Hugging Face ships a WebGPU kernel loader and 207 versioned kernels for browser inference

Hugging Face's WebAI team published @huggingface/kernels, described as a minimal library for loading and running optimised WebGPU kernels from the Hugging Face Hub, together with an initial collection of 207 kernels hosted at huggingface.co/webgpu-kernels. The kernels are Apache-2.0 licensed. The post is credited to Nico Martin and Joshua (Xenova) and is dated September 1, 2026.

The clearest difference from the previous arrangement is packaging rather than raw shader code. Each kernel has its own repository and kernel card documenting the operation's semantics, inputs, outputs, attributes, supported data types, source files, and a ready-to-run example. Inside, manifest.json acts as the source of truth for the operation contract, defining inputs, outputs, attributes, type constraints and shape derivation; metadata.json records the kernel identifier, digests and provenance; test.json holds correctness cases; bench.json holds benchmark and tuning cases; and *.wgsl.jinja files hold the parameterised WGSL implementations. The post's own framing is that this turns a shader into a reusable software artifact with an inspectable interface, so published versions can be loaded explicitly rather than depending on an unversioned file URL.

The stated motivation is that portability is not performance. Two shaders can implement the same operation and produce the same output while behaving completely differently across accelerators, and the source lists workgroup sizes, memory access patterns, vectorisation, data types and fusion strategies as factors — adding that the best choice can change with input shape, device, browser and available WebGPU features. Higher-level runtimes, in its framing, can only be as efficient as the operations they dispatch.

Access terms are limited in the material given. Installation is via npm install @huggingface/kernels@preview, and running the kernels requires a browser with WebGPU support, which the post says depends on the browser, operating system, GPU and driver, and can be detected in JavaScript with "gpu" in navigator. The collection also covers operations across a wide variety of machine learning architectures and workloads, including matrix multiplications, normalisations, convolutions, attention primitives, quantisation operations and data-layout transformations.

Fleet is the accompanying browser-based benchmarking and testing suite that runs and scores kernels on the reader's own hardware. The post says that with consent, every run adds private evidence that can help find failures such as incorrect results and pathologically slow cases, improve kernel variants, and inform optimisation decisions across real-world hardware. It positions this as a way to contribute evidence from devices a conventional test lab would not cover.

A bounded first step: install the preview package, confirm WebGPU presence in the target browser, run Fleet once on your own machine, and load a specific published kernel version explicitly rather than a floating URL. Treat both the preview tag and crowdsourced benchmark evidence as provisional until correctness cases and repeat runs agree.