Hugging Face has announced the release of @huggingface/kernels, a minimal library designed to load and execute optimized WebGPU kernels from the Hugging Face Hub. This release includes an initial collection of 207 kernels, covering a wide range of machine learning operations and workloads. Each kernel is published as a complete, versioned package, including its interface, shader templates, correctness cases, benchmark cases, and usage instructions. Additionally, Hugging Face has launched Fleet, an in-browser GPU benchmarking and testing suite, which allows users to contribute performance and correctness evidence from their devices. This foundational effort aims to make browser inference faster and more user-friendly, with plans to connect these kernels to higher-level model tooling and expand operation coverage.
Source: Hugging Face
