Kernel Design Agents (KDA) provides an agent-centric workflow for automating the research, implementation, verification, and iteration of performance-sensitive CUDA kernel tasks. It targets engineers and researchers seeking to streamline the development of high-performance kernels, offering a structured approach to complex optimization challenges.

How It Works

KDA employs coding agents to manage the kernel development lifecycle. The workflow involves defining a task contract (objective, constraints, validation), initiating an agent session, drafting and refining an executable plan, and implementing the kernel in iterative cycles. Each iteration includes verification, benchmarking, and profiling, with results meticulously recorded to ensure transparency and reproducibility. The system leverages modular "skills" (e.g., ncu-report-skill, KernelWiki) and prompt templates, allowing flexibility across different benchmarks and hardware targets.

Quick Start & Requirements

Installation involves cloning the repository with submodules (git clone --recurse-submodules) and linking necessary skills (like ncu-report-skill and KernelWiki) to ~/.claude/skills. A key prerequisite is installing the humanize Claude Code plugin (/plugin marketplace add PolyArch/humanize). Users are directed to set up separate implementation workspaces, following a recommended directory structure for recording runs, outputs, and profiling data.

Highlighted Details

• This project is associated with HAN Lab's solutions that ranked #1-3 at the MLSys Kernel Contest.
• It facilitates an agent-driven, iterative process for CUDA kernel optimization.
• The workflow emphasizes detailed recording of implementation context, validation, and performance metrics.

Maintenance & Community

The project is described as an "early research prototype" and is "still under active development," actively seeking community feedback. It originates from the HAN Lab.

Licensing & Compatibility

No specific license information is provided in the README. Compatibility for commercial use or closed-source linking cannot be determined from the given text.

Limitations & Caveats

As an early research prototype, KDA is under active development, implying potential instability or incomplete features. Its functionality is tied to the Claude Code environment and specific plugin installations.