AKO4ALL

Run `npx skills add TongmingLAIC/AKO4ALL --all -g -y` in your terminal to install the skill globally.

Place your kernel in a working directory and ask your coding agent to optimize it. AKO4ALL profiles the baseline, rewrites the kernel, benchmarks each change, and keeps iterating until performance stops improving.

You need a coding agent like Claude Code, an NVIDIA GPU with CUDA-enabled PyTorch, Python 3.10 or later, and Git. Nsight Compute (ncu) is recommended for profiling but the loop works without it.

Yes. Every iteration is checked against a reference implementation with fresh randomized inputs, and the built-in evaluator flags suspicious speedups to prevent reward hacking.

Yes. You only provide the kernel and test inputs; the agent handles profiling, rewriting, benchmarking, and logging based on your plain-English request.