GitHub releases KernelOPT for agentic GPU kernel optimization
GitHub user giveen released KernelOPT, an open-source tool that automatically optimizes GPU kernels using agentic search. It works on CUDA, HIP/ROCm, and frameworks like PyTorch (Triton), ninfer, and llama.cpp. KernelOPT profiles kernels, generates optimized versions, and verifies correctness before applying changes in an isolated git worktree—preserving the original codebase. The tool uses LLM…
Key points
- KernelOPT automates GPU kernel optimization via agentic search, supporting CUDA, HIP/ROCm, and frameworks like PyTorch, ninfer, and llama.cpp
- Uses LLM agents (Planner/Executor/Summarizer) with profiling tools (nsys/ncu, Graphsignal) to generate, verify, and apply optimized kernels without touching the original repo
- Five correctness gates ensure speedups are measurable, plausible, and reproducible; runs on cloud LLMs to avoid GPU resource conflicts
Key features include five gates for correctness (compile, multi-seed, model-level, performance, and shape correctness) and support for cloud-based LLM optimization (e.g., OpenAI, OpenRouter) to avoid GPU contention. Users can run it on NVIDIA/AMD GPUs with CUDA/ROCm toolkits, Rust, Python, and CMake. The project is licensed under Apache 2.0 and builds on research from arXiv:2609.30059. Documentation and setup instructions are available in the repo.
GitHub - giveen/KernelOPT: Dispatch-aware agentic GPU kernel optimization
github.com · 29 September 2026Loading the full article…
This text was published by github.com. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
Coverage and discussion
1source- Reddit discussionreddit.com
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Researcher tests GPT-6 Astra in WoW with agent-wow framework · 1 src
- Trillium Labs launches open AI research nonprofit with $30M training budget · 1 src
- Karpathy shares how to make AI outputs like aircraft manuals · 1 src
- Allen Institute for AI open-sources AstaBrief 8B for faster scientific reports · 1 src
- Datalab releases OmniExtractBench to audit PDF extraction accuracy · 1 src
Comments
via GitHub Discussions