Liger Kernel
LinkedInLiger Kernel is a set of Triton kernels for large language model training, replacing fused operations such as RMSNorm, RoPE, SwiGLU and cross-entropy, and adding memory-efficient post-training loss functions for DPO and ORPO. It patches into Hugging Face model classes so existing training scripts pick it up without restructuring. LinkedIn develops it.
Openness
5 high confidence- license
- BSD-2-Clause(OSI)
- source
- public
- core features withheld
- no
The LICENSE file contains the BSD-2-Clause text. The repository is public and the README describes no paid tier, enterprise edition or license-gated build, so the core is ungated.
- https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/docs/license.md recorded 2026-09-30
License doc states the project is "licensed under the [BSD 2-CLAUSE]" license and "It also includes components from projects licensed under:" Apache License 2.0 and MIT, all OSI-approved.
- https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/LICENSE recorded 2026-09-30
LICENSE file opens "BSD 2-CLAUSE LICENSE" with "Copyright 2024 LinkedIn Corporation" and grants "Redistribution and use in source and binary forms, with or" without modification under listed conditions.
- https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/README.md recorded 2026-09-30
README says Liger Kernel "is a collection of Triton kernels designed specifically for LLM training." and that it "has minimal dependencies, requiring only Torch and Triton—no extra libraries needed!" Searched the body for enterprise, license key, paid, commercial, proprietary and closed: none.
- https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/setup.py recorded 2026-09-30
setup.py installs the product from the repository with plain package requirements through "install_requires=get_default_dependencies()," and "extras_require=get_optional_dependencies(),". Searched the body for enterprise, license key, paid, commercial, proprietary and closed: none.
Adoption
3 high confidenceAdoption is measured from PyPI downloads of the `liger-kernel` package.
- https://pypistats.org/api/packages/liger-kernel/recent recorded 2026-08-18
last_month downloads = 584,477 for liger-kernel
Capability
4 medium confidenceLiger Kernel is a set of transformer training operators rather than a single kernel, and Hugging Face TRL declares it as an optional liger extra in its pyproject. Its scope is a fixed list of training operations, which keeps it one step below Apache TVM.
- https://raw.githubusercontent.com/huggingface/trl/main/pyproject.toml recorded 2026-09-30
TRL pyproject declares a Liger extra, "liger = [" with "liger-kernel>=0.8.2", and a test comment "Triggered by the DPO and KTO Liger tests".
- https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/README.md recorded 2026-09-30
README documents "We support losses like DPO, CPO, ORPO, SimPO, KTO, JSD, and many more." and says users can "Simply patch your Hugging Face model with one line of code" using RMSNorm, RoPE, SwiGLU and CrossEntropy kernels.
- https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/README.md recorded 2026-09-30
README lists a "Trainer Framework Integration" line naming Axolotl, LLaMa-Factory, SFTTrainer, Hugging Face Trainer, SWIFT and oumi, and says "Compatible with multi-GPU setups (PyTorch FSDP, DeepSpeed, DDP, etc.)."
Verified 2026-09-24