Shypd.aiThe largest open index of AI toolsLiger-Kernel is an Open Source & Models tool that provides efficient Triton kernels for Large Language Model (LLM) training. It significantly increases multi-GPU training throughput by 20% and reduces memory usage by 60%.
Trending
Also listed in
About
Liger-Kernel is an open-source collection of Triton kernels specifically engineered to optimize Large Language Model (LLM) training. Developed by LinkedIn, this tool boasts a 20% increase in multi-GPU training throughput and a 60% reduction in memory usage, enabling longer context lengths, larger batch sizes, and massive vocabularies. It offers optimized Post-Training kernels, including DPO, ORPO, CPO, and SimPO, which can deliver up to 80% memory savings for alignment and distillation tasks. Liger-Kernel is designed for ease of use, allowing users to patch Hugging Face models with a single line of code or compose custom models using its modules. It is compatible with multi-GPU setups like PyTorch FSDP, DeepSpeed, and DDP, and integrates with popular trainer frameworks such as Axolotl and Hugging Face Trainer. The kernels are exact, ensuring computational accuracy with rigorous unit tests and convergence testing.
Capabilities
Pricing & Plans
Open Source
Free
FAQs
Trending
