{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-07-23T06:40:50.084Z","headline":"小红书与北大开源 UltraEP：面向大规模 MoE 训推的实时负载均衡方案","description":"小红书与北大提出 UltraEP，首次将基于精确路由信息的实时负载均衡引入生产系统，在每个 microbatch 和每一层动态复制热点专家。在 Qwen3-235B 等模型上，训练吞吐平均达到理想性能的 94.6%，相比 Megatron-LM 提升 42%；推理 prefill 吞吐相比 SGLang 提升 1.56 倍。","url":"https://www.aioga.com/news/cmrsy2z9e03wpbitlqgh25fol/","mainEntityOfPage":"https://www.aioga.com/news/cmrsy2z9e03wpbitlqgh25fol/","datePublished":"2026-07-20T08:00:45.000Z","dateModified":"2026-07-20T08:00:45.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://mp.weixin.qq.com/s/rAoF65ywi5trWbI-heJieg","https://aihot.virxact.com/items/cmrsy2z9e03wpbitlqgh25fol"],"canonicalUrl":"https://www.aioga.com/news/cmrsy2z9e03wpbitlqgh25fol/","directAnswer":{"@type":"Answer","text":"小红书与北大提出 UltraEP，将基于精确路由信息的实时负载均衡用于生产系统，并按每个 microbatch、每一层动态复制热点专家，以改善大规模 MoE 训练与推理的负载分配。","url":"https://www.aioga.com/news/cmrsy2z9e03wpbitlqgh25fol/","dateCreated":"2026-07-20T08:00:45.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"公众号：小红书技术（dots.llm） source article","url":"https://mp.weixin.qq.com/s/rAoF65ywi5trWbI-heJieg","datePublished":"2026-07-20T08:00:45.000Z","provider":{"@type":"Organization","name":"公众号：小红书技术（dots.llm）","url":"https://mp.weixin.qq.com/s/rAoF65ywi5trWbI-heJieg"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmrsy2z9e03wpbitlqgh25fol","datePublished":"2026-07-20T08:00:45.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmrsy2z9e03wpbitlqgh25fol"}}],"aggregationSource":"公众号：小红书技术（dots.llm）","originalPublisher":{"name":"公众号：小红书技术（dots.llm）","url":"https://mp.weixin.qq.com/s/rAoF65ywi5trWbI-heJieg"},"article":{"id":"cmrsy2z9e03wpbitlqgh25fol","slug":"cmrsy2z9e03wpbitlqgh25fol","url":"https://www.aioga.com/news/cmrsy2z9e03wpbitlqgh25fol/","title":"小红书与北大开源 UltraEP：面向大规模 MoE 训推的实时负载均衡方案","title_en":"小红书、北大开源 UltraEP：面向大规模 MoE 训推的「最优」负载均衡方案","summary":"小红书与北大提出 UltraEP，首次将基于精确路由信息的实时负载均衡引入生产系统，在每个 microbatch 和每一层动态复制热点专家。在 Qwen3-235B 等模型上，训练吞吐平均达到理想性能的 94.6%，相比 Megatron-LM 提升 42%；推理 prefill 吞吐相比 SGLang 提升 1.56 倍。","source":"公众号：小红书技术（dots.llm）","sourceUrl":"https://mp.weixin.qq.com/s/rAoF65ywi5trWbI-heJieg","aiHotUrl":"https://aihot.virxact.com/items/cmrsy2z9e03wpbitlqgh25fol","publishedAt":"2026-07-20T08:00:45.000Z","category":"产品更新","score":68,"selected":true,"articleBody":["小红书与北大提出 UltraEP，首次将基于精确路由信息的实时负载均衡引入生产系统，在每个 microbatch 和每一层动态复制热点专家。","在 Qwen3-235B 等模型上，训练吞吐平均达到理想性能的 94.6%，相比 Megatron-LM 提升 42%；","推理 prefill 吞吐相比 SGLang 提升 1.56 倍。"],"articleImages":[],"mediaStatus":"none","articleBodyZh":[],"translationStatus":"","bodyOrigin":"summary-fallback","editorial":{"summary":"小红书与北大提出 UltraEP，将基于精确路由信息的实时负载均衡用于生产系统，并按每个 microbatch、每一层动态复制热点专家，以改善大规模 MoE 训练与推理的负载分配。","background":"公开材料称，该方案面向大规模 MoE 训练与推理场景。在 Qwen3-235B 等模型上，其训练吞吐平均达到理想性能的 94.6%，相比 Megatron-LM 提升 42%。","viewpoint":"Aioga 判断，UltraEP 的核心关注点不是静态配置专家副本，而是依据精确路由信息进行细粒度、实时调整。值得关注的是，这种机制已被材料明确描述为引入生产系统。","implications":"公开材料还称，UltraEP 的推理 prefill 吞吐相比 SGLang 提升 1.56 倍。Aioga 判断，这组结果显示实时复制热点专家可能同时影响训练吞吐与推理 prefill 阶段的效率。","nextStep":"值得关注的是，现有材料未披露测试环境、硬件配置、模型范围及对比参数。后续应结合完整技术报告核对“理想性能”的定义，并确认不同负载条件下结果是否一致。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-07-23T01:50:14.766Z","sourceHash":"5aca9dc05a36d4f2","review":{"approved":true,"groundedness":96,"clarity":92,"duplicationRisk":18,"blockingIssues":[],"notes":["“核心关注点不是静态配置专家副本”属于基于机制描述作出的概括性判断，候选内容已用“Aioga 判断”明确标示为观点。","“可能同时影响训练吞吐与推理 prefill 阶段的效率”措辞审慎，并有来源中的两组性能数据支撑。","“现有材料未披露测试环境、硬件配置、模型范围及对比参数”应理解为所提供的摘要和正文摘录未披露这些信息，不宜扩大为对完整原文或技术报告的断言。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","low-source-overlap","no-html","independent-ai-review"]}},"tags":["产品更新","公众号：小红书技术（dots.llm）"],"translations":{"zh-CN":{"title":"小红书与北大开源 UltraEP：面向大规模 MoE 训推的实时负载均衡方案","summary":"小红书与北大提出 UltraEP，首次将基于精确路由信息的实时负载均衡引入生产系统，在每个 microbatch 和每一层动态复制热点专家。在 Qwen3-235B 等模型上，训练吞吐平均达到理想性能的 94.6%，相比 Megatron-LM 提升 42%；推理 prefill 吞吐相比 SGLang 提升 1.56 倍。","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"小红书与北大开源 UltraEP：面向大规模 MoE 训推的实时负载均衡方案 - Aioga AI资讯","description":"小红书与北大提出 UltraEP，首次将基于精确路由信息的实时负载均衡引入生产系统，在每个 microbatch 和每一层动态复制热点专家。在 Qwen3-235B 等模型上，训练吞吐平均达到理想性能的 94.6%，相比 Megatron-LM 提升 42%；推理 prefill 吞吐相比 SGLang 提升 1.56 倍。","url":"https://www.aioga.com/news/cmrsy2z9e03wpbitlqgh25fol/"},"en":{"title":"Xiaohongshu and Peking University Open Source UltraEP: A Real-Time Load Balancing Solution for Large-Scale MoE Training and Inference","summary":"Xiaohongshu and Peking University proposed UltraEP, introducing for the first time real-time load balancing based on precise routing information into production systems, dynamically replicating hot experts at each microbatch and each layer. On models like Qwen3-235B, the training throughput reaches on average 94.6% of the ideal performance, an improvement of 42% compared to Megatron-LM; inference prefill throughput is 1.56 times higher compared to SGLang.","category":"Products","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu and Peking University Open Source UltraEP: A Real-Time Load Balancing Solution for Large-Scale MoE Training and Inference - Aioga AI News","description":"Xiaohongshu and Peking University proposed UltraEP, introducing for the first time real-time load balancing based on precise routing information into production systems, dynamicall...","url":"https://www.aioga.com/en/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:25:14.142Z"},"ja":{"title":"小红書と北京大学のオープンソース UltraEP：大規模 MoE 訓練と推論向けのリアルタイム負荷分散ソリューション","summary":"小红書と北大は UltraEP を提案し、初めて正確なルーティング情報に基づくリアルタイム負荷分散を生産システムに導入し、各マイクロバッチと各層でホットスポットの専門家を動的に複製します。Qwen3-235B などのモデルでは、訓練スループットは平均で理想的な性能の 94.6% に達し、Megatron-LM と比べて 42% 向上しています；推論 prefill スループットは SGLang と比べて 1.56 倍向上しています。","category":"製品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"小红書と北京大学のオープンソース UltraEP：大規模 MoE 訓練と推論向けのリアルタイム負荷分散ソリューション - Aioga AIニュース","description":"小红書と北大は UltraEP を提案し、初めて正確なルーティング情報に基づくリアルタイム負荷分散を生産システムに導入し、各マイクロバッチと各層でホットスポットの専門家を動的に複製します。Qwen3-235B などのモデルでは、訓練スループットは平均で理想的な性能の 94.6% に達し、Megatron-LM と比べて 42% 向上しています；推論 pref...","url":"https://www.aioga.com/ja/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:25:34.173Z"},"ko":{"title":"샤오홍슈와 베이징대 오픈소스 UltraEP: 대규모 MoE 학습 및 추론을 위한 실시간 부하 균형 솔루션","summary":"샤오홍슈와 베이징대는 UltraEP를 제안했으며, 최초로 정확한 라우팅 정보를 기반으로 한 실시간 부하 균형을 생산 시스템에 도입하여 각 마이크로배치와 각 층에서 핫스팟 전문가를 동적으로 복제했습니다. Qwen3-235B 등의 모델에서 훈련 처리량은 평균 94.6%의 이상적인 성능에 도달했으며, Megatron-LM 대비 42% 향상되었습니다; 추론 prefill 처리량은 SGLang 대비 1.56배 향상되었습니다.","category":"제품 업데이트","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"샤오홍슈와 베이징대 오픈소스 UltraEP: 대규모 MoE 학습 및 추론을 위한 실시간 부하 균형 솔루션 - Aioga AI 뉴스","description":"샤오홍슈와 베이징대는 UltraEP를 제안했으며, 최초로 정확한 라우팅 정보를 기반으로 한 실시간 부하 균형을 생산 시스템에 도입하여 각 마이크로배치와 각 층에서 핫스팟 전문가를 동적으로 복제했습니다. Qwen3-235B 등의 모델에서 훈련 처리량은 평균 94.6%의 이상적인 성능에 도달했으며, Megatron-LM 대비...","url":"https://www.aioga.com/ko/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:26:18.916Z"},"es":{"title":"Xiaohongshu y la Universidad de Pekín lanzan UltraEP de código abierto: Solución de balanceo de carga en tiempo real para entrenamiento e inferencia a gran escala de MoE","summary":"Xiaohongshu y la Universidad de Pekín propusieron UltraEP, introduciendo por primera vez el equilibrio de carga en tiempo real basado en información de enrutamiento precisa en sistemas de producción, replicando dinámicamente expertos en hotspots en cada microbatch y en cada capa. En modelos como Qwen3-235B, el rendimiento promedio de entrenamiento alcanzó el 94,6% del rendimiento ideal, un aumento del 42% en comparación con Megatron-LM; el throughput de prellenado de inferencia aumentó 1,56 veces en comparación con SGLang.","category":"Productos","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu y la Universidad de Pekín lanzan UltraEP de código abierto: Solución de balanceo de carga en tiempo real para entrenamiento e inferencia a gran escala de MoE - Aioga Noticias de IA","description":"Xiaohongshu y la Universidad de Pekín propusieron UltraEP, introduciendo por primera vez el equilibrio de carga en tiempo real basado en información de enrutamiento precisa en sist...","url":"https://www.aioga.com/es/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:26:15.635Z"},"fr":{"title":"Xiaohongshu et l'Open Source de l'Université de Pékin UltraEP : Une solution d'équilibrage de charge en temps réel pour l'entraînement et l'inférence à grande échelle de MoE","summary":"Xiaohongshu et l'Université de Pékin ont présenté UltraEP, introduisant pour la première fois l'équilibrage de charge en temps réel basé sur des informations de routage précises dans les systèmes de production, en répliquant dynamiquement les experts très sollicités à chaque micro-lot et à chaque couche. Sur des modèles tels que Qwen3-235B, le débit d'entraînement atteint en moyenne 94,6 % des performances idéales, avec une amélioration de 42 % par rapport à Megatron-LM ; le débit de pré-remplissage pour l'inférence est 1,56 fois supérieur à celui de SGLang.","category":"Produits","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu et l'Open Source de l'Université de Pékin UltraEP : Une solution d'équilibrage de charge en temps réel pour l'entraînement et l'inférence à grande échelle de MoE - Aioga Actualités IA","description":"Xiaohongshu et l'Université de Pékin ont présenté UltraEP, introduisant pour la première fois l'équilibrage de charge en temps réel basé sur des informations de routage précises da...","url":"https://www.aioga.com/fr/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:27:08.477Z"},"de":{"title":"Xiaohongshu und Peking-Universität Open Source UltraEP: Echtzeit-Lastenausgleichslösung für groß angelegte MoE-Training und Inferenz","summary":"Xiaohongshu und Peking University haben UltraEP vorgeschlagen, das erstmals ein auf präzisen Routing-Informationen basierendes Echtzeit-Load-Balancing in Produktionssysteme einführt und in jedem Microbatch sowie in jeder Schicht dynamisch Hotspot-Experten dupliziert. Bei Modellen wie Qwen3-235B erreicht der Trainingsthroughput im Durchschnitt 94,6 % der idealen Leistung, was im Vergleich zu Megatron-LM eine Steigerung von 42 % bedeutet; der Inferenz-Prefill-Throughput steigt im Vergleich zu SGLang um das 1,56-fache.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu und Peking-Universität Open Source UltraEP: Echtzeit-Lastenausgleichslösung für groß angelegte MoE-Training und Inferenz - Aioga KI-News","description":"Xiaohongshu und Peking University haben UltraEP vorgeschlagen, das erstmals ein auf präzisen Routing-Informationen basierendes Echtzeit-Load-Balancing in Produktionssysteme einführ...","url":"https://www.aioga.com/de/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:27:00.487Z"},"pt-BR":{"title":"Xiaohongshu e Peking University Open Source UltraEP: Solução de balanceamento de carga em tempo real para treinamento e inferência de MoE em grande escala","summary":"Xiaohongshu e a Universidade de Pequim propuseram o UltraEP, introduzindo pela primeira vez o balanceamento de carga em tempo real baseado em informações de roteamento precisas em sistemas de produção, replicando dinamicamente especialistas de hotspot em cada microbatch e em cada camada. Em modelos como Qwen3-235B, a taxa média de processamento durante o treinamento alcançou 94,6% do desempenho ideal, um aumento de 42% em comparação ao Megatron-LM; a taxa de preenchimento durante a inferência aumentou 1,56 vezes em comparação ao SGLang.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu e Peking University Open Source UltraEP: Solução de balanceamento de carga em tempo real para treinamento e inferência de MoE em grande escala - Aioga Notícias de IA","description":"Xiaohongshu e a Universidade de Pequim propuseram o UltraEP, introduzindo pela primeira vez o balanceamento de carga em tempo real baseado em informações de roteamento precisas em...","url":"https://www.aioga.com/pt-BR/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:27:50.338Z"},"ru":{"title":"Xiaohongshu и Пекинский университет открывают исходный код UltraEP: решение для балансировки нагрузки в реальном времени для масштабного обучения и推推 MoE","summary":"Xiaohongshu и Пекинский университет предложили UltraEP, впервые внедрив в производственные системы балансировку нагрузки в реальном времени на основе точной информации о маршрутизации, динамически реплицируя горячие эксперты на каждом микропакете и на каждом слое. На моделях типа Qwen3-235B средняя производительность тренировки достигает 94,6% от идеальной, что на 42% выше по сравнению с Megatron-LM; пропускная способность предзаполнения для инференса по сравнению с SGLang увеличена в 1,56 раза.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu и Пекинский университет открывают исходный код UltraEP: решение для балансировки нагрузки в реальном времени для масштабного обучения и推推 MoE - Aioga Новости ИИ","description":"Xiaohongshu и Пекинский университет предложили UltraEP, впервые внедрив в производственные системы балансировку нагрузки в реальном времени на основе точной информации о маршрутиза...","url":"https://www.aioga.com/ru/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:27:56.105Z"},"ar":{"title":"Xiaohongshu وجامعة بكين تطلق UltraEP مفتوح المصدر: حل التوازن الحي للحمل لتدريب ودفع MoE على نطاق واسع","summary":"قدمت Xiaohongshu وجامعة بكين UltraEP، والتي أدخلت لأول مرة التوازن الفوري للحمل المبني على معلومات التوجيه الدقيقة في نظام الإنتاج، مع نسخ الخبراء الساخنين ديناميكيًا في كل microbatch وكل طبقة. في نماذج مثل Qwen3-235B، بلغ متوسط معدل الإرسال خلال التدريب 94.6٪ من الأداء المثالي، بزيادة 42٪ مقارنة بـ Megatron-LM؛ وبالنسبة لمعدل الإرسال أثناء prefill في الاستدلال، ارتفع بمقدار 1.56 ضعف مقارنة بـ SGLang.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu وجامعة بكين تطلق UltraEP مفتوح المصدر: حل التوازن الحي للحمل لتدريب ودفع MoE على نطاق واسع - Aioga أخبار الذكاء الاصطناعي","description":"قدمت Xiaohongshu وجامعة بكين UltraEP، والتي أدخلت لأول مرة التوازن الفوري للحمل المبني على معلومات التوجيه الدقيقة في نظام الإنتاج، مع نسخ الخبراء الساخنين ديناميكيًا في كل microba...","url":"https://www.aioga.com/ar/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:28:43.219Z"},"hi":{"title":"शाओहोंगशू और बीजिंग विश्वविद्यालय ने ओपन सोर्स UltraEP लॉन्च किया: बड़े पैमाने पर MoE प्रशिक्षण और 推送 के लिए रियल-टाइम लोड बैलेंसिंग समाधान","summary":"शाओहोंगशु और बेइजिंग यूनिवर्सिटी ने UltraEP पेश किया, जो पहली बार सटीक रूटिंग जानकारी पर आधारित रीयल-टाइम लोड बैलेंसिंग को प्रोडक्शन सिस्टम में लाता है, प्रत्येक माइक्रोबैच और प्रत्येक लेयर में हॉटस्पॉट विशेषज्ञों की डायनामिक नकल करता है। Qwen3-235B जैसी मॉडलों पर, ट्रेनिंग थ्रूपुट औसतन आदर्श प्रदर्शन का 94.6% तक पहुंचता है, Megatron-LM की तुलना में 42% सुधार; इंफरेंस प्रीफिल थ्रूपुट SGLang की तुलना में 1.56 गुना बढ़ गया।","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"शाओहोंगशू और बीजिंग विश्वविद्यालय ने ओपन सोर्स UltraEP लॉन्च किया: बड़े पैमाने पर MoE प्रशिक्षण और 推送 के लिए रियल-टाइम लोड बैलेंसिंग समाधान - Aioga AI समाचार","description":"शाओहोंगशु और बेइजिंग यूनिवर्सिटी ने UltraEP पेश किया, जो पहली बार सटीक रूटिंग जानकारी पर आधारित रीयल-टाइम लोड बैलेंसिंग को प्रोडक्शन सिस्टम में लाता है, प्रत्येक माइक्रोबैच और प्रत...","url":"https://www.aioga.com/hi/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:28:53.529Z"},"it":{"title":"Xiaohongshu e l'Open Source di Peking University UltraEP: Soluzione di bilanciamento del carico in tempo reale per l'addestramento e l'inferenza su larga scala di MoE","summary":"Xiaohongshu e l'Università di Pechino hanno proposto UltraEP, introducendo per la prima volta il bilanciamento del carico in tempo reale basato su informazioni di routing precise nei sistemi di produzione, replicando dinamicamente gli esperti hotspot in ogni microbatch e in ogni layer. Su modelli come Qwen3-235B, la velocità di addestramento media ha raggiunto il 94,6% delle prestazioni ideali, con un aumento del 42% rispetto a Megatron-LM; la velocità di prefill per l'inferenza è stata superiore di 1,56 volte rispetto a SGLang.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu e l'Open Source di Peking University UltraEP: Soluzione di bilanciamento del carico in tempo reale per l'addestramento e l'inferenza su larga scala di MoE - Aioga Notizie IA","description":"Xiaohongshu e l'Università di Pechino hanno proposto UltraEP, introducendo per la prima volta il bilanciamento del carico in tempo reale basato su informazioni di routing precise n...","url":"https://www.aioga.com/it/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:29:48.842Z"},"nl":{"title":"Xiaohongshu en Peking University Open Source UltraEP: Een realtime load-balancing oplossing voor grootschalige MoE training en inference","summary":"Xiaohongshu en Peking University hebben UltraEP voorgesteld, waarmee voor het eerst real-time load balancing op basis van nauwkeurige routeringsinformatie in productiesystemen wordt geïntroduceerd, met dynamische replicatie van hot-experts in elke microbatch en op elk niveau. Bij modellen zoals Qwen3-235B bereikt de trainingsdoorvoer gemiddeld 94,6% van de ideale prestaties, een verbetering van 42% vergeleken met Megatron-LM; bij inference prefill doorvoer is er een verbetering van 1,56 keer ten opzichte van SGLang.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu en Peking University Open Source UltraEP: Een realtime load-balancing oplossing voor grootschalige MoE training en inference - Aioga AI-nieuws","description":"Xiaohongshu en Peking University hebben UltraEP voorgesteld, waarmee voor het eerst real-time load balancing op basis van nauwkeurige routeringsinformatie in productiesystemen word...","url":"https://www.aioga.com/nl/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:29:38.265Z"},"tr":{"title":"Xiaohongshu ve Peking Üniversitesi Açık Kaynak UltraEP: Büyük Ölçekli MoE Eğitim ve İnferans için Gerçek Zamanlı Yük Dengeleme Çözümü","summary":"Xiaohongshu ve Peking Üniversitesi UltraEP’yi önerdi, ilk kez hassas yönlendirme bilgilerine dayalı gerçek zamanlı yük dengelemesini üretim sistemine getirdi ve her mikrobatch ile her katmanda sıcak uzmanları dinamik olarak kopyaladı. Qwen3-235B gibi modellerde, eğitim verimliliği ortalama olarak ideal performansın %94,6’sına ulaştı ve Megatron-LM’ye kıyasla %42 artış sağladı; çıkarım prefill verimliliği ise SGLang’e kıyasla 1,56 kat artırıldı.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu ve Peking Üniversitesi Açık Kaynak UltraEP: Büyük Ölçekli MoE Eğitim ve İnferans için Gerçek Zamanlı Yük Dengeleme Çözümü - Aioga AI Haberleri","description":"Xiaohongshu ve Peking Üniversitesi UltraEP’yi önerdi, ilk kez hassas yönlendirme bilgilerine dayalı gerçek zamanlı yük dengelemesini üretim sistemine getirdi ve her mikrobatch ile...","url":"https://www.aioga.com/tr/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:30:44.783Z"},"vi":{"title":"Xiaohongshu và Đại học Bắc Kinh mở nguồn UltraEP: Giải pháp cân bằng tải thời gian thực cho đào tạo và推送 MoE quy mô lớn","summary":"Xiaohongshu và Bắc Đại đề xuất UltraEP, lần đầu tiên đưa cân bằng tải thời gian thực dựa trên thông tin định tuyến chính xác vào hệ thống sản xuất, sao chép các chuyên gia điểm nóng một cách động trong mỗi microbatch và mỗi lớp. Trên các mô hình như Qwen3-235B, thông lượng huấn luyện trung bình đạt 94,6% hiệu suất lý tưởng, tăng 42% so với Megatron-LM; thông lượng prefill suy luận tăng 1,56 lần so với SGLang.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu và Đại học Bắc Kinh mở nguồn UltraEP: Giải pháp cân bằng tải thời gian thực cho đào tạo và推送 MoE quy mô lớn - Tin tức AI Aioga","description":"Xiaohongshu và Bắc Đại đề xuất UltraEP, lần đầu tiên đưa cân bằng tải thời gian thực dựa trên thông tin định tuyến chính xác vào hệ thống sản xuất, sao chép các chuyên gia điểm nón...","url":"https://www.aioga.com/vi/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:30:41.463Z"},"id":{"title":"Xiaohongshu dan Peking University membuka UltraEP: Solusi penyeimbangan beban waktu nyata untuk pelatihan dan inferensi MoE skala besar","summary":"Xiaohongshu dan Peking University mengusulkan UltraEP, pertama kali memperkenalkan penyeimbangan beban waktu nyata berbasis informasi routing yang tepat ke dalam sistem produksi, secara dinamis menggandakan pakar hotspot di setiap microbatch dan setiap lapisan. Pada model seperti Qwen3-235B, throughput pelatihan rata-rata mencapai 94,6% dari performa ideal, meningkat 42% dibandingkan Megatron-LM; throughput prefill inferensi meningkat 1,56 kali dibandingkan SGLang.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu dan Peking University membuka UltraEP: Solusi penyeimbangan beban waktu nyata untuk pelatihan dan inferensi MoE skala besar - Berita AI Aioga","description":"Xiaohongshu dan Peking University mengusulkan UltraEP, pertama kali memperkenalkan penyeimbangan beban waktu nyata berbasis informasi routing yang tepat ke dalam sistem produksi, s...","url":"https://www.aioga.com/id/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:31:30.130Z"},"th":{"title":"Xiaohongshu และ ปักกิ่งมหาวิทยาลัย Open Source UltraEP: โซลูชันการปรับสมดุลโหลดแบบเรียลไทม์สำหรับการฝึกและ推 推 MoE ขนาดใหญ่","summary":"Xiaohongshu และ Peking University ได้นำเสนอ UltraEP เป็นครั้งแรกที่นำการปรับสมดุลโหลดแบบเรียลไทม์ที่อิงกับข้อมูลการกำหนดเส้นทางที่แม่นยำเข้าสู่ระบบการผลิต โดยทำการคัดลอกผู้เชี่ยวชาญยอดนิยมแบบไดนามิกในแต่ละ microbatch และแต่ละชั้น ในโมเดลเช่น Qwen3-235B ประสิทธิภาพการฝึกอบรมเฉลี่ยสูงถึง 94.6% ของประสิทธิภาพที่ต้องการ เทียบกับ Megatron-LM เพิ่มขึ้น 42%; ประสิทธิภาพการ prefill ในการคาดคะเนเพิ่มขึ้น 1.56 เท่าเมื่อเทียบกับ SGLang.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu และ ปักกิ่งมหาวิทยาลัย Open Source UltraEP: โซลูชันการปรับสมดุลโหลดแบบเรียลไทม์สำหรับการฝึกและ推 推 MoE ขนาดใหญ่ - ข่าว AI Aioga","description":"Xiaohongshu และ Peking University ได้นำเสนอ UltraEP เป็นครั้งแรกที่นำการปรับสมดุลโหลดแบบเรียลไทม์ที่อิงกับข้อมูลการกำหนดเส้นทางที่แม่นยำเข้าสู่ระบบการผลิต โดยทำการคัดลอกผู้เชี่ยวชา...","url":"https://www.aioga.com/th/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:31:37.259Z"},"pl":{"title":"Xiaohongshu i Uniwersytet Pekin Open Source UltraEP: rozwiązanie do równoważenia obciążenia w czasie rzeczywistym dla treningu i推推 dużej skali MoE","summary":"Xiaohongshu i Uniwersytet Pekiński zaproponowały UltraEP, po raz pierwszy wprowadzając do systemów produkcyjnych równoważenie obciążenia w czasie rzeczywistym oparte na dokładnych informacjach o trasowaniu, dynamicznie replikując gorące eksperckie w każdym mikropartii i każdej warstwie. W modelach takich jak Qwen3-235B wydajność treningu osiąga średnio 94,6% idealnej, co stanowi wzrost o 42% w porównaniu z Megatron-LM; przepustowość prefill w wnioskowaniu wzrosła 1,56 razy w porównaniu z SGLang.","category":"产品更新","source":"公众号：小红书技术（dots.llm）","aggregationSource":"公众号：小红书技术（dots.llm）","pageTitle":"Xiaohongshu i Uniwersytet Pekin Open Source UltraEP: rozwiązanie do równoważenia obciążenia w czasie rzeczywistym dla treningu i推推 dużej skali MoE - Aioga Wiadomości AI","description":"Xiaohongshu i Uniwersytet Pekiński zaproponowały UltraEP, po raz pierwszy wprowadzając do systemów produkcyjnych równoważenie obciążenia w czasie rzeczywistym oparte na dokładnych...","url":"https://www.aioga.com/pl/news/cmrsy2z9e03wpbitlqgh25fol/","contentTranslated":true,"sourceHash":"bc3cff117545f449","translatedAt":"2026-07-22T17:32:23.425Z"}}}}