NVlabs/kda

Kernel Design Agents (KDA) is a agent-centric workflow to write high-performance CUDA Kernels.

What it solves

Kernel Design Agents (KDA) provides a structured workflow for using coding agents to automate the research, implementation, and optimization of performance-sensitive CUDA kernels. It streamlines the process of creating high-performance GPU kernels by integrating research, verification, and iterative refinement.

How it works

KDA is an agent-centric workflow that guides a coding agent through a specific lifecycle: defining a task contract (objectives and constraints), drafting a plan, and implementing the kernel in small, verified iterations. It utilizes prompt templates and specialized "skills" (such as ncu-report-skill and KernelWiki) to provide the agent with the agent-specific knowledge and profiling tools needed for CUDA development.

Who it’s for

Developers and researchers focusing on GPU performance optimization and CUDA kernel development, particularly those targeting NVIDIA B200 and B300 GPUs.

Highlights

  • Iterative Verification: Focuses on small, incremental changes followed by immediate verification to ensure correctness and performance.
  • Specialized Skills: Integrates with tools like ncu-report-skill for profiling and KernelWiki for domain knowledge.
  • Agent-Centric Workflow: Provides a structured approach (from task contract to final promotion) that is independent of any single benchmark harness or hardware target.
  • Community Wishlist: Maintains a community-driven list of kernels that need optimization requests.

Related

  • Project
    valkor-ai/loom

    An open delivery harness that provides a stateful, resumable loop for coding agents to complete large software tasks without losing context or state.

  • Project
    alumnium-hq/alumnium

    An AI-native library and MCP for end-to-end testing that allows users to perform actions and verify assertions in applications using natural language via Appium, Playwright, or Selenium.

  • Project
    Thinking-Space/Rethinking-OPD

    A research project and toolkit for analyzing and improving On-Policy Distillation (OPD) in LLMs, providing strategies to recover failing distillation and understanding its token-level mechanisms.

  • Project
    johannesjo/parallel-code

    A GUI for dispatching multiple AI coding agents in parallel, using git worktrees to isolate each task into its own branch and directory for concurrent development.