liger-kernel-perf

What problem does it solve?

The skill automates the end‑to‑end tuning of Liger Kernel Triton kernels, eliminating the manual effort of profiling, diagnosing bottlenecks, generating variant implementations, and validating correctness.

Core Features & Use Cases

• Profiling: Detects whether a kernel is memory‑bound or compute‑bound and records a detailed optimization profile.
• Variant Generation: Systematically sweeps parameters and applies diagnosis‑driven strategies from a catalog of optimizations.
• Safety Gates: Runs smoke tests, full test suites, and checkstyle checks before any code change is applied.
• Finalization: Commits the winning variant, produces comparison plots, and creates a PR with only the necessary changes.
• Use Cases: Speed up inference, lower memory footprint, or balance both for kernels like rms_norm, cross_entropy, or any custom Liger kernel on NVIDIA GPUs.

Quick Start

Ask the skill to optimize the rms_norm kernel for faster inference on an Ampere GPU.