Introducing @huggingface/kernels: 200+ WebGPU Kernels for Local AI

Hugging Face’s WebAI team has put the first layer of its browser inference plan on the table: @huggingface/kernels, a small JavaScript package for loading optimized WebGPU kernels from the Hugging Face Hub. Alongside it comes an initial batch of 207 kernels, each published as its own repository in the webgpu-kernels organization. The pitch is simple. If browser AI is going to be fast, the GPU operations underneath it need to be versioned, testable, and easier to tune across different devices and browsers.

That matters because browser inference is not one big problem. It’s matrix multiplies, normalizations, convolutions, attention pieces, quantization, and data shuffling, all chained together. WebGPU and WGSL give those operations a portable home, but portability alone doesn’t make them quick. Two shaders can do the same job and still behave very differently depending on workgroup size, memory access, vectorization, the device, the browser, and which WebGPU features are available. Hugging Face is betting that the kernel layer deserves first-class treatment, not just whatever code falls out of a runtime.

Each kernel comes with its own repository and kernel card. The card spells out the operation’s semantics, inputs, outputs, data types, and source files, and it includes a ready-to-run example for @huggingface/kernels. Under that, the repository bundles the contract and evidence: manifest.json for the operation definition, metadata.json for provenance, test.json for correctness cases, bench.json for benchmark and tuning cases, and *.wgsl.jinja files for parameterized WGSL templates. The idea is to make a shader feel less like a loose file and more like a reusable software artifact.

The examples are meant to be straightforward. A call to getKernel points at a Hub repository ID and a contract version, then runs typed inputs through the loaded kernel. Hugging Face uses ai.onnx.Add as the simplest demo, including broadcasting and automatic output-shape allocation. Even there, the library can pick between variants such as equal-shape, vectorized broadcasting, scalar processing, and general broadcasting without changing the application-facing API. That separation between the published contract and the implementation is the whole point.

Hugging Face also launched Fleet, a browser-based benchmarking and testing suite. It runs the kernels on your hardware, and with consent, privately contributes evidence back to the project so the team can spot failures, slow paths, and better variant choices across devices that a normal test lab would miss. On an Apple M4 GPU, the project says its kernels beat ORT WebGPU across 809 matched cases, with a 2.57x geometric-mean speedup and 1.90x at the median. The standout numbers are wild: one bilinear Einsum case was more than 10,000x faster, and a row-wise CumSum was 301x faster. Those are edge cases, not a promise, but they show why specialized kernels still matter.