auto-gpu-kernel

Automate iterative profiling and benchmarking for Triton-based sparse attention kernels.

4|2|Updated Feb 19, 2026
One-click install
npx skills add https://github.com/drunkcoding/AgentSkillsArxiv --skill auto-gpu-kernel
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: auto-gpu-kernel
Source: https://github.com/drunkcoding/AgentSkillsArxiv/tree/main/skills/auto-gpu-kernel
Command: npx skills add https://github.com/drunkcoding/AgentSkillsArxiv --skill auto-gpu-kernel

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires modal, flashinfer-bench, and includes references (resource) components.

What problem does it solve?

Automates end-to-end autonomous kernel optimization for Triton-based DSA sparse attention workloads using a templated project structure.

Core Features & Use Cases

  • Scaffold a complete kernel-optimization project from a bundled template, including CLAUDE.md, config.toml, and solution/triton skeletons.
  • Run an autonomous optimization loop with sub-agents (profiler, workload-inspector, research) that communicate through on-disk artifacts to guide iterative improvements.
  • Leverage a Modal-based benchmarking harness and a paired A/B workflow to evaluate changes under realistic workloads while preserving strict experimental discipline.
  • Apply the workflow across different kernels and adapt the template to new FlashInfer-Bench definitions, hazard considerations, and customization guidance.
  • Use the workflow to document experiments, capture insights in summaries, and progressively reach improved kernel configurations without manual guesswork.

Quick Start

Scaffold a new kernel-optimization project from the bundled template and start the autonomous optimization loop using CLAUDE.md workflow.

Frequently Asked Questions about auto-gpu-kernel

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

FAQPage Schema
How do I automate Triton kernel optimization for sparse attention workloads?

Automate Triton kernel optimization by running an autonomous agent loop that iteratively profiles, plans, implements, and benchmarks DSA sparse attention kernels using a templated project structure and Modal-based benchmarking harness.

What is autonomous kernel optimization and how does it work with FlashInfer-Bench?

Autonomous kernel optimization uses sub-agents for profiling, workload inspection, and research that communicate through on-disk artifacts to guide iterative improvements, evaluating changes against FlashInfer-Bench definitions within a strict A/B workflow.

How do I scaffold a kernel optimization project from a bundled template?

Scaffold a kernel optimization project by generating a complete template structure containing CLAUDE.md, config.toml, and solution/triton skeletons with baseline and fused stubs, then start the autonomous optimization loop defined in the CLAUDE.md workflow.

Do I need Modal to run the autonomous kernel benchmarking workflow?

Yes, Modal is required to run the autonomous kernel benchmarking workflow, providing the cloud-based harness needed to evaluate Triton kernel changes under realistic workloads during iterative optimization.

Can I adapt the Triton kernel optimization template to new FlashInfer-Bench definitions?

Yes, you can adapt the templated project structure across different kernels and new FlashInfer-Bench definitions, applying customization guidance and hazard considerations to target new sparse attention workloads.

What are the limitations of using autonomous agents for Triton kernel development?

Autonomous Triton kernel development requires strict experimental discipline and depends on predefined config.toml settings and baseline stubs; agents communicate through on-disk artifacts, so workflows break if templated structures are improperly modified.