arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2602.16052cs.LG
收藏

MoE-Spec:专家预算用于高效推测解码

MoE-Spec: Expert Budgeting for Efficient Speculative Decoding

Franklin and Marshall College(弗兰克林与马歇尔学院);Meta Reality Labs(Meta现实实验室)

Bradley McDanel, Steven Li, Sruthikesh Surineni, Harshit Khaitan

AI总结:

MoE-Spec通过专家预算管理提升推测解码效率,在保持质量的同时提高吞吐量并灵活调整精度与延迟。

摘要:

推测解码通过并行验证多个草稿标记来加速大型语言模型(LLM)的推理过程。然而,对于混合专家(MoE)模型,这种并行性引入了一个严重的瓶颈:大规模的草稿树会激活许多独特的专家,显著增加内存压力,并削弱推测解码相对于自回归解码的速度提升。先前的方法在MoE验证变得昂贵时减少推测深度。我们提出了MoE-Spec,一种无需训练的验证时间专家预算方法,通过在每一层强制执行固定专家容量限制,将推测深度与内存成本解耦。该方法只加载对验证最有贡献的专家,并丢弃很少使用的专家的长尾,这些专家会驱动带宽开销。在多个模型规模和数据集上的实验表明,该方法在与现有最佳推测解码基线(EAGLE-3)相当的质量下,能够实现10-30%的吞吐量提升,并且可以通过更紧的预算灵活地在精度和进一步的延迟减少之间进行权衡。

英文摘要:

Speculative decoding accelerates Large Language Model (LLM) inference by verifying multiple drafted tokens in parallel. However, for Mixture-of-Experts (MoE) models, this parallelism introduces a severe bottleneck: large draft trees activate many unique experts, significantly increasing memory pressure and diminishing speedups from speculative decoding relative to autoregressive decoding. Prior methods reduce speculation depth when MoE verification becomes expensive. We propose MoE-Spec, a training-free verification-time expert budgeting method that decouples speculation depth from memory cost by enforcing a fixed expert capacity limit at each layer, loading only the experts that contribute most to verification and dropping the long tail of rarely used experts that drive bandwidth overhead. Experiments across multiple model scales and datasets show that this method yields 10--30\% higher throughput than state-of-the-art speculative decoding baselines (EAGLE-3) at comparable quality, with flexibility to trade accuracy for further latency reductions through tighter budgets.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑