KDA Wishlist is a community program for turning reproducible kernel definitions into optimized kernel solutions. A kernel is a performance-critical program that runs on accelerated hardware, such as a graphics processing unit (GPU), and implements the core computation of a machine learning operation.
Live site: https://docs.humanfia.ai/KDA-wishlist/
KDA stands for Kernel Design Agents: a workflow in which coding agents research, implement, verify, profile, and iterate on performance-sensitive kernel tasks.
This website introduces the program, helps contributors prepare reproducible kernel definitions and workloads, and directs them to GitHub Issues. GitHub Issues is the repository's public request tracker, where contributors submit kernel needs, add useful context, and signal demand with thumbs-up reactions.
The project requires Node.js 22.13.0 or later.
Install the dependencies and start the development server:
npm install
npm run devThen open http://localhost:3000.
To run the production build locally:
npm run build
npm run start -- --port 3000To create the static export used by GitHub Pages:
GITHUB_PAGES=true GITHUB_REPOSITORY=humanfia/KDA-wishlist npm run build:pagesThe exported site is written to out/.
npm run build
npm run lint
npm auditThe production build targets Cloudflare Workers, Cloudflare's distributed edge runtime. The logical hosting configuration lives in .openai/hosting.json.
Pushes to main also run .github/workflows/deploy-pages.yml, which builds the static export and deploys it to GitHub Pages.
- Define the task in FlashInfer Trace, a reproducible format that describes the reference implementation, input and output contract, correctness requirements, and representative workloads.
- Open a GitHub issue with the repository's Kernel request form.
- Community members add thumbs-up reactions to the top-level issue and use comments to contribute new workload evidence or implementation context.
- The team reviews the task. Selected requests enter a measured loop of research, implementation, correctness validation, performance profiling, and candidate selection.
- Completed tasks may return an optimized kernel, benchmark comparisons, reproduction instructions, environment details, design notes, known limitations, and an upstream-ready contribution.
Submission does not guarantee selection. The program prioritizes tasks that affect real systems, can be evaluated automatically, produce publicly reproducible results, benefit multiple projects, and have a realistic path to upstream adoption.
app/contains the page structure, copy, metadata, and styles.public/og.pngis the branded social-sharing preview image..github/ISSUE_TEMPLATE/contains the structured kernel request form and issue settings..openai/hosting.jsoncontains the logical website-hosting configuration.