tensormuxOfficial·10 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, INT8/FP8 quantization, KV cache, sampling, and LLM inference kernels with strict correctness and performance rules. Eliminates vague prompts, silent numerical bugs, and fake speedup claims by enforcing measured, expert-grade reasoning.

54|7

↗

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

Available:

Instructs the agent on the repository's mission, quality standards, and required SKILL.md structure so it produces correct, performance-aware kernel code instead of generic optimization advice.

All Skills in This Repository (10)

Pure Emerald Level Indicators

📦 In Repo

tensormux

Choose Tile Size and Work Partitioning

Determine tile sizes and grid layouts for CUDA or Triton kernels.

Write CUDA LayerNorm Kernel

Write CUDA LayerNorm kernels with Welford reduction and FP32 accumulation.

Choose CUDA Launch Configuration

Determine optimal CUDA launch configurations by analyzing resource usage and hardware limits.

Avoid Warp Divergence

Classify warp divergence in CUDA kernels and estimate cost with Nsight Compute metrics.

Write Triton LayerNorm Kernel

Write a Triton LayerNorm kernel with RMSNorm support and masking.

Optimize Triton Block Parameters

Autotune Triton GEMM block sizes and launch parameters for GPU throughput.

Write an FP8 Kernel

Design FP8 compute kernels with E4M3/E5M2 scaling for NVIDIA Hopper and Ada GPUs.

Port CUDA Kernel to HIP

Port CUDA kernels to HIP for ROCm-enabled AMD GPUs.

Write a Triton RMSNorm Kernel

Implement a numerically stable RMSNorm kernel in Triton with fp32 accumulation.

Write a Triton Dequant Kernel (int4 / int8 → fp16 / bf16)

Implements Triton dequantization of int4/int8 weights to fp16/bf16 with AWQ/GPTQ/NF4 support.

Official

Advanced

Frequently Asked Questions

FAQPage Schema

How to install kernel-skills?▼

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

What does kernel-skills do?▼

It provides structured SKILL.md playbooks that guide coding agents to write correct, fast CUDA, Triton, and quantized GPU kernels. Each skill enforces constraint gathering, numerical stability, and honest performance reasoning.

Which kernels and frameworks are covered?▼

Skills cover CUDA and Triton GEMM, softmax, layernorm, attention, reductions, INT8/FP8 quantization, and LLM inference kernels like RoPE, RMSNorm, KV cache, and sampling. Portability skills cover CUDA-to-Triton and CUDA-to-HIP migrations.

Does kernel-skills work with Claude Code and Cursor?▼

Yes. The skills are plain Markdown files that work with Claude Code, ChatGPT, Cursor, Gemini CLI, Codex, and any agent that accepts skill files.

Is there proof these skills improve kernel quality?▼

Yes. The proof directory contains measured before-and-after results, such as fixing naive softmax failures across all tested shapes and achieving 2.6-8.6x speedups on reduction and GEMM kernels.