Kernel-Smith: A Unified Recipe for Evolutionary Kernel Optimization

2026-03-30

AI Summary

This work addresses the challenges of unstable search, poor generalization, and limited cross-platform adaptability in automatic GPU kernel generation by introducing a unified evolutionary optimization framework. For the first time, a large language model is integrated not as a one-shot generator but as a powerful local improver within the evolutionary loop. The approach synergistically combines population-based evolutionary search, structured execution feedback—encompassing compilation success, correctness, and speedup—and post-training fine-tuning, leveraging long-term evolutionary trajectories to generate step-level supervision and reinforcement learning signals. The resulting model, Kernel-Smith-235B-RL, achieves state-of-the-art performance on KernelBench, outperforming Gemini-3.0-Pro and Claude-4.6-Opus in average speedup. Its MetaX variant, Kernel-Smith-MACA-30B, also significantly surpasses DeepSeek-V3.2-Think and has been deployed in production systems such as SGLang and LMDeploy.

Abstract

We present Kernel-Smith, a framework for high-performance GPU kernel and operator generation that combines a stable evaluation-driven evolutionary agent with an evolution-oriented post-training recipe. On the agent side, Kernel-Smith maintains a population of executable candidates and iteratively improves them using an archive of top-performing and diverse programs together with structured execution feedback on compilation, correctness, and speedup. To make this search reliable, we build backend-specific evaluation services for Triton on NVIDIA GPUs and Maca on MetaX GPUs. On the training side, we convert long-horizon evolution trajectories into step-centric supervision and reinforcement learning signals by retaining correctness-preserving, high-gain revisions, so that the model is optimized as a strong local improver inside the evolutionary loop rather than as a one-shot generator. Under a unified evolutionary protocol, Kernel-Smith-235B-RL achieves state-of-the-art overall performance on KernelBench with Nvidia Triton backend, attaining the best average speedup ratio and outperforming frontier proprietary models including Gemini-3.0-pro and Claude-4.6-opus. We further validate the framework on the MetaX MACA backend, where our Kernel-Smith-MACA-30B surpasses large-scale counterparts such as DeepSeek-V3.2-think and Qwen3-235B-2507-think, highlighting potential for seamless adaptation across heterogeneous platforms. Beyond benchmark results, the same workflow produces upstream contributions to production systems including SGLang and LMDeploy, demonstrating that LLM-driven kernel optimization can transfer from controlled evaluation to practical deployment.

Problem

• GPU kernel optimization
• evolutionary search
• code generation
• heterogeneous platforms
• performance portability

Innovation

• evolutionary kernel optimization
• GPU kernel generation
• LLM-driven code optimization
• backend-specific evaluation
• reinforcement learning for compilation