AKO4ALL skill
A Claude Code skill for GPU kernel tuning: the agent profiles, edits and benchmarks a kernel in a loop until speedup stalls.
- 383 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The agent runs and rewrites kernel code on your machine, loading the GPU with long benchmarks
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add skills/ako4allDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The agent runs and rewrites kernel code on your machine, loading the GPU with long benchmarks
This is third-party code. Review the repository files before installing.
What it does
AKO4ALL, a single skill with its protocol in `SKILL.md`, starts from a working directory holding a kernel. The agent builds a workspace beside it, checks results against a correctness reference and repeats the loop: profile, edit, benchmark, record the iteration. Kernels in Triton, CUDA, C++, TileLang, CuTe DSL and Python work, as does any language that can be measured. A kernel and at least one set of inputs are required. You can add your own benchmark script, a reference implementation, algorithm notes and a `HINTS.md` file with constraints. Without a benchmark script the skill uses the built-in KernelBench evaluator.
Who it is for. Engineers who write and tune GPU kernels for inference and want an agent to take over the routine search for optimizations.
Good fit when
- An attention or normalization kernel is slower than the reference and needs a series of verified optimizations.
- A reference implementation exists for the agent to check each edit against.
- You must interrupt the session and resume later from a saved iteration log.
- You want a fast estimate of how much speedup an automated loop gives on your kernel.
Not a fit when
- There is no GPU or input data for measurement.
- You need a long campaign on one operator with memory across runs, for which the author points to the sibling project AKO4X.
- The kernel cannot be run on the developer machine.
Example request
Optimize `kernel.py` in this folder, with the reference in `reference.py`, at most ten iterations and no web search.Limitations
The protocol is static: the skill does not learn between runs. Speedups in the README were measured on NVIDIA B200 in the authors' own tests and will differ on your hardware. The agent runs code and benchmarks on your machine; the README has a permissions section. A GPU and an environment where the kernel runs are needed. The agent can occupy the machine for a long time, so cap the iteration count in the request. If something misbehaves, check the README and the repository's open discussions.
How to disable. Remove `~/.claude/skills/ako4all` (or the symlink). Delete workspaces and iteration logs created inside your projects manually.
FAQ
How does AKO4ALL differ from AKO4X?
AKO4ALL optimizes one kernel in one continuous run. AKO4X, per the README, builds a separate environment for a multi-round campaign on one operator.
What inputs are needed?
The kernel itself and at least one set of inputs: in code, in a benchmark script, in a `get_inputs()` file or in `.npz` and `.bin`.
How does the skill guard against overfitting to the test?
The README has an Anti-Cheat section. Read its rules before running on your own benchmarks.
Does it work in other agents?
The skill is written for Claude Code. The protocol sits in `SKILL.md`, so other agents can read it after manual installation.
Can a run be interrupted?
Yes. The README says optimization can be stopped at any point and resumed in another session.
Discussion
No questions or reviews yet.
Related
SkillsForgeKernel is a CLI that tunes a GPU kernel for hours: Claude Code writes variants, timing and profiling drop slower ones.
An Alibaba framework for tuning GPU kernels with coding agents: a Supervisor checks correctness and speed first
Skills for writing, debugging, measuring and optimizing TileLang and CUDA programs from source, lowering and profiler evidence
Holds write-ups of GPU kernel optimizations for Blackwell and Hopper so an agent answers from sources, not model memory.
Three Intel skills for Linux perf: profiling, finding bottlenecks in code and perf output, and tuning Phoronix benchmarks
Claude Code skill reading Nsight Compute reports on Blackwell: from a CUDA kernel profile to a diagnosis and an optimization plan.
README in short
The skill drives a kernel optimization loop in one working folder and points to a report comparing speedups against FlashInfer's best kernels.