{"categories":[{"label":"Artificial Intelligence","url":"https://skillfed.io/packages/category/scientific-engineering-artificial-intelligence/5"}],"enrichment":{"capability":"Liger Kernel provides optimized Triton kernels for LLM training, implementing fused operations like RMSNorm, RoPE, SwiGLU, and CrossEntropy to reduce memory usage and increase training throughput.","skillfed_tags":["gpu-optimization","llm-training","kernel-fusion"],"use_cases":["Fine-tune large language models like LLaMA 3-8B on consumer-grade or enterprise GPU clusters with reduced memory footprint and faster training time","Implement post-training alignment tasks (DPO, ORPO, CPO) with up to 80% memory savings for distillation and preference optimization","Train vision-language models (e.g., Qwen2-VL) on multi-GPU setups using FSDP or DeepSpeed with kernel-fused operations","Extend context length or batch size in existing training pipelines by reducing per-layer memory overhead through kernel fusion","Integrate optimized kernels into custom training loops or frameworks that already use torch and triton"],"what_it_does":"Liger Kernel is a collection of Triton-based GPU kernels optimized specifically for large language model training. It implements fused operations for common LLM layers\u2014RMSNorm, RoPE, SwiGLU, CrossEntropy, and post-training losses like DPO, ORPO, and CPO\u2014that reduce memory overhead and increase computational efficiency through kernel fusion and in-place operations. The package requires only torch and triton as dependencies and integrates with PyTorch FSDP, DeepSpeed, and popular training frameworks like Hugging Face Trainer and Lightning.\n\nThe package is designed to work out of the box: you can patch a Hugging Face model with a single line of code or compose custom models using Liger's module APIs. It supports multi-GPU training across CUDA, ROCm, and Ascend NPU platforms. All computations are exact\u2014no approximations\u2014with forward and backward passes validated against standard implementations. The project is actively maintained by LinkedIn, with recent releases adding support for new loss functions and hardware platforms.","worth_installing":"Yes, if you are training LLMs on GPU hardware. The package has low install friction, active maintenance, permissive licensing, no known vulnerabilities, and integrates cleanly with standard PyTorch training workflows. Start with a trial on a small model to verify the stated performance gains apply to your setup; the exact speedup and memory savings depend on your hardware, model size, and training configuration."},"id":"liger-kernel","links":{"html":"https://skillfed.io/packages/liger-kernel","md":"https://skillfed.io/packages/liger-kernel.md","pypi":"https://pypi.org/project/liger-kernel/"},"maintenance":{"status":"active"},"meta":{"latest_release":"2026-07-23","license_spdx":null,"license_treatment":"permissive","name":"liger-kernel","python_support":"unspecified","summary":"Efficient Triton kernels for LLM Training"},"popularity":{"monthly_downloads":620759,"position":5723,"tier":"top_15000"},"security":{"n_vulnerabilities":0},"version":"0.8.1"}