[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-iree-org--iree":3,"tool-iree-org--iree":64},[4,17,27,35,43,56],{"id":5,"name":6,"github_repo":7,"description_zh":8,"stars":9,"difficulty_score":10,"last_commit_at":11,"category_tags":12,"status":16},3808,"stable-diffusion-webui","AUTOMATIC1111\u002Fstable-diffusion-webui","stable-diffusion-webui 是一个基于 Gradio 构建的网页版操作界面，旨在让用户能够轻松地在本地运行和使用强大的 Stable Diffusion 图像生成模型。它解决了原始模型依赖命令行、操作门槛高且功能分散的痛点，将复杂的 AI 绘图流程整合进一个直观易用的图形化平台。\n\n无论是希望快速上手的普通创作者、需要精细控制画面细节的设计师，还是想要深入探索模型潜力的开发者与研究人员，都能从中获益。其核心亮点在于极高的功能丰富度：不仅支持文生图、图生图、局部重绘（Inpainting）和外绘（Outpainting）等基础模式，还独创了注意力机制调整、提示词矩阵、负向提示词以及“高清修复”等高级功能。此外，它内置了 GFPGAN 和 CodeFormer 等人脸修复工具，支持多种神经网络放大算法，并允许用户通过插件系统无限扩展能力。即使是显存有限的设备，stable-diffusion-webui 也提供了相应的优化选项，让高质量的 AI 艺术创作变得触手可及。",162132,3,"2026-04-05T11:01:52",[13,14,15],"开发框架","图像","Agent","ready",{"id":18,"name":19,"github_repo":20,"description_zh":21,"stars":22,"difficulty_score":23,"last_commit_at":24,"category_tags":25,"status":16},1381,"everything-claude-code","affaan-m\u002Feverything-claude-code","everything-claude-code 是一套专为 AI 编程助手（如 Claude Code、Codex、Cursor 等）打造的高性能优化系统。它不仅仅是一组配置文件，而是一个经过长期实战打磨的完整框架，旨在解决 AI 代理在实际开发中面临的效率低下、记忆丢失、安全隐患及缺乏持续学习能力等核心痛点。\n\n通过引入技能模块化、直觉增强、记忆持久化机制以及内置的安全扫描功能，everything-claude-code 能显著提升 AI 在复杂任务中的表现，帮助开发者构建更稳定、更智能的生产级 AI 代理。其独特的“研究优先”开发理念和针对 Token 消耗的优化策略，使得模型响应更快、成本更低，同时有效防御潜在的攻击向量。\n\n这套工具特别适合软件开发者、AI 研究人员以及希望深度定制 AI 工作流的技术团队使用。无论您是在构建大型代码库，还是需要 AI 协助进行安全审计与自动化测试，everything-claude-code 都能提供强大的底层支持。作为一个曾荣获 Anthropic 黑客大奖的开源项目，它融合了多语言支持与丰富的实战钩子（hooks），让 AI 真正成长为懂上",140436,2,"2026-04-05T23:32:43",[13,15,26],"语言模型",{"id":28,"name":29,"github_repo":30,"description_zh":31,"stars":32,"difficulty_score":23,"last_commit_at":33,"category_tags":34,"status":16},2271,"ComfyUI","Comfy-Org\u002FComfyUI","ComfyUI 是一款功能强大且高度模块化的视觉 AI 引擎，专为设计和执行复杂的 Stable Diffusion 图像生成流程而打造。它摒弃了传统的代码编写模式，采用直观的节点式流程图界面，让用户通过连接不同的功能模块即可构建个性化的生成管线。\n\n这一设计巧妙解决了高级 AI 绘图工作流配置复杂、灵活性不足的痛点。用户无需具备编程背景，也能自由组合模型、调整参数并实时预览效果，轻松实现从基础文生图到多步骤高清修复等各类复杂任务。ComfyUI 拥有极佳的兼容性，不仅支持 Windows、macOS 和 Linux 全平台，还广泛适配 NVIDIA、AMD、Intel 及苹果 Silicon 等多种硬件架构，并率先支持 SDXL、Flux、SD3 等前沿模型。\n\n无论是希望深入探索算法潜力的研究人员和开发者，还是追求极致创作自由度的设计师与资深 AI 绘画爱好者，ComfyUI 都能提供强大的支持。其独特的模块化架构允许社区不断扩展新功能，使其成为当前最灵活、生态最丰富的开源扩散模型工具之一，帮助用户将创意高效转化为现实。",107662,"2026-04-03T11:11:01",[13,14,15],{"id":36,"name":37,"github_repo":38,"description_zh":39,"stars":40,"difficulty_score":23,"last_commit_at":41,"category_tags":42,"status":16},3704,"NextChat","ChatGPTNextWeb\u002FNextChat","NextChat 是一款轻量且极速的 AI 助手，旨在为用户提供流畅、跨平台的大模型交互体验。它完美解决了用户在多设备间切换时难以保持对话连续性，以及面对众多 AI 模型不知如何统一管理的痛点。无论是日常办公、学习辅助还是创意激发，NextChat 都能让用户随时随地通过网页、iOS、Android、Windows、MacOS 或 Linux 端无缝接入智能服务。\n\n这款工具非常适合普通用户、学生、职场人士以及需要私有化部署的企业团队使用。对于开发者而言，它也提供了便捷的自托管方案，支持一键部署到 Vercel 或 Zeabur 等平台。\n\nNextChat 的核心亮点在于其广泛的模型兼容性，原生支持 Claude、DeepSeek、GPT-4 及 Gemini Pro 等主流大模型，让用户在一个界面即可自由切换不同 AI 能力。此外，它还率先支持 MCP（Model Context Protocol）协议，增强了上下文处理能力。针对企业用户，NextChat 提供专业版解决方案，具备品牌定制、细粒度权限控制、内部知识库整合及安全审计等功能，满足公司对数据隐私和个性化管理的高标准要求。",87618,"2026-04-05T07:20:52",[13,26],{"id":44,"name":45,"github_repo":46,"description_zh":47,"stars":48,"difficulty_score":23,"last_commit_at":49,"category_tags":50,"status":16},2268,"ML-For-Beginners","microsoft\u002FML-For-Beginners","ML-For-Beginners 是由微软推出的一套系统化机器学习入门课程，旨在帮助零基础用户轻松掌握经典机器学习知识。这套课程将学习路径规划为 12 周，包含 26 节精炼课程和 52 道配套测验，内容涵盖从基础概念到实际应用的完整流程，有效解决了初学者面对庞大知识体系时无从下手、缺乏结构化指导的痛点。\n\n无论是希望转型的开发者、需要补充算法背景的研究人员，还是对人工智能充满好奇的普通爱好者，都能从中受益。课程不仅提供了清晰的理论讲解，还强调动手实践，让用户在循序渐进中建立扎实的技能基础。其独特的亮点在于强大的多语言支持，通过自动化机制提供了包括简体中文在内的 50 多种语言版本，极大地降低了全球不同背景用户的学习门槛。此外，项目采用开源协作模式，社区活跃且内容持续更新，确保学习者能获取前沿且准确的技术资讯。如果你正寻找一条清晰、友好且专业的机器学习入门之路，ML-For-Beginners 将是理想的起点。",84991,"2026-04-05T10:45:23",[14,51,52,53,15,54,26,13,55],"数据工具","视频","插件","其他","音频",{"id":57,"name":58,"github_repo":59,"description_zh":60,"stars":61,"difficulty_score":10,"last_commit_at":62,"category_tags":63,"status":16},3128,"ragflow","infiniflow\u002Fragflow","RAGFlow 是一款领先的开源检索增强生成（RAG）引擎，旨在为大语言模型构建更精准、可靠的上下文层。它巧妙地将前沿的 RAG 技术与智能体（Agent）能力相结合，不仅支持从各类文档中高效提取知识，还能让模型基于这些知识进行逻辑推理和任务执行。\n\n在大模型应用中，幻觉问题和知识滞后是常见痛点。RAGFlow 通过深度解析复杂文档结构（如表格、图表及混合排版），显著提升了信息检索的准确度，从而有效减少模型“胡编乱造”的现象，确保回答既有据可依又具备时效性。其内置的智能体机制更进一步，使系统不仅能回答问题，还能自主规划步骤解决复杂问题。\n\n这款工具特别适合开发者、企业技术团队以及 AI 研究人员使用。无论是希望快速搭建私有知识库问答系统，还是致力于探索大模型在垂直领域落地的创新者，都能从中受益。RAGFlow 提供了可视化的工作流编排界面和灵活的 API 接口，既降低了非算法背景用户的上手门槛，也满足了专业开发者对系统深度定制的需求。作为基于 Apache 2.0 协议开源的项目，它正成为连接通用大模型与行业专有知识之间的重要桥梁。",77062,"2026-04-04T04:44:48",[15,14,13,26,54],{"id":65,"github_repo":66,"name":67,"description_en":68,"description_zh":69,"ai_summary_zh":69,"readme_en":70,"readme_zh":71,"quickstart_zh":72,"use_case_zh":73,"hero_image_url":74,"owner_login":75,"owner_name":76,"owner_avatar_url":77,"owner_bio":78,"owner_company":79,"owner_location":79,"owner_email":79,"owner_twitter":79,"owner_website":80,"owner_url":81,"languages":82,"stars":121,"forks":122,"last_commit_at":123,"license":124,"difficulty_score":10,"env_os":125,"env_gpu":126,"env_ram":126,"env_deps":127,"category_tags":132,"github_topics":133,"view_count":10,"oss_zip_url":79,"oss_zip_packed_at":79,"status":16,"created_at":145,"updated_at":146,"faqs":147,"releases":176},767,"iree-org\u002Firee","iree","A retargetable MLIR-based machine learning compiler and runtime toolkit.","IREE（发音同 eerie）是一个基于 MLIR 构建的端到端机器学习编译器与运行时工具包。它的核心使命是让机器学习模型能够高效地运行在各种硬件平台上。\n\n在实际应用中，开发者常面临模型部署的难题：不同厂商的芯片架构各异，从云端数据中心到移动端边缘设备，适配成本高昂。IREE 通过生成统一的中间表示（IR），屏蔽了底层硬件差异，实现了高效的跨平台部署方案。无论是训练好的大模型还是轻量级应用，IREE 都能根据目标设备的约束进行优化，确保性能表现。\n\n这个项目非常适合 AI 工程师、系统开发人员以及致力于模型落地的研究者。其技术亮点在于强大的可重定向能力，支持包括 CPU、GPU、NPU 在内的多种后端加速，并已成功加入 Linux 基金会 AI 与数据基金会。如果你正在寻找解决跨平台模型推理性能问题的方案，IREE 提供了一个成熟且活跃的开源选择。","# IREE: Intermediate Representation Execution Environment\n\n\u003Cp>\u003Cimg src=\"docs\u002Fwebsite\u002Fdocs\u002Fassets\u002Fimages\u002FIREE_Logo_Icon_Color.svg\" width=\"48px\">\u003C\u002Fp>\n\nIREE (**I**ntermediate **R**epresentation **E**xecution **E**environment,\npronounced as \"eerie\") is an [MLIR](https:\u002F\u002Fmlir.llvm.org\u002F)-based end-to-end\ncompiler and runtime that lowers Machine Learning (ML) models to a unified IR\nthat scales up to meet the needs of the datacenter and down to satisfy the\nconstraints and special considerations of mobile and edge deployments.\n\nSee [our website](https:\u002F\u002Firee.dev\u002F) for project details, user\nguides, and instructions on building from source.\n\n[![IREE Discord Status](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Firee-org_iree_readme_c7504c3b128c.png)]([https:\u002F\u002Fdiscord.gg\u002FwEWh6Z9nMU](https:\u002F\u002Fdiscord.gg\u002FwEWh6Z9nMU))\n[![pre-commit](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fpre--commit-enabled-brightgreen?logo=pre-commit)](https:\u002F\u002Fgithub.com\u002Fpre-commit\u002Fpre-commit)\n[![OpenSSF Best Practices](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Firee-org_iree_readme_b1e164492b48.png)](https:\u002F\u002Fwww.bestpractices.dev\u002Fprojects\u002F8738)\n\n## Project news\n\n* 2025-04-02:\n[AMD submitted an IREE-based SDXL implementation to the MLPerf benchmark suite](https:\u002F\u002Frocm.blogs.amd.com\u002Fartificial-intelligence\u002Fmi325x-accelerates-mlperf-inference\u002FREADME.html#stable-diffusion-xl-sdxl-text-to-image-mlperf-inference-benchmark)\n* 2024-05-23:\n[IREE joins the LF AI & Data Foundation as a sandbox-stage project](https:\u002F\u002Flfaidata.foundation\u002Fblog\u002F2024\u002F05\u002F23\u002Fannouncing-iree-a-new-initiative-for-machine-learning-deployment\u002F)\n\n## Project status\n\n### Release status\n\nReleases notes are\n[published on GitHub releases](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Freleases?q=prerelease%3Afalse).\n\n| Package | Release status |\n| -- | -- |\nGitHub release (stable) | [![GitHub Release](https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fv\u002Frelease\u002Firee-org\u002Firee)](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Freleases\u002Flatest)\nGitHub release (nightly) | [![GitHub Release](https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fv\u002Frelease\u002Firee-org\u002Firee?include_prereleases)](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Freleases)\n`iree-base-compiler` | [![PyPI version](https:\u002F\u002Fbadge.fury.io\u002Fpy\u002Firee-base-compiler.svg)](https:\u002F\u002Fpypi.org\u002Fproject\u002Firee-base-compiler)\n`iree-base-runtime` | [![PyPI version](https:\u002F\u002Fbadge.fury.io\u002Fpy\u002Firee-base-runtime.svg)](https:\u002F\u002Fpypi.org\u002Fproject\u002Firee-base-runtime)\n\nFor more details on the release process, see\nhttps:\u002F\u002Firee.dev\u002Fdevelopers\u002Fgeneral\u002Frelease-management\u002F.\n\n### Build status\n\n[![CI](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci.yml\u002Fbadge.svg?query=branch%3Amain+event%3Apush)](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci.yml?query=branch%3Amain+event%3Apush)\n[![PkgCI](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fpkgci.yml\u002Fbadge.svg?query=branch%3Amain+event%3Apush)](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fpkgci.yml?query=branch%3Amain+event%3Apush)\n\n#### Nightly build status\n\n| Operating system | Build status |\n| -- | --: |\nLinux | [![CI - Linux arm64 clang](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci_linux_arm64_clang.yml\u002Fbadge.svg?query=branch%3Amain+event%3Aschedule)](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci_linux_arm64_clang.yml?query=branch%3Amain+event%3Aschedule)\nmacOS | [![CI - macOS x64 clang](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci_macos_x64_clang.yml\u002Fbadge.svg?query=branch%3Amain+event%3Aschedule)](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci_macos_x64_clang.yml?query=branch%3Amain+event%3Aschedule)\nmacOS | [![CI - macOS arm64 clang](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci_macos_arm64_clang.yml\u002Fbadge.svg?query=branch%3Amain+event%3Aschedule)](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci_macos_arm64_clang.yml?query=branch%3Amain+event%3Aschedule)\n\nFor the full list of workflows see\nhttps:\u002F\u002Firee.dev\u002Fdevelopers\u002Fgeneral\u002Fgithub-actions\u002F.\n\n## Communication channels\n\n*   [GitHub issues](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fissues): Feature requests,\n    bugs, and other work tracking\n*   [IREE Discord server](https:\u002F\u002Fdiscord.gg\u002FwEWh6Z9nMU): Daily development\n    discussions with the core team and collaborators\n*   (New) [iree-announce email list](https:\u002F\u002Flists.lfaidata.foundation\u002Fg\u002Firee-announce):\n    Announcements\n*   (New) [iree-technical-discussion email list](https:\u002F\u002Flists.lfaidata.foundation\u002Fg\u002Firee-technical-discussion):\n    General and low-priority discussion\n*   (Legacy) [iree-discuss email list](https:\u002F\u002Fgroups.google.com\u002Fforum\u002F#!forum\u002Firee-discuss):\n    Announcements, general and low-priority discussion\n\n### Related project channels\n\n*   [MLIR topic within LLVM Discourse](https:\u002F\u002Fllvm.discourse.group\u002Fc\u002Fllvm-project\u002Fmlir\u002F31):\n    IREE is enabled by and heavily relies on [MLIR](https:\u002F\u002Fmlir.llvm.org). IREE\n    sometimes is referred to in certain MLIR discussions. Useful if you are also\n    interested in MLIR evolution.\n\n## Architecture overview\n\n\u003C!-- TODO(scotttodd): switch to \u003Cpicture> once better supported? https:\u002F\u002Fgithub.blog\u002Fchangelog\u002F2022-05-19-specify-theme-context-for-images-in-markdown-beta\u002F -->\n![IREE Architecture](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Firee-org_iree_readme_7168b08bfc2e.png)\n![IREE Architecture](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Firee-org_iree_readme_1a09ec6ff4b4.png)\n\nSee [our website](https:\u002F\u002Firee.dev\u002F) for more information.\n\n## Presentations and talks\n\nCommunity meeting recordings: [IREE YouTube channel](https:\u002F\u002Fwww.youtube.com\u002F@iree4356)\n\nDate | Title | Recording | Slides\n---- | ----- | --------- | ------\n2025-06-10 | Data-Tiling in IREE: Achieving High Performance Through Compiler Design (AsiaLLVM) | [recording](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=iANJWUL_SOo) | [slides](https:\u002F\u002Fllvm.org\u002Fdevmtg\u002F2025-06\u002Fslides\u002Ftechnical-talk\u002Fwang-data-tilling.pdf)\n2025-05-17 | Introduction to GPU architecture and IREE's GPU CodeGen Pipeline | [recording](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=9Fy2jxj0ARE) | [slides](https:\u002F\u002Fdrive.google.com\u002Ffile\u002Fd\u002F1xbABUy3kQxxBzOUb3WjBOFCSY_sQYdGo\u002Fview)\n2025-02-12 | The Long Tail of AI: SPIR-V in IREE and MLIR (Vulkanised) | [recording](https:\u002F\u002Fyoutu.be\u002F0zwfc6UkxeE) | [slides](https:\u002F\u002Fwww.vulkan.org\u002Fuser\u002Fpages\u002F09.events\u002Fvulkanised-2025\u002FT12-Jakub-Kuderski-AMD-IREE-MLIR.pdf)\n2024-10-01 | Unveiling the Inner Workings of IREE: An MLIR-Based Compiler for Diverse Hardware | [recording](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=a3T74I9gGH8) |\n2021-06-09 | IREE Runtime Design Tech Talk | [recording](https:\u002F\u002Fdrive.google.com\u002Ffile\u002Fd\u002F1p0DcysaIg8rC7ErKYEgutQkOJGPFCU3s\u002Fview) | [slides](https:\u002F\u002Fdrive.google.com\u002Ffile\u002Fd\u002F1ikgOdZxnMz1ExqwrAiuTY9exbe3yMWbB\u002Fview?usp=sharing)\n2020-08-20 | IREE CodeGen (MLIR Open Design Meeting) | [recording](https:\u002F\u002Fdrive.google.com\u002Ffile\u002Fd\u002F1325zKXnNIXGw3cdWrDWJ1-bp952wvC6W\u002Fview?usp=sharing) | [slides](https:\u002F\u002Fdocs.google.com\u002Fpresentation\u002Fd\u002F1NetHjKAOYg49KixY5tELqFp6Zr2v8_ujGzWZ_3xvqC8\u002Fedit)\n2020-03-18 | Interactive HAL IR Walkthrough | [recording](https:\u002F\u002Fdrive.google.com\u002Ffile\u002Fd\u002F1_sWDgAPDfrGQZdxAapSA90AD1jVfhp-f\u002Fview?usp=sharing) |\n2020-01-31 | End-to-end MLIR Workflow in IREE (MLIR Open Design Meeting) | [recording](https:\u002F\u002Fdrive.google.com\u002Fopen?id=1os9FaPodPI59uj7JJI3aXnTzkuttuVkR) | [slides](https:\u002F\u002Fdrive.google.com\u002Fopen?id=1RCQ4ZPQFK9cVgu3IH1e5xbrBcqy7d_cEZ578j84OvYI)\n\n## License\n\nIREE is licensed under the terms of the Apache 2.0 License with LLVM Exceptions.\nSee [LICENSE](LICENSE) for more information.\n","# IREE：中间表示执行环境\n\n\u003Cp>\u003Cimg src=\"docs\u002Fwebsite\u002Fdocs\u002Fassets\u002Fimages\u002FIREE_Logo_Icon_Color.svg\" width=\"48px\">\u003C\u002Fp>\n\nIREE（**I**ntermediate **R**epresentation **E**xecution **E**environment，读音为\"eerie\"）是一个基于 [MLIR](https:\u002F\u002Fmlir.llvm.org\u002F)（机器学习中间表示）的端到端编译器和运行时，它将机器学习（ML）模型转换为统一的 IR（中间表示），该 IR 既能向上扩展以满足数据中心的需求，也能向下适配以满足移动和边缘部署的约束及特殊考量。\n\n有关项目详情、用户指南以及从源代码构建的说明，请参阅 [我们的网站](https:\u002F\u002Firee.dev\u002F)。\n\n[![IREE Discord Status](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Firee-org_iree_readme_c7504c3b128c.png)]([https:\u002F\u002Fdiscord.gg\u002FwEWh6Z9nMU](https:\u002F\u002Fdiscord.gg\u002FwEWh6Z9nMU))\n[![pre-commit](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fpre--commit-enabled-brightgreen?logo=pre-commit)](https:\u002F\u002Fgithub.com\u002Fpre-commit\u002Fpre-commit)\n[![OpenSSF Best Practices](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Firee-org_iree_readme_b1e164492b48.png)](https:\u002F\u002Fwww.bestpractices.dev\u002Fprojects\u002F8738)\n\n## 项目新闻\n\n* 2025-04-02:\n[AMD 向 MLPerf 基准测试套件提交了一个基于 IREE 的 SDXL 实现](https:\u002F\u002Frocm.blogs.amd.com\u002Fartificial-intelligence\u002Fmi325x-accelerates-mlperf-inference\u002FREADME.html#stable-diffusion-xl-sdxl-text-to-image-mlperf-inference-benchmark)\n* 2024-05-23:\n[IREE 作为沙盒阶段项目加入 LF AI & Data 基金会](https:\u002F\u002Flfaidata.foundation\u002Fblog\u002F2024\u002F05\u002F23\u002Fannouncing-iree-a-new-initiative-for-machine-learning-deployment\u002F)\n\n## 项目状态\n\n### 发布状态\n\n发布说明已发布在 [GitHub releases](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Freleases?q=prerelease%3Afalse)。\n\n| 包 | 发布状态 |\n| -- | -- |\nGitHub 发布（稳定版） | [![GitHub Release](https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fv\u002Frelease\u002Firee-org\u002Firee)](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Freleases\u002Flatest)\nGitHub 发布（夜间构建版） | [![GitHub Release](https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fv\u002Frelease\u002Firee-org\u002Firee?include_prereleases)](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Freleases)\n`iree-base-compiler` | [![PyPI version](https:\u002F\u002Fbadge.fury.io\u002Fpy\u002Firee-base-compiler.svg)](https:\u002F\u002Fpypi.org\u002Fproject\u002Firee-base-compiler)\n`iree-base-runtime` | [![PyPI version](https:\u002F\u002Fbadge.fury.io\u002Fpy\u002Firee-base-runtime.svg)](https:\u002F\u002Fpypi.org\u002Fproject\u002Firee-base-runtime)\n\n关于发布流程的更多详情，请参阅\nhttps:\u002F\u002Firee.dev\u002Fdevelopers\u002Fgeneral\u002Frelease-management\u002F。\n\n### 构建状态\n\n[![CI](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci.yml\u002Fbadge.svg?query=branch%3Amain+event%3Apush)](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci.yml?query=branch%3Amain+event%3Apush)\n[![PkgCI](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fpkgci.yml\u002Fbadge.svg?query=branch%3Amain+event%3Apush)](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fpkgci.yml?query=branch%3Amain+event%3Apush)\n\n#### 夜间构建状态\n\n| 操作系统 | 构建状态 |\n| -- | --: |\nLinux | [![CI - Linux arm64 clang](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci_linux_arm64_clang.yml\u002Fbadge.svg?query=branch%3Amain+event%3Aschedule)](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci_linux_arm64_clang.yml?query=branch%3Amain+event%3Aschedule)\nmacOS | [![CI - macOS x64 clang](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci_macos_x64_clang.yml\u002Fbadge.svg?query=branch%3Amain+event%3Aschedule)](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci_macos_x64_clang.yml?query=branch%3Amain+event%3Aschedule)\nmacOS | [![CI - macOS arm64 clang](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci_macos_arm64_clang.yml\u002Fbadge.svg?query=branch%3Amain+event%3Aschedule)](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Factions\u002Fworkflows\u002Fci_macos_arm64_clang.yml?query=branch%3Amain+event%3Aschedule)\n\n有关工作流程的完整列表，请参阅\nhttps:\u002F\u002Firee.dev\u002Fdevelopers\u002Fgeneral\u002Fgithub-actions\u002F。\n\n## 沟通渠道\n\n*   [GitHub 问题](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fissues)：功能请求、错误报告及其他工作跟踪\n*   [IREE Discord 服务器](https:\u002F\u002Fdiscord.gg\u002FwEWh6Z9nMU)：与核心团队及协作者进行的日常开发讨论\n*   (新) [iree-announce 邮件列表](https:\u002F\u002Flists.lfaidata.foundation\u002Fg\u002Firee-announce)：公告\n*   (新) [iree-technical-discussion 邮件列表](https:\u002F\u002Flists.lfaidata.foundation\u002Fg\u002Firee-technical-discussion)：一般性及低优先级讨论\n*   (遗留) [iree-discuss 邮件列表](https:\u002F\u002Fgroups.google.com\u002Fforum\u002F#!forum\u002Firee-discuss)：公告、一般性及低优先级讨论\n\n### 相关项目渠道\n\n*   [LLVM Discourse 中的 MLIR 主题](https:\u002F\u002Fllvm.discourse.group\u002Fc\u002Fllvm-project\u002Fmlir\u002F31)：\n    IREE 由 [MLIR](https:\u002F\u002Fmlir.llvm.org) 支持并高度依赖它。在某些 MLIR 讨论中有时会提到 IREE。如果您也对 MLIR 的演进感兴趣，这将很有用。\n\n## 架构概述\n\n\u003C!-- TODO(scotttodd): 切换到 \u003Cpicture> 标签，待支持更完善后？https:\u002F\u002Fgithub.blog\u002Fchangelog\u002F2022-05-19-specify-theme-context-for-images-in-markdown-beta\u002F -->\n![IREE 架构](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Firee-org_iree_readme_7168b08bfc2e.png)\n![IREE 架构](https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Firee-org_iree_readme_1a09ec6ff4b4.png)\n\n有关更多信息，请参阅 [我们的网站](https:\u002F\u002Firee.dev\u002F)。\n\n## 演讲与报告\n\n社区会议录像：[IREE YouTube 频道](https:\u002F\u002Fwww.youtube.com\u002F@iree4356)\n\n日期 | 标题 | 录像 | 幻灯片\n---- | ----- | --------- | ------\n2025-06-10 | IREE 中的数据分块：通过编译器设计实现高性能 (AsiaLLVM) | [recording](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=iANJWUL_SOo) | [slides](https:\u002F\u002Fllvm.org\u002Fdevmtg\u002F2025-06\u002Fslides\u002Ftechnical-talk\u002Fwang-data-tilling.pdf)\n2025-05-17 | GPU 架构与 IREE GPU CodeGen (代码生成) 流水线介绍 | [recording](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=9Fy2jxj0ARE) | [slides](https:\u002F\u002Fdrive.google.com\u002Ffile\u002Fd\u002F1xbABUy3kQxxBzOUb3WjBOFCSY_sQYdGo\u002Fview)\n2025-02-12 | AI 长尾效应：IREE 和 MLIR (多级别中间表示) 中的 SPIR-V (Vulkanised) | [recording](https:\u002F\u002Fyoutu.be\u002F0zwfc6UkxeE) | [slides](https:\u002F\u002Fwww.vulkan.org\u002Fuser\u002Fpages\u002F09.events\u002Fvulkanised-2025\u002FT12-Jakub-Kuderski-AMD-IREE-MLIR.pdf)\n2024-10-01 | 揭秘 IREE 内部工作原理：面向多样化硬件的基于 MLIR 的编译器 | [recording](https:\u002F\u002Fwww.youtube.com\u002Fwatch?v=a3T74I9gGH8) | \n2021-06-09 | IREE 运行时设计技术讲座 | [recording](https:\u002F\u002Fdrive.google.com\u002Ffile\u002Fd\u002F1p0DcysaIg8rC7ErKYEgutQkOJGPFCU3s\u002Fview) | [slides](https:\u002F\u002Fdrive.google.com\u002Ffile\u002Fd\u002F1ikgOdZxnMz1ExqwrAiuTY9exbe3yMWbB\u002Fview?usp=sharing)\n2020-08-20 | IREE CodeGen (MLIR 开放设计会议) | [recording](https:\u002F\u002Fdrive.google.com\u002Ffile\u002Fd\u002F1325zKXnNIXGw3cdWrDWJ1-bp952wvC6W\u002Fview?usp=sharing) | [slides](https:\u002F\u002Fdocs.google.com\u002Fpresentation\u002Fd\u002F1NetHjKAOYg49KixY5tELqFp6Zr2v8_ujGzWZ_3xvqC8\u002Fedit)\n2020-03-18 | 交互式 HAL (硬件抽象层) IR (中间表示) 导览 | [recording](https:\u002F\u002Fdrive.google.com\u002Ffile\u002Fd\u002F1_sWDgAPDfrGQZdxAapSA90AD1jVfhp-f\u002Fview?usp=sharing) | \n2020-01-31 | IREE 中的端到端 MLIR 工作流 (MLIR 开放设计会议) | [recording](https:\u002F\u002Fdrive.google.com\u002Fopen?id=1os9FaPodPI59uj7JJI3aXnTzkuttuVkR) | [slides](https:\u002F\u002Fdrive.google.com\u002Fopen?id=1RCQ4ZPQFK9cVgu3IH1e5xbrBcqy7d_cEZ578j84OvYI)\n\n## 许可证\n\nIREE 根据 Apache 2.0 许可证条款及 LLVM 例外条款进行许可。\n有关更多信息，请参阅 [LICENSE](LICENSE)。","# IREE 快速上手指南\n\n## 简介\nIREE (**I**ntermediate **R**epresentation **E**xecution **E**nvironment，发音为 \"eerie\") 是一个基于 [MLIR](https:\u002F\u002Fmlir.llvm.org\u002F) 的端到端编译器和运行时。它将机器学习（ML）模型降低到统一的中间表示（IR），支持从数据中心扩展到移动和边缘部署的各种需求。\n\n## 环境准备\n根据构建状态，IREE 主要支持以下操作系统架构：\n- **操作系统**: Linux (arm64\u002Fx64), macOS (x64\u002Farm64)\n- **语言环境**: Python 3.x\n- **其他依赖**: 若需从源码构建，需要 CMake、Ninja 等工具；通过 PyPI 安装则仅需 Python 环境。\n\n## 安装步骤\n推荐使用 PyPI 包进行快速安装，包含编译器 (`iree-base-compiler`) 和运行时 (`iree-base-runtime`)。\n\n### 1. 基础安装\n```bash\npip install iree-base-compiler iree-base-runtime\n```\n\n### 2. 国内加速建议\n为了提升下载速度，建议使用国内镜像源（如清华源）：\n```bash\npip install -i https:\u002F\u002Fpypi.tuna.tsinghua.edu.cn\u002Fsimple iree-base-compiler iree-base-runtime\n```\n\n### 3. 源码构建（可选）\n如需获取最新功能或定制版本，请参考官方文档构建源码：\n- 项目地址：[GitHub Releases](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Freleases)\n- 构建指南：[iree.dev](https:\u002F\u002Firee.dev\u002F)\n\n## 基本使用\n安装完成后，可通过命令行工具对模型进行编译和运行。以下是典型的使用流程概念：\n\n### 1. 编译模型\n使用 `iree-compile` 将 ML 模型转换为 IREE 可执行格式（VMBC 或 HAL）。\n```bash\niree-compile --iree-hal-target-backends=vmvx your_model.mlir -o output.vmfb\n```\n\n### 2. 运行模型\n使用 `iree-run-module` 加载编译后的模块并执行推理。\n```bash\niree-run-module --module=output.vmfb --input=\"tensor\u003C1x3x224x224xf32>\"\n```\n\n> **注意**：具体参数和后端支持（如 GPU、Vulkan、Metal 等）请查阅 [官方网站用户指南](https:\u002F\u002Firee.dev\u002F)。\n\n## 社区与支持\n- **官方网站**: [iree.dev](https:\u002F\u002Firee.dev\u002F)\n- **Discord 社区**: [加入讨论](https:\u002F\u002Fdiscord.gg\u002FwEWh6Z9nMU)\n- **问题反馈**: [GitHub Issues](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fissues)\n- **技术邮件列表**: [iree-technical-discussion](https:\u002F\u002Flists.lfaidata.foundation\u002Fg\u002Firee-technical-discussion)","某自动驾驶初创团队急需将训练好的 YOLOv8 目标检测模型同时高效部署到云端 GPU 服务器和车载嵌入式芯片上。\n\n### 没有 iree 时\n- 需要为不同硬件维护多套推理引擎代码，开发和维护成本极高且易出错。\n- 模型转换格式不统一，TensorFlow 和 PyTorch 需分别适配特定后端环境。\n- 边缘设备性能优化困难，难以自动发挥 NPU 或 GPU 的底层硬件特性。\n- 跨平台迁移时容易出现精度丢失，导致反复调试严重延长产品上线周期。\n\n### 使用 iree 后\n- iree 提供统一的中间表示，一套编译流程即可生成多种后端可执行文件。\n- 自动针对目标硬件进行算子融合与内存优化，显著提升推理速度与能效。\n- 支持从云端数据中心到移动端的无缝部署，大幅减少重复适配工作量。\n- 保持模型精度高度一致，简化了跨设备测试验证流程，加快业务迭代节奏。\n\niree 通过统一编译架构显著降低了多端部署的复杂度并提升了运行效率。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002Firee-org_iree_112eb146.png","iree-org","IREE","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002Firee-org_e1fb00a9.png","",null,"iree.dev","https:\u002F\u002Fgithub.com\u002Firee-org",[83,87,91,95,99,103,107,111,115,118],{"name":84,"color":85,"percentage":86},"C++","#f34b7d",41.1,{"name":88,"color":89,"percentage":90},"MLIR","#5EC8DB",26.3,{"name":92,"color":93,"percentage":94},"C","#555555",23.7,{"name":96,"color":97,"percentage":98},"CMake","#DA3434",2.8,{"name":100,"color":101,"percentage":102},"Python","#3572A5",2.7,{"name":104,"color":105,"percentage":106},"Starlark","#76d275",2.2,{"name":108,"color":109,"percentage":110},"Shell","#89e051",0.5,{"name":112,"color":113,"percentage":114},"Objective-C","#438eff",0.4,{"name":116,"color":79,"percentage":117},"Linker Script",0.1,{"name":119,"color":120,"percentage":117},"Java","#b07219",3689,873,"2026-04-04T00:02:24","Apache-2.0","Linux, macOS","未说明",{"notes":128,"python":126,"dependencies":129},"基于 MLIR 的端到端编译器和运行时，支持从数据中心到移动边缘设备的模型部署。CI 构建状态显示支持 Linux 和 macOS。支持多种硬件后端（如 GPU、Vulkan、SPIR-V、ROCm）。采用 Apache 2.0 许可证。",[130,131,88],"iree-base-compiler","iree-base-runtime",[13],[134,135,136,137,138,139,140,141,142,143,144],"mlir","vulkan","tensorflow","spirv","cuda","jax","pytorch","compiler","machine-learning","runtime","onnx","2026-03-27T02:49:30.150509","2026-04-06T11:31:10.235560",[148,153,157,162,166,171],{"id":149,"question_zh":150,"answer_zh":151,"source_url":152},3292,"如何验证 IREE 的数值计算差异是否由工具链引起？","在不同硬件环境（如 A2-highgpu-1g 和工作站）上运行相同测试。如果最大差异为 0，则非 IREE 问题。应检查 TF 实现生成的输出是否正确，或与真实黄金值（golden values）对比。若发现关联重排序等问题，请单独提交 Bug 跟踪。","https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fissues\u002F9536",{"id":154,"question_zh":155,"answer_zh":156,"source_url":152},3293,"遇到模型数值不正确时应如何处理？","此类问题可能源于黄金值（golden values）本身而非 IREE。请确保 TF 实现生成良好输出。若是特定的关联重排序问题，建议创建独立 Issue 跟踪（如 issue #9716），原问题可关闭。",{"id":158,"question_zh":159,"answer_zh":160,"source_url":161},3294,"为什么 Llama 模型在 CPU 和 HIP 目标上运行会给出不同数值？","这是由于 mfma 指令的重新关联导致 fp16 累加越过边界造成的精度差异。这是一个已知问题，不应在 f16 中累积。","https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fissues\u002F19347",{"id":163,"question_zh":164,"answer_zh":165,"source_url":161},3295,"如何解决 fp16 累加导致的精度不一致问题？","参考修复方案 #19519，该修复解决了从 mfma 重新关联带来的进一步 fp16 累加问题。请确保编译配置避免在 f16 中累积，并检查前端是否有相关变更。",{"id":167,"question_zh":168,"answer_zh":169,"source_url":170},3296,"如何降低不支持的 MHLO 操作（如 sort）？","可使用 `linalg_ext` 模块，其结构与 `linalg` 相似便于复用。通过类型转换器（type converter）并插入 `convertType` 来处理这些 MHLO 特殊情况（参考 ConvertMHLOToLinalgExt.cpp 第 65-76 行）。","https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fissues\u002F6154",{"id":172,"question_zh":173,"answer_zh":174,"source_url":175},3297,"如何在 ARM64 设备上提升 IREE 性能以接近 TFLite？","需启用微内核（microkernels）支持并使用特定补丁版本编译。建议在编译命令中添加 `--iree-llvmcpu-enable-microkernels` 标志，并更新到包含 PR #16615 支持的版本。","https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fissues\u002F15399",[177,182,186,190,194,198,202,206,210,214,218,222,226,230,234,238,242,246,251,255],{"id":178,"version":179,"summary_zh":180,"released_at":181},112520,"iree-3.12.0rc20260405","Automatic candidate release of iree.","2026-04-05T10:52:06",{"id":183,"version":184,"summary_zh":180,"released_at":185},112521,"iree-3.12.0rc20260404","2026-04-04T10:48:52",{"id":187,"version":188,"summary_zh":180,"released_at":189},112522,"iree-3.12.0rc20260403","2026-04-03T10:49:41",{"id":191,"version":192,"summary_zh":180,"released_at":193},112523,"iree-3.12.0rc20260402","2026-04-02T10:28:21",{"id":195,"version":196,"summary_zh":180,"released_at":197},112524,"iree-3.12.0rc20260401","2026-04-01T11:01:17",{"id":199,"version":200,"summary_zh":180,"released_at":201},112525,"iree-3.12.0rc20260331","2026-03-31T10:52:54",{"id":203,"version":204,"summary_zh":180,"released_at":205},112526,"iree-3.12.0rc20260330","2026-03-30T11:05:20",{"id":207,"version":208,"summary_zh":180,"released_at":209},112527,"iree-3.12.0rc20260329","2026-03-29T10:51:22",{"id":211,"version":212,"summary_zh":180,"released_at":213},112528,"iree-3.12.0rc20260328","2026-03-28T10:45:26",{"id":215,"version":216,"summary_zh":180,"released_at":217},112529,"iree-3.12.0rc20260327","2026-03-27T10:51:53",{"id":219,"version":220,"summary_zh":180,"released_at":221},112530,"iree-3.12.0rc20260326","2026-03-26T10:50:27",{"id":223,"version":224,"summary_zh":180,"released_at":225},112531,"iree-3.12.0rc20260325","2026-03-25T10:47:39",{"id":227,"version":228,"summary_zh":180,"released_at":229},112532,"iree-3.12.0rc20260324","2026-03-24T10:41:20",{"id":231,"version":232,"summary_zh":180,"released_at":233},112533,"iree-3.11.0rc20260323","2026-03-23T10:58:59",{"id":235,"version":236,"summary_zh":180,"released_at":237},112534,"iree-3.11.0rc20260322","2026-03-22T10:49:11",{"id":239,"version":240,"summary_zh":180,"released_at":241},112535,"iree-3.11.0rc20260321","2026-03-21T10:40:45",{"id":243,"version":244,"summary_zh":180,"released_at":245},112536,"iree-3.11.0rc20260320","2026-03-20T10:41:20",{"id":247,"version":248,"summary_zh":249,"released_at":250},112537,"v3.11.0","# IREE v3.11.0 Release Notes\r\n\r\n\r\n**Release Candidate:** `iree-3.11.0rc20260316`\r\n**Commits:** ~539 commits since v3.10.0\r\n**VMFB Bytecode Version:** 17.0 (unchanged from v3.10.0)\r\n\r\n---\r\n\r\n## Highlights\r\n\r\n- **New async I\u002FO infrastructure**: Proactor-based async I\u002FO with causal frontier scheduling, enabling cross-process shared memory support\r\n- **Streaming tokenizer**: Full HuggingFace-compatible tokenizer with tiktoken format support for OpenAI BPE vocabularies (click [here](https:\u002F\u002Fwww.linkedin.com\u002Fposts\u002Fstellalaurenzo_fastest-llm-tokenizer-on-the-internet-authored-share-7435479755161915392-dj1K?utm_source=share&utm_medium=member_android&rcm=ACoAAAIKjF8BNq6yjA-sYlihYIO1enJHJNT8RCE) for more info)\r\n- **Python 3.10+ requirement**: Minimum Python version bumped to 3.10; Python 3.12+ supported via Stable ABI (abi3).\r\n- **ROCm flag rename**: `iree-hip-*` compiler flags renamed to `iree-rocm-*` (old names deprecated with warnings)\r\n- **Enhanced vector distribution**: Refactored 2-phase forward\u002Fbackward layout analysis with improved transfer_gather support\r\n\r\n---\r\n\r\n## Breaking Changes\r\n\r\n### VMFB Compatibility\r\n\r\n- **VMFB bytecode version unchanged (17.0)** - VMFBs compiled with `v3.10.0` remain compatible with `v3.11.0` runtime\r\n  - No recompilation needed when upgrading from v3.10.0\r\n\r\n### Python Version Requirement\r\n- Minimum Python version is now 3.10 ([#23591](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23591))\r\n\r\n\r\n### Compiler Flag Renames\r\n- **`iree-hip-*` flags renamed to `iree-rocm-*`** ([#23420](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23420))\r\n  - Old flag names emit deprecation warnings but still work\r\n  - CMake: `IREE_HIP_TEST_TARGET_CHIP` → `IREE_ROCM_TEST_TARGET_CHIP`\r\n\r\n### Build System Changes\r\n- **Minimum CMake version bumped to 3.26** ([#23607](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23607))\r\n  - Required for Python Stable ABI support\r\n\r\n### API Changes\r\n- `map_gather`\u002F`map_scatter` ops renamed to `map_load`\u002F`map_store` in LinalgExt ([#23481](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23481))\r\n\r\n---\r\n\r\n## What's New\r\n\r\n### 1. Compiler\r\n\r\n#### 1.1 Async Infrastructure & Tokenizers\r\n\r\nMajor new infrastructure for async I\u002FO and text processing:\r\n\r\n- Added proactor-based async I\u002FO with causal frontier scheduling (`iree\u002Fasync\u002F`) ([#23527](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23527))\r\n- Added streaming tokenizer with full HuggingFace compatibility (`iree\u002Ftokenizer\u002F`) ([#23528](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23528))\r\n- Graceful degradation for io_uring slab registration on RLIMIT_MEMLOCK ([#23654](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23654))\r\n- Added tiktoken format loader for OpenAI BPE vocabularies ([#23663](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23663))\r\n- Added async infrastructure for cross-process shared memory ([#23688](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23688))\r\n\r\n#### 1.2 Codegen & Vector Distribution\r\n\r\nSignificant improvements to vector distribution and code generation:\r\n\r\n- Added support for `shape_cast` in vector distribution ([#23307](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23307))\r\n- Support for padding integer attention masks ([#23430](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23430))\r\n- Added `arg_compare` operation to VectorExt ([#23386](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23386))\r\n- Refactored `transfer_gather` to use unified `indexing_maps` ([#23510](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23510))\r\n- Added distribution pattern for `iree_codegen.inner_tiled` ([#23483](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23483))\r\n- Added vectorization support for `iree_linalg_ext.arg_compare` ([#23440](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23440))\r\n- Added `transfer_gather` unrolling ([#23517](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23517))\r\n- Support multi-batch gather vectorization to `transfer_gather` ([#23552](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23552))\r\n- Added `transfer_gather` canonicalizations for masking ([#23565](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23565))\r\n- Refactored `VectorLayoutAnalysis` into 2-phase forward\u002Fbackward design ([#23611](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23611))\r\n- Added `TransferScatterOp` definition and verifier ([#23666](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23666))\r\n- Introduced `VectorizableOpInterface` and migrated all ops ([#23653](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23653), [#23656](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23656), [#23658](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23658), [#23662](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23662), [#23712](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23712), [#23713](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23713), [#23767](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23767))\r\n- Added `iree_map` dialect with `PackMapAttr` and `VectorLayoutInterface` ([#23671](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23671), [#23672](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23672))\r\n- Added `TransferScatterOp` bufferization support ([#23719](https:\u002F\u002Fgithub.com\u002Firee-org\u002Firee\u002Fpull\u002F23719))\r\n- Materialize vector masking on `VectorDistribute` pipeline ([#23679](https:","2026-03-19T23:25:34",{"id":252,"version":253,"summary_zh":180,"released_at":254},112538,"iree-3.11.0rc20260319","2026-03-19T10:44:39",{"id":256,"version":257,"summary_zh":180,"released_at":258},112539,"iree-3.11.0rc20260318","2026-03-18T10:20:38"]