SDFP: Speculative Decoding with FIT-Pruned Models for Training-Free and Plug-and-Play LLM Acceleration

TL;DR

SDFP introduces a training-free, plug-and-play method for accelerating large language model decoding by pruning layers based on Fisher Information Trace, achieving significant speedups without retraining or complex tuning.

Contribution

It presents a novel, training-free layer pruning approach using Fisher Information Trace to create efficient draft models for speculative decoding.

Findings

Achieves 1.32x-1.5x decoding speedup on benchmarks.

No additional training or hyperparameter tuning required.

Maintains output distribution fidelity of the original model.

Abstract

Large language models (LLMs) underpin interactive multimedia applications such as captioning, retrieval, recommendation, and creative content generation, yet their autoregressive decoding incurs substantial latency. Speculative decoding reduces latency using a lightweight draft model, but deployment is often limited by the cost and complexity of acquiring, tuning, and maintaining an effective draft model. Recent approaches usually require auxiliary training or specialization, and even training-free methods incur costly search or optimization. We propose SDFP, a fully training-free and plug-and-play framework that builds the draft model via Fisher Information Trace (FIT)-based layer pruning of a given LLM. Using layer sensitivity as a proxy for output perturbation, SDFP removes low-impact layers to obtain a compact draft while preserving compatibility with the original model for standard speculative verification. SDFP needs no additional training, hyperparameter tuning, or separately maintained drafts, enabling rapid, deployment-friendly draft construction. Across benchmarks, SDFP delivers 1.32x-1.5x decoding speedup without altering the target model's output distribution, supporting low-latency multimedia applications.