Poster
in
Workshop: Deep Learning for Code: Towards Human-Centered Coding Agents

Kernel-Smith: A Unified Recipe for Evolutionary Kernel Optimization

He Du ⋅ Qiming Ge ⋅ Jiakai Hu ⋅ Aijun Yang ⋅ Zheng Cai ⋅ Zixian Huang ⋅ Sheng Yuan ⋅ QinXiu Cheng ⋅ XIE Xinchen ⋅ Yicheng Chen ⋅ Yining Li ⋅ Jiaxing xie ⋅ Huanan Dong ⋅ Yaguang Wu ⋅ Xiangjun Huang ⋅ jian yang ⋅ Hui Wang ⋅ Bowen Zhou ⋅ Bowen Li ⋅ Qipeng Guo ⋅ Kai Chen

Project Page [ OpenReview]

Abstract

High-performance GPU kernel generation is critical for large-scale systems and scientific computing, yet existing LLM approaches lack reliable iterative optimization. We introduce Kernel-Smith, a framework for high-performance operator generation combining an evaluation-driven evolutionary agent with an evolution-oriented post-training strategy. The agent maintains a population of executable candidates, iteratively refining programs using a diverse archive and structured feedback regarding compilation, correctness, and latency. For robust search, we implement dedicated evaluation backends for Triton on NVIDIA and Maca on MetaX GPUs. During training, we transform long-horizon evolution trajectories into step-centric supervision and reinforcement learning signals by distilling high-gain, correctness-preserving revisions. This optimizes the model as an effective local improver within the evolutionary loop. Evaluated on KernelBench, Kernel-Smith-235B-RL achieves state-of-the-art performance on the NVIDIA Triton backend, surpassing frontier models such as Gemini-3.0-pro and Claude-4.6-opus in average speedup. Furthermore, Kernel-Smith-MACA-30B outperforms larger counterparts including DeepSeek-V3.2-think and Qwen3-235B on the MetaX MACA backend, demonstrating cross-platform adaptability. Beyond benchmarks, our framework has generated upstream contributions to production systems such as SGLang and LMDeploy, validating the practical utility of LLM-driven kernel optimization.

Chat is not available.