skills.rest
slowlyC
slowlyCCOMMUNITY
Explore all skills by this author →
GitHub

cutlass-skill

Develop, debug, and optimize CUTLASS and CuTeDSL GPU kernels for NVIDIA architectures.

154|18|Updated Mar 1, 2026
Search Tags:#cuda#gpu kernels#gemm#cute#cutlass#cutedsl#hopper
One-click install
npx skills add https://github.com/slowlyC/agent-gpu-skills --skill cutlass-skill-slowlyc
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: cutlass-skill
Source: https://github.com/slowlyC/agent-gpu-skills/tree/main/cutlass_skill
Command: npx skills add https://github.com/slowlyC/agent-gpu-skills --skill cutlass-skill-slowlyc

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Developing high-performance CUTLASS and CuTeDSL GPU kernels requires deep expertise in NVIDIA GPU architectures, CUTLASS template metaprogramming, and scattered official documentation, making kernel development, debugging, and optimization time-consuming and error-prone for GPU engineers.

Core Features & Use Cases

  • CUTLASS/CuTeDSL Kernel Development: Write, debug, and optimize GEMM, attention, sparse, and custom GPU kernels using either CUTLASS C++ templates or the CuTeDSL Python DSL, with support for Ampere, Hopper, and Blackwell GPU architectures.
  • Local Resource Access: Instantly access local CUTLASS source code, architecture-specific examples (including FP8 GEMM, grouped GEMM, MoE GEMM, and flash attention implementations), and CuTe/CUTLASS header references without searching online.
  • Use Case: For example, if you need to implement a TMA-based GEMM kernel for Hopper GPUs with epilogue fusion, this skill lets you pull relevant collective builder examples, pipeline configurations, and epilogue visitor tree patterns directly from the local CUTLASS repository.

Quick Start

Use the cutlass-skill to implement a grouped FP8 GEMM kernel for Blackwell GPUs using the CUTLASS collective builder and epilogue fusion patterns.

Frequently Asked Questions about cutlass-skill

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I write a CUTLASS GEMM kernel with epilogue fusion for Hopper GPUs?▼

To write a CUTLASS GEMM kernel with epilogue fusion for Hopper GPUs, use collective builder examples and epilogue visitor tree patterns from the local CUTLASS repository to configure TMA pipelines and fuse output operations.

Can I use CuTeDSL to build GPU kernels for Blackwell architectures?▼

Yes, you can use CuTeDSL to build GPU kernels for Blackwell architectures, with support for writing and debugging custom GEMM, attention, and sparse kernels through the Python DSL.

What is the best way to debug CUTLASS template compilation errors?▼

The best way to debug CUTLASS template compilation errors is to resolve them using local source code access and API references, allowing you to look up collective builders, TMA pipelines, and epilogue fusion patterns directly.

Does this support implementing FP8 grouped GEMM kernels?▼

Yes, it supports implementing FP8 grouped GEMM kernels by providing access to architecture-specific examples, including grouped GEMM, MoE GEMM, and flash attention implementations for Ampere, Hopper, and Blackwell GPUs.

How do I find collective builder patterns for TMA pipelines in CUTLASS?▼

To find collective builder patterns for TMA pipelines in CUTLASS, query the local CUTLASS repository to retrieve relevant source code, architecture-specific examples, and CuTe header references without searching online.

Why does my CUTLASS kernel fail to compile when using custom epilogue visitors?▼

CUTLASS kernels using custom epilogue visitors may fail to compile due to template metaprogramming errors; resolve them by checking local API references and epilogue visitor tree patterns to ensure correct configuration.

Related Skills

Explore More Software Engineering Skills→
Harmeet10000Harmeet10000
7

mojo-gpu-fundamentals

Explain CUDA-free Mojo GPU programming with LayoutTensor and shared memory patterns.

Community
Advanced
dotrurudotruru

cutlass-skill

Automates development, debugging, and optimization of CUTLASS and CuTeDSL GPU kernels using local sources.

Community
Advanced
leonling-llleonling-ll

kernel-perf-analysis

Collect timing, hardware counters, ATT traces, and memory-hierarchy data from AMD GPU kernels.

Community
Advanced
guqiong96guqiong96
105

add-jit-kernel

Add custom CUDA kernels to the SGLang JIT kernel module.

Community
Advanced