LinkedIn's engineering team develops and open-sourced Liger-Kernel.

A collection of Triton GPU kernels that accelerate LLM training and reduce memory usage for common ops like RMSNorm, RoPE, and cross-entropy.

Installed as a Python package and used with PyTorch training pipelines on CUDA GPUs.

Open-sourced in 2024 and actively maintained through 2026.

Increases multi-GPU training throughput by up to 20% and cuts memory use by up to 60 to 80% for alignment and distillation.

Works with Flash Attention, PyTorch FSDP, and DeepSpeed, and integrates into Axolotl, LLaMA-Factory, SFTTrainer, and ms-swift.

Open-source (BSD-2-Clause license) on GitHub from LinkedIn, self-hosted, purely a local compute library with no data transmission.