[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84657-en":3,"doc-seo-84657-105":29,"detail-sidebar-cat-0-en-105":90},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":4,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},84657,4810365810221,"Aurora","https://ap-avatar.wpscdn.com/davatar_155a257f0dc6eb9ab79c44ca47cae57d",8,"Research & Report","Multi-Turn Distributed Inference with Mixture of Experts for 6G Edge Cloud Networks","Mixture-of-Experts (MoE) models are increasingly used in 6G edge–cloud networks, where sparse activation lowers per-inference computation by engaging only a subset of experts. In edge–cloud settings, MoE inference faces a conflict between keeping Key-Value (KV) cache state local and dispatching expert computation elastically across the network. This conflict grows in multi-turn dialogue as KV state must persist and expand across turns. StateFlow separates persistent KV state from transient sparse computation by pinning KV at a sticky serving site and optimizing expert dispatch and aggregation placement, increasing stable dialogue concurrency by over 2× and cutting p95 turn latency by 53%.","Multi-Turn Distributed Inference with Mixture of Experts for 6G Edge–Cloud Networks  \nBo Liu, Haiyuan Li, Yuelin Liu, Yulei Wu, Rasheed Hussain, Shadi Moazzeni, Dimitra Simeonidou Smart Internet Lab, School of Electrical, Electronic and Mechanical Engineering (EEME), University of Bristol, U.K. E-mail: {bo.liu, [ocean.h.li](ocean.h.li), y.l.wu, yuelin.liu, rasheed.hussain, shadi.moazzeni, [dimitra.simeonidou](dimitra.simeonidou}@bristol.ac.uk)[}](dimitra.simeonidou}@bristol.ac.uk)[@bristol.ac.uk](dimitra.simeonidou}@bristol.ac.uk)  \narXiv :2607 .02522v1 [ cs .DC] 6 May 2026  \nAbstract—Mixture-of-Experts (MoE) architectures are increasingly deployed across 6G edge–cloud networks, where sparse activation reduces the computational footprint of each inference to only a fraction of the full expert set. However, MoE inference in edge-cloud networks creates a tension between KV state locality and elastic expert dispatch. KV state relocation incurs substantial transfer overhead, while expert computation benefits from spreading across the network to exploit available capacity. This tension is amplified in multi-turn inference, where each turn extends the KV state that must persist across the dialogue. To this end, we present StateFlow, a distributed inference policy that decouples persistent KV state from transient sparse computation. StateFlow pins KV state at a sticky serving site for cross-turn reuse and jointly optimizes expert dispatch and aggregation placement across the network. We further implement a real-world testbed with kernel-level network emulation and experimental results show that StateFlow sustains more than 2 × higher stable dialogue concurrency than the distributed baseline solutions and reduces turn-level p95 latency under multi-turn inference by 53.0% .  \nIndex Terms—6G networks, edge–cloud computing, mixture of experts, distributed inference, large language models  \nI. INTRODUCTION  \nThe emergence of 6G edge-cloud networks, spanning access edges, edge clouds, and remote clouds, creates new opportunities for distributing generative-AI inference closer to end users [1], [2] . Mixture-of-Experts (MoE) architectures take this further by replacing the feed-forward component of each transformer layer with a set of parallel expert subnetworks, of which only a small subset is activated per token through sparse routing. This reduces the computational footprint of each inference and allows the full expert set to be distributed across the network. As experts reside on different network sites, each inference exchanges activations with the selected experts and gathers outputs back at every transformer layer.  \nThis cross-site execution introduces a tension between two competing requirements. The Key-Value (KV) cache produced during inference must reside at a stable location, as relocating it incurs substantial transfer overhead proportional to the accumulated context length. Meanwhile, expert computation benefits from elastic dispatch across the network to exploit available capacity beyond the hosting site.  \nFor dialogue-oriented workloads such as conversational assistants, copilots, and agents, this tension is compounded across multiple turns, each corresponding to one user query and model response that extends the shared KV state. Successive turns within the same dialogue share an ever-growing  \nKV state that must persist across the full dialogue lifecycle. Treating each turn as an independent request risks repeated dialogue migration, KV state reconstruction, and redundant cross-site transfers, all consuming the limited bandwidth and latency budget of the edge–cloud network.  \nIn response, existing work approaches this problem from three directions. Collaborative edge inference [3]–[5] partitions dense models across device-edge-cloud tiers but targets stateless pipelines without persistent dialogue state. LLM serving systems [6]–[9] improve KV-cache management and multi-turn reuse [10], [11], but assume centralized executi","cbCaibfBZY32RU4H","https://ap.wps.com/l/cbCaibfBZY32RU4H","pdf",926600,1,6,"English","en",105,"# Introduction\n## Multi-turn distributed MoE inference tension\n## Proposed solution: StateFlow\n# Related Work\n## Collaborative edge inference\n## LLM serving systems\n## Communication-aware MoE schemes","[{\"question\":\"What problem does StateFlow address in multi-turn MoE inference over 6G edge–cloud networks?\",\"answer\":\"StateFlow targets the tension between KV state locality and elastic expert dispatch. In multi-turn dialogue, KV state must persist across turns while experts are distributed, amplifying communication and latency overhead.\"},{\"question\":\"How does StateFlow decouple KV cache from expert computation?\",\"answer\":\"StateFlow pins KV state at a sticky serving site for cross-turn reuse, while dispatching sparse expert branches across the network. It also jointly optimizes expert dispatch and aggregation placement.\"},{\"question\":\"What performance improvements does the real-world evaluation report?\",\"answer\":\"The evaluation shows StateFlow sustains more than 2× higher stable dialogue concurrency than distributed baseline solutions. It also reduces turn-level p95 latency by 53.0% under multi-turn inference.\"}]",1784197515,15,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":85,"head_meta":87,"extra_data":89,"updated_unix":27},"multi-turn-distributed-inference-with-mixture-of-experts-for-6g-edge-cloud-networks","",{"@graph":35,"@context":84},[36,53,67],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/multi-turn-distributed-inference-with-mixture-of-experts-for-6g-edge-cloud-networks/84657/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":61,"encodingFormat":60,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":4},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"What problem does StateFlow address in multi-turn MoE inference over 6G edge–cloud networks?","Question",{"text":74,"@type":75},"StateFlow targets the tension between KV state locality and elastic expert dispatch. In multi-turn dialogue, KV state must persist across turns while experts are distributed, amplifying communication and latency overhead.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How does StateFlow decouple KV cache from expert computation?",{"text":79,"@type":75},"StateFlow pins KV state at a sticky serving site for cross-turn reuse, while dispatching sparse expert branches across the network. It also jointly optimizes expert dispatch and aggregation placement.",{"name":81,"@type":72,"acceptedAnswer":82},"What performance improvements does the real-world evaluation report?",{"text":83,"@type":75},"The evaluation shows StateFlow sustains more than 2× higher stable dialogue concurrency than distributed baseline solutions. It also reduces turn-level p95 latency by 53.0% under multi-turn inference.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":91},[92,96,100,104,109,113,118,121,126,129,133],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":21,"doc_module":4,"doc_module_name":45,"category_name":110,"show_sort_weight":111,"slug":112},"Technology",50,"technology",{"id":114,"doc_module":4,"doc_module_name":45,"category_name":115,"show_sort_weight":116,"slug":117},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":119,"slug":120},30,"research-report",{"id":122,"doc_module":4,"doc_module_name":45,"category_name":123,"show_sort_weight":124,"slug":125},9,"Religion & Spirituality",20,"religion-spirituality",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":127,"show_sort_weight":124,"slug":128},"World Cup","world-cup",{"id":130,"doc_module":4,"doc_module_name":45,"category_name":131,"show_sort_weight":130,"slug":132},10,"Lifestyle","lifestyle",{"id":134,"doc_module":4,"doc_module_name":45,"category_name":135,"show_sort_weight":105,"slug":136},19,"General","general"]