optimize runtime speed and GPU memory management concurrently. By employing a scheduled speculative execution, SeeD efficiently handles multiple iterations for the thought generation and the state evaluation, leveraging a rounds-scheduled strategy to manage draft model dispatching. Extensive experimental evaluations on three reasoning datasets demonstrate superior speedup performance of SeeD, providing a viable path for batched inference in training-free speculative decoding.