ai-agent

Liger Kernel

Efficient Triton Kernels for LLM Training

6,623 stars604 forksPythonUpdated 9/18/2026100% free · open source

What it does

Provides drop‑in, high‑performance Triton kernels that speed up the matrix‑multiply layers of large‑language‑model training.

When to use it

•You are training a transformer‑based LLM on GPUs and need > 10% speed‑up on the linear layers.
•Your code already uses PyTorch nn.Linear and you want a minimal‑code change to get faster kernels.
•You have a CUDA‑compatible GPU (11.8+) and want to squeeze more FLOPs out without rewriting the whole training loop.

Ready-to-paste prompt

python examples/train_gpt2.py --model-name gpt2-medium --batch-size 16 --seq-length 1024 --epochs 5

Heads up: Liger‑Kernel only works on GPUs with compute capability ≥ 8.0 (e.g., A100, H100, RTX 4090) and requires CUDA 11.8+; mismatched CUDA/Triton versions will cause a runtime import error.

Saves to your device

Use with Claude

New

Skip the builder — one click puts this in Claude, Cursor, Antigravity and more.

✅ Light setup

Installs with a command or two; your AI agent can do it for you.

Try it instantly — no install

Claude Code

mkdir -p ~/.claude/skills/liger-kernel && curl -fsSL https://workflowstacks.com/api/skills/liger-kernel/claude-skill -o ~/.claude/skills/liger-kernel/SKILL.md

Open in another AI app

Opens the app with this repo with the prompt ready to go — no copy-paste needed.

Connect the whole catalog (MCP)

claude mcp add --transport http workflowstacks https://workflowstacks.com/api/mcp

Adds a WorkflowStacks connector to Claude Code: search and load any skill here by chatting.

How Liger Kernel works

Codeflow

Free to inspect

Liger Kernel is a very large Python project (~104k lines across 372 code files, plus 129 test files). Setup is light: installs like a normal app. Reading the code is optional. Last commit this month, BSD-2-Clause license, has a test suite.

Size

Very large codebase

~104k lines · 372 code files · days to read — use, don't read

Setup

One-command install

Installs like a normal app. Reading the code is optional.

Runs on

Python

No API keys detected

Python 61%Cuda 33%NASL 4%C++ 1%CMake 1%

Where to start reading

1
README.md
Start here — what it does and how to install it
2
AGENTS.md
The instructions the AI actually follows
3
pyproject.toml
Dependencies and the commands it exposes
4
src/liger_kernel/__init__.py
Inside src/ — the main logic begins here
5
examples/alignment/accelerate_config.yaml
A worked example — copy this to get going

What's in each folder

src/Core code — the actual logic282 files

examples/Examples you can copy41 files

docs/Documentation15 files

liger_cute_kernels/Folder225 files

benchmark/Evaluations & benchmarks60 files

licenses/Folder5 files

dev/Folder3 files

test/Tests — proof it works131 files

READMEHas testsDocumentedCI checksExamples includedBSD-2-Clause licenseUpdated this month