{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-09-01T18:00:49.064Z","headline":"Hugging Face 发布 @huggingface/kernels，提供 207 个 WebGPU 内核用于浏览器本地 AI 推理","description":"Hugging Face WebAI 团队发布 @huggingface/kernels 库及 207 个以独立仓库形式托管在 Hub 上的 WebGPU 内核（Apache-2.0），每个内核带 manifest、正确性测试、基准用例和 WGSL 着色器模板。","url":"https://www.aioga.com/news/cmtitozyq04j2ro9ydv3oxgsd/","mainEntityOfPage":"https://www.aioga.com/news/cmtitozyq04j2ro9ydv3oxgsd/","datePublished":"2026-09-01T00:00:00.000Z","dateModified":"2026-09-01T00:00:00.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://huggingface.co/blog/webgpu-kernels","https://aihot.virxact.com/items/cmtitozyq04j2ro9ydv3oxgsd"],"canonicalUrl":"https://www.aioga.com/news/cmtitozyq04j2ro9ydv3oxgsd/","directAnswer":{"@type":"Answer","text":"Hugging Face WebAI 团队发布 @huggingface/kernels 库，并在 Hugging Face Hub 上以独立仓库形式提供 207 个 WebGPU 内核，采用 Apache-2.0 许可。","url":"https://www.aioga.com/news/cmtitozyq04j2ro9ydv3oxgsd/","dateCreated":"2026-09-01T00:00:00.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"huggingface.co source article","url":"https://huggingface.co/blog/webgpu-kernels","datePublished":"2026-09-01T00:00:00.000Z","provider":{"@type":"Organization","name":"huggingface.co","url":"https://huggingface.co/blog/webgpu-kernels"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmtitozyq04j2ro9ydv3oxgsd","datePublished":"2026-09-01T00:00:00.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmtitozyq04j2ro9ydv3oxgsd"}}],"aggregationSource":"Hugging Face：Blog（RSS）","originalPublisher":{"name":"huggingface.co","url":"https://huggingface.co/blog/webgpu-kernels"},"geoDeepAnswer":null,"article":{"id":"cmtitozyq04j2ro9ydv3oxgsd","slug":"cmtitozyq04j2ro9ydv3oxgsd","url":"https://www.aioga.com/news/cmtitozyq04j2ro9ydv3oxgsd/","title":"Hugging Face 发布 @huggingface/kernels，提供 207 个 WebGPU 内核用于浏览器本地 AI 推理","title_en":"","summary":"Hugging Face WebAI 团队发布 @huggingface/kernels 库及 207 个以独立仓库形式托管在 Hub 上的 WebGPU 内核（Apache-2.0），每个内核带 manifest、正确性测试、基准用例和 WGSL 着色器模板。","source":"Hugging Face：Blog（RSS）","sourceUrl":"https://huggingface.co/blog/webgpu-kernels","aiHotUrl":"https://aihot.virxact.com/items/cmtitozyq04j2ro9ydv3oxgsd","publishedAt":"2026-09-01T00:00:00.000Z","category":"行业动态","score":72,"selected":true,"articleBody":["TL;DR ：#tldr Why start with kernels? ：#why-start-with-kernels A kernel repository, not just a shader ：#a-kernel-repository-not-just-a-shader Loading a kernel from the Hub ：#loading-a-kernel-from-the-hub How fast are the kernels? ：#how-fast-are-the-kernels From one device to a fleet ：#from-one-device-to-a-fleet Building a shared foundation for WebAI ：#building-a-shared-foundation-for-webai One of our biggest goals on the WebAI team at Hugging Face is to make browser inference as fast and as user-friendly as possible. Getting there is a multi-layer effort: models need browser-friendly representations, runtimes need to build efficient execution plans, and the individual GPU operations at the bottom of the stack need to make the most of many different devices and browser implementations.","Today, we are releasing the first layer of that effort: @huggingface/kernels ：https://www.npmjs.com/package/@huggingface/kernels, a minimal library for loading and running optimized WebGPU kernels from the Hugging Face Hub, together with an initial collection of 207 kernels at huggingface.co/webgpu-kernels：https://huggingface.co/webgpu-kernels.","The collection covers operations used across a wide variety of machine learning architectures and workloads. More importantly, each kernel is published as a complete, versioned package: its interface, shader templates, correctness cases, benchmark cases, and usage instructions all live together on the Hub.","We are also launching Fleet：https://webgpu-kernels-fleet.hf.space/, an in-browser GPU benchmarking and testing suite that runs and scores the kernels on your hardware. Beyond the results for your own machine, Fleet gives the community a way to contribute performance and correctness evidence from devices we could never cover in a conventional test lab. With your consent, every run adds private evidence that can help us find failures (incorrect results, pathologically slow cases, etc.), improve kernel variants, and make better optimization decisions across real-world hardware.","A model running in the browser eventually becomes a sequence of GPU operations: matrix multiplications, normalizations, convolutions, attention primitives, quantization operations, data-layout transformations, and many more. WebGPU makes these operations available across modern browsers through a portable API, while WGSL provides a common language for the shaders that execute them.","Portability, however, does not automatically mean performance. Two shaders can implement the same operation and produce the same output while behaving completely differently across different accelerators. Workgroup sizes, memory access patterns, vectorization, data types, and fusion strategies can all affect performance. The best choice can also change with the input shape, device, browser, and available WebGPU features.","This is why kernels form a foundational layer of fast browser inference. Higher-level runtimes can only be as efficient as the operations they dispatch. By making those operations individually discoverable, testable, benchmarkable, and versioned, we can improve the foundation independently while keeping a stable contract for the layers above it.","Each kernel in the collection has its own repository and kernel card. The card documents the operation's semantics, inputs, outputs, attributes, supported data types, source files, and a ready-to-run @huggingface/kernels example.","For example, ai.onnx.Add ：https://huggingface.co/webgpu-kernels/ai.onnx.Add implements elementwise addition with multidirectional broadcasting. It is one of the simplest operations in a neural network, used everywhere from residual connections to adding a bias. Its card documents the two inputs, the broadcasted output shape, supported data types, and the variants available for different shapes and devices.","Behind the card, the repository contains the artifacts needed to understand and evaluate the implementation:","This structure turns a shader into a reusable software artifact. The interface is inspectable without reading WGSL, correctness and performance cases travel with the implementation, and published versions can be loaded explicitly rather than depending on an unversioned file URL. Our kernels can also serve as reference implementations for developers building custom WebGPU kernels or integrating these operations into their own runtimes.","Running these kernels requires a browser with WebGPU support：https://developer.mozilla.org/en-US/docs/Web/API/WebGPU_API. WebGPU availability depends on the browser, operating system, GPU, and driver. You can check for it in JavaScript with \"gpu\" in navigator .","@huggingface/kernels provides the bridge between a kernel repository and your application. Call getKernel with a Hub repository ID and a contract version, then invoke the returned function with typed input data and tensor shapes. Here is a small bias-add example:","The second input is broadcast across the first dimension, producing an output with shape [2, 3] . The loader derives that output shape and logical data type from the manifest contract and the inputs, then allocates c automatically.","Addition on six floats is deliberately the smallest possible demo. At this size, the GPU round trip costs far more than the math. The point is the call pattern: it stays exactly the same for the heavyweight operations where optimized kernels actually pay off, such as matrix multiplication ( ai.onnx.MatMul ). Only the repository ID and the inputs change.","Even this elementary operation illustrates why kernels need variants. Equal-shape addition can use a direct vectorized path, while broadcasted inputs need different indexing logic. The published Add kernel includes variants for equal shapes, vectorized broadcasting, scalar processing, and general broadcasting. The runtime can select an implementation that fits the current call and device without changing the application-facing API.","The version: 1 option selects version 1 of the published kernel contract . It is separate from an ONNX opset, an operator's since_version , or a model revision. Keeping those concepts separate lets applications depend on a stable JavaScript-facing contract while kernel implementations evolve behind it.","So, how much of a difference do optimized kernels actually make? We put our collection head-to-head with ORT WebGPU on an Apple M4 GPU, using ONNX Runtime Web 1.30.0-dev.20260826-b1f76d586a . We started with 1,756 test cases across all 207 operations and kept the 809 cases where both sides produced matching outputs and reliable timings.","Across those comparisons, our kernels were 2.57x faster by geometric mean and 1.90x faster at the median , with 629 wins, 176 losses, and 4 ties. Here is a closer look at four familiar operations:","Some individual wins were much bigger. A particularly difficult bilinear Einsum case ( i,ij,j with size 4096) ran in 0.136 ms with our kernel versus 1,396 ms with ORT WebGPU: more than 10,000x faster . A row-wise CumSum over [256, 4096] was 301x faster , at 0.016 ms versus 4.784 ms. These are unusual cases rather than the speedups you should expect everywhere, but they show how much a specialized kernel can help when a general implementation hits a slow path.","We timed the work done on the GPU itself, leaving out setup such as loading kernels, creating sessions, uploading inputs, compiling shaders, and reading outputs back. Very short workloads are naturally harder to measure, and small cases can benefit from the GPU cache, so these numbers are best read as a useful comparison rather than a promise for every application.","They are also results for individual operations, not complete models. Exact performance will change across GPUs and browsers, which is why Fleet is so important for building a broader picture.","We are also working with the ONNX Runtime team to upstream these improvements so they can benefit the broader ONNX Runtime Web ecosystem.","WebGPU performance varies across GPUs, browsers, and drivers, so results from one machine only tell part of the story. Fleet：https://webgpu-kernels-fleet.hf.space/ lets anyone run correctness and performance checks in the browser and see how the kernels behave on their hardware.","With consent, each run privately contributes evidence that helps us spot device-specific failures, compare variants, and improve selection rules. The goal is simple: use broad, real-world coverage to make the kernels faster and more reliable for everyone.","The initial 207 kernels are a starting point, not the end state. Publishing kernels independently on the Hub gives us a common place to inspect contracts, compare implementations, reproduce correctness checks, and improve performance without embedding every shader directly into every runtime.","The collection is also part of the Hub's broader kernel ecosystem: on the Kernels page：https://huggingface.co/kernels?platform=webgpu&sort=trending, the WebGPU kernels sit alongside kernels for CUDA, ROCm, Metal, and other platforms, and can be filtered, sorted, and explored like any other artifact on the Hub.","This is the low-level foundation for the next steps in our browser inference stack. We are excited to connect these kernels to higher-level model tooling, continue expanding operation coverage, and make fast local inference easier to use across the WebAI ecosystem.","Explore the WebGPU kernel collection：https://huggingface.co/webgpu-kernels, try @huggingface/kernels ：https://www.npmjs.com/package/@huggingface/kernels, and join the Fleet：https://webgpu-kernels-fleet.hf.space/ to contribute evidence from your device and help us make the kernels better for everyone."],"articleImages":[{"sourceUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/68df176c403a7bf9e8ae85a8/mffYA6B8aumvl9flbiw5Y.png","alt":"","afterParagraph":0,"url":"/media/articles/cmtitozyq04j2ro9ydv3oxgsd/4ecc4fa004c13004.webp"},{"sourceUrl":"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/webgpu-kernels/ai-onnx-add.png","alt":"Files in the ai.onnx.Add WebGPU kernel repository","afterParagraph":8,"url":"/media/articles/cmtitozyq04j2ro9ydv3oxgsd/af1fd3d4dcfebe8a.png"},{"sourceUrl":"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/webgpu-kernels/kernels.png","alt":"The Hub Kernels page filtered to the WebGPU platform, listing the 207 published kernels","afterParagraph":26,"url":"/media/articles/cmtitozyq04j2ro9ydv3oxgsd/f9e8571eef8d921e.png"}],"mediaStatus":"ok","articleBodyZh":["TL;DR：#tldr 为什么从内核开始？：#why-start-with-kernels 内核仓库，而不仅仅是着色器：#a-kernel-repository-not-just-a-shader 从 Hub 加载内核：#loading-a-kernel-from-the-hub 内核有多快？：#how-fast-are-the-kernels 从单个设备到设备群：#from-one-device-to-a-fleet 为 WebAI 打造共享基础：#building-a-shared-foundation-for-webai Hugging Face 的 WebAI 团队的一个最大目标是让浏览器推理尽可能快且用户友好。实现这一目标需要多层努力：模型需要浏览器友好的表示方式，运行时需要构建高效的执行计划，而堆栈底层的各个 GPU 操作需要充分利用不同设备和浏览器实现的性能。","今天，我们发布了这项工作第一层：@huggingface/kernels：https://www.npmjs.com/package/@huggingface/kernels，这是一个用于从 Hugging Face Hub 加载和运行优化 WebGPU 内核的最小化库，同时提供了 huggingface.co/webgpu-kernels 上的初始 207 个内核合集：https://huggingface.co/webgpu-kernels。","该合集涵盖了用于各种机器学习架构和工作负载的操作。更重要的是，每个内核都作为完整的、版本化的包发布：其接口、着色器模板、正确性测试、基准测试以及使用说明都集中存放在 Hub 上。","我们还推出了 Fleet：https://webgpu-kernels-fleet.hf.space/，这是一个在浏览器中运行的 GPU 基准测试和测试套件，可在你的硬件上运行和评分内核。除了获取你自己机器的结果外，Fleet 还为社区提供了一种途径，可以贡献来自我们在传统测试实验室无法覆盖的设备的性能和正确性证据。在你的同意下，每次运行都会增加私密证据，帮助我们发现失败（结果错误、病态缓慢情况等）、改进内核变体，并在实际硬件上做出更好的优化决策。","在浏览器中运行的模型最终会变成一系列 GPU 操作：矩阵乘法、归一化、卷积、注意力原语、量化操作、数据布局变换，以及更多操作。WebGPU 通过一个便携的 API 在现代浏览器中提供这些操作，而 WGSL 提供了一个用于执行这些操作的着色器的通用语言。","然而，可移植性并不自动意味着高性能。两个着色器可以实现同样的操作并产生相同的输出，但在不同加速器上表现完全不同。工作组大小、内存访问模式、向量化、数据类型和融合策略都可能影响性能。最佳选择也会随输入形状、设备、浏览器以及可用的 WebGPU 功能而变化。","这就是为什么内核构成了快速浏览器推理的基础层。高层运行时的效率只能与其调度的操作效率相匹配。通过使这些操作可单独发现、可测试、可基准测试和版本化，我们可以独立改进基础，同时为上层保持稳定契约。","集合中的每个内核都有自己的仓库和内核卡。内核卡记录了操作的语义、输入、输出、属性、支持的数据类型、源文件，以及一个可运行的 @huggingface/kernels 示例。","例如，ai.onnx.Add：https://huggingface.co/webgpu-kernels/ai.onnx.Add 实现了带有多向广播的元素级加法。它是神经网络中最简单的操作之一，从残差连接到加偏置都广泛使用。其内核卡记录了两个输入、广播后的输出形状、支持的数据类型，以及针对不同形状和设备的可用变体。","在内核卡背后，仓库包含了理解和评估实现所需的制品：","这个结构将着色器转化为可重用的软件成果。接口可以在不阅读 WGSL 的情况下进行检查，正确性和性能测试会随实现一起传播，已发布的版本可以显式加载，而不是依赖未版本化的文件 URL。我们的内核也可以作为开发者构建自定义 WebGPU 内核或将这些操作集成到其自身运行时的参考实现。","运行这些内核需要支持 WebGPU 的浏览器：https://developer.mozilla.org/en-US/docs/Web/API/WebGPU_API。WebGPU 的可用性取决于浏览器、操作系统、GPU 和驱动程序。你可以在 JavaScript 中通过 \"gpu\" in navigator 来检查。","@huggingface/kernels 提供了内核存储库与应用程序之间的桥梁。使用 Hub 存储库 ID 和合约版本调用 getKernel，然后用类型化的输入数据和张量形状调用返回的函数。下面是一个小型偏置加法示例：","第二个输入被沿第一个维度广播，生成形状为 [2, 3] 的输出。加载器从清单合约和输入中推导输出形状和逻辑数据类型，然后自动分配 c。","对六个浮点数的加法故意是最小的演示。在这个规模下，GPU 往返的开销远大于计算本身。重点是调用模式：对于优化内核真正有优势的重量级操作（如矩阵乘法 ai.onnx.MatMul），调用模式完全相同。只有存储库 ID 和输入会改变。","即使是这个基础操作也说明了为什么内核需要变体。相同形状的加法可以使用直接向量化路径，而广播输入则需要不同的索引逻辑。已发布的 Add 内核包括针对相同形状、向量化广播、标量处理和通用广播的变体。运行时可以选择适合当前调用和设备的实现，而无需更改面向应用的 API。","version: 1 选项选择已发布内核合约的版本 1。它独立于 ONNX opset、操作符的 since_version 或模型修订。将这些概念分开使应用程序可以依赖稳定的面向 JavaScript 的合约，同时内核实现可以在其后演进。","那么，优化后的内核到底能带来多大差异呢？我们将自己的内核集合与 Apple M4 GPU 上的 ORT WebGPU 进行了正面比较，使用的是 ONNX Runtime Web 1.30.0-dev.20260826-b1f76d586a。我们从涵盖 207 个操作的 1,756 个测试用例开始，并保留了 809 个在两边都产生匹配输出且时间可靠的用例。","在这些比较中，我们的内核在几何平均上快 2.57 倍，按中位数计算快 1.90 倍，共有 629 胜、176 负和 4 平。以下是对四个熟悉操作的详细观察：","一些单独的胜利差距更大。一个特别困难的双线性 Einsum 用例（i, ij, j，大小为 4096）使用我们的内核运行时间为 0.136 毫秒，而 ORT WebGPU 则为 1,396 毫秒：快了超过 10,000 倍。对 [256, 4096] 的按行 CumSum 操作快了 301 倍，为 0.016 毫秒，而 ORT WebGPU 为 4.784 毫秒。这些都是不寻常的情况，而不是你在所有地方都应该预期的加速，但它们显示了当通用实现遇到慢路径时，专用内核能提供多少帮助。","我们计时的是 GPU 本身完成的工作，不包括加载内核、创建会话、上传输入、编译着色器和读回输出等设置步骤。非常短的工作负载自然更难测量，小用例可能受益于 GPU 缓存，所以这些数字最好作为有用的比较，而不是对每个应用程序的承诺。","这些结果也是针对单个操作的，而不是完整模型的。不同 GPU 和浏览器上的确切性能会有所不同，这也是 Fleet 对建立更广泛性能图谱的重要性所在。","我们还在与 ONNX Runtime 团队合作，将这些改进向上游贡献，以便更广泛的 ONNX Runtime Web 生态系统都能受益。","WebGPU 性能因 GPU、浏览器和驱动而异，因此单台机器的结果只能说明部分情况。Fleet：https://webgpu-kernels-fleet.hf.space/ 允许任何人在浏览器中运行正确性和性能检查，并查看这些内核在其硬件上的表现。","经同意，每次运行都会私下贡献证据，帮助我们发现设备特定的失效情况、比较不同版本，并改进选择规则。目标很简单：利用广泛的真实世界覆盖，让内核对所有人更加快速和可靠。","最初的 207 个内核只是一个起点，而不是最终状态。在 Hub 上独立发布内核为我们提供了一个共同的空间来检查合约、比较实现、复现正确性检查，并提升性能，而无需将每个着色器直接嵌入到每个运行时中。","该集合也是 Hub 更广泛内核生态系统的一部分：在 Kernels 页面：https://huggingface.co/kernels?platform=webgpu&sort=trending 上，WebGPU 内核与 CUDA、ROCm、Metal 及其他平台的内核并列展示，并且可以像对 Hub 上的其他工件一样进行筛选、排序和探索。","这是我们浏览器推理堆栈下一步的低级基础。我们很高兴将这些内核与更高级的模型工具连接起来，继续扩展操作覆盖范围，并让快速本地推理在 WebAI 生态系统中更易于使用。","探索 WebGPU 内核集合：https://huggingface.co/webgpu-kernels，尝试 @huggingface/kernels：https://www.npmjs.com/package/@huggingface/kernels，并加入 Fleet：https://webgpu-kernels-fleet.hf.space/ 从您的设备贡献证据，帮助我们让内核更好地服务于每个人。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Hugging Face WebAI 团队发布 @huggingface/kernels 库，并在 Hugging Face Hub 上以独立仓库形式提供 207 个 WebGPU 内核，采用 Apache-2.0 许可。","background":"每个内核包含 manifest、正确性测试、基准用例和 WGSL 着色器模板；内核卡还记录操作语义、输入输出、属性、支持的数据类型、源文件及运行示例。官方同时推出浏览器内 GPU 测试与基准套件 Fleet。","viewpoint":"Aioga 判断：该发布把浏览器推理所需的部分 GPU 操作整理为可发现、可测试、可基准评估和可版本管理的独立组件，显示出建设共享 WebAI 内核基础的产品方向。","implications":"可能影响：WebGPU 推理应用在底层算子选择、正确性验证和性能评估方面可能获得更完整的组件材料；但内核表现仍可能受输入形状、设备、浏览器及 WebGPU 特性影响，不代表跨设备性能一致。","nextStep":"后续观察：需要关注这些内核在不同硬件上的正确性与性能表现，以及 Fleet 经用户同意后收集的私有证据如何用于发现错误或异常慢速情况、改进内核变体和优化决策。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-09-01T15:48:42.652Z","sourceHash":"36d29872ceaf6459","review":{"approved":true,"groundedness":97,"clarity":92,"duplicationRisk":18,"blockingIssues":[],"notes":["“Aioga 判断”已明确标示为观点，且其判断与来源中“建设共享 WebAI 基础”的表述一致。","“可能影响”和“后续观察”均使用审慎措辞，没有将潜在效果或未来结果表述为既成事实。","如需进一步提升可读性，可简要说明 Aioga 的身份或将其改为“编辑判断”，但这不构成事实性问题。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":1,"checks":["schema","length","source-attribution","editorial-labels","inference-boundary","low-source-overlap","no-html","independent-ai-review"]}},"tags":["行业动态","Hugging Face：Blog（RSS）"],"translations":{"zh-CN":{"title":"Hugging Face 发布 @huggingface/kernels，提供 207 个 WebGPU 内核用于浏览器本地 AI 推理","summary":"Hugging Face WebAI 团队发布 @huggingface/kernels 库及 207 个以独立仓库形式托管在 Hub 上的 WebGPU 内核（Apache-2.0），每个内核带 manifest、正确性测试、基准用例和 WGSL 着色器模板。","category":"行业动态","source":"huggingface.co","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Face 发布 @huggingface/kernels，提供 207 个 WebGPU 内核用于浏览器本地 AI 推理 - Aioga AI资讯","description":"Hugging Face WebAI 团队发布 @huggingface/kernels 库及 207 个以独立仓库形式托管在 Hub 上的 WebGPU 内核（Apache-2.0），每个内核带 manifest、正确性测试、基准用例和 WGSL 着色器模板。","url":"https://www.aioga.com/news/cmtitozyq04j2ro9ydv3oxgsd/","articleBody":["TL;DR：#tldr 为什么从内核开始？：#why-start-with-kernels 内核仓库，而不仅仅是着色器：#a-kernel-repository-not-just-a-shader 从 Hub 加载内核：#loading-a-kernel-from-the-hub 内核有多快？：#how-fast-are-the-kernels 从单个设备到设备群：#from-one-device-to-a-fleet 为 WebAI 打造共享基础：#building-a-shared-foundation-for-webai Hugging Face 的 WebAI 团队的一个最大目标是让浏览器推理尽可能快且用户友好。实现这一目标需要多层努力：模型需要浏览器友好的表示方式，运行时需要构建高效的执行计划，而堆栈底层的各个 GPU 操作需要充分利用不同设备和浏览器实现的性能。","今天，我们发布了这项工作第一层：@huggingface/kernels：https://www.npmjs.com/package/@huggingface/kernels，这是一个用于从 Hugging Face Hub 加载和运行优化 WebGPU 内核的最小化库，同时提供了 huggingface.co/webgpu-kernels 上的初始 207 个内核合集：https://huggingface.co/webgpu-kernels。","该合集涵盖了用于各种机器学习架构和工作负载的操作。更重要的是，每个内核都作为完整的、版本化的包发布：其接口、着色器模板、正确性测试、基准测试以及使用说明都集中存放在 Hub 上。","我们还推出了 Fleet：https://webgpu-kernels-fleet.hf.space/，这是一个在浏览器中运行的 GPU 基准测试和测试套件，可在你的硬件上运行和评分内核。除了获取你自己机器的结果外，Fleet 还为社区提供了一种途径，可以贡献来自我们在传统测试实验室无法覆盖的设备的性能和正确性证据。在你的同意下，每次运行都会增加私密证据，帮助我们发现失败（结果错误、病态缓慢情况等）、改进内核变体，并在实际硬件上做出更好的优化决策。","在浏览器中运行的模型最终会变成一系列 GPU 操作：矩阵乘法、归一化、卷积、注意力原语、量化操作、数据布局变换，以及更多操作。WebGPU 通过一个便携的 API 在现代浏览器中提供这些操作，而 WGSL 提供了一个用于执行这些操作的着色器的通用语言。","然而，可移植性并不自动意味着高性能。两个着色器可以实现同样的操作并产生相同的输出，但在不同加速器上表现完全不同。工作组大小、内存访问模式、向量化、数据类型和融合策略都可能影响性能。最佳选择也会随输入形状、设备、浏览器以及可用的 WebGPU 功能而变化。","这就是为什么内核构成了快速浏览器推理的基础层。高层运行时的效率只能与其调度的操作效率相匹配。通过使这些操作可单独发现、可测试、可基准测试和版本化，我们可以独立改进基础，同时为上层保持稳定契约。","集合中的每个内核都有自己的仓库和内核卡。内核卡记录了操作的语义、输入、输出、属性、支持的数据类型、源文件，以及一个可运行的 @huggingface/kernels 示例。","例如，ai.onnx.Add：https://huggingface.co/webgpu-kernels/ai.onnx.Add 实现了带有多向广播的元素级加法。它是神经网络中最简单的操作之一，从残差连接到加偏置都广泛使用。其内核卡记录了两个输入、广播后的输出形状、支持的数据类型，以及针对不同形状和设备的可用变体。","在内核卡背后，仓库包含了理解和评估实现所需的制品：","这个结构将着色器转化为可重用的软件成果。接口可以在不阅读 WGSL 的情况下进行检查，正确性和性能测试会随实现一起传播，已发布的版本可以显式加载，而不是依赖未版本化的文件 URL。我们的内核也可以作为开发者构建自定义 WebGPU 内核或将这些操作集成到其自身运行时的参考实现。","运行这些内核需要支持 WebGPU 的浏览器：https://developer.mozilla.org/en-US/docs/Web/API/WebGPU_API。WebGPU 的可用性取决于浏览器、操作系统、GPU 和驱动程序。你可以在 JavaScript 中通过 \"gpu\" in navigator 来检查。","@huggingface/kernels 提供了内核存储库与应用程序之间的桥梁。使用 Hub 存储库 ID 和合约版本调用 getKernel，然后用类型化的输入数据和张量形状调用返回的函数。下面是一个小型偏置加法示例：","第二个输入被沿第一个维度广播，生成形状为 [2, 3] 的输出。加载器从清单合约和输入中推导输出形状和逻辑数据类型，然后自动分配 c。","对六个浮点数的加法故意是最小的演示。在这个规模下，GPU 往返的开销远大于计算本身。重点是调用模式：对于优化内核真正有优势的重量级操作（如矩阵乘法 ai.onnx.MatMul），调用模式完全相同。只有存储库 ID 和输入会改变。","即使是这个基础操作也说明了为什么内核需要变体。相同形状的加法可以使用直接向量化路径，而广播输入则需要不同的索引逻辑。已发布的 Add 内核包括针对相同形状、向量化广播、标量处理和通用广播的变体。运行时可以选择适合当前调用和设备的实现，而无需更改面向应用的 API。","version: 1 选项选择已发布内核合约的版本 1。它独立于 ONNX opset、操作符的 since_version 或模型修订。将这些概念分开使应用程序可以依赖稳定的面向 JavaScript 的合约，同时内核实现可以在其后演进。","那么，优化后的内核到底能带来多大差异呢？我们将自己的内核集合与 Apple M4 GPU 上的 ORT WebGPU 进行了正面比较，使用的是 ONNX Runtime Web 1.30.0-dev.20260826-b1f76d586a。我们从涵盖 207 个操作的 1,756 个测试用例开始，并保留了 809 个在两边都产生匹配输出且时间可靠的用例。","在这些比较中，我们的内核在几何平均上快 2.57 倍，按中位数计算快 1.90 倍，共有 629 胜、176 负和 4 平。以下是对四个熟悉操作的详细观察：","一些单独的胜利差距更大。一个特别困难的双线性 Einsum 用例（i, ij, j，大小为 4096）使用我们的内核运行时间为 0.136 毫秒，而 ORT WebGPU 则为 1,396 毫秒：快了超过 10,000 倍。对 [256, 4096] 的按行 CumSum 操作快了 301 倍，为 0.016 毫秒，而 ORT WebGPU 为 4.784 毫秒。这些都是不寻常的情况，而不是你在所有地方都应该预期的加速，但它们显示了当通用实现遇到慢路径时，专用内核能提供多少帮助。","我们计时的是 GPU 本身完成的工作，不包括加载内核、创建会话、上传输入、编译着色器和读回输出等设置步骤。非常短的工作负载自然更难测量，小用例可能受益于 GPU 缓存，所以这些数字最好作为有用的比较，而不是对每个应用程序的承诺。","这些结果也是针对单个操作的，而不是完整模型的。不同 GPU 和浏览器上的确切性能会有所不同，这也是 Fleet 对建立更广泛性能图谱的重要性所在。","我们还在与 ONNX Runtime 团队合作，将这些改进向上游贡献，以便更广泛的 ONNX Runtime Web 生态系统都能受益。","WebGPU 性能因 GPU、浏览器和驱动而异，因此单台机器的结果只能说明部分情况。Fleet：https://webgpu-kernels-fleet.hf.space/ 允许任何人在浏览器中运行正确性和性能检查，并查看这些内核在其硬件上的表现。","经同意，每次运行都会私下贡献证据，帮助我们发现设备特定的失效情况、比较不同版本，并改进选择规则。目标很简单：利用广泛的真实世界覆盖，让内核对所有人更加快速和可靠。","最初的 207 个内核只是一个起点，而不是最终状态。在 Hub 上独立发布内核为我们提供了一个共同的空间来检查合约、比较实现、复现正确性检查，并提升性能，而无需将每个着色器直接嵌入到每个运行时中。","该集合也是 Hub 更广泛内核生态系统的一部分：在 Kernels 页面：https://huggingface.co/kernels?platform=webgpu&sort=trending 上，WebGPU 内核与 CUDA、ROCm、Metal 及其他平台的内核并列展示，并且可以像对 Hub 上的其他工件一样进行筛选、排序和探索。","这是我们浏览器推理堆栈下一步的低级基础。我们很高兴将这些内核与更高级的模型工具连接起来，继续扩展操作覆盖范围，并让快速本地推理在 WebAI 生态系统中更易于使用。","探索 WebGPU 内核集合：https://huggingface.co/webgpu-kernels，尝试 @huggingface/kernels：https://www.npmjs.com/package/@huggingface/kernels，并加入 Fleet：https://webgpu-kernels-fleet.hf.space/ 从您的设备贡献证据，帮助我们让内核更好地服务于每个人。"]},"en":{"title":"Hugging Face released @huggingface/kernels, providing 207 WebGPU cores for browser-native AI inference","summary":"The Hugging Face WebAI team released the @huggingface/kernels library and 207 WebGPU cores (Apache-2.0) hosted on Hub as separate repositories, each kernel with manifests, correctness tests, benchmark use cases, and WGSL shader templates.","category":"Industry","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Face released @huggingface/kernels, providing 207 WebGPU cores for browser-native AI inference - Aioga AI News","description":"The Hugging Face WebAI team released the @huggingface/kernels library and 207 WebGPU cores (Apache-2.0) hosted on Hub as separate repositories, each kernel with manifests, correctn...","url":"https://www.aioga.com/en/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:43:32.983Z"},"ja":{"title":"Hugging Faceは@huggingface/カーネルをリリースし、ブラウザネイティブAI推論用の207個のWebGPUコアを提供しました","summary":"Hugging Face WebAIチームは、@huggingface/カーネルライブラリと207個のWebGPUコア(Apache-2.0)をHub上で個別リポジトリとして公開し、各カーネルにはマニフェスト、正確性テスト、ベンチマークユースケース、WGSLシェーダーテンプレートが搭載されています。","category":"業界動向","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Faceは@huggingface/カーネルをリリースし、ブラウザネイティブAI推論用の207個のWebGPUコアを提供しました - Aioga AIニュース","description":"Hugging Face WebAIチームは、@huggingface/カーネルライブラリと207個のWebGPUコア(Apache-2.0)をHub上で個別リポジトリとして公開し、各カーネルにはマニフェスト、正確性テスト、ベンチマークユースケース、WGSLシェーダーテンプレートが搭載されています。","url":"https://www.aioga.com/ja/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:43:33.422Z"},"ko":{"title":"Hugging Face는 브라우저 네이티브 AI 추론을 위한 207개의 WebGPU 코어를 제공하는 @huggingface/커널을 출시했습니다","summary":"Hugging Face WebAI 팀은 @huggingface/커널 라이브러리와 Hub에 호스팅된 207개의 WebGPU 코어(Apache-2.0)를 별도의 저장소로 공개했으며, 각 커널에는 매니페스트, 정확성 테스트, 벤치마크 사용 사례, WGSL 셰이더 템플릿이 포함되어 있습니다.","category":"업계 동향","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Face는 브라우저 네이티브 AI 추론을 위한 207개의 WebGPU 코어를 제공하는 @huggingface/커널을 출시했습니다 - Aioga AI 뉴스","description":"Hugging Face WebAI 팀은 @huggingface/커널 라이브러리와 Hub에 호스팅된 207개의 WebGPU 코어(Apache-2.0)를 별도의 저장소로 공개했으며, 각 커널에는 매니페스트, 정확성 테스트, 벤치마크 사용 사례, WGSL 셰이더 템플릿이 포함되어 있습니다.","url":"https://www.aioga.com/ko/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:43:42.025Z"},"es":{"title":"Hugging Face lanzó @huggingface/kernels, proporcionando 207 núcleos WebGPU para la inferencia de IA nativa del navegador","summary":"El equipo de WebAI de Hugging Face lanzó la biblioteca @huggingface/kernels y 207 núcleos WebGPU (Apache-2.0) alojados en Hub como repositorios separados, cada núcleo con manifiestos, pruebas de corrección, casos de uso de benchmarks y plantillas de shaders WGSL.","category":"Industria","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Face lanzó @huggingface/kernels, proporcionando 207 núcleos WebGPU para la inferencia de IA nativa del navegador - Aioga Noticias de IA","description":"El equipo de WebAI de Hugging Face lanzó la biblioteca @huggingface/kernels y 207 núcleos WebGPU (Apache-2.0) alojados en Hub como repositorios separados, cada núcleo con manifiest...","url":"https://www.aioga.com/es/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:43:42.733Z"},"fr":{"title":"Hugging Face a publié des @huggingface/kernels, fournissant 207 cœurs WebGPU pour l’inférence IA native du navigateur","summary":"L’équipe de WebAI de Hugging Face a publié la bibliothèque @huggingface/noyaux et 207 cœurs WebGPU (Apache-2.0) hébergés sur Hub comme dépôts séparés, chaque noyau comprenant manifestes, tests de correction, cas d’utilisation de benchmarks et modèles de shaders WGSL.","category":"Industrie","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Face a publié des @huggingface/kernels, fournissant 207 cœurs WebGPU pour l’inférence IA native du navigateur - Aioga Actualités IA","description":"L’équipe de WebAI de Hugging Face a publié la bibliothèque @huggingface/noyaux et 207 cœurs WebGPU (Apache-2.0) hébergés sur Hub comme dépôts séparés, chaque noyau comprenant manif...","url":"https://www.aioga.com/fr/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:43:51.765Z"},"de":{"title":"Hugging Face veröffentlichte @huggingface/kernels, die 207 WebGPU-Kerne für browsernative KI-Inferenz bereitstellten","summary":"Das Hugging Face WebAI-Team veröffentlichte die @huggingface/kernels-Bibliothek und 207 WebGPU-Kerne (Apache-2.0), die auf Hub gehostet wurden, als separate Repositories, wobei jeder Kernel Manifests, Korrektheitstests, Benchmark-Anwendungsfälle und WGSL-Shader-Vorlagen enthielt.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Face veröffentlichte @huggingface/kernels, die 207 WebGPU-Kerne für browsernative KI-Inferenz bereitstellten - Aioga KI-News","description":"Das Hugging Face WebAI-Team veröffentlichte die @huggingface/kernels-Bibliothek und 207 WebGPU-Kerne (Apache-2.0), die auf Hub gehostet wurden, als separate Repositories, wobei jed...","url":"https://www.aioga.com/de/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:43:51.822Z"},"pt-BR":{"title":"O Hugging Face lançou @huggingface/kernels, fornecendo 207 núcleos WebGPU para inferência de IA nativa do navegador","summary":"A equipe da WebAI da Hugging Face lançou a biblioteca @huggingface/kernels e 207 núcleos WebGPU (Apache-2.0) hospedados no Hub como repositórios separados, cada kernel com manifestos, testes de correção, casos de uso de benchmark e templates de shader WGSL.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"O Hugging Face lançou @huggingface/kernels, fornecendo 207 núcleos WebGPU para inferência de IA nativa do navegador - Aioga Notícias de IA","description":"A equipe da WebAI da Hugging Face lançou a biblioteca @huggingface/kernels e 207 núcleos WebGPU (Apache-2.0) hospedados no Hub como repositórios separados, cada kernel com manifest...","url":"https://www.aioga.com/pt-BR/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:44:00.866Z"},"ru":{"title":"Hugging Face выпустила @huggingface/ядра, предоставив 207 ядер WebGPU для браузерного AI-вывода","summary":"Команда Hugging Face WebAI выпустила библиотеку @huggingface/kernels и 207 ядер WebGPU (Apache-2.0), размещённые на Hub в виде отдельных репозиториев, каждое ядро с манифестами, тестами корректности, сценариями использования бенчмарков и шаблонами шейдеров WGSL.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Face выпустила @huggingface/ядра, предоставив 207 ядер WebGPU для браузерного AI-вывода - Aioga Новости ИИ","description":"Команда Hugging Face WebAI выпустила библиотеку @huggingface/kernels и 207 ядер WebGPU (Apache-2.0), размещённые на Hub в виде отдельных репозиториев, каждое ядро с манифестами, те...","url":"https://www.aioga.com/ru/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:44:00.893Z"},"ar":{"title":"أصدرت شركة Hugging Face @huggingface/نوى، موفرة 207 نوى WebGPU لاستنتاج الذكاء الاصطناعي الأصلي في المتصفح","summary":"أصدر فريق Hugging Face WebAI مكتبة @huggingface/النواة و207 أنوية WebGPU (Apache-2.0) المستضافة على Hub كمستودعات منفصلة، كل نواة تحتوي على بيانات، واختبارات الصحة، وحالات استخدام المعيار، وقوالب ظلال WGSL.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"أصدرت شركة Hugging Face @huggingface/نوى، موفرة 207 نوى WebGPU لاستنتاج الذكاء الاصطناعي الأصلي في المتصفح - Aioga أخبار الذكاء الاصطناعي","description":"أصدر فريق Hugging Face WebAI مكتبة @huggingface/النواة و207 أنوية WebGPU (Apache-2.0) المستضافة على Hub كمستودعات منفصلة، كل نواة تحتوي على بيانات، واختبارات الصحة، وحالات استخدام...","url":"https://www.aioga.com/ar/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:44:09.859Z"},"hi":{"title":"हगिंग फेस ने @huggingface/कर्नेल जारी किए, जो ब्राउज़र-नेटिव एआई अनुमान के लिए 207 वेबजीपीयू कोर प्रदान करता है","summary":"हगिंग फेस वेबएआई टीम ने हब पर होस्ट की गई @huggingface/कर्नेल लाइब्रेरी और 207 वेबजीपीयू कोर (अपाचे-2.0) को अलग-अलग रिपॉजिटरी के रूप में जारी किया, प्रत्येक कर्नेल में मैनिफेस्ट, शुद्धता परीक्षण, बेंचमार्क उपयोग के मामले और डब्ल्यूजीएसएल शेडर टेम्प्लेट थे।","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"हगिंग फेस ने @huggingface/कर्नेल जारी किए, जो ब्राउज़र-नेटिव एआई अनुमान के लिए 207 वेबजीपीयू कोर प्रदान करता है - Aioga AI समाचार","description":"हगिंग फेस वेबएआई टीम ने हब पर होस्ट की गई @huggingface/कर्नेल लाइब्रेरी और 207 वेबजीपीयू कोर (अपाचे-2.0) को अलग-अलग रिपॉजिटरी के रूप में जारी किया, प्रत्येक कर्नेल में मैनिफेस्ट, श...","url":"https://www.aioga.com/hi/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:44:09.737Z"},"it":{"title":"Hugging Face ha rilasciato @huggingface/kernel, fornendo 207 core WebGPU per l'inferenza AI nativa del browser","summary":"Il team di Hugging Face WebAI ha rilasciato la libreria @huggingface/kernels e 207 core WebGPU (Apache-2.0) ospitati su Hub come repository separati, ciascun kernel con manifesti, test di correttezza, casi d'uso benchmark e template shader WGSL.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Face ha rilasciato @huggingface/kernel, fornendo 207 core WebGPU per l'inferenza AI nativa del browser - Aioga Notizie IA","description":"Il team di Hugging Face WebAI ha rilasciato la libreria @huggingface/kernels e 207 core WebGPU (Apache-2.0) ospitati su Hub come repository separati, ciascun kernel con manifesti,...","url":"https://www.aioga.com/it/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:44:17.917Z"},"nl":{"title":"Hugging Face bracht @huggingface/kernels uit, met 207 WebGPU-cores voor browser-native AI-inferentie","summary":"Het Hugging Face WebAI-team bracht de @huggingface/kernels-bibliotheek en 207 WebGPU-cores (Apache-2.0) die op Hub werden gehost als aparte repositories uit, elke kernel met manifesten, correctheidstests, benchmark-gebruiksgevallen en WGSL-shadertemplates.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Face bracht @huggingface/kernels uit, met 207 WebGPU-cores voor browser-native AI-inferentie - Aioga AI-nieuws","description":"Het Hugging Face WebAI-team bracht de @huggingface/kernels-bibliotheek en 207 WebGPU-cores (Apache-2.0) die op Hub werden gehost als aparte repositories uit, elke kernel met manife...","url":"https://www.aioga.com/nl/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:44:18.925Z"},"tr":{"title":"Hugging Face, tarayıcı tabanlı yapay zeka çıkarımı için 207 WebGPU çekirdeği sağlayan @huggingface/çekirdekler yayımladı","summary":"Hugging Face WebAI ekibi, @huggingface/çekirdek kütüphanesini ve Hub'da bulunan 207 WebGPU çekirdeği (Apache-2.0) ayrı depolar olarak yayınladı; her çekirdek manifesto, doğruluk testleri, kıyaslama kullanım senaryoları ve WGSL shader şablonlarıyla birlikte.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Face, tarayıcı tabanlı yapay zeka çıkarımı için 207 WebGPU çekirdeği sağlayan @huggingface/çekirdekler yayımladı - Aioga AI Haberleri","description":"Hugging Face WebAI ekibi, @huggingface/çekirdek kütüphanesini ve Hub'da bulunan 207 WebGPU çekirdeği (Apache-2.0) ayrı depolar olarak yayınladı; her çekirdek manifesto, doğruluk te...","url":"https://www.aioga.com/tr/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:44:27.967Z"},"vi":{"title":"Hugging Face đã phát hành @huggingface/kernel, cung cấp 207 lõi WebGPU để suy luận AI gốc trên trình duyệt","summary":"Nhóm Hugging Face WebAI đã phát hành thư viện @huggingface/kernel và 207 lõi WebGPU (Apache-2.0) được lưu trữ trên Hub như các kho riêng biệt, mỗi kernel chứa manifests, kiểm tra độ chính xác, trường hợp sử dụng benchmark và mẫu shader WGSL.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Face đã phát hành @huggingface/kernel, cung cấp 207 lõi WebGPU để suy luận AI gốc trên trình duyệt - Tin tức AI Aioga","description":"Nhóm Hugging Face WebAI đã phát hành thư viện @huggingface/kernel và 207 lõi WebGPU (Apache-2.0) được lưu trữ trên Hub như các kho riêng biệt, mỗi kernel chứa manifests, kiểm tra đ...","url":"https://www.aioga.com/vi/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:44:27.304Z"},"id":{"title":"Hugging Face merilis @huggingface/kernel, menyediakan 207 inti WebGPU untuk inferensi AI native browser","summary":"Tim Hugging Face WebAI merilis pustaka @huggingface/kernels dan 207 inti WebGPU (Apache-2.0) yang dihosting di Hub sebagai repositori terpisah, masing-masing kernel dilengkapi manifest, tes kebenaran, kasus penggunaan benchmark, dan template shader WGSL.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Face merilis @huggingface/kernel, menyediakan 207 inti WebGPU untuk inferensi AI native browser - Berita AI Aioga","description":"Tim Hugging Face WebAI merilis pustaka @huggingface/kernels dan 207 inti WebGPU (Apache-2.0) yang dihosting di Hub sebagai repositori terpisah, masing-masing kernel dilengkapi mani...","url":"https://www.aioga.com/id/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:44:36.873Z"},"th":{"title":"Hugging Face ได้ปล่อย @huggingface/kernel ที่ให้คอร์ WebGPU จํานวน 207 คอร์สําหรับการอนุมาน AI แบบเนทีฟของเบราว์เซอร์","summary":"ทีม Hugging Face WebAI ได้ปล่อยไลบรารี @huggingface/kernels และ 207 WebGPU cores (Apache-2.0) ที่โฮสต์บน Hub เป็นที่เก็บแยกกัน โดยแต่ละเคอร์เนลมี manifests, การทดสอบความถูกต้อง, กรณีการใช้งานเบนช์มาร์ก และเทมเพลตเชดเดอร์ WGSL","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Face ได้ปล่อย @huggingface/kernel ที่ให้คอร์ WebGPU จํานวน 207 คอร์สําหรับการอนุมาน AI แบบเนทีฟของเบราว์เซอร์ - ข่าว AI Aioga","description":"ทีม Hugging Face WebAI ได้ปล่อยไลบรารี @huggingface/kernels และ 207 WebGPU cores (Apache-2.0) ที่โฮสต์บน Hub เป็นที่เก็บแยกกัน โดยแต่ละเคอร์เนลมี manifests, การทดสอบความถูกต้อง, กร...","url":"https://www.aioga.com/th/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:44:36.729Z"},"pl":{"title":"Hugging Face wypuścił @huggingface/jądra, dostarczając 207 rdzeni WebGPU do wnioskowania AI natywnie w przeglądarce","summary":"Zespół Hugging Face WebAI udostępnił bibliotekę @huggingface/kernels oraz 207 rdzeni WebGPU (Apache-2.0) hostowanych na Hubie jako osobne repozytoria, z każdym kernelem z manifestami, testami poprawności, przypadkami użycia benchmarków oraz szablonami shaderów WGSL.","category":"行业动态","source":"Hugging Face：Blog（RSS）","aggregationSource":"Hugging Face：Blog（RSS）","pageTitle":"Hugging Face wypuścił @huggingface/jądra, dostarczając 207 rdzeni WebGPU do wnioskowania AI natywnie w przeglądarce - Aioga Wiadomości AI","description":"Zespół Hugging Face WebAI udostępnił bibliotekę @huggingface/kernels oraz 207 rdzeni WebGPU (Apache-2.0) hostowanych na Hubie jako osobne repozytoria, z każdym kernelem z manifesta...","url":"https://www.aioga.com/pl/news/cmtitozyq04j2ro9ydv3oxgsd/","contentTranslated":true,"sourceHash":"048ebb60c227842d","translatedAt":"2026-09-01T15:44:45.878Z"}},"evidenceTier":"verified-news","reviewStatus":"editorial-selected","indexable":true,"editorialCover":""}}