KDA: Kernel Design Agents for High-Performance CUDA Kernel Development

Introduction

Kernel Design Agents (KDA) is an agent-centric workflow developed by NVlabs, designed to leverage coding agents for researching, implementing, verifying, and iterating on performance-sensitive CUDA kernel tasks. This repository presents an early research prototype that is under active development, welcoming community feedback and contributions to further its evolution.

Why Use KDA and Its Benefits

KDA offers a structured and agent-driven approach to a traditionally complex area: high-performance CUDA kernel development. Its key benefits include:

• Agent-Centric Workflow: Automates and streamlines various stages of kernel development, from initial research to final verification, using intelligent coding agents.
• Performance Focus: Specifically tailored for tasks where performance is critical, helping developers achieve optimized CUDA kernels.
• Community Kernel Wishlist: Provides a unique platform for users to submit requests for kernel optimizations, currently supporting NVIDIA B200 and B300 GPUs, fostering collaborative improvement.
• Flexible and Independent: The workflow is designed to be independent of any single benchmark harness or hardware target, allowing for broad applicability.
• Structured Development: Promotes a recommended workspace layout that ensures comprehensive documentation of the agent's work, including drafts, plans, runs, outputs, and benchmarks, making the development process transparent and reproducible.

Installation

To get started with Kernel Design Agents, follow these steps to clone the repository and set up the necessary skills:

git clone --recurse-submodules https://github.com/mit-han-lab/kernel-design-agents.git
cd kernel-design-agents

# Link skills
mkdir -p ~/.claude/skills
ln -s "$(pwd)/skills/ncu-report-skill" ~/.claude/skills/ncu-report-skill
ln -s "$(pwd)/skills/KernelWiki" ~/.claude/skills/KernelWiki

Alternatively, you can clone the independently licensed ncu-report-skill directly:

mkdir -p ~/.claude/skills && cd ~/.claude/skills
git clone https://github.com/mit-han-lab/ncu-report-skill.git

Additionally, install the humanize Claude Code plugin from the Claude Code plugin UI:

/plugin marketplace add PolyArch/humanize
/plugin install humanize@PolyArch

Examples

KDA provides a clear "Minimal Flow" and a "Recommended Workspace Layout" to guide users in leveraging the agent-centric approach effectively.

Minimal Flow

1. Create a separate implementation workspace for your target task.
2. Define the task contract: objective, constraints, validation command, and promotion criteria.
3. Start an agent session in the implementation workspace.
4. Provide the agent with prompts/basic-flow.md, filled with task-specific details.
5. Ask the agent to draft a short plan to docs/draft.md in the implementation workspace.
6. Convert the draft into an executable plan, either manually or using a planning tool like Humanize.
7. Implement in small iterations, verifying after each meaningful change.
8. Record candidates, benchmark or evaluation results, profiling evidence, and final promotion decisions.

Recommended Workspace Layout

For optimal organization and reproducibility, it is recommended to use a workspace layout similar to this, keeping implementation work separate from the KDA reference material:

task-workspace/
  docs/
    draft.md
    plan.md
  runs/
  outputs/
  profile/
  benchmark.csv
  candidates.jsonl

This layout ensures that enough context is recorded for another engineer to understand the process, validated outcomes, and the rationale behind the final candidate selection.

Links

• GitHub Repository: https://github.com/NVlabs/kda
• Project Website: https://nvlabs.github.io/kda/
• Contributing Guide: https://github.com/NVlabs/kda/blob/main/CONTRIBUTING.md