Hugging Face Releases WebGPU Kernels

Hugging Face's WebAI team has released the @huggingface/kernels library and 207 WebGPU kernels to improve the speed and usability of browser-based AI inference. The goal is to optimize GPU operations at the lowest layer of the browser inference stack, enabling higher-level runtimes to establish more efficient execution plans.

Kernel Composition and Features

The released kernels include operations used in various ML architectures, such as matrix multiplication, normalization, convolution, and attention. Each kernel is published on the Hub as a versioned independent package, rather than just simple shader code.

• Complete Contract: Each repository includes a manifest.json defining the operation's inputs/outputs, data types, and shape rules, accuracy test cases (test.json), benchmark cases (bench.json), and parameterized WGSL shader templates.
• Reproducibility and Verification: Interfaces can be verified without reading the WGSL code, and accuracy and performance data are delivered alongside the implementation, making them usable as reliable software artifacts.

Fleet: Crowdsourced Benchmarking Tool

Hugging Face also launched Fleet, a tool for testing GPU performance and accuracy directly in the browser. Data executed with user consent is used to collect cases of errors or performance degradation on real hardware environments, allowing the community to contribute to kernel optimization across diverse devices.