[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_344":345},"routes/_layout",{"_11":272},"routes/entity.$type.$id",{"_11":12,"_13":14,"_15":16,"_17":18,"_19":20,"_21":22,"_23":24,"_25":26,"_27":28},"dehydratedState",{"_60":61,"_62":63},"type","person","id","person_A5100738931","section","articles","entity",{"_15":16,"_29":30,"_31":-5,"_32":33,"_34":-5,"_35":-5,"_36":37,"_38":39,"_40":-5,"_41":42,"_43":44,"_45":46,"_47":-5,"_48":-5,"_49":-5,"_50":-5,"_51":-5,"_52":-5,"_53":-5,"_54":-5,"_55":-5,"_56":-5,"_57":-5,"_58":30,"_59":-5},"canonical","https://www.hugmap.com/entity/person/person_A5100738931/articles","siteBaseUrl","https://www.hugmap.com","sectionCount",1,"jsonLd","{\"@context\":\"https://schema.org\",\"@graph\":[{\"name\":\"En Li\",\"url\":\"https://www.hugmap.com/entity/person/person_A5100738931/articles\",\"@type\":\"Person\",\"jobTitle\":null},{\"@type\":\"CollectionPage\",\"name\":\"En Li 相关文章\",\"description\":\"En Li 在 HugMap 中的相关文章，共 1 条，含结构化关联信息与详情链接。\",\"url\":\"https://www.hugmap.com/entity/person/person_A5100738931/articles\",\"inLanguage\":\"zh-CN\"},{\"@type\":\"BreadcrumbList\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"首页\",\"item\":\"https://www.hugmap.com/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"人物\",\"item\":\"https://www.hugmap.com/search?type=person\"},{\"@type\":\"ListItem\",\"position\":3,\"name\":\"En Li\",\"item\":\"https://www.hugmap.com/entity/person/person_A5100738931\"},{\"@type\":\"ListItem\",\"position\":4,\"name\":\"相关文章\",\"item\":\"https://www.hugmap.com/entity/person/person_A5100738931/articles\"}]}]}","name","En Li","nameZh","aliases",[],"description","descriptionZh","status","active","externalIds","{\"openalex\":\"A5100738931\"}","properties","source","openalex,graphrag","createdAt","2026-05-14T04:22:15.379243872","updatedAt","2026-05-29T09:35:23.892047837","orcid","title","affiliation","researchArea","paperCount","citationCount","homepage","linkedin","twitter","github","hindex","displayName","displayDescription","mutations",[],"queries",[64],{"_65":66,"_67":68,"_69":70,"_71":72},"dehydratedAt",1784848011922,"state",{"_74":75,"_76":26,"_77":78,"_79":-5,"_80":81,"_82":81,"_83":81,"_84":-5,"_85":-5,"_86":87,"_36":88,"_89":90},"queryKey",[73,19,14,16],"queryHash","[\"portal\",\"entity\",\"person\",\"person_A5100738931\"]","portal","data",{"_19":20,"_91":92,"_93":94,"_95":96,"_97":98,"_99":100,"_101":102,"_18":103},"dataUpdateCount","dataUpdatedAt",1784848011921,"error","errorUpdateCount",0,"errorUpdatedAt","fetchFailureCount","fetchFailureReason","fetchMeta","isInvalidated",false,"success","fetchStatus","idle","entityType","Person","taxonomies",[252],"papers",[166,167,168,169,170,171],"companies",[121,122,123],"events",[],"patents",[],[104],{"_15":105,"_29":106,"_48":106,"_107":108,"_109":110,"_111":112,"_113":114,"_115":116,"_117":118,"_119":120},"article_rss_www_infoq_cn_article_B36ZgoaReVDs3l05yw0z","超越 TurboQuant！OSCAR：面向真实 Serving 的 2-bit KV Cache量化","articleType","BLOG","publishedDate","2026-05-29","sourceSite","InfoQ 中文","author","Zhongzhu Zhou等","url","https://www.infoq.cn/article/B36ZgoaReVDs3l05yw0z","summary","作者 | Zhongzhu Zhou, Donglin Zhuang, Jisen Li, Ziyan Chen, Shuaiwen Leon Song, Ben Athiwaratkun, Xiaoxia Wu从 KV Cache 瓶颈说起长上下文模型的能力还在往前走，但在线推理服务遇到的压力，很多时候已经不只是计算量本身。每生成一个新 token，系统都要反复访问越来越长的历史 Key 和 Value；上下文拉长、batch 放大之后，KV Cache 同时吞掉显存容量和显存带宽。最新论文 OSCAR: Offline Spectral Covariance-Aware Rotation for 2-bit KV Cache Quantization 直接瞄准这个痛点，给出了一套面向长上下文 serving 的近 2-bit KV Cache 方案，并且已经接入 SGLang，可用于真实服务链路。相比之前量化的工作，比如TurboQuant ，压缩的是向量，但忽略了真正影响模型的是attention的质量, OSCAR 保留的是 attention 真正会读的方向。朴素 INT2 和全模型层的3-bit K/V TurboQuant 都会在困难推理任务上明显掉分；OSCAR 在约 2.28 effective bits per KV element 下仍能接近 BF16，并在 Qwen3-4B-Thinking 上相对3-bit K/V TurboQuant 最高提升 40.1 分。如果历史 KV 能稳定压到 2-bit，理论上历史缓存的存储开销可以接近缩小 8 倍。不过，低比特 KV Cache 真正困难的地方从来不是“能不能压”，而是压缩之后模型推理质量不能塌，系统实现也不能停留在离线实验脚本里。OSCAR 的价值正是在这两个问题上同时给出了答案。2-bit INT 只有 4 个离散等级，而 KV activation 里经常会出现少量幅值很大的 outlier channel。一旦量化尺度被这些极端通道牵着走，大部分正常值就会挤在很窄的有效区间内，attention 分布也随之偏移。普通 Hadamard 旋转可以把 outlier 打散，却没有区分模型在 attention 里真正使用的方向。","language","zh",{"_15":156,"_29":157,"_31":158,"_32":159,"_34":-5,"_35":-5,"_36":37,"_38":160,"_40":-5,"_41":129,"_43":161,"_45":162,"_132":163,"_134":-5,"_135":164,"_137":138,"_139":-5,"_140":-5,"_141":-5,"_142":-5,"_143":165,"_145":-5,"_146":-5,"_147":-5,"_58":158,"_59":-5},{"_15":148,"_29":149,"_31":-5,"_32":150,"_34":-5,"_35":-5,"_36":37,"_38":151,"_40":-5,"_41":129,"_43":152,"_45":153,"_132":154,"_134":-5,"_135":-5,"_137":138,"_139":-5,"_140":-5,"_141":-5,"_142":-5,"_143":-5,"_145":-5,"_146":-5,"_147":-5,"_58":149,"_59":-5},{"_15":124,"_29":125,"_31":126,"_32":127,"_34":-5,"_35":-5,"_36":37,"_38":128,"_40":-5,"_41":129,"_43":130,"_45":131,"_132":133,"_134":-5,"_135":136,"_137":138,"_139":-5,"_140":-5,"_141":-5,"_142":-5,"_143":144,"_145":-5,"_146":-5,"_147":-5,"_58":126,"_59":-5},"company_I19820366","Chinese Academy of Sciences","中国科学院",[125,126],"{\"ror\":\"034t30j35\",\"openalex\":\"I19820366\"}","openalex","2026-05-14T00:31:27.338078372","2026-07-01T04:38:29.877113219","companyType","COMPANY","foundedDate","headquarters","Beijing, China","country","CN","employeeCount","fundingStage","totalFunding","valuation","website","https://www.cas.cn","stockSymbol","uscc","crunchbaseUrl","company_I4210094879","Shandong Institute of Automation",[149,155],"{\"ror\":\"00qdtba35\",\"openalex\":\"I4210094879\"}","2026-05-14T01:04:19.322851188","2026-06-01T06:07:17.364767143","RESEARCH_INSTITUTE","山东省科学院自动化研究所成","company_I157773358","Sun Yat-sen University","中山大学",[157,158],"{\"ror\":\"0064kty71\",\"openalex\":\"I157773358\"}","2026-05-14T00:32:16.223592967","2026-06-01T04:18:31.281743573","UNIVERSITY","Guangzhou, China","https://www.sysu.edu.cn",{"_15":249,"_29":242,"_31":-5,"_32":-5,"_34":-5,"_35":-5,"_36":37,"_38":-5,"_40":244,"_41":245,"_43":250,"_45":251,"_177":-5,"_179":-5,"_180":-5,"_48":-5,"_182":-5,"_109":-5,"_185":-5,"_187":-5,"_52":-5,"_189":-5,"_190":-5,"_192":-5,"_194":-5,"_58":242,"_59":-5},{"_15":241,"_29":242,"_31":-5,"_32":243,"_34":-5,"_35":-5,"_36":37,"_38":-5,"_40":244,"_41":245,"_43":246,"_45":247,"_177":-5,"_179":-5,"_180":-5,"_48":-5,"_182":-5,"_109":-5,"_185":-5,"_187":-5,"_52":-5,"_189":-5,"_190":-5,"_192":-5,"_194":-5,"_58":242,"_59":-5},{"_15":227,"_29":228,"_31":-5,"_32":-5,"_34":-5,"_35":-5,"_36":37,"_38":229,"_40":-5,"_41":129,"_43":230,"_45":231,"_177":232,"_179":-5,"_180":233,"_48":228,"_182":234,"_109":235,"_185":222,"_187":236,"_52":237,"_189":-5,"_190":238,"_192":239,"_194":87,"_58":228,"_59":-5},{"_15":213,"_29":214,"_31":-5,"_32":-5,"_34":-5,"_35":-5,"_36":37,"_38":215,"_40":-5,"_41":129,"_43":216,"_45":217,"_177":218,"_179":-5,"_180":219,"_48":214,"_182":220,"_109":221,"_185":222,"_187":223,"_52":224,"_189":-5,"_190":225,"_192":-5,"_194":87,"_58":214,"_59":-5},{"_15":198,"_29":199,"_31":-5,"_32":-5,"_34":-5,"_35":-5,"_36":37,"_38":200,"_40":-5,"_41":129,"_43":201,"_45":202,"_177":203,"_179":-5,"_180":204,"_48":199,"_182":205,"_109":206,"_185":207,"_187":208,"_52":209,"_189":-5,"_190":210,"_192":-5,"_194":87,"_58":199,"_59":-5},{"_15":172,"_29":173,"_31":-5,"_32":-5,"_34":-5,"_35":-5,"_36":37,"_38":174,"_40":-5,"_41":129,"_43":175,"_45":176,"_177":178,"_179":-5,"_180":181,"_48":173,"_182":183,"_109":184,"_185":186,"_187":-5,"_52":188,"_189":-5,"_190":191,"_192":193,"_194":87,"_58":173,"_59":-5},"paper_W2809251854","Edge Intelligence","{\"openalex\":\"W2809251854\",\"doi\":\"10.1145/3229556.3229562\"}","2026-05-19T21:36:10.284987721","2026-05-19T21:37:12.189563472","doi","10.1145/3229556.3229562","arxivId","openalexId","W2809251854","abstractText","As the backbone technology of machine learning, deep neural networks (DNNs) have have quickly ascended to the spotlight. Running DNNs on resource-constrained mobile devices is, however, by no means trivial, since it incurs high performance and energy overhead. While offloading DNNs to the cloud for execution suffers unpredictable performance, due to the uncontrolled long wide-area network latency. To address these challenges, in this paper, we propose Edgent, a collaborative and on-demand DNN co-inference framework with device-edge synergy. Edgent pursues two design knobs: (1) DNN partitioning that adaptively partitions DNN computation between device and edge, in order to leverage hybrid computation resources in proximity for real-time DNN inference. (2) DNN right-sizing that accelerates DNN inference through early-exit at a proper intermediate DNN layer to further reduce the computation latency. The prototype implementation and extensive evaluations based on Raspberry Pi demonstrate Edgent's effectiveness in enabling on-demand low-latency edge intelligence.","2018-08-01","publishedYear",2018,"venue",377,"authors","keywords",[195,196,197],"pdfUrl","https://dl.acm.org/doi/pdf/10.1145/3229556.3229562","topVenue","IoT and Edge/Fog Computing","Advanced Neural Network Applications","Context-Aware Activity Recognition Systems","paper_W3098650625","A Review on State-of-the-Art Power Line Inspection Techniques","{\"openalex\":\"W3098650625\",\"doi\":\"10.1109/tim.2020.3031194\"}","2026-05-19T17:09:30.399403105","2026-05-19T17:11:17.196243944","10.1109/tim.2020.3031194","W3098650625","With the fast development of smart grid, the power line mileage and power equipments get rapid growth. The contradiction between the large number of maintenance equipments and the small number of maintenance workers becomes increasingly prominent. Meanwhile, the traditional maintenance mode has the disadvantages of over or under maintenance, which will lead to the increasing failure risk of power transmission system. Faced with these issues, to enhance the intelligence and automation level of power line inspection, many researchers have devoted much effort to the research of automatic power line inspection and some state-of-the-art techniques about power line inspection are proposed to improve the inspection efficiency and quality, such as unmanned aerial vehicle (UAV), image processing, deep architecture, and so on. In this article, we analyzed and summarized the state-of-the-art techniques on power line inspection to provide a valuable reference for the researchers engaged in the smart grid. First, the common inspection tasks on power line inspection are reviewed. Second, the existing inspection platforms are examined in this article. Also, the advantages and disadvantages of different platforms are analyzed accordingly. Third, faced with different inspection tasks, different inspection sensors are integrated into the inspection platforms for data collection. Therefore, the common sensors on inspection platforms are discussed. Finally, to realize automatic power line inspection, different inspection methods are proposed or improved, and these advanced inspection methods are surveyed and discussed.","2020-10-15",2020,"IEEE Transactions on Instrumentation and Measurement",394,[211,212,196],"Power Line Inspection Robots","Remote Sensing and LiDAR Applications","paper_W2980856918","Edge AI: On-Demand Accelerating Deep Neural Network Inference via Edge Computing","{\"openalex\":\"W2980856918\",\"doi\":\"10.1109/twc.2019.2946140\"}","2026-05-14T20:02:51.429246588","2026-05-20T20:28:38.211656395","10.1109/twc.2019.2946140","W2980856918","As a key technology of enabling Artificial Intelligence (AI) applications in 5G era, Deep Neural Networks (DNNs) have quickly attracted widespread attention. However, it is challenging to run computation-intensive DNN-based tasks on mobile devices due to the limited computation resources. What’s worse, traditional cloud-assisted DNN inference is heavily hindered by the significant wide-area network latency, leading to poor real-time performance as well as low quality of user experience. To address these challenges, in this paper, we propose Edgent , a framework that leverages edge computing for DNN collaborative inference through device-edge synergy. Edgent exploits two design knobs: (1) DNN partitioning that adaptively partitions computation between device and edge for purpose of coordinating the powerful cloud resource and the proximal edge resource for real-time DNN inference; (2) DNN right-sizing that further reduces computing latency via early exiting inference at an appropriate intermediate DNN layer. In addition, considering the potential network fluctuation in real-world deployment, Edgent is properly design to specialize for both static and dynamic network environment. Specifically, in a static environment where the bandwidth changes slowly, Edgent derives the best configurations with the assist of regression-based prediction models, while in a dynamic environment where the bandwidth varies dramatically, Edgent generates the best execution plan through the online change point detection algorithm that maps the current bandwidth state to the optimal configuration. We implement Edgent prototype based on the Raspberry Pi and the desktop PC and the extensive experimental evaluations demonstrate Edgent ’s effectiveness in enabling on-demand low-latency edge intelligence.","2019-10-18",2019,"IEEE Transactions on Wireless Communications",844,[195,196,226],"Blockchain Technology Applications and Security","paper_W2950865323","Edge Intelligence: Paving the Last Mile of Artificial Intelligence With Edge Computing","{\"openalex\":\"W2950865323\",\"doi\":\"10.1109/jproc.2019.2918951\"}","2026-05-14T04:22:15.241805924","2026-05-21T06:20:11.184718077","10.1109/jproc.2019.2918951","W2950865323","With the breakthroughs in deep learning, the recent years have witnessed a booming of artificial intelligence (AI) applications and services, spanning from personal assistant to recommendation systems to video/audio surveillance. More recently, with the proliferation of mobile computing and Internet of Things (IoT), billions of mobile and IoT devices are connected to the Internet, generating zillions bytes of data at the network edge. Driving by this trend, there is an urgent need to push the AI frontiers to the network edge so as to fully unleash the potential of the edge big data. To meet this demand, edge computing, an emerging paradigm that pushes computing tasks and services from the network core to the network edge, has been widely recognized as a promising solution. The resulted new interdiscipline, edge AI or edge intelligence (EI), is beginning to receive a tremendous amount of interest. However, research on EI is still in its infancy stage, and a dedicated venue for exchanging the recent advances of EI is highly desired by both the computer system and AI communities. To this end, we conduct a comprehensive survey of the recent research efforts on EI. Specifically, we first review the background and motivation for AI running at the network edge. We then provide an overview of the overarching architectures, frameworks, and emerging key technologies for deep learning model toward training/inference at the network edge. Finally, we discuss future research opportunities on EI. We believe that this survey will elicit escalating attentions, stimulate fruitful discussions, and inspire further research ideas on EI.","2019-06-12","Proceedings of the IEEE",2118,[195,196,240],"https://ieeexplore.ieee.org/ielx7/5/8789751/08736011.pdf","Privacy-Preserving Technologies in Data","e4e3bcf0-c228-41cd-be62-498c04e6be32","OSCAR: Offline Spectral Covariance-Aware Rotation for 2-bit KV Cache Quantization",[248],"{}","graphrag","2026-05-29T09:35:29.364069319","2026-05-29T09:35:29.364073302","OSCAR","4cb8c2f2-cfe0-4324-84fe-09c5f8e83327","2026-05-29T09:34:03.328698504","2026-05-29T09:34:03.328700328",{"_15":253,"_29":254,"_31":-5,"_32":255,"_34":256,"_35":-5,"_36":37,"_38":-5,"_40":-5,"_41":257,"_43":-5,"_45":-5,"_258":259,"_260":-5,"_261":26,"_262":263,"_264":265,"_266":81,"_267":87,"_268":81,"_269":-5,"_58":254,"_59":256},"tax_ind_ai","人工智能",[270,271],"Artificial Intelligence","manual","taxonomyType","INDUSTRY","code","level","parentId","tax_ind_root","path","科技行业/人工智能","sortOrder","isLeaf","entityCount","children","AI","AI 技术",{"_60":273,"_62":274},[],[275],{"_65":276,"_67":277,"_69":278,"_71":279},1784848011121,{"_74":281,"_76":26,"_77":276,"_79":-5,"_80":81,"_82":81,"_83":81,"_84":-5,"_85":-5,"_86":87,"_36":88,"_89":90},[73,280],"[\"portal\",\"industries\"]","industries",[282,283,284,285,286,287,288,289,290,291,292,293],{"_15":253,"_29":254,"_31":-5,"_32":343,"_34":256,"_35":-5,"_36":37,"_38":-5,"_40":-5,"_41":257,"_43":-5,"_45":-5,"_258":259,"_260":-5,"_261":26,"_262":263,"_264":265,"_266":81,"_267":87,"_268":81,"_269":-5,"_58":254,"_59":256},{"_15":339,"_29":340,"_31":-5,"_32":-5,"_34":341,"_35":-5,"_36":37,"_38":-5,"_40":-5,"_41":257,"_43":-5,"_45":-5,"_258":259,"_260":-5,"_261":26,"_262":263,"_264":342,"_266":81,"_267":87,"_268":81,"_269":-5,"_58":340,"_59":341},{"_15":333,"_29":334,"_31":-5,"_32":335,"_34":336,"_35":-5,"_36":37,"_38":-5,"_40":-5,"_41":257,"_43":-5,"_45":-5,"_258":259,"_260":-5,"_261":26,"_262":263,"_264":337,"_266":81,"_267":87,"_268":81,"_269":-5,"_58":334,"_59":336},{"_15":329,"_29":330,"_31":-5,"_32":-5,"_34":331,"_35":-5,"_36":37,"_38":-5,"_40":-5,"_41":257,"_43":-5,"_45":-5,"_258":259,"_260":-5,"_261":26,"_262":263,"_264":332,"_266":81,"_267":87,"_268":81,"_269":-5,"_58":330,"_59":331},{"_15":325,"_29":326,"_31":-5,"_32":-5,"_34":327,"_35":-5,"_36":37,"_38":-5,"_40":-5,"_41":257,"_43":-5,"_45":-5,"_258":259,"_260":-5,"_261":26,"_262":263,"_264":328,"_266":81,"_267":87,"_268":81,"_269":-5,"_58":326,"_59":327},{"_15":319,"_29":320,"_31":-5,"_32":321,"_34":322,"_35":-5,"_36":37,"_38":-5,"_40":-5,"_41":257,"_43":-5,"_45":-5,"_258":259,"_260":-5,"_261":26,"_262":263,"_264":323,"_266":81,"_267":87,"_268":81,"_269":-5,"_58":320,"_59":322},{"_15":314,"_29":315,"_31":-5,"_32":-5,"_34":316,"_35":-5,"_36":37,"_38":-5,"_40":-5,"_41":257,"_43":-5,"_45":-5,"_258":259,"_260":-5,"_261":26,"_262":263,"_264":317,"_266":81,"_267":318,"_268":81,"_269":-5,"_58":315,"_59":316},{"_15":310,"_29":311,"_31":-5,"_32":-5,"_34":312,"_35":-5,"_36":37,"_38":-5,"_40":-5,"_41":257,"_43":-5,"_45":-5,"_258":259,"_260":-5,"_261":26,"_262":263,"_264":313,"_266":81,"_267":87,"_268":81,"_269":-5,"_58":311,"_59":312},{"_15":306,"_29":307,"_31":-5,"_32":-5,"_34":308,"_35":-5,"_36":37,"_38":-5,"_40":-5,"_41":257,"_43":-5,"_45":-5,"_258":259,"_260":-5,"_261":26,"_262":263,"_264":309,"_266":81,"_267":87,"_268":81,"_269":-5,"_58":307,"_59":308},{"_15":302,"_29":303,"_31":-5,"_32":-5,"_34":304,"_35":-5,"_36":37,"_38":-5,"_40":-5,"_41":257,"_43":-5,"_45":-5,"_258":259,"_260":-5,"_261":26,"_262":263,"_264":305,"_266":81,"_267":87,"_268":81,"_269":-5,"_58":303,"_59":304},{"_15":298,"_29":299,"_31":-5,"_32":-5,"_34":300,"_35":-5,"_36":37,"_38":-5,"_40":-5,"_41":257,"_43":-5,"_45":-5,"_258":259,"_260":-5,"_261":26,"_262":263,"_264":301,"_266":81,"_267":87,"_268":81,"_269":-5,"_58":299,"_59":300},{"_15":294,"_29":295,"_31":-5,"_32":-5,"_34":296,"_35":-5,"_36":37,"_38":-5,"_40":-5,"_41":257,"_43":-5,"_45":-5,"_258":259,"_260":-5,"_261":26,"_262":263,"_264":297,"_266":81,"_267":87,"_268":81,"_269":-5,"_58":295,"_59":296},"tax_ind_smartphone","智能手机","Smartphone","科技行业/智能手机","tax_ind_robot","机器人","Robotics","科技行业/机器人","tax_ind_wearable","智能穿戴","Smart Wearables","科技行业/智能穿戴","tax_ind_energy","新能源与智能制造","New Energy & Smart Manufacturing","科技行业/新能源与智能制造","tax_ind_biotech","生物科技","Biotech","科技行业/生物科技","tax_ind_quantum","量子计算","Quantum Computing","科技行业/量子计算",true,"tax_ind_blockchain","区块链",[324],"Blockchain","科技行业/区块链","Web3","tax_ind_enterprise","企业软件","Enterprise Software","科技行业/企业软件","tax_ind_security","网络安全","Cybersecurity","科技行业/网络安全","tax_ind_semi","半导体",[338],"Semiconductor","科技行业/半导体","芯片","tax_ind_cloud","云计算","Cloud Computing","科技行业/云计算",[270,271],"env",{"_346":24},"SITE_BASE_URL"]

En Li 人物

En Li · 相关文章

相关文章