[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f1q5niakbv5ems":3},{"slug":4,"category":5,"publishedAt":6,"titleZh":7,"titleEn":8,"summaryZh":9,"summaryEn":10,"models":11,"hardwares":12,"bodyZh":13,"bodyEn":14,"sourceName":15,"sourceUrl":16,"updatedAt":17},"executorch-webgpu-browser-local-inference-2026-11-11","RUNTIME","2026-11-10T23:08:59.386Z","ExecuTorch 新增浏览器 WebGPU 路径：本地推理不再必须转到另一套格式","ExecuTorch adds a browser WebGPU path without switching model formats","PyTorch 11月10日介绍 WebGPU 后端：模型导出为 .pte，通过浏览器本地执行。它减少跨工具链转换，但并不取消导出、权重下载与设备兼容验证；尚无本站独立实测。","PyTorch introduced its WebGPU backend on November 10: export to .pte and execute locally in the browser. This removes a cross-toolchain conversion, not export, downloads or device validation. AraoAI has not independently benchmarked it.",[],[],"11月10日，PyTorch 官方介绍 ExecuTorch 的 WebGPU 后端。文章元数据发布时间为22:15 UTC，落在本次采集窗口内。开发者通过 torch.export 与 WebGPUPartitioner 导出 .pte 文件，再由浏览器中的 WebAssembly 运行时调用 WebGPU。新意不是“模型不用处理就能打开”，而是不用先转入另一套推理框架的格式。\n\n1. 同一份产物，不等于所有设备表现相同。官方介绍其以 WGSL 计算着色器执行，并按平台映射到 Metal、Vulkan 或 D3D12。浏览器、驱动、算子与数据类型的支持仍决定实际能否运行。交付物至少包括模型文件、导出配方和对应模型卡，不能只留下一个演示链接就称可复现。\n\n2. 本地计算，不等于首次访问零成本。官方 Playground 将提示、图片和音频留在设备上处理；模型文件仍需下载，页面资源也需要正常加载。离线能力应在缓存完成后断网验证，隐私判断则要覆盖页面其他网络请求，不能把推理后端的设计承诺扩大为任意第三方页面都不传数据。\n\n3. 支持清单不是性能榜单。官方列举文本、视觉、语音及端侧微调示例，同时明确后端仍处早期开发阶段。本文不搬用官方速度倍数作为选型结论：其对照针对特定设备、浏览器和模型，CPU与GPU的激活精度也不同，不能外推成原生GPU引擎的替代成绩。\n\n实际部署先固定浏览器、设备、模型仓库与导出版本，再确认结果正确、首次加载时间、内存峰值及长请求稳定性。运行时开源不替代模型许可；逐个检查权重与应用依赖。本次确认的是一条正式介绍的浏览器本地部署路径，没有独立社区复现数据，不新增速度记录。\n\n来源：https:\u002F\u002Fpytorch.org\u002Fblog\u002Fexecutorch-webgpu-pytorch-models-on-the-gpu-in-the-browser","PyTorch introduced the ExecuTorch WebGPU backend on November 10. The page metadata records publication at 22:15 UTC, inside this reporting window. Developers export a .pte file using torch.export and WebGPUPartitioner, then execute it through a WebAssembly runtime and WebGPU in the browser. This avoids converting to another framework's format; it does not eliminate export.\n\n1. One artifact does not guarantee identical behavior everywhere. The backend uses WGSL compute shaders and maps to Metal, Vulkan or D3D12. Actual usability still depends on browser, driver, operator and data-type support. A reproducible package needs the model file, export recipe and model card, not only a demo URL.\n\n2. Local computation is not a zero-cost first visit. The official Playground processes prompts, images and audio locally, but model files must still be downloaded and page assets loaded. Test offline use after caching and inspect other page requests before making a blanket privacy claim about any third-party application.\n\n3. A support list is not a performance ranking. The article lists text, vision, speech and on-device fine-tuning examples while describing the backend as early development. We do not turn the publisher's speed ratios into a hardware recommendation: they concern specific devices, browsers and models, and CPU and GPU activation precision differs. They do not establish parity with native GPU engines.\n\nFor deployment, pin browser, device, model repository and export version; then assess correct outputs, first-load time, peak memory and long-request stability. Runtime openness does not replace model licensing. This is a documented local deployment path, not an independently replicated benchmark, and adds no speed record.\n\nSource: https:\u002F\u002Fpytorch.org\u002Fblog\u002Fexecutorch-webgpu-pytorch-models-on-the-gpu-in-the-browser","PyTorch 官方博客","https:\u002F\u002Fpytorch.org\u002Fblog\u002Fexecutorch-webgpu-pytorch-models-on-the-gpu-in-the-browser","2026-11-10T23:08:59.387Z"]