Tilekernels Gpu Kernels

This gives Claude fluency in TileKernels, a library of hand-optimized GPU kernels for LLM operations like MoE routing, FP8 quantization, and transposes, all written in TileLang for Hopper and Blackwell architectures. You'd reach for this when implementing custom LLM components that need to squeeze performance out of H100s or B200s, like fused SwiGLU quantization or Sinkhorn-normalized Manifold HyperConnections. The library includes both low-level kernels and PyTorch autograd wrappers, plus reference implementations for validation. It's pretty specialized hardware (SM90+ only) and assumes you're comfortable writing or modifying GPU kernels, but if you're building production LLM infrastructure and need to beat standard PyTorch ops, this is the kind of tooling that matters.

skills.sh installs: 256
GitHub stars: 26
last updated: Jun 3, 2026

View on GitHub

Install

$ npx -y skills add aradotso/trending-skills --skill tilekernels-gpu-kernels --agent claude-code

Installs into .claude/skills of the current project.

Files

TileKernels GPU Kernel Library

Skill by ara.so — Daily 2026 Skills collection.

TileKernels is a high-performance GPU kernel library for LLM operations (MoE routing, FP8/FP4 quantization, transpose, engram gating, Manifold HyperConnection) written in TileLang — a Python DSL for expressing GPU kernels with automatic optimization. Kernels target NVIDIA SM90/SM100 (Hopper/Blackwell) architectures and approach hardware performance limits.

Requirements

Python 3.10+
PyTorch 2.10+
TileLang 0.1.9+
NVIDIA SM90 or SM100 GPU (H100/H200/B100/B200)
CUDA Toolkit 13.1+

Installation

# Development install (recommended for extending/modifying kernels)
pip install -e ".[dev]"

# Release install
pip install tile-kernels

Project Structure

SKILL.md on GitHub

Details

Repository: github.com/aradotso/trending-skills
First seen: Jun 3, 2026