[{"data":1,"prerenderedAt":511},["ShallowReactive",2],{"content-doc-\u002Fzh\u002Fnews\u002F20260605":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":7,"date":9,"author":10,"tags":11,"body":14,"_type":505,"_id":506,"_source":507,"_file":508,"_stem":509,"_extension":510},"\u002Fzh\u002Fnews\u002F20260605","news",false,"","HiFloat 开源代码仓正式上线：从技术规范走向开发实践，共建高效 AI 计算生态","2026\u002F06\u002F05","全球计算联盟GCC",[12,13],"开源","代码仓",{"type":15,"children":16,"toc":483},"root",[17,28,42,58,72,79,84,89,109,114,120,127,141,146,151,157,170,175,180,186,192,205,210,215,221,226,232,245,250,255,261,274,279,284,290,303,308,313,319,332,337,343,356,361,366,372,385,390,396,409,414,420,425,430,435,468,473,478],{"type":18,"tag":19,"props":20,"children":21},"element","p",{},[22],{"type":18,"tag":23,"props":24,"children":27},"img",{"alt":25,"src":26},"0","\u002Fnews\u002F20260605\u002F0.gif",[],{"type":18,"tag":19,"props":29,"children":30},{},[31,34,40],{"type":32,"value":33},"text","近日，",{"type":18,"tag":35,"props":36,"children":37},"strong",{},[38],{"type":32,"value":39},"HiFloat 开源代码仓正式上线",{"type":32,"value":41},"。围绕 HiFloat 低位宽浮点数据格式，社区已陆续开放量化工具、模型适配代码、开发框架和可复现方案，为开发者提供从技术理解、量化验证到模型部署的实践入口。",{"type":18,"tag":19,"props":43,"children":44},{},[45,50],{"type":18,"tag":35,"props":46,"children":47},{},[48],{"type":32,"value":49},"HiFloat 开源代码仓：",{"type":18,"tag":51,"props":52,"children":56},"a",{"href":53,"rel":54},"https:\u002F\u002Fgithub.com\u002FGCC-HiFloat",[55],"nofollow",[57],{"type":32,"value":53},{"type":18,"tag":19,"props":59,"children":60},{},[61,66],{"type":18,"tag":35,"props":62,"children":63},{},[64],{"type":32,"value":65},"HiFloat 社区网站：",{"type":18,"tag":51,"props":67,"children":70},{"href":68,"rel":69},"https:\u002F\u002Fhifloat.gccorg.com\u002F",[55],[71],{"type":32,"value":68},{"type":18,"tag":73,"props":74,"children":76},"h2",{"id":75},"从看懂技术到跑通代码",[77],{"type":32,"value":78},"从“看懂技术”到“跑通代码”",{"type":18,"tag":19,"props":80,"children":81},{},[82],{"type":32,"value":83},"开发者可以通过HiFloat 社区网站上的白皮书、标准规范和技术文档了解 HiF8、HiF4 等低位宽浮点格式。开源代码仓进一步补齐了从技术资料走向实际开发的关键环节。开发者不仅可以了解 HiFloat 的设计思路，还可以直接查看代码、运行示例、开展模型量化实验，并根据自己的业务场景探索新的适配方式。",{"type":18,"tag":19,"props":85,"children":86},{},[87],{"type":32,"value":88},"目前，HiFloat 开源代码仓包括以下三类资源：",{"type":18,"tag":90,"props":91,"children":92},"ul",{},[93,99,104],{"type":18,"tag":94,"props":95,"children":96},"li",{},[97],{"type":32,"value":98},"面向 HiF4 和 HiF8 的基础量化仿真工具；",{"type":18,"tag":94,"props":100,"children":101},{},[102],{"type":32,"value":103},"面向模型量化、导出、评测和推理部署的开发框架；",{"type":18,"tag":94,"props":105,"children":106},{},[107],{"type":32,"value":108},"面向文生视频、图生视频和语言模型等场景的模型适配方案。",{"type":18,"tag":19,"props":110,"children":111},{},[112],{"type":32,"value":113},"无论你是第一次接触低位宽量化，还是已经在开展大模型训练、推理优化和工程部署，都可以在代码仓中找到适合的切入点。",{"type":18,"tag":73,"props":115,"children":117},{"id":116},"基础工具快速体验-hifloat-量化能力",[118],{"type":32,"value":119},"基础工具：快速体验 HiFloat 量化能力",{"type":18,"tag":121,"props":122,"children":124},"h3",{"id":123},"_1-hifloat4-quantization_library",[125],{"type":32,"value":126},"1. HiFloat4-Quantization_Library",{"type":18,"tag":19,"props":128,"children":129},{},[130,135],{"type":18,"tag":35,"props":131,"children":132},{},[133],{"type":32,"value":134},"仓库链接：",{"type":18,"tag":51,"props":136,"children":139},{"href":137,"rel":138},"https:\u002F\u002Fgithub.com\u002FGCC-HiFloat\u002FHiFloat4-Quantization_Library",[55],[140],{"type":32,"value":137},{"type":18,"tag":19,"props":142,"children":143},{},[144],{"type":32,"value":145},"面向 HiFloat4 的基础量化库。",{"type":18,"tag":19,"props":147,"children":148},{},[149],{"type":32,"value":150},"提供 Float4 量化与仿真能力，支持开发者在不同硬件后端上开展低位宽实验。通过伪量化方式，开发者可以观察模型使用 HiF4 数据格式后的表现变化，并进一步分析精度、效率和模型效果之间的关系。仓库目前提供 NVIDIA CUDA 和华为昇腾 NPU 相关支持，适合作为了解 HiFloat4、开展 W4A4 等低位宽量化实验的起点。",{"type":18,"tag":121,"props":152,"children":154},{"id":153},"_2-hifloat8-quantization_library",[155],{"type":32,"value":156},"2. HiFloat8-Quantization_Library",{"type":18,"tag":19,"props":158,"children":159},{},[160,164],{"type":18,"tag":35,"props":161,"children":162},{},[163],{"type":32,"value":134},{"type":18,"tag":51,"props":165,"children":168},{"href":166,"rel":167},"https:\u002F\u002Fgithub.com\u002FGCC-HiFloat\u002FHiFloat8-Quantization_Library",[55],[169],{"type":32,"value":166},{"type":18,"tag":19,"props":171,"children":172},{},[173],{"type":32,"value":174},"面向 HiFloat8 的基础量化库。",{"type":18,"tag":19,"props":176,"children":177},{},[178],{"type":32,"value":179},"提供 Float8 量化和仿真场景，支持在 CUDA 与昇腾 NPU 等环境中进行验证。对于希望研究 W8A8 量化训练、推理优化或不同数据格式精度表现的开发者，快速建立基础实验环境。",{"type":18,"tag":73,"props":181,"children":183},{"id":182},"通用框架进一步探索模型量化与推理部署",[184],{"type":32,"value":185},"通用框架：进一步探索模型量化与推理部署",{"type":18,"tag":121,"props":187,"children":189},{"id":188},"_3-hifp8",[190],{"type":32,"value":191},"3. HiFP8",{"type":18,"tag":19,"props":193,"children":194},{},[195,199],{"type":18,"tag":35,"props":196,"children":197},{},[198],{"type":32,"value":134},{"type":18,"tag":51,"props":200,"children":203},{"href":201,"rel":202},"https:\u002F\u002Fgithub.com\u002FGCC-HiFloat\u002FHiFP8",[55],[204],{"type":32,"value":201},{"type":18,"tag":19,"props":206,"children":207},{},[208],{"type":32,"value":209},"HiFP8 是面向 HiFloat8 的量化框架。",{"type":18,"tag":19,"props":211,"children":212},{},[213],{"type":32,"value":214},"与基础量化仿真库相比，HiFP8 更关注完整的模型量化与部署流程。仓库支持 BF16 伪量化和 uint8 真量化两种模式，并提供模型导出、标准评测、SmoothQuant、MoE 模型支持以及 vLLM 推理部署相关能力。对于希望进一步把 HiFloat8 应用到真实大模型推理场景中的开发者，HiFP8 提供了更加完整的实践路径。",{"type":18,"tag":73,"props":216,"children":218},{"id":217},"模型实践覆盖视频生成与语言模型等应用场景",[219],{"type":32,"value":220},"模型实践：覆盖视频生成与语言模型等应用场景",{"type":18,"tag":19,"props":222,"children":223},{},[224],{"type":32,"value":225},"除了基础工具和通用框架，HiFloat 开源代码仓还汇聚了多项模型量化实践。展示了 HiF4、HiF8 在不同任务、不同模型和不同技术路线中的应用方式。",{"type":18,"tag":121,"props":227,"children":229},{"id":228},"_4-hisqrot4_wan22",[230],{"type":32,"value":231},"4. HiSQRot4_Wan2.2",{"type":18,"tag":19,"props":233,"children":234},{},[235,239],{"type":18,"tag":35,"props":236,"children":237},{},[238],{"type":32,"value":134},{"type":18,"tag":51,"props":240,"children":243},{"href":241,"rel":242},"https:\u002F\u002Fgithub.com\u002FGCC-HiFloat\u002FHiSQRot4_Wan2.2",[55],[244],{"type":32,"value":241},{"type":18,"tag":19,"props":246,"children":247},{},[248],{"type":32,"value":249},"面向 Wan2.2 文生视频模型，提供基于 HiFloat4 的 W4A4 训练后量化方案。方案将 SmoothQuant 和 Rotation 等思路用于模型量化，在尽可能保持原始生成流程的基础上，对目标线性层进行低位宽替换。",{"type":18,"tag":19,"props":251,"children":252},{},[253],{"type":32,"value":254},"这个仓库适合希望研究视频生成模型量化、低比特推理和效果保持能力的开发者参考。",{"type":18,"tag":121,"props":256,"children":258},{"id":257},"_5-ptq_wan22",[259],{"type":32,"value":260},"5. PTQ_Wan2.2",{"type":18,"tag":19,"props":262,"children":263},{},[264,268],{"type":18,"tag":35,"props":265,"children":266},{},[267],{"type":32,"value":134},{"type":18,"tag":51,"props":269,"children":272},{"href":270,"rel":271},"https:\u002F\u002Fgithub.com\u002FGCC-HiFloat\u002FPTQ_Wan2.2",[55],[273],{"type":32,"value":270},{"type":18,"tag":19,"props":275,"children":276},{},[277],{"type":32,"value":278},"面向 Wan2.2 图生视频模型，提供基于 HiFloat4 的训练后量化方案。",{"type":18,"tag":19,"props":280,"children":281},{},[282],{"type":32,"value":283},"仓库包含量化代码、运行示例、技术报告和模型资源链接，为开发者提供较为完整的复现入口。如果开发者希望了解如何将 HiFloat4 用于图生视频任务，可以通过该仓库快速查看量化流程和实验配置。",{"type":18,"tag":121,"props":285,"children":287},{"id":286},"_6-percentile_wan22",[288],{"type":32,"value":289},"6. Percentile_Wan2.2",{"type":18,"tag":19,"props":291,"children":292},{},[293,297],{"type":18,"tag":35,"props":294,"children":295},{},[296],{"type":32,"value":134},{"type":18,"tag":51,"props":298,"children":301},{"href":299,"rel":300},"https:\u002F\u002Fgithub.com\u002FGCC-HiFloat\u002FPercentile_Wan2.2",[55],[302],{"type":32,"value":299},{"type":18,"tag":19,"props":304,"children":305},{},[306],{"type":32,"value":307},"面向 Wan2.2 文生视频模型，提供基于分位数校准的 HiFloat4 W4A4 训练后量化方案。",{"type":18,"tag":19,"props":309,"children":310},{},[311],{"type":32,"value":312},"仓库开放了推理代码、配置文件和复现说明，重点展示如何通过校准方法改善低比特量化表现。对于关注量化校准策略、模型效果对比和可复现流程的开发者，具有较强的参考价值。",{"type":18,"tag":121,"props":314,"children":316},{"id":315},"_7-llmc_wan22",[317],{"type":32,"value":318},"7. LLMC_Wan2.2",{"type":18,"tag":19,"props":320,"children":321},{},[322,326],{"type":18,"tag":35,"props":323,"children":324},{},[325],{"type":32,"value":134},{"type":18,"tag":51,"props":327,"children":330},{"href":328,"rel":329},"https:\u002F\u002Fgithub.com\u002FGCC-HiFloat\u002FLLMC_Wan2.2",[55],[331],{"type":32,"value":328},{"type":18,"tag":19,"props":333,"children":334},{},[335],{"type":32,"value":336},"提供了LLMC与Wan2.2相结合的一体化实践流程，覆盖模型量化、视频生成以及VBench评测。适合希望了解完整实验链路和工程实践的开发者，包括如何准备环境、执行量化、生成视频并完成效果评测。",{"type":18,"tag":121,"props":338,"children":340},{"id":339},"_8-qat_ptq_wan21",[341],{"type":32,"value":342},"8. QAT_PTQ_Wan2.1",{"type":18,"tag":19,"props":344,"children":345},{},[346,350],{"type":18,"tag":35,"props":347,"children":348},{},[349],{"type":32,"value":134},{"type":18,"tag":51,"props":351,"children":354},{"href":352,"rel":353},"https:\u002F\u002Fgithub.com\u002FGCC-HiFloat\u002FQAT_PTQ_Wan2.1",[55],[355],{"type":32,"value":352},{"type":18,"tag":19,"props":357,"children":358},{},[359],{"type":32,"value":360},"该仓库面向Wan2.1文生视频模型，展示了基于HiFloat8 W8A8的训练后量化和量化感知训练方案。",{"type":18,"tag":19,"props":362,"children":363},{},[364],{"type":32,"value":365},"其中：PTQ，即训练后量化，适合在已有模型基础上进行快速量化；QAT，即量化感知训练，在训练过程中考虑量化带来的影响，以进一步改善模型效果。适合希望对比不同量化路线、研究视频生成模型W8A8训练与评测流程的开发者。",{"type":18,"tag":121,"props":367,"children":369},{"id":368},"_9-qat_wan21",[370],{"type":32,"value":371},"9. QAT_Wan2.1",{"type":18,"tag":19,"props":373,"children":374},{},[375,379],{"type":18,"tag":35,"props":376,"children":377},{},[378],{"type":32,"value":134},{"type":18,"tag":51,"props":380,"children":383},{"href":381,"rel":382},"https:\u002F\u002Fgithub.com\u002FGCC-HiFloat\u002FQAT_Wan2.1",[55],[384],{"type":32,"value":381},{"type":18,"tag":19,"props":386,"children":387},{},[388],{"type":32,"value":389},"该仓库围绕多模态模型训练和 Wan2.1 相关场景提供工程实践资源，并包含 HiF8 相关模块、训练脚本、模型配置和评测文件。适合具备一定训练经验的开发者进一步探索多模态大模型训练、昇腾环境适配和量化训练流程。",{"type":18,"tag":121,"props":391,"children":393},{"id":392},"_10-qat_pangu",[394],{"type":32,"value":395},"10. QAT_Pangu",{"type":18,"tag":19,"props":397,"children":398},{},[399,403],{"type":18,"tag":35,"props":400,"children":401},{},[402],{"type":32,"value":134},{"type":18,"tag":51,"props":404,"children":407},{"href":405,"rel":406},"https:\u002F\u002Fgithub.com\u002FGCC-HiFloat\u002FQAT_Pangu",[55],[408],{"type":32,"value":405},{"type":18,"tag":19,"props":410,"children":411},{},[412],{"type":32,"value":413},"该仓库面向 OpenPangu-Embedded-1B 模型，展示了基于 HiFloat8 W8A8 的量化感知训练方案。包括模型评测、复现说明和量化模型权重入口，适合希望了解 HiFloat8 在语言模型训练和量化部署中应用方式的开发者。",{"type":18,"tag":73,"props":415,"children":417},{"id":416},"不只是代码发布更是开放共建的开始",[418],{"type":32,"value":419},"不只是代码发布，更是开放共建的开始",{"type":18,"tag":19,"props":421,"children":422},{},[423],{"type":32,"value":424},"HiFloat 开源代码仓的上线，不只是一次资源汇总，也意味着社区建设进入了新的阶段。",{"type":18,"tag":19,"props":426,"children":427},{},[428],{"type":32,"value":429},"低比特计算是一项需要持续迭代的系统工程。它不仅涉及数据格式本身，也涉及模型算法、训练框架、推理引擎、硬件适配、评测工具和实际应用场景。",{"type":18,"tag":19,"props":431,"children":432},{},[433],{"type":32,"value":434},"不同开发者可以从不同方向参与进来：",{"type":18,"tag":90,"props":436,"children":437},{},[438,443,448,453,458,463],{"type":18,"tag":94,"props":439,"children":440},{},[441],{"type":32,"value":442},"运行示例，验证 HiFloat 在更多模型上的效果；",{"type":18,"tag":94,"props":444,"children":445},{},[446],{"type":32,"value":447},"提交 PR，完善工具、文档和适配代码；",{"type":18,"tag":94,"props":449,"children":450},{},[451],{"type":32,"value":452},"分享实验结果，为其他开发者提供参考；",{"type":18,"tag":94,"props":454,"children":455},{},[456],{"type":32,"value":457},"适配新的模型、算子和推理框架；",{"type":18,"tag":94,"props":459,"children":460},{},[461],{"type":32,"value":462},"参与技术讨论，共同探索低比特 AI 计算的更多可能性。",{"type":18,"tag":94,"props":464,"children":465},{},[466],{"type":32,"value":467},"开源自己的HiFloat项目并联系我们合入！",{"type":18,"tag":19,"props":469,"children":470},{},[471],{"type":32,"value":472},"无论你是算法工程师、系统开发者、科研人员、高校学生，还是关注 AI 算力效率的产业伙伴，都欢迎加入 HiFloat 开发工作。",{"type":18,"tag":19,"props":474,"children":475},{},[476],{"type":32,"value":477},"HiFloat 社区与开源代码仓将持续完善代码仓和技术文档，推动更多模型适配、工具升级和实践案例开放，为开发者提供更加清晰、易用、可复现的技术资源。",{"type":18,"tag":19,"props":479,"children":480},{},[481],{"type":32,"value":482},"让我们一起推动低比特 AI 计算技术从开放走向共建，从实验走向应用。",{"title":7,"searchDepth":484,"depth":484,"links":485},2,[486,487,492,495,504],{"id":75,"depth":484,"text":78},{"id":116,"depth":484,"text":119,"children":488},[489,491],{"id":123,"depth":490,"text":126},3,{"id":153,"depth":490,"text":156},{"id":182,"depth":484,"text":185,"children":493},[494],{"id":188,"depth":490,"text":191},{"id":217,"depth":484,"text":220,"children":496},[497,498,499,500,501,502,503],{"id":228,"depth":490,"text":231},{"id":257,"depth":490,"text":260},{"id":286,"depth":490,"text":289},{"id":315,"depth":490,"text":318},{"id":339,"depth":490,"text":342},{"id":368,"depth":490,"text":371},{"id":392,"depth":490,"text":395},{"id":416,"depth":484,"text":419},"markdown","app:zh:news:20260605.md","app","zh\u002Fnews\u002F20260605.md","zh\u002Fnews\u002F20260605","md",1785483339219]