SEED: Accelerating Reasoning Tree Construction via Scheduled Speculative Decoding
This work addresses the problem of slow inference in complex reasoning tasks for LLM users, offering an incremental improvement in efficiency.
The paper tackles the high inference latency of tree-search-based reasoning methods for LLMs by introducing SeeD, a framework that uses scheduled speculative decoding to accelerate reasoning tree construction, achieving significant speedups on three reasoning datasets.
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.