KrxGuCommunity·7 Agent Skills Included

kernel-skills

Expert playbooks for writing fast, correct GPU kernels

Guides coding agents through writing, optimizing, debugging, and porting CUDA, Triton, and quantized GPU kernels. Covers GEMM, softmax, layernorm, attention, KV cache, sampling, and LLM inference kernels with strict correctness and performance rules. Eliminates vague prompts, silent numerical bugs, and fake speedup claims with measured before-and-after proof for every skill.

54|7

↗

npx skills add KrxGu/kernel-skills --all -g -y

Available:

Tells the agent the repository's mission, quality bar, and required SKILL.md structure so it produces correctness-first, performance-aware kernel code instead of generic optimization advice.

All Skills in This Repository (7)

Pure Emerald Level Indicators

📦 In Repo

KrxGu

Write a Numerically Stable Kernel

Identify and mitigate numerical instability in GPU compute kernels with dtype-annotated stabilizations.

Write CUDA Reduction Kernel

Design CUDA reduction kernels with warp and block reductions.

Write CUDA Softmax Kernel

Implement numerically stable CUDA softmax kernels with fp32 accumulation and masking.

Write CUDA GEMM Kernel

Design and implement a CUDA GEMM kernel with FP32 accumulation and tiling.

Write a Triton Softmax Kernel

Implement a numerically stable Triton softmax kernel with fp32 accumulation and multi-block online rescaling.

Write an INT8 Quantized Kernel

Design INT8 quantized GEMM kernels with dp4a and INT32 accumulation for NVIDIA GPUs.

Write a Backend-Agnostic Kernel Plan

Plan backend-agnostic compute kernels for NVIDIA, AMD, and CPU targets.

Community

Advanced

Frequently Asked Questions

FAQPage Schema

How to install kernel-skills?▼

Run `npx skills add KrxGu/kernel-skills --all -g -y` in your terminal to install all skills globally.

What does kernel-skills do?▼

It gives your coding agent expert playbooks for writing CUDA, Triton, quantized, and LLM inference kernels. Each skill forces correct constraint gathering, numerical stability, and honest performance reasoning.

Does kernel-skills actually improve kernel code?▼

Yes. The repo includes measured proofs: for example, the softmax skill fixed 16/16 failing shapes and the GEMM skill produced 7.7-8.6x speedups over naive agent output.

Which agents work with kernel-skills?▼

The SKILL.md files work with Claude Code, ChatGPT, Cursor, Gemini CLI, Codex, and any tool that accepts Markdown skill files.

Do I need GPU expertise to use kernel-skills?▼

No. You paste or reference a skill and the agent follows its step-by-step reasoning, though the content targets engineers working on GPU performance problems.