This repository serves as a curated catalog of openly available Triton kernels, aiming to save practitioners time and provide learners with high-quality examples. It also identifies community-needed kernels to guide development efforts and offer entry points for new contributors.

How It Works

The project collects Triton kernels from various sources, organizing them into a searchable overview within the /kernels directory. Each entry is documented in a Markdown file, facilitating easy discovery and learning. This centralized approach leverages Triton's accessibility for Python and PyTorch users compared to traditional CUDA.

Quick Start & Requirements

• Install: No explicit installation instructions are provided for the catalog itself; it's a collection of code and documentation.
• Prerequisites: Access to the Triton programming language and potentially PyTorch for running the kernels.
• Resources: Viewing the catalog requires standard web browsing. Running kernels will depend on individual kernel requirements, likely including a GPU.
• Links:
  • Searchable overview: [link to overview]
  • A Practioner's Guide to Triton: [link to guide]
  • Triton Tutorials: [link to tutorials]

Highlighted Details

• Catalogs kernels from prominent libraries like Flash Attention, Unsloth, xformers, Applied AI, and AO.
• Includes kernels for advanced techniques such as MoE, GaLoRe, HQQ, DoRA, and various attention mechanisms.
• Highlights torch.compile's ability to codegenerate Triton kernels.
• Provides a structured contribution process via Markdown templates.

Maintenance & Community

Initiated by Hailey and Umer, with contributions welcomed from the community.

Licensing & Compatibility

The repository itself does not specify a license. Individual kernels within the catalog will carry their own licenses, which must be checked for compatibility with commercial or closed-source use.

Limitations & Caveats

The catalog's primary function is aggregation; it does not guarantee the performance, correctness, or licensing compliance of individual kernels. Users must independently verify these aspects for each kernel they intend to use.