liger-kernel-perf

To optimize Triton kernels for better GPU performance, you can profile existing kernels to detect memory or compute bottlenecks, generate parameter variants, and benchmark them to select the fastest configuration for your specific NVIDIA GPU.

To reduce memory usage of Liger kernels during inference, you can systematically sweep kernel parameters and apply diagnosis-driven optimization strategies to generate variants with a lower memory footprint while ensuring correctness.

Profiling identifies whether a Triton kernel is memory-bound or compute-bound by recording a detailed optimization profile during execution, which then drives the generation of targeted variant implementations to overcome the specific bottleneck.

Yes, you can tune Liger kernels for specific NVIDIA GPU architectures like Ampere, Hopper, or Blackwell by profiling the kernel on your target hardware and selecting the best-performing configuration variant from the generated sweep.

Yes, you need a CUDA-capable GPU and a development environment with Triton and Python installed to run Triton kernel benchmarking and validate the correctness of the optimized variants.

The best way to automate Triton kernel tuning without breaking existing tests is to run safety gates that execute smoke tests, full test suites, and checkstyle checks before applying any code changes to the winning variant.