Hugging Face launches @huggingface/kernels with 207 WebGPU kernels for browser-based AI inference

Hugging Face’s WebAI team introduced @huggingface/kernels, a minimal JavaScript library designed to load and execute optimized WebGPU kernels directly from the Hugging Face Hub. The initial release includes 207 kernels covering operations common to many machine learning architectures, such as matrix multiplications, normalizations, convolutions, attention primitives, quantization, and data-layout transformations.

Each kernel is published as a versioned, self-contained package on the Hugging Face Hub, including its interface definition, WGSL shader templates, correctness tests, benchmark cases, and usage instructions. For example, the ai.onnx.Add kernel repository documents inputs, outputs, supported data types, and multiple implementation variants for equal-shape, vectorized, scalar, and general broadcasting scenarios.

The team also launched Fleet, a browser-based benchmarking and testing suite that scores kernel performance on user hardware. With user consent, Fleet crowdsources correctness and performance evidence from real-world devices, helping identify failures, slow cases, and optimization opportunities across diverse GPUs and browser implementations.

The @huggingface/kernels loader provides a stable API to retrieve and invoke kernels by repository ID and contract version. The library handles tensor shape derivation, data type selection, and memory allocation, allowing developers to focus on application logic rather than low-level GPU details.

The project emphasizes portability and performance by making kernel variants discoverable and testable. For instance, the Add kernel includes multiple implementation paths, and the runtime selects the appropriate variant based on input shapes and device capabilities without requiring changes to the calling code.