liger-kernel-dev

To convert PyTorch operations into Triton kernels, provide the operation description to initiate a three-stage pipeline that analyzes the logic, generates the kernel, and validates it with unit tests and benchmarks.

The Triton kernel development workflow uses a three-stage pipeline: Analyze, Generate, and Validate. Human review is required between each stage to ensure the final implementation meets production-ready standards.

Liger Kernel Triton development strictly enforces NVIDIA GPUs only. The automated generation pipeline does not support AMD GPUs or other hardware platforms for kernel validation and benchmarking.

The best way to generate unit tests and benchmarks for custom Triton kernels is to use an automated generation pipeline that creates module wrappers and optional performance plots alongside the kernel implementation.

You can update existing Triton kernels in Liger Kernel as well as create new ones. The development pipeline supports modifying current implementations and converting PyTorch operations into new kernels.

Generated Triton kernels require human review between the Analyze, Generate, and Validate stages to ensure quality control. This step verifies the production-ready implementation before finalizing the unit tests and module wrappers.