TechPrufer
Get started
Hugging Face
Hugging Face logoHugging Face · Feature trail

WebGPU Kernels for Browser Inference

3 posts · Jan 2026 – Aug 2026

What this trail covers

Hugging Face is building a foundational layer for fast and user-friendly browser inference by releasing `@huggingface/kernels`, a library for loading and running optimized WebGPU kernels from the Hugging Face Hub. This initial release includes 207 kernels covering a wide range of machine learning operations, each published as a versioned package with explicit contracts, correctness tests, and benchmark cases. A browser-based benchmarking suite, Fleet, is also launched to crowdsource performance.

All Hugging Face trails
  1. Introducing @huggingface/kernels: 200+ WebGPU Kernels for Local AIThis post introduces `@huggingface/kernels`, a library for loading and running optimized WebGPU kernels from the Hugging Face Hub, along with an initial collection of 207 kernels. It details the structure of kernel repositories, including manifest, metadata, correctness, and benchmark cases, and demonstrates loading and executing kernels with a JavaScript example. Performance comparisons against ORT WebGPU show significant speedups for individual operations. The Fleet benchmarking tool is also launched to gather real-world performance and correctness evidence.
  2. 🤗 Kernels: Major UpdatesThis post introduces a new 'kernel' repository type on the Hugging Face Hub for packaging and distributing custom compute kernels, enhancing discoverability and providing metadata on supported hardware and backends. It details significant security enhancements including 'trusted kernel publishers' (requiring explicit opt-in for non-trusted sources) and kernel signing using Sigstore's cosign for integrity verification. The CLIs for `kernels` and `kernel-builder` have been revamped for better separation of concerns and agent optimization. Support for Torch Stable ABI and Apache TVM FFI has been added, enabling broader framework interoperability. The post also outlines how these updates lay the foundation for agentic kernel development by providing structured scaffolding, reproducible builds, and integrated benchmarking capabilities via HF Jobs.
  3. We Got Claude to Build CUDA Kernels and teach open models!This post details how Claude was used to generate CUDA kernels for diffusers. The generated kernels are then used to teach open models, with a focus on enabling them to run efficiently, potentially for browser inference scenarios.