1.中国科学院计算机网络信息中心,中国 北京 100083
2.中国科学院大学,中国 北京 100190
3.中国移动通信集团有限公司,中国 北京 100032
宋俊平,中国科学院计算机网络信息中心高级工程师、中科院青年创新促进会会员、中国通信标准化协会TC630语义智简通信技术与标准推进委员会资源及模型管理组副组长;主要研究方向为算力网络、人工智能;获省部级奖励1项;发表论文30余篇,申请国家发明专利24项。
黄天一,中国科学院计算机网络信息中心在读硕士研究生;主要研究方向为分布式推理、算网融合。
周旭,中国科学院计算机网络信息中心研究员,担任中国通信标准化协会TC614标准推进委员会副主席、中国通信学会算力网络委员会委员、中国通信学会信息通信网络技术委员会委员、重庆大学通信学院兼职教授/博导等学术职务,国家重点研发计划项目首席科学家,国家级人才计划入选者;主要研究方向为计算机网络体系结构、人工智能、5G/6G移动网络;近年来作为项目负责人完成重大项目10余项,获得部级科学技术奖一等奖4次、二等奖2次;发表论文80余篇,申请专利60余项(其中已授权25项),撰写专著1本,主持/参与制定国际标准/行业标准15项。
李智,中国移动通信有限公司网络事业部高级技术专家,高级工程师;主要研究方向为互联网技术、CDN、算力网络、边缘计算等;获得北京市科技进步奖1项,中国通信学会科学技术奖1项,工业和信息化部中国信息通信研究院科技进步奖9项,互联网、CDN、云、人工智能产业协会等科技进步奖4项,中国移动科技进步奖2项;发表论文18篇,拥有专利36项,参与研发标准24项。
吴茜,中国移动通信有限公司网络事业部高级工程师;主要研究方向为CDN的云化、智能化关键技术;获2020年中国通信学会科学技术奖三等奖、2023年工业和信息化部中国信息通信研究院“华彩杯”算力应用创新大赛北区总决赛一等奖、2024年全国信息通信领域安全生产和网络运行安全成果一等奖等。
收稿:2026-04-25,
网络首发:2026-08-07,
纸质出版:2026-06-25
移动端阅览
宋俊平, 黄天一, 周旭, 等. 面向分布式推理的大模型键值缓存分发网络 [J]. 中兴通讯技术, 2026, 32(3): 45-50.
Song J P, Huang T Y, Zhou X, et al. Large model key-value cache distribution network for distributed inference services [J]. ZTE technology journal, 2026, 32(3): 45-50.
宋俊平, 黄天一, 周旭, 等. 面向分布式推理的大模型键值缓存分发网络 [J]. 中兴通讯技术, 2026, 32(3): 45-50. DOI: 10.12142/ZTETJ.202603009
Song J P, Huang T Y, Zhou X, et al. Large model key-value cache distribution network for distributed inference services [J]. ZTE technology journal, 2026, 32(3): 45-50. DOI: 10.12142/ZTETJ.202603009
大模型推理已成为智能算力体系的核心,传统云端集中式推理面临带宽压力大、推理延迟高、算力资源紧张等瓶颈。现有云边协同方案在通信开销与键值缓存(KV-Cache)复用效率方面仍存在明显不足。为此,提出一种基于内容分发网络(CDN)的大模型KV-Cache分发网络(KVCDN),将传统CDN升级为具备张量感知与状态管理能力的分布式缓存架构。该架构采用端侧语义规范化、边缘热点分级缓存与核心全局去重编排3层协同结构,并结合提示词(Prompt)标准化、分级存储预放置、全局去重融合与语义路由4项关键技术,从而显著提升KV-Cache复用率,降低冗余计算与推理延迟。
Large language model inference has become a core component of intelligent computing systems. Traditional cloud-centric inference approaches face critical bottlenecks such as high bandwidth consumption
excessive latency
and intensive computational demands. Existing cloud-edge collaborative schemes still suffer from significant communication overhead and low cache reuse efficiency of Key-Value Caches (KV-Cache). To address these issues
a KV-Cache distribution network based on content delivery network (CDN)
termed KVCDN
is proposed. In this architecture
conventional CDN nodes are upgraded into a distributed caching system with tensor-aware perception and state management capabilities. A three-layer collaborative structure is adopted
consisting of semantic normalization at the end side
hotspot-aware hierarchical caching at the edge
and global deduplication with orchestration at the core. Combined with four key enabling techniques
i.e.
prompt standardization
tiered storage and proactive pre-placement
global deduplication and fusion
and semantic-aware routing
the KVCDN significantly improves KV-Cache reuse ratio
while reducing redundant computations and inference latency.
Qu G Q, Chen Q Y, Wei W, et al. Mobile edge intelligence for large language models: a contemporary survey [PP/OL]. arXiv (2025-03-20)[2026-04-20]. https://arxiv.org/abs/2407.18921
Kafetzis D, Khalili R, Koutsopoulos I. Large language model partitioning for low-latency inference at the edge [C]//Proceedings of the 23rd International Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOpt). IEEE, 2025: 1-8
Cunningham M. Privacy-aware split inference with speculative decoding for large language models over wide-area networks [PP/OL]. arXiv (2026-02-18)[2026-04-22]. https://arxiv.org/abs/2602.16760
Sung M, Palakonda V, Im S, et al. Memory- and latency-constrained inference of large language models via adaptive split computing [PP/OL]. arXiv (2025-11-06)[2026-04-20]. https://arxiv.org/abs/2511.04002
Feng Z D, Lu L, Li Q, et al. Distributed inference optimization for large language model in edge-cloud collaborative networks [C]//Proceedings of ICC 2025 - IEEE International Conference on Communications. IEEE, 2025: 6161-6166. DOI: 10.1109/icc52391.2025.11160773
Jin H P, Wu Y Z. CE-CoLLM: efficient and adaptive large language models through cloud-edge collaboration [C]//Proceedings of IEEE International Conference on Web Services (ICWS). IEEE, 2025: 316-323. DOI: 10.1109/icws67624.2025.00046
Patel P, Choukse E, Zhang C J, et al. Splitwise: efficient generative LLM inference using phase splitting [C]//Proceedings of ACM/IEEE 51st Annual International Symposium on Computer Architecture (ISCA). IEEE, 2024: 118-132. DOI: 10.1109/ISCA59077.2024.00019
Younesi A, Shabrang M A, Oustad E, et al. Splitwise: collaborative edge–cloud inference for LLMs via Lyapunov-assisted DRL [C]//Proceedings of the 18th IEEE/ACM International Conference on Utility and Cloud Computing. ACM, 2025: 1-11. DOI: 10.1145/3773274.3774267
Ye S Y, Ouyang B, Zeng L K, et al. Jupiter: fast and resource-efficient collaborative inference of generative LLMs on edge devices [C]//Proceedings of IEEE INFOCOM 2025 - IEEE Conference on Computer Communications. IEEE, 2025: 1-10. DOI: 10.1109/infocom55648.2025.11044734
Chen C, Borgeaud S, Irving G, et al. Accelerating large language model decoding with speculative sampling [PP/OL]. arXiv (2023-02-02)[2026-04-18]. https://arxiv.org/abs/2302.01318
Leviathan Y, Kalman M, Matias Y. Fast inference from transformers via speculative decoding [C]//Proceedings of the 40th International Conference on Machine Learning. ACM, 2023: 19274-19286. DOI: 10.5555/3618408.3619203
Ning J, Zheng C, Yang T. DSSD: Efficient edge-device LLM deployment and collaborative inference via distributed split speculative decoding [PP/OL]. arXiv (2025-07-16)[2026-04-18]. https://arxiv.org/abs/2507.12000
Cheng Y, Du K, Yao J, et al. Do large language models need a content delivery network [PP/OL]. arXiv (2024-09-16)[2026-04-15]. https://arxiv.org/abs/2409.13761
Liu Y, Cheng Y, Yao J, et al. LMCache: an efficient KV cache layer for enterprise-scale LLM inference [PP/OL]. arXiv (2025-12-05)[2026-04-18]. https://arxiv.org/abs/2510.09665
Team TAib. AIBrix: towards scalable, cost-effective large language model inference infrastructure [PP/OL]. arXiv (2025-02-22)[2026-04-18]. https://arxiv.org/html/2504.03648v1
0
浏览量
0
下载量
0
CSCD
关联资源
相关文章
相关作者
相关机构
京公网安备11010602201714号