[{"data":1,"prerenderedAt":213},["ShallowReactive",2],{"content-query-xx8H01cQkp":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":207,"_id":208,"_source":209,"_file":210,"_stem":211,"_extension":212},"/news/zh/3832","zh",false,"","昇思亮相GOSIM 2025，分享SGLang支持方案与算子自动生成等最新技术进展","未来，昇思社区将继续深耕AI框架技术，推动开源生态繁荣发展。","2025-09-14","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/09/19/3df0f83fc0274fcab40009c2cb2b0c1e.png","news",{"type":14,"children":15,"toc":204},"root",[16,24,38,54,64,72,77,97,102,117,127,134,139,162,167,182,187,194,199],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"昇思亮相gosim-2025分享sglang支持方案与算子自动生成等最新技术进展",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28,30,36],{"type":23,"value":29},"2025年9月13-14日，",{"type":17,"tag":31,"props":32,"children":33},"strong",{},[34],{"type":23,"value":35},"GOSIM HANGZHOU 2025",{"type":23,"value":37},"在杭州成功举办。本次大会汇聚了1500多名开发者和100多位海内外专家，围绕开源与AI技术展开深入交流。昇思MindSpore开源社区在此次盛会上精彩亮相，分享技术创新成果，展示昇思全栈能力，获得与会者的关注。",{"type":17,"tag":18,"props":39,"children":41},{"id":40},"_01-sglang与mindspore-ai框架的深度融合",[42,47,49],{"type":17,"tag":31,"props":43,"children":44},{},[45],{"type":23,"value":46},"# 01",{"type":23,"value":48}," ",{"type":17,"tag":31,"props":50,"children":51},{},[52],{"type":23,"value":53},"SGLang与MindSpore AI框架的深度融合",{"type":17,"tag":25,"props":55,"children":56},{},[57,62],{"type":17,"tag":31,"props":58,"children":59},{},[60],{"type":23,"value":61},"华为人工智能技术专家王天策",{"type":23,"value":63},"分享了《SGLang使能MindSpore框架方案》，介绍了如何将MindSpore高效集成至SGLang这一大语言模型服务框架中，以提升LLM推理性能与开发效率。",{"type":17,"tag":25,"props":65,"children":66},{},[67],{"type":17,"tag":68,"props":69,"children":71},"img",{"alt":7,"src":70},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/09/19/e5681c6b57fc46a7952fb6401db719cb.png",[],{"type":17,"tag":25,"props":73,"children":74},{},[75],{"type":23,"value":76},"在SGLang中集成MindSpore的关键方案：",{"type":17,"tag":78,"props":79,"children":80},"ul",{},[81,87,92],{"type":17,"tag":82,"props":83,"children":84},"li",{},[85],{"type":23,"value":86},"最小化接口设计：通过新增mindspore_runner.py（提供分布式能力）和MindsporeForCausalLM类封装模型，保持与HuggingFace Transformers类似的使用体验，用户仅需添加一行配置即可启用MindSpore后端。",{"type":17,"tag":82,"props":88,"children":89},{},[90],{"type":23,"value":91},"跨框架资源共享：通过DLPack实现MindSpore与PyTorch间的Tensor无缝转换，大幅减少内存开销与数据传输延迟。",{"type":17,"tag":82,"props":93,"children":94},{},[95],{"type":23,"value":96},"通信组复用：SGLang初始化时创建的通信组可直接被MindSpore复用，避免重复构建，提升分布式推理效率。",{"type":17,"tag":25,"props":98,"children":99},{},[100],{"type":23,"value":101},"该集成方案目前已支持Qwen3系列的稠密模型与MOE模型和Tensor Parallel等关键特性，未来还将支持DeepSeek V3模型以及DP Attention、MoE专家并行、PD分离、量化、投机解码和LoRA等SGLang优化特性，进一步拓展MindSpore在LLM推理与服务领域的生态影响力。",{"type":17,"tag":18,"props":103,"children":105},{"id":104},"_02-大模型驱动的算子自动生成技术",[106,111,112],{"type":17,"tag":31,"props":107,"children":108},{},[109],{"type":23,"value":110},"# 02",{"type":23,"value":48},{"type":17,"tag":31,"props":113,"children":114},{},[115],{"type":23,"value":116},"大模型驱动的算子自动生成技术",{"type":17,"tag":25,"props":118,"children":119},{},[120,125],{"type":17,"tag":31,"props":121,"children":122},{},[123],{"type":23,"value":124},"MindSpore AKG SIG Maintainer张任伟",{"type":23,"value":126},"分享了题为《AIKG：大模型驱动的算子生成技术探索》的议题，系统介绍了基于大语言模型（LLM）的下一代算子自动生成技术AIKG（AI Driven Kernel Generator）。",{"type":17,"tag":25,"props":128,"children":129},{},[130],{"type":17,"tag":68,"props":131,"children":133},{"alt":7,"src":132},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/09/19/ee43e1ff1ccf42828a199b050dc33cc6.png",[],{"type":17,"tag":25,"props":135,"children":136},{},[137],{"type":23,"value":138},"与传统AI编译技术不同，AIKG构建了一套以LLM为核心、多智能体（Multi-Agent）协同的算子生成框架。该框架通过“文档驱动”接入机制，支持多种前端框架、DSL和硬件后端的灵活扩展，显著降低定制化开发与维护成本。其核心创新包括：",{"type":17,"tag":78,"props":140,"children":141},{},[142,147,152,157],{"type":17,"tag":82,"props":143,"children":144},{},[145],{"type":23,"value":146},"Unified Sketch设计：将算子生成过程拆分为“草图设计”与“代码生成”两阶段，提升生成可控性与优化效果；",{"type":17,"tag":82,"props":148,"children":149},{},[150],{"type":23,"value":151},"Conductor智能调度：通过中央调度Agent动态管理任务流，提升复杂场景下的生成效率与成功率；",{"type":17,"tag":82,"props":153,"children":154},{},[155],{"type":23,"value":156},"多层级检索增强生成（RAG）：结合算子特征提取与混合检索策略，增强模型对历史优化方案的理解与复用；",{"type":17,"tag":82,"props":158,"children":159},{},[160],{"type":23,"value":161},"搜索优化框架：AIKG引入了搜索优化支持，通过多轮迭代、LLM自主分析优化方向的方式，持续优化算子代码性能表现，部分场景性能提升400%。",{"type":17,"tag":25,"props":163,"children":164},{},[165],{"type":23,"value":166},"目前，AIKG已在多个硬件平台上实现高效算子生成。在KernelBench评测中，AIKG在算子生成任务上表现稳定，并构建了覆盖230+用例的AIKG-Bench评测体系，支持静态/动态Shape等多种场景，为大模型、推荐系统、科学计算等AI应用提供高性能算子支持。",{"type":17,"tag":18,"props":168,"children":170},{"id":169},"_03-展览互动昇思社区成果与进展",[171,176,177],{"type":17,"tag":31,"props":172,"children":173},{},[174],{"type":23,"value":175},"# 03",{"type":23,"value":48},{"type":17,"tag":31,"props":178,"children":179},{},[180],{"type":23,"value":181},"展览互动：昇思社区成果与进展",{"type":17,"tag":25,"props":183,"children":184},{},[185],{"type":23,"value":186},"在大会展览区，昇思社区设置了专门的展示交流区，通过技术展示和面对面交流，全面呈现了框架的最新成果与社区成果。展台吸引了开发者驻足交流，通过深度交流了解昇思MindSpore AI框架的技术优势和应用价值。",{"type":17,"tag":25,"props":188,"children":189},{},[190],{"type":17,"tag":68,"props":191,"children":193},{"alt":7,"src":192},"https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2025/09/19/1eefd675ab7c4a4daf5913aea5b4bb04.png",[],{"type":17,"tag":25,"props":195,"children":196},{},[197],{"type":23,"value":198},"会议期间，昇思MindSpore开源社区与开发者、企业代表和学术界专家就AI框架发展趋势和开源社区建设进行深度交流。",{"type":17,"tag":25,"props":200,"children":201},{},[202],{"type":23,"value":203},"通过本次GOSIM大会，昇思社区展示了在AI框架领域的技术实力，与全球开源社区建立了更加紧密的联系。AIKG与SGLang两大技术方案的分享，展示了昇思MindSpore在性能优化、生态扩展和开发者体验方面的持续创新。未来，昇思社区将继续深耕AI框架技术，推动开源生态繁荣发展。",{"title":7,"searchDepth":205,"depth":205,"links":206},4,[],"markdown","content:news:zh:3832.md","content","news/zh/3832.md","news/zh/3832","md",1776506090508]