README.md

---
title: Kernels Community
emoji: 🔥
colorFrom: gray
colorTo: green
sdk: static
pinned: false
---

<div align="center">
<img
  src="https://github.com/user-attachments/assets/4b5175f3-1d60-455b-8664-43b2495ee1c3"
  alt="Kernels Community"
  width="200"
/>
</div>

# Kernels Community

The **Kernel Hub** allows Python libraries and applications to **load optimized compute kernels directly from the Hugging Face Hub**.

You can think of it as the Model Hub, but for low-level, high-performance code (kernels) that accelerate specific operations, often on GPUs.

Instead of manually managing complex dependencies, dealing with compilation flags, or building libraries like Triton or CUTLASS from source, the `kernels` library lets you fetch and run **pre-compiled, optimized kernels** on demand.

---

## Repos

The Kernel Hub team maintains two core repos:

### `kernels`

The main repository containing:
- A deterministic kernel builder
- A Python library for loading and executing kernels

Documentation:  
https://huggingface.co/docs/kernels/

### `kernels-community`

A repository that contains the source code for all of the kernels-community kernels

Source code:  
https://github.com/huggingface/kernels-community

---

## Compliant Kernels

Kernels published on the Hub are designed to be:

- **Portable** — loadable from paths outside `PYTHONPATH`
- **Isolated** — multiple kernel versions can run in the same process
- **Compatible** — support different Python versions, PyTorch builds, and C++ ABIs

---

Learn more about the Kernel Hub and the `kernels` library by reading the docs:    
https://huggingface.co/docs/kernels/index