KernelSeek
A question-driven CUDA kernel optimization agent that, according to its paper, produced dependency-free FlashAttention forward kernels faster than FA3 and FA4 on the authors' suites.
The paper (arXiv:2611.03517) compares a tool-rich autonomous baseline that plateaued at 0.50× FA3 with KernelSeek reaching 1.13× on the same 26-configuration evaluator, then reports fa3_hopper.cu at 1.16× FA3 (182/182) and fa4_blackwell.cu at 1.46× FA4 on paired scenarios. Sessions used Claude Code and Cursor. The paper says the kernels, harness, question bank, and traces will be released, but no public repository URL was found during this research pass.
Human involvement
Person directed, agent wrote the code. The paper says the human wrote no kernel code and steered with short questions, directives, and a stop-hook goal. Sixteen archived sessions contained 162 human messages and 11,532 tool calls.
Evidence
- Paper
We release both kernels under Apache 2.0 with the evaluator, the PTX probes, the search scripts, the question bank and skill, and the complete agent traces of both approaches.
Notes
Benchmark numbers are the paper's, not independently reproduced here. A public GitHub URL was not found on arXiv or the Hugging Face paper page.