{"name":"KDA: Kernel Design Agents for High-Performance CUDA Kernel Development","description":"Kernel Design Agents (KDA) offers an agent-centric workflow designed to streamline the research, implementation, verification, and iteration of performance-sensitive CUDA kernel tasks. This innovative approach leverages coding agents to accelerate the development of high-performance kernels. It is an early research prototype from NVlabs, welcoming community feedback and contributions.","github":"https://github.com/NVlabs/kda","url":"https://osrepos.com/repo/nvlabs-kda","source":"osrepos.com","sourceDescription":"This repository profile is provided by osrepos.com, an open source repository discovery platform.","repositoryProfile":"https://osrepos.com/repo/nvlabs-kda","generatedFor":"open source discovery and AI-assisted research","markdown":"https://osrepos.com/repo/nvlabs-kda.md","json":"https://osrepos.com/repo/nvlabs-kda.json","topics":["CUDA","AI Agents","High Performance Computing","GPU Programming","Kernel Optimization","NVlabs","Workflow Automation"],"keywords":["CUDA","AI Agents","High Performance Computing","GPU Programming","Kernel Optimization","NVlabs","Workflow Automation"],"stars":null,"summary":"Kernel Design Agents (KDA) offers an agent-centric workflow designed to streamline the research, implementation, verification, and iteration of performance-sensitive CUDA kernel tasks. This innovative approach leverages coding agents to accelerate the development of high-performance kernels. It is an early research prototype from NVlabs, welcoming community feedback and contributions.","content":"## Introduction\n\nKernel 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.\n\n## Why Use KDA and Its Benefits\n\nKDA offers a structured and agent-driven approach to a traditionally complex area: high-performance CUDA kernel development. Its key benefits include:\n\n*   **Agent-Centric Workflow**: Automates and streamlines various stages of kernel development, from initial research to final verification, using intelligent coding agents.\n*   **Performance Focus**: Specifically tailored for tasks where performance is critical, helping developers achieve optimized CUDA kernels.\n*   **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.\n*   **Flexible and Independent**: The workflow is designed to be independent of any single benchmark harness or hardware target, allowing for broad applicability.\n*   **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.\n\n## Installation\n\nTo get started with Kernel Design Agents, follow these steps to clone the repository and set up the necessary skills:\n\nbash\ngit clone --recurse-submodules https://github.com/mit-han-lab/kernel-design-agents.git\ncd kernel-design-agents\n\n# Link skills\nmkdir -p ~/.claude/skills\nln -s \"$(pwd)/skills/ncu-report-skill\" ~/.claude/skills/ncu-report-skill\nln -s \"$(pwd)/skills/KernelWiki\" ~/.claude/skills/KernelWiki\n\n\nAlternatively, you can clone the independently licensed `ncu-report-skill` directly:\n\nbash\nmkdir -p ~/.claude/skills && cd ~/.claude/skills\ngit clone https://github.com/mit-han-lab/ncu-report-skill.git\n\n\nAdditionally, install the `humanize` Claude Code plugin from the Claude Code plugin UI:\n\ntext\n/plugin marketplace add PolyArch/humanize\n/plugin install humanize@PolyArch\n\n\n## Examples\n\nKDA provides a clear \"Minimal Flow\" and a \"Recommended Workspace Layout\" to guide users in leveraging the agent-centric approach effectively.\n\n### Minimal Flow\n\n1.  Create a separate implementation workspace for your target task.\n2.  Define the task contract: objective, constraints, validation command, and promotion criteria.\n3.  Start an agent session in the implementation workspace.\n4.  Provide the agent with `prompts/basic-flow.md`, filled with task-specific details.\n5.  Ask the agent to draft a short plan to `docs/draft.md` in the implementation workspace.\n6.  Convert the draft into an executable plan, either manually or using a planning tool like Humanize.\n7.  Implement in small iterations, verifying after each meaningful change.\n8.  Record candidates, benchmark or evaluation results, profiling evidence, and final promotion decisions.\n\n### Recommended Workspace Layout\n\nFor optimal organization and reproducibility, it is recommended to use a workspace layout similar to this, keeping implementation work separate from the KDA reference material:\n\ntext\ntask-workspace/\n  docs/\n    draft.md\n    plan.md\n  runs/\n  outputs/\n  profile/\n  benchmark.csv\n  candidates.jsonl\n\n\nThis layout ensures that enough context is recorded for another engineer to understand the process, validated outcomes, and the rationale behind the final candidate selection.\n\n## Links\n\n*   **GitHub Repository**: [https://github.com/NVlabs/kda](https://github.com/NVlabs/kda)\n*   **Project Website**: [https://nvlabs.github.io/kda/](https://nvlabs.github.io/kda/)\n*   **Contributing Guide**: [https://github.com/NVlabs/kda/blob/main/CONTRIBUTING.md](https://github.com/NVlabs/kda/blob/main/CONTRIBUTING.md)","metrics":{"detailViews":0,"githubClicks":0},"dates":{"published":null,"modified":"2026-09-26T08:25:23.000Z"}}