@huggingface/kernels

Load and run WebGPU compute kernels from the Hugging Face Hub.

Early preview — the API may still change and kernel coverage is growing. Updates will follow regularly.

Install

npm install @huggingface/kernels@preview

Usage

import { getKernel } from "@huggingface/kernels";

const relu = await getKernel("webgpu-kernels/ai.onnx.Relu", { version: 1 });
const { y } = await relu({ x: { data: new Float32Array(100), shape: [10, 10] } });
console.log(y.data, y.shape, y.dtype);

Every load names what it wants: version: 1 follows the repository's v1 branch, which moves as fixes land. Pass a 40-character commit revision instead to pin bytes that can never change. Each kernel's own README shows the call for that kernel.

Inputs take the natural typed array — Float32Array, Uint8Array, Float16Array and the rest are widened to their WebGPU storage form for you. Results come back the same way: float16 as a Float16Array where the browser provides one, otherwise as raw binary16 words in a Uint16Array. Pass { output: "gpu" } to keep results on the device and feed them straight into the next call; you own those and release them with destroy().

Kernels load from the webgpu-kernels organization by default. A kernel is code that runs on your GPU, so loading one from another publisher is opt-in — pass { trustRemoteCode: true } if you trust them.

Runs in the browser, and needs WebGPU (navigator.gpu).

License

Apache-2.0. Bundled third-party code is listed in THIRD-PARTY-NOTICES.md.