arXiv 2511.14102 · cs.LG · 2025-11-18
Specter: Fast MoE Inference via Quantization-Enabled Multi-Step Speculative Offloading
Wenfeng Wang, Xiaofeng Hou, Xinfeng Xia, Xinkai Wang, Mingxuan Zhang, Peng Tang, Jiacheng Liu, Chao Li
机构待核实
一句话要点
Specter通过量化与投机解码结合,实现专家模型的高效离载推理,在单请求场景下平均提速1.5倍,显著提升内存受限GPU上MoE模型的推理效率。
Specter achieves up to 1.88x TPOT speedup via quantized speculative offloading for memory-constrained MoE inference on single GPUs.
要点
- MoE模型因专家路由依赖数据而造成内存溢出和PCIe带宽瓶颈,传统离载方案通信开销大。
- 引入量化轻量级草案模型进行多步投机解码,预测未来专家调用,提前规划专家迁移与执行顺序。
- 在单请求推理中平均提速1.5倍,最高达1.88倍,显著降低通信延迟和设备内存压力。
- 使单GPU上内存受限的MoE模型在消费级硬件上更实用;未涉及多请求并发场景下的性能表现。
English overview
Specter introduces a quantization-enabled speculative offloading framework for MoE models, combining lightweight draft models with multi-step lookahead to predict expert usage. This enables efficient expert movement and execution planning ahead of verification, reducing I/O bottlenecks. It achieves up to 1.88x speedup in token throughput on single-GPU systems with limited memory.
只根据 arXiv 官方摘要,由开源模型整理,经数字/专名自动核对(必须在摘要中出现),以论文原文为准。
阅读 arXiv 原文 PDF Hugging Face 讨论
主题与学科
同主题:高效训练与推理
- The Teacher Is a Direction, Not a Destination: Extrapolating RL-Induced Representation Residuals in On-Policy Distillation ▲511
- Scaling Properties of Same-Family On-Policy Distillation ▲327
- Beyond Domain-Routed Supervision: Cross-Teacher Guarding for Multi-Teacher On-Policy Distillation ▲250
- On-Policy or Off-Policy Learning? A Systematic Study of Distillation Dynamics ▲205
- Beyond Teacher Assignment: Domain-Normalized Multi-Teacher On-Policy Distillation ▲181
同主题:代码与软件工程
- MaLiang-Harness: A Programmable Path to Image and Video Generation 2026-09-28 · ▲367
- OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use 2026-10-15 · ▲325
- HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents 2026-10-16 · ▲294
- AgentSWE: Can Coding Agents Build the Agent You Actually Want? 2026-10-09 · ▲233
- CodeMidas: Scaling Agentic Coding RL Environments from Code Itself 2026-09-18 · ▲138
数据说明
元数据来自 arXiv API(已核实);点赞来自 Hugging Face 每日论文;引用来自 Semantic Scholar。最近刷新 2025-11-18。
要点由本站根据摘要自写,不转载摘要原文,请阅读原文。