EdgeSpec: Distributed Speculative Decoding for Large Language Models at Edge

Abstract

The growing demand for deploying large language models (LLMs) in applications requiring low latency and strong privacy has brought increasing attention to edge inference. While distributing computation across edge devices helps alleviate resource bottlenecks, the sequential and layer-wise nature of autoregressive decoding still limits real-time deployment, leading to device underutilization, increased latency, and workload imbalance. Existing methods rely on heavyweight drafts and static policies, which cannot adapt to device heterogeneity and bandwidth variation, leading to limited parallelism and elevated latency. To bridge this gap, we propose EdgeSpec, a hierarchical speculative decoding framework at edges. Instead of relying on last-layer speculation, EdgeSpec enables intermediate transformer layers to emit draft tokens, which are validated and used to predict the next token in parallel across devices via attention masking. The system combines offline entropy profiling with an online Kalman filter to adaptively select confident trigger layers based on runtime load and input complexity. To further improve efficiency, EdgeSpec jointly optimizes model partitioning, entropy thresholds, and candidate sizes through a two-stage strategy that integrates offline planning with online adaptation. Experiments on LLaMA2 showed that EdgeSpec reduced inference latency by up to 17.1× while preserving output quality comparable to standard decoding.