Auto GPU Kernel
github.com/dogacel/auto-gpu-kernel- Category
- AI Agents
- Rank
- No. 1174Tools index
Previous survey · No. 1154 ·
- Pricing
- Open Source
- Type
- AGENT
- Use case
- Coding
- Builder
- @dogacel0
- GitHub
- 317 stars
- Latest release
- v1.0.0
- Date
About
Auto GPU Kernel is an autonomous agent system that discovers and optimizes GPU kernels, evolving both the kernel code and its benchmarking harness rather than relying on a human-authored one. It won the agent-only track of the MLSys 2026 FlashInfer AI Kernel Generation Contest for DeepSeek Sparse Attention, achieving an average 34.93x speedup, with its kopt tool driving optimization and kbench handling validation and benchmarking.
What it does
The project page describes an agent loop for making GPU code faster. One command drives a model through repeated optimization attempts while a second runs correctness checks, timing and paired comparisons. It targets contest kernel definitions or any Git repository described in a short plain-language task file, where a setup agent first writes the checks.
Stated on the product site
- License
- Apache-2.0
- Requirements
- Python 3.11 or later, the uv package manager and OMP
- Interfaces
- Two command-line tools: kopt for the optimization agent, kbench for validation and benchmarking
- Task formats
- FlashInfer kernel definitions, or a natural-language task against a Git repository
- Execution backends
- Local Linux machine with a GPU, or Modal cloud GPUs
- Platform support
- Bundled MLX example runs the task loop on Apple Silicon
- Contest result (self-reported)
- First in the agent-only DeepSeek Sparse Attention track of the MLSys 2026 FlashInfer contest
Not stated on the site
- The page does not link an independent record of the contest ranking or show how the stated average speedup was measured.
- Which model providers work beyond the Anthropic model used in every example is not stated.
- The page gives no estimate of model usage or cost for a typical optimization run.
- Generated-task adapters are described as running on the local machine for now; remote execution for that mode is not described.
Written from the product site at github.com.
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.