Liger Kernel
Project record
Liger Kernel is an open-source collection of Triton GPU kernels for training large language models, developed at LinkedIn. It provides Hugging Face-compatible implementations of layers such as RMSNorm, RoPE, SwiGLU, and cross-entropy (including a fused linear cross-entropy), plus memory-efficient losses for post-training methods such as DPO, ORPO, and KTO.27
- Repository: Repository (external site: github.com)
- Documentation: Documentation (external site: linkedin.github.io)
- License: LICENSE (external site: raw.githubusercontent.com)
- Release notes: Releases (external site: github.com)
- Paper: Liger Kernel: Efficient Triton Kernels for LLM Training (arXiv 2410.10989) (external site: arxiv.org)
Availability and license
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| Source codeIs the source code publicly readable? | Public | 1 |
| DocumentationIs user documentation published? | Public | A documentation site covers getting started, examples, and high- and low-level APIs.4 |
| InstallationAre installation instructions or packages publicly available? | Public | Installable from PyPI (pip install liger-kernel), as a nightly package, or from source.26 |
| Supported platformsAre supported operating systems or hardware documented? | Public | The README lists dependencies for NVIDIA CUDA (torch 2.1.2 or later, Triton 2.3.1 or later), AMD ROCm (torch 2.5.0 or later, Triton 3.0.0 or later), and Ascend NPU, plus optional CUDA-only cuTile and CuTe DSL backends (the latter targeting Hopper and Blackwell GPUs). It notes that the kernels inherit Triton's hardware compatibility.2 |
| Release statusAre versioned releases published? | Public | Versioned releases are published on GitHub and PyPI; v0.8.3 was released on September 16, 2026.56 |
What it is useful for
Reducing GPU memory use and increasing throughput when training or fine-tuning language models, either by patching a Hugging Face Transformers model with one line of code or by composing models from its modules. It works with PyTorch FSDP, DeepSpeed, and DDP, and is integrated in trainers such as Axolotl, LLaMA-Factory, TRL's SFTTrainer, and the Hugging Face Trainer.2
Organization context
U.S. eligibility
Eligible · basis: U.S.-governed project
The project is hosted in LinkedIn's GitHub organization, its LICENSE names LinkedIn Corporation as copyright holder, and LinkedIn Engineering describes it as LinkedIn's open-source project. LinkedIn Corporation is listed as a U.S. subsidiary in Exhibit 21 to Microsoft's Form 10-K for fiscal 2026 (see the LinkedIn record).1378
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.