Utility-Driven Speculative Decoding for Mixture-of-Experts

TL;DR

This paper introduces Cascade, a dynamic framework that selectively enables speculative decoding in Mixture-of-Experts models, significantly improving throughput and avoiding slowdowns caused by naive speculation methods.

Contribution

Cascade is a utility-driven, dynamic approach that tunes speculation parameters in MoE models, making speculative decoding practical and efficient.

Findings

Cascade reduces slowdowns to 5% from 1.5x.

It improves throughput by 7-14% over static K.

Cascade effectively adapts to different tasks and models.

Abstract

GPU memory bandwidth is the main bottleneck for low-latency Large Language Model (LLM) inference. Speculative decoding leverages idle GPU compute by using a lightweight drafter to propose K tokens, which the LLM verifies in parallel, boosting token throughput. In conventional dense LLMs, all model weights are fetched each iteration, so speculation adds no latency overhead. Emerging Mixture of Experts (MoE) models activate only a subset of weights per token, greatly reducing data movement. However, we show that speculation is ineffective for MoEs: draft tokens collectively activate more weights, increasing data movement and verification time by 2-3x. When token throughput gains fail to offset this overhead, speculation causes slowdowns up to 1.5x, making it infeasible. Even when useful, the optimal K varies by task, model, and even between requests and iterations. Thus, despite widespread use in dense LLMs, speculation remains impractical in leading MoEs. We present Cascade, a utility-driven framework that selectively enables speculation to avoid slowdowns and dynamically tunes K to accelerate MoE serving. Cascade uses a lightweight metric, speculation utility, the ratio of token gains to verification cost, which shows iteration-level locality, enabling periodic decisions via short test and longer set phases. For each request, Cascade disables speculation if utility drops below one during testing, and when utility exceeds one, tests multiple K-values to choose the utility-maximizing K for the set phase. We implement Cascade in vLLM and evaluate it on five popular MoEs with workloads spanning code, math, extraction, and mixed tasks. Cascade limits slowdown to 5% (vs. 1.5x) and improves throughput by 7-14% over static K, making speculative decoding practical for MoEs.