Liger-Kernel

Overview

Liger-Kernel is a collection of efficient Triton kernels designed to accelerate large language model training. It provides drop-in replacements for common operations like attention and normalization, reducing memory usage and improving throughput. The kernels are implemented in Python using OpenAI's Triton language.

Best for

Best for
Developers training large language models who want to optimize throughput and memory without rewriting their training loops.

Use cases

• Speed up LLM training by replacing standard PyTorch operations with optimized kernels
• Reduce GPU memory consumption during training of large transformer models
• Integrate into existing training pipelines with minimal code changes

Notes

Liger-Kernel is a collection of efficient Triton kernels designed to accelerate large language model training. It provides drop-in replacements for common operations like attention and normalization, reducing memory usage and improving throughput. The kernels are implemented in Python using OpenAI’s Triton language.

6,400 stars on GitHub. Last updated 2026-06-01. Licensed BSD-2-Clause.

Use cases

• Speed up LLM training by replacing standard PyTorch operations with optimized kernels
• Reduce GPU memory consumption during training of large transformer models
• Integrate into existing training pipelines with minimal code changes

Pros

• Significant performance gains with simple drop-in replacement
• Open source with 6,400 GitHub stars indicating community trust
• Reduces memory footprint, enabling larger batch sizes or models

Cons

• Requires Triton compiler and compatible GPU hardware
• Limited to operations covered by the provided kernels
• May not support all model architectures or custom layers

Indexed from awesome-llm and enriched against its public facts.