Kernels

Kernels is a Hugging Face open-source project that lets Python libraries and applications load compute kernels directly from the Hugging Face Hub. It consists of the kernels Python package for loading kernels and the kernel-builder tooling for building, packaging, and distributing them. The documentation lists v0.18.1 as the latest stable version.

What It Is

The Kernel Hub treats kernels as a first-class repository type on the Hub, with dedicated pages that surface supported hardware and versions. Hub kernels differ from traditional Python kernel packages: a kernel can be loaded from paths outside PYTHONPATH, multiple versions of the same kernel can be loaded in one Python process, and kernels must support recent Python versions and various PyTorch build configurations (different CUDA versions and C++ ABIs), along with older C library versions.

How the Workflow Works

Users install the package with pip (requires torch>=2.5 and CUDA in the quickstart), then call get_kernel with a Hub repository such as kernels-community/activation and a version. The returned module exposes the kernel functions, which run on tensors such as CUDA tensors. Kernel authors use kernel-builder to write, build with Nix, develop locally, and publish kernels. The docs also cover locking kernel versions, using layers, a kernels CLI (cache, info, lock, download, benchmark, verify-signature), and guides for developing kernels with agents.

Hardware Support

The README groups hardware into tiers. CUDA and CPU are tier 1 and validated in CI. ROCm, XPU, Metal, and TPU are tier 2. Huawei NPU and Neuron are tier 3. Neuron and TPU support are described as experimental.