H100 Diffusers Skill for Claude Code

Optimized CUDA kernels for H100 GPUs targeting the HuggingFace diffusers library, with a Claude Code skill for guided kernel development.

What's Included

• CUDA Kernels: Optimized implementations for RMSNorm, RoPE (1D/3D), GEGLU, SwiGLU, and AdaLN
• Python API: Drop-in replacements for diffusers operations via ltx_kernels
• Claude Code Skill: Expert guidance for writing custom H100 kernels

Quick Start

Installation

# Install the package
pip install -e .

# With diffusers support
pip install -e ".[diffusers]"

Build Kernels (requires CUDA)

# Using Docker (recommended)
docker run --rm --mount type=bind,source=$(pwd),target=/kernelcode \
  -w /kernelcode ghcr.io/huggingface/kernel-builder:main build

# Or with Nix
nix run .#build-and-copy --max-jobs 2 --cores 8 -L

Using the Claude Code Skill

This repository includes a Claude Code skill that provides expert guidance for developing optimized CUDA kernels targeting H100 GPUs.

Activating the Skill

The skill activates automatically when you ask Claude Code about:

• Writing CUDA kernels for diffusion models
• Optimizing attention, normalization, or activation layers
• Integrating custom kernels with diffusers pipelines
• H100-specific optimizations

Example Prompts

Writing a new kernel:

Write a fused RMSNorm + residual kernel optimized for H100

Optimizing existing code:

Help me optimize this attention kernel for H100's 192KB shared memory

Integration questions:

How do I add a custom AdaLN kernel to the LTX-Video transformer?

Architecture guidance:

What block sizes should I use for flash attention on H100?

What the Skill Provides

1. H100 Architecture Reference: SM count, shared memory, memory bandwidth, warp size
2. Kernel Templates: Complete CUDA implementations for common operations
3. Block Size Guidelines: Optimal configurations for different kernel types
4. PyTorch Integration Patterns: C++ bindings and Python API examples
5. Performance Profiling: Commands for nsys and ncu analysis

Skill Files

The skill documentation is in .claude/skills/h100-diffusers-kernels/: