Kernel-Smith: A Unified Recipe for Evolutionary Kernel Optimization

Kernel-Smith combines a stable evaluation-driven evolutionary agent with an evolution-oriented post-training recipe for high-performance GPU kernel and operator generation. Long-horizon evolution trajectories are converted into step-centric supervision and reinforcement learning signals, optimizing the model as a strong local improver inside the evolutionary loop rather than as a one-shot generator.

Kernel-Smith-235B-RL reaches state-of-the-art overall performance on KernelBench with the NVIDIA Triton backend, outperforming frontier proprietary models. The framework also transfers to the MetaX MACA backend, and the same workflow produced upstream contributions to production systems including SGLang and LMDeploy.