[{"data":1,"prerenderedAt":511},["ShallowReactive",2],{"content-query-BApauoeIxR":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"category":13,"body":14,"_type":505,"_id":506,"_source":507,"_file":508,"_stem":509,"_extension":510},"/technology-blogs/zh/2026-4-7","zh",false,"","AKG Agent：大模型驱动的算子自动生成技术探索","希望沉淀下来的经验能够服务更广泛的场景：算子迁移、性能调优、子图分析、算子拆分与融合……","2026-4-7","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2024/11/28/8e0e0150508a4c5ba4287fa3bec8ea3f.png","technology-blogs","技术解读",{"type":15,"children":16,"toc":479},"root",[17,25,31,36,43,48,55,60,65,70,75,80,85,91,96,107,113,119,124,129,134,139,144,149,155,160,165,170,175,180,185,191,196,201,206,211,216,221,226,231,236,242,247,253,260,266,271,276,281,286,292,299,304,309,314,319,324,330,335,340,345,350,355,361,368,373,378,384,389,394,404,409,416,422,427,432,437,442,448,453,458,469,474],{"type":18,"tag":19,"props":20,"children":22},"element","h1",{"id":21},"akg-agent大模型驱动的算子自动生成技术探索",[23],{"type":24,"value":8},"text",{"type":18,"tag":26,"props":27,"children":28},"p",{},[29],{"type":24,"value":30},"在AI模型与软硬件快速演进的今天，高质量算子已成为制约AI性能提升的关键瓶颈。我们看到一个有趣的趋势：算子自动生成技术正从\"学术研究\"走向\"工程实用\"。特别是2025年以来，随着大模型代码生成能力的显著提升，业界开始系统性地探索如何将 LLM 应用于高性能算子开发这一垂直领域。各大高校、AI厂商纷纷投入研发，期望通过智能化手段解决算子开发的效率和可移植性问题。",{"type":18,"tag":26,"props":32,"children":33},{},[34],{"type":24,"value":35},"AKG 团队自2025年初起持续投入这一方向，并取得了一系列成果（昇思MindSpore AKG大模型驱动算子生成）。与此同时，我们的思路也在不断演进——与其为单一具体问题打造一个专用工具，不如构建一个通用的多Agent协作框架？从初期的只聚焦算子生成，到如今的 AKG Agent 定位为通用AI Infra多Agent协作平台，我们希望沉淀下来的经验能够服务更广泛的场景：算子迁移、性能调优、子图分析、算子拆分与融合……",{"type":18,"tag":37,"props":38,"children":40},"h2",{"id":39},"_01akg-agent-框架简介",[41],{"type":24,"value":42},"01AKG Agent 框架简介",{"type":18,"tag":26,"props":44,"children":45},{},[46],{"type":24,"value":47},"AKG Agent 是一个面向 AI Infra 与高性能计算场景的 LLM 多 Agent 协作框架，通过智能 Agent 协同，将高性能代码的开发与优化效率提升到新的高度。",{"type":18,"tag":49,"props":50,"children":52},"h3",{"id":51},"_11设计理念",[53],{"type":24,"value":54},"1.1设计理念",{"type":18,"tag":26,"props":56,"children":57},{},[58],{"type":24,"value":59},"AKG Agent 采用了通用的架构设计：构建一个可复用的多 Agent 协作基础框架，让不同领域的 AI 工作都能受益于智能化的代码生成与优化。",{"type":18,"tag":26,"props":61,"children":62},{},[63],{"type":24,"value":64},"在这个设计理念下，AKG Agent 具备以下特点：",{"type":18,"tag":26,"props":66,"children":67},{},[68],{"type":24,"value":69},"通用性：核心框架不绑定具体场景，可扩展至算子迁移、性能调优、子图分析与算子融合等更多领域",{"type":18,"tag":26,"props":71,"children":72},{},[73],{"type":24,"value":74},"可组合性：Agent、Skill、Tools 可灵活组合，适应不同任务需求",{"type":18,"tag":26,"props":76,"children":77},{},[78],{"type":24,"value":79},"可观测性：完整的Trace追踪系统，让AI决策过程透明可控",{"type":18,"tag":26,"props":81,"children":82},{},[83],{"type":24,"value":84},"可扩展性：通过Skill系统和Agent注册机制，社区可轻松扩展能力",{"type":18,"tag":49,"props":86,"children":88},{"id":87},"_12核心架构",[89],{"type":24,"value":90},"1.2核心架构",{"type":18,"tag":26,"props":92,"children":93},{},[94],{"type":24,"value":95},"AKG Agent 提供了完整的 Agent 基础设施，整体架构如下：",{"type":18,"tag":97,"props":98,"children":100},"div",{"style":99},"text-align: center;",[101],{"type":18,"tag":102,"props":103,"children":106},"img",{"src":104,"style":105,"alt":7},"/category/information/technology-blogs/banner/2026-4-7/1.jpg","display: block;margin: 0 auto;max-width:60%",[],{"type":18,"tag":37,"props":108,"children":110},{"id":109},"_02关键设计",[111],{"type":24,"value":112},"02关键设计",{"type":18,"tag":49,"props":114,"children":116},{"id":115},"_21skill-系统让-agent-学会按需取用",[117],{"type":24,"value":118},"2.1Skill 系统：让 Agent 学会\"按需取用\"",{"type":18,"tag":26,"props":120,"children":121},{},[122],{"type":24,"value":123},"早期方案把所有知识文档一股脑塞给 LLM，生成简单算子和复杂算子喂的是同一份“大部头”——就像考每一科都带着全部课本，信息过载反而找不到重点。",{"type":18,"tag":26,"props":125,"children":126},{},[127],{"type":24,"value":128},"AKG Agent 的 Skill 系统改变了这一做法。我们将领域知识拆解为一个个独立的 Skill 单元（以 SKILL.md 格式存储），每个 Skill 聚焦于一个特定主题——比如\"triton-ascend-basics\"。当任务到来时，系统先根据元数据（目标硬件、DSL 类型等）快速缩小候选范围，再由 LLM 根据当前任务的具体特点，智能挑选最相关的少数几个 Skill 注入到上下文中。",{"type":18,"tag":26,"props":130,"children":131},{},[132],{"type":24,"value":133},"这带来了显著的好处：",{"type":18,"tag":26,"props":135,"children":136},{},[137],{"type":24,"value":138},"Prompt 长度大幅缩短，Agent 能够更专注于当前任务真正需要的知识；",{"type":18,"tag":26,"props":140,"children":141},{},[142],{"type":24,"value":143},"不同场景、不同硬件自动加载不同的知识组合；",{"type":18,"tag":26,"props":145,"children":146},{},[147],{"type":24,"value":148},"成功经验可以沉淀为新 Skill ，社区开发者也能通过提交 SKILL.md 轻松贡献知识。",{"type":18,"tag":49,"props":150,"children":152},{"id":151},"_22trace-系统给-ai-决策装上行车记录仪",[153],{"type":24,"value":154},"2.2Trace 系统：给 AI 决策装上\"行车记录仪\"",{"type":18,"tag":26,"props":156,"children":157},{},[158],{"type":24,"value":159},"AKG Agent 的 Trace 系统用一棵树状结构记录了 AI 的完整决策历程。每当 React Agent 做出一个决策——选择了哪个工具、收到了什么反馈——都会作为一个节点挂在这棵树上。与线性日志不同，树状结构天然支持\"分叉\"：当某一步有多个可能的方向时，你可以在同一个节点上探索不同的分支，就像游戏里的存档读档。",{"type":18,"tag":26,"props":161,"children":162},{},[163],{"type":24,"value":164},"这就意味着：",{"type":18,"tag":26,"props":166,"children":167},{},[168],{"type":24,"value":169},"开发者可以随时回溯到任意历史节点，查看当时的完整状态（包括代码快照、Agent 的输入输出等）",{"type":18,"tag":26,"props":171,"children":172},{},[173],{"type":24,"value":174},"从某个节点重新出发，尝试不同的策略。",{"type":18,"tag":26,"props":176,"children":177},{},[178],{"type":24,"value":179},"程序中途崩溃也不要紧，所有状态都持久化在文件系统中，恢复后可以从上次中断的地方继续。",{"type":18,"tag":26,"props":181,"children":182},{},[183],{"type":24,"value":184},"对于复杂算子的生成与优化，这种\"时光机\"般的能力让调试效率有了质的提升。",{"type":18,"tag":49,"props":186,"children":188},{"id":187},"_23agent-与工作流灵活性与稳定性的平衡",[189],{"type":24,"value":190},"2.3Agent 与工作流：灵活性与稳定性的平衡",{"type":18,"tag":26,"props":192,"children":193},{},[194],{"type":24,"value":195},"用 LLM 来驱动复杂任务，有一个永恒的矛盾：你既希望 AI 足够灵活，能根据情况自主判断下一步做什么；又希望关键流程足够稳定，不会因为模型的一次\"脑洞大开\"而跑偏。AKG Agent 通过两套互补的执行机制来解决这个问题。",{"type":18,"tag":26,"props":197,"children":198},{},[199],{"type":24,"value":200},"ReAct Agent 代表的是灵活的一面。",{"type":18,"tag":26,"props":202,"children":203},{},[204],{"type":24,"value":205},"实现了经典的 “推理-行动”循环：观察状态 → 推理下一步 → 选择并执行工具 → 根据结果决定后续动作。整个过程由 LLM 自主驱动，没有预设路径。",{"type":18,"tag":26,"props":207,"children":208},{},[209],{"type":24,"value":210},"适用场景：探索性、开放性的任务。比如在 OpenCode 工作台中，用户说“帮我优化这个模型的性能”，ReAct Agent 可以自主决定：先做环境检测 → 再分析瓶颈 → 最后选择优化策略——每一步都灵活应变。",{"type":18,"tag":26,"props":212,"children":213},{},[214],{"type":24,"value":215},"基于 LangGraph 的工作流 则代表稳定的一面。",{"type":18,"tag":26,"props":217,"children":218},{},[219],{"type":24,"value":220},"将多 Agent 协作的流程预先编排为一张有向图：哪个 Agent 先执行、结果传给谁、失败了怎么重试、最多重试几次——这些都是确定性的。比如算子生成的 Default 工作流：Designer → Coder → Verifier 的流转路径是固定的，不会因为 LLM 的随机性而跳过验证环节。路由函数还内置了步数限制和防循环机制，确保流程不会失控。",{"type":18,"tag":26,"props":222,"children":223},{},[224],{"type":24,"value":225},"💡 一句话总结\n工作流：适合有明确范式的任务（如算子生成、代码迁移），保证流程的可靠性和可复现性；",{"type":18,"tag":26,"props":227,"children":228},{},[229],{"type":24,"value":230},"ReAct Agent 适合需要自主决策的场景（如交互式问题排查、开放式优化探索），充分发挥 LLM 的推理能力。",{"type":18,"tag":26,"props":232,"children":233},{},[234],{"type":24,"value":235},"两者在 AKG Agent 中并不互斥——实际上，工作流中的每个节点本身就可以是一个 ReAct Agent，在预定义的流程框架内保留局部的灵活性。",{"type":18,"tag":37,"props":237,"children":239},{"id":238},"_03落地场景ai-算子自动生成",[240],{"type":24,"value":241},"03落地场景：AI 算子自动生成",{"type":18,"tag":26,"props":243,"children":244},{},[245],{"type":24,"value":246},"AKG Agent 的首个落地场景是 AI 算子代码生成——这正是原先项目的核心能力。在AKG Agent 框架下，我们重构了算子生成流程，使其更加清晰、可维护、可扩展。",{"type":18,"tag":49,"props":248,"children":250},{"id":249},"_31支持的后端与-dsl",[251],{"type":24,"value":252},"3.1支持的后端与 DSL",{"type":18,"tag":97,"props":254,"children":255},{"style":99},[256],{"type":18,"tag":102,"props":257,"children":259},{"src":258,"style":105,"alt":7},"/category/information/technology-blogs/banner/2026-4-7/2.jpg",[],{"type":18,"tag":49,"props":261,"children":263},{"id":262},"_32工作流",[264],{"type":24,"value":265},"3.2工作流",{"type":18,"tag":26,"props":267,"children":268},{},[269],{"type":24,"value":270},"算子生产Agent 围绕算子生成的不同阶段需求，提供了三种预定义工作流。",{"type":18,"tag":26,"props":272,"children":273},{},[274],{"type":24,"value":275},"Default 工作流是最完整的流程，走\"设计→生成→验证\"的全链路——先由 Designer 分析需求并生成高层次的算法草图，再由 Coder 将草图转化为具体 DSL 代码，最后由 Verifier 编译运行并与参考实现对比验证，验证失败时 Conductor 会分析错误并引导 Coder 自动修正，如此迭代直到通过。",{"type":18,"tag":26,"props":277,"children":278},{},[279],{"type":24,"value":280},"KernelGenOnly 工作流则更加轻量，跳过独立的设计阶段，由 KernelGen 直接基于 Skill 知识生成代码后进入验证-修复循环，适合模式较为明确、不需要复杂设计推理的算子。",{"type":18,"tag":26,"props":282,"children":283},{},[284],{"type":24,"value":285},"VerifierOnly 工作流不涉及任何代码生成，仅对用户传入的已有代码执行正确性验证与性能测量，适合手写算子的快速检验或性能基准测试。",{"type":18,"tag":49,"props":287,"children":289},{"id":288},"_33算子优化策略",[290],{"type":24,"value":291},"3.3算子优化策略",{"type":18,"tag":293,"props":294,"children":296},"h4",{"id":295},"_331evolve-算法详解",[297],{"type":24,"value":298},"3.3.1Evolve 算法详解",{"type":18,"tag":26,"props":300,"children":301},{},[302],{"type":24,"value":303},"Evolve 采用进化算法思想，通过多轮迭代和变异操作，持续探索更优的算子实现：",{"type":18,"tag":26,"props":305,"children":306},{},[307],{"type":24,"value":308},"核心特点：",{"type":18,"tag":26,"props":310,"children":311},{},[312],{"type":24,"value":313},"同步迭代：每一轮需要等待所有并行任务全部完成后才进入下一轮选择-变异，轮间存在等待开销",{"type":18,"tag":26,"props":315,"children":316},{},[317],{"type":24,"value":318},"岛屿隔离模型：将种群划分为多个独立岛屿，各岛屿独立进化并定期迁移精英个体，保持种群多样性，避免过早收敛",{"type":18,"tag":26,"props":320,"children":321},{},[322],{"type":24,"value":323},"精英池机制：维护全局精英池，从多个岛屿汇聚最优个体，变异来源更丰富",{"type":18,"tag":49,"props":325,"children":327},{"id":326},"_332adaptive-search-算法详解",[328],{"type":24,"value":329},"3.3.2Adaptive Search 算法详解",{"type":18,"tag":26,"props":331,"children":332},{},[333],{"type":24,"value":334},"Adaptive Search 采用UCB（Upper Confidence Bound）策略，实现自适应的异步流水线搜索：",{"type":18,"tag":26,"props":336,"children":337},{},[338],{"type":24,"value":339},"核心优势：",{"type":18,"tag":26,"props":341,"children":342},{},[343],{"type":24,"value":344},"即时递补：任意任务一完成，立即基于当前最优父代启动新任务，无需等待同批次其他任务结束，最大化资源利用率",{"type":18,"tag":26,"props":346,"children":347},{},[348],{"type":24,"value":349},"UCB 智能选择：每次扩展时基于 UCB（Upper Confidence Bound）策略选择最有潜力的父代进化，自动平衡探索与利用，趋向极致性能",{"type":18,"tag":26,"props":351,"children":352},{},[353],{"type":24,"value":354},"收敛更快：持续滚动式搜索，相同时间内可完成更多轮迭代，更快找到高性能实现",{"type":18,"tag":37,"props":356,"children":358},{"id":357},"_04opencode-集成从工具到工作台",[359],{"type":24,"value":360},"04OpenCode 集成：从工具到工作台",{"type":18,"tag":97,"props":362,"children":363},{"style":99},[364],{"type":18,"tag":102,"props":365,"children":367},{"src":366,"style":105,"alt":7},"/category/information/technology-blogs/banner/2026-4-7/3.jpg",[],{"type":18,"tag":26,"props":369,"children":370},{},[371],{"type":24,"value":372},"AKG Agent 与 OpenCode 深度集成，构建了一个完整的 AI Infra 智能工作台。OpenCode 通过内置的 Agent 和 Skills 机制，实现了对 AKG Agent 的无缝调用。用户在 OpenCode 工作台中，只需用自然语言描述需求，OpenCode 会自动识别任务类型，并通过 akg_cli 将算子相关的专有任务委托给 AKG Agent 框架处理。整个过程对用户透明——环境检测、任务构建、工作流调度、结果验证都在后台自动完成。",{"type":18,"tag":26,"props":374,"children":375},{},[376],{"type":24,"value":377},"这意味着开发者有了两种互补的使用方式：akg_cli 命令行直接调用AKG Agent 执行算子相关任务；OpenCode 工作台则提供交互式的对话体验，间歇性调用AKG能力，适合探索性的优化任务和快速原型验证。两种入口调用的是同一套 AKG Agent 框架，能力完全一致，开发者可以根据场景自由选择。",{"type":18,"tag":37,"props":379,"children":381},{"id":380},"_05akg-cli开箱即用",[382],{"type":24,"value":383},"05AKG CLI：开箱即用",{"type":18,"tag":26,"props":385,"children":386},{},[387],{"type":24,"value":388},"AKG Agent 提供了 akg_cli 命令行工具，让开发者可以快速上手。",{"type":18,"tag":26,"props":390,"children":391},{},[392],{"type":24,"value":393},"具体安装配置：",{"type":18,"tag":395,"props":396,"children":398},"pre",{"code":397},"# 1. 创建环境\nconda create -n akg_agents python=3.11\nconda activate akg_agents\n# 2. 克隆仓库\ngit clone https://gitcode.com/mindspore/akg.git -b br_agents\ncd akg\n# 3. 安装依赖\npip install -r akg_agents/requirements.txt\npip install -e ./akg_agents --no-build-isolation\n# 4. 快速配置llm\nexport AKG_AGENTS_BASE_URL=\"https://api.deepseek.com/beta/\"\nexport AKG_AGENTS_API_KEY=\"your key\"\nexport AKG_AGENTS_MODEL_NAME=\"deepseek-chat\"\nexport AKG_AGENTS_MODEL_ENABLE_THINK=\"disabled\"\n",[399],{"type":18,"tag":400,"props":401,"children":402},"code",{"__ignoreMap":7},[403],{"type":24,"value":397},{"type":18,"tag":26,"props":405,"children":406},{},[407],{"type":24,"value":408},"实机演示：",{"type":18,"tag":97,"props":410,"children":411},{"style":99},[412],{"type":18,"tag":102,"props":413,"children":415},{"src":414,"style":105,"alt":7},"/category/information/technology-blogs/banner/2026-4-7/4.jpg",[],{"type":18,"tag":37,"props":417,"children":419},{"id":418},"_06结语",[420],{"type":24,"value":421},"06结语",{"type":18,"tag":26,"props":423,"children":424},{},[425],{"type":24,"value":426},"AKG Agent 项目是 AKG 团队在 Agent 时代的一次重要尝试。从 2025 到 2026，我们的思路也在不断演进：",{"type":18,"tag":26,"props":428,"children":429},{},[430],{"type":24,"value":431},"我们之前专注于解决算子生成这一具体问题，探索了 Unified Sketch、Conductor、文档驱动式接入等关键技术方案，取得了不错的成果。而现在，我们将这些经验沉淀为通用的多 Agent 协作框架，希望让更多 AI Infra 相关场景都能受益于智能化的代码生成与优化。",{"type":18,"tag":26,"props":433,"children":434},{},[435],{"type":24,"value":436},"无论是哪个硬件平台，快速迭代、高效开发与持续优化算子始终是刚需。随着模型架构日新月异、硬件能力不断演进，算子开发的效率和质量直接决定了AI系统的落地速度。而 Agent + 通用框架的路线，目前看来有望从根本上提升这一环节的生产力。",{"type":18,"tag":26,"props":438,"children":439},{},[440],{"type":24,"value":441},"AKG Agent 会持续在这个方向上快速迭代演进，希望在 AI 民主化、算子民主化的过程中做积极的贡献。",{"type":18,"tag":37,"props":443,"children":445},{"id":444},"_07实战赛事邀请",[446],{"type":24,"value":447},"07实战赛事邀请",{"type":18,"tag":26,"props":449,"children":450},{},[451],{"type":24,"value":452},"当前 AKG Agent 已作为 AKG 项目子仓开源（br_agents分支），相关最新进展也将会在 AKG SIG 例会、知乎专栏上进行分享，欢迎大家使用、评论及参与共建。",{"type":18,"tag":26,"props":454,"children":455},{},[456],{"type":24,"value":457},"项目地址：",{"type":18,"tag":26,"props":459,"children":460},{},[461],{"type":18,"tag":462,"props":463,"children":467},"a",{"href":464,"rel":465},"https://gitcode.com/mindspore/akg/tree/br_agents",[466],"nofollow",[468],{"type":24,"value":464},{"type":18,"tag":26,"props":470,"children":471},{},[472],{"type":24,"value":473},"相关文档：项目提供了详细的架构、Agent 体系、Skill 系统、工作流等文档。",{"type":18,"tag":26,"props":475,"children":476},{},[477],{"type":24,"value":478},"纸上得来终觉浅，想要亲手验证 AKG Agent 的硬核实力？不妨报名参加第八届 CCF 开源创新大赛，用这套技术栈冲击 10万奖金池！",{"title":7,"searchDepth":480,"depth":480,"links":481},4,[482,488,493,501,502,503,504],{"id":39,"depth":483,"text":42,"children":484},2,[485,487],{"id":51,"depth":486,"text":54},3,{"id":87,"depth":486,"text":90},{"id":109,"depth":483,"text":112,"children":489},[490,491,492],{"id":115,"depth":486,"text":118},{"id":151,"depth":486,"text":154},{"id":187,"depth":486,"text":190},{"id":238,"depth":483,"text":241,"children":494},[495,496,497,500],{"id":249,"depth":486,"text":252},{"id":262,"depth":486,"text":265},{"id":288,"depth":486,"text":291,"children":498},[499],{"id":295,"depth":480,"text":298},{"id":326,"depth":486,"text":329},{"id":357,"depth":483,"text":360},{"id":380,"depth":483,"text":383},{"id":418,"depth":483,"text":421},{"id":444,"depth":483,"text":447},"markdown","content:technology-blogs:zh:2026-4-7.md","content","technology-blogs/zh/2026-4-7.md","technology-blogs/zh/2026-4-7","md",1776506119814]