[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-0049e060-f142-4b32-b655-5734b67b346f":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"0049e060-f142-4b32-b655-5734b67b346f","Kernels 大重构:把 GPU kernel 升级为 Hub 一等公民,LLM 基础设施开始标准化","Hugging Face 重构 🤗 Kernels 项目,把 GPU kernel 从零散脚本升级为 Hub 一等公民:新增 kernel 仓库类型,所有 kernel 挂在 huggingface.co\u002Fkernels 下,可直接看到加速器(CUDA\u002FROCm\u002FXPU)、OS、后端版本支持矩阵,源码 Git SHA1 内嵌编译产物,搭配 Nix hermetic build,任何发布版本可被独立复现,告别\"我编出来跟你不一样\"的玄学问题。安全升到协议层:默认只允许 trusted publishers 发布,同时用 Sigstore cosign 内核签名 + GitHub Actions workflow 双重验证,即便账号泄漏攻击者也无法签出恶意版本——把 OCI 供应链安全经验搬到了 ML kernel 生态。为 Agent 工程化铺路:`kernels` 与 `kernel-builder` 两个 CLI 拆分干净,后者刻意做成\"agent-optimized\"(非交互、输出可解析),配合后端 skills 文件,Agent 可端到端脚手架、构建、benchmark 并迭代优化一个 flash-attention 类内核;新增 Torch Stable ABI 向下兼容约 2 年,Apache TVM FFI 让同一份 kernel 跨 PyTorch \u002F JAX \u002F CuPy 运行。flash-attention、Fused MoE、MLA decoding 等高频内核从此 install 即用、可信可复现,不必再\"上传 PyPI + 配 CMake + 写 gotcha 文档\"。Agent 自动生成并验证 kernel 的闭环有了真实落地底座,2026 年下半年 LLM 训练推理栈的\"基础设施标准化\"才刚开跑。","https:\u002F\u002Fhuggingface.co\u002Fblog\u002Frevamped-kernels","24d5c6c5-6573-4180-a1fd-f1459842d1af",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source","2026-07-10T04:01:00Z","2026-07-10T04:09:45.499895Z","2026-07-10T04:09:45.499904Z",true,"agent",2]