huggingface/kernels

Build compute kernels and load them from the Hub.

What it solves

It provides a way for Python libraries and applications to dynamically load optimized compute kernels directly from the Hugging Face Hub, bypassing the need for traditional Python package installation via PYTHONPATH.

How it works

The project consists of a Python package called kernels that allows users to to download and execute kernels from the Hub. For developers, it includes a kernel-builder tool to create these portable, versioned kernels. These kernels are designed to be portable across different Python versions, PyTorch build configurations (including various CUDA versions and C++ ABIs), and older C library versions.

Who it’s for

  • AI Developers: Those who need high-performance compute kernels for their applications without managing complex installation processes.
  • Kernel Authors: Developers who want to build and distribute optimized compute kernels that are compatible across various environments.

Highlights

  • Dynamic Loading: Load kernels directly from the Hugging Face Hub.
  • Portability: Kernels can be loaded from paths outside the standard Python path.
  • Version Control: Support for loading multiple versions of the same kernel within a single Python process.
  • Compatibility: Broad support for recent Python versions and various PyTorch/CUDA configurations.

Related

  • Dispatch
    Hugging Face kernel-builder: A Guide to Building and Scaling Production-Ready CUDA Kernels

    Hugging Face introduces the kernel-builder library to simplify the development, multi-architecture compilation, and distribution of production-ready CUDA kernels via the Hugging Face Hub.

  • Dispatch
    Hugging Face Kernel Hub Release

    Hugging Face has introduced the Kernel Hub, a centralized repository for loading pre-compiled, optimized compute kernels directly into Python applications to accelerate GPU operations without local compilation.

  • Dispatch
    Hugging Face Kernels Major Updates

    Hugging Face announced major updates to its Kernels project, introducing a new kernel repository type, trusted publishers and code signing, revamped CLIs, broader framework support, and foundations for agentic kernel development.

  • Project
    huggingface/huggingface_hub

    The official Python client and CLI for the Hugging Face Hub, enabling users to programmatically download, upload, and manage ML models, datasets, and demos.

  • Dispatch
    Easily Build and Share ROCm Kernels with Hugging Face

    Hugging Face introduces a streamlined workflow using the kernels library and kernel-builder to build, package, and share ROCm-compatible GPU kernels on the Hugging Face Hub.