aiqbench analysis
This research addresses a critical bottleneck in deploying large Mixture-of-Experts models on end devices by optimizing how expert weights are loaded from CPU memory to accelerators. The proposed DraftExpert framework significantly improves decode throughput for latency-critical applications without compromising the accuracy of final token verification. By balancing draft model size against loading overhead, this work advances practical inference efficiency for resource-constrained AI hardware.
DraftExpert: Expansion-Aware Self-Speculative Decoding for End-Device MoE Inference
arXiv:2607.24434v1 Announce Type: cross Abstract: Large Mixture-of-Experts (MoE) language models are attractive for end-device deployment because only a small subset of experts is active per token, but their routed expert weights often exceed accelerator memory. We target latency-critical single-user settings where routed experts are staged on demand from CPU memory to a GPU or from Flash to a mobile NPU. In this setting, self-speculative decoding faces a new bottleneck: increasing the draft expert set improves accuracy but triggers extra expert loading, while cheap small-footprint drafts have low acceptance; moreover, verifying a multi-token block activates the union of target experts and is no longer close to one target step. We propose DraftExpert, an expansion-aware self-speculative decoding framework for expert-offloaded MoE inference. DraftExpert trains one lightweight accelerator-resident draft expert per layer by self-distilling residual, logit/token, and router-agreement signals from the frozen target MoE. At inference time, it uses a fixed-footprint shared+top-1+draft-expert drafter together with confidence--expansion truncation and target-expert prefetching, while final tokens are still exactly verified by the target model. On DeepSeek-V2-Lite and Moonlight-16B-A3B across CPU-GPU and Flash-NPU offload, DraftExpert improves decode throughput by 1.45x on average, raises draft acceptance to 84~87%, and achieves 86~88% prefetch hit rates.
Abstract:Large Mixture-of-Experts (MoE) language models are attractive for end-device deployment because only a small subset of experts is active per token, but their routed expert weights often exceed accelerator memory. We target latency-critical single-user settings where routed experts are staged on demand from CPU memory to a GPU or from Flash to a mobile NPU. In this setting, self-speculative decoding faces a new bottleneck: increasing the draft expert set improves accuracy but triggers extra expert loading, while cheap small-footprint drafts have low acceptance; moreover, verifying a multi-token block activates the union of target experts and is no longer close to one target step. We propose DraftExpert, an expansion-aware self-speculative decoding framework for expert-offloaded MoE inference. DraftExpert trains one lightweight accelerator-resident draft expert per layer by self-distilling residual, logit/token, and router-agreement signals from the frozen target MoE. At inference time, it uses a fixed-footprint shared+top-1+draft-expert drafter together with confidence--expansion truncation and target-expert prefetching, while final tokens are still exactly verified by the target model. On DeepSeek-V2-Lite and Moonlight-16B-A3B across CPU-GPU and Flash-NPU offload, DraftExpert improves decode throughput by 1.45x on average, raises draft acceptance to 84~87%, and achieves 86~88% prefetch hit rates.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2607.24434 [cs.LG]
(or arXiv:2607.24434v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2607.24434 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 288 words
Related coverage
S2-MoE: Enabling Efficient Self-Speculative Decoding for Mixture-of-Experts on Edge Devices
arXiv · Aug 18, 2026
[AINews] Meta Muse Wonder & Wonder Full: MSL’s Claude 5.5-tier Frontier Model
Latent Space · Oct 13, 2026
[AINews] DeepSeek v4.1-Flash: 763B-P8B-D16B novel causal Encoder–Decoder architecture with vision marks the Return of the Whale
Latent Space · Sep 12, 2026
[AINews] OpenAI DevDay 2026: Dots, 6.1 Sol, Ultrafast, Decisions API, Agents API, Spaces, Marketplace, and 1.2 Billion ChatGPT WAU
Latent Space · Sep 30, 2026