CoSine: Enhancing LLM Serving via Collaborative and Decoupled Speculative Inference

摘要

Speculative inference accelerates Large Language Model (LLM) serving by leveraging small speculative models (SSMs) as drafters to generate draft tokens, which are subsequently verified in parallel by the LLM. However, the practical effectiveness of this approach is frequently hindered by the tight coupling of compute-intensive LLMs and memory-intensive SSMs within a single node (e.g., single-GPU device). This coupling induces resource contention, thereby significantly lowering draft acceptance rates and system throughput. To overcome these obstacles, we present CoSine, a novel speculative inference system that decouples sequential speculation from parallel verification. Specifically, CoSine distributes multiple specialized drafters across the heterogeneous nodes and coordinates them via a verification server, enabling cost-effective collaboration to mitigate inference bottlenecks. First, CoSine dynamically routes requests to the proper drafters and incorporates a confidence-based token fusion mechanism to synthesize high-quality drafts. Then, CoSine orchestrates the verification pipeline featuring prioritized draft queueing and adaptive batch management to balance the decoupled workflows and minimize pipeline bubbles. Experimental evaluation demonstrates that CoSine significantly outperforms state-of-the-art approaches. Under equivalent resource constraints, CoSine achieves latency reductions of up to 27.1% and throughput improvements from 1.24× to 1.62×.