AutoKernel: Autonomous GPU Kernel Optimization via Iterative Agent-Driven Search (arxiv.org)

Researchers have introduced AutoKernel, an innovative open-source framework designed for optimizing GPU kernels autonomously, specifically for PyTorch models. Utilizing an agent-driven search mechanism, AutoKernel identifies and addresses computational bottlenecks by profiling models and refining kernel implementations in Triton or CUDA C++ through a series of automated experiments. Importantly, this system incorporates a rigorous five-stage correctness harness that validates each kernel candidate, ensuring efficiency gains are legitimate and not at the cost of performance integrity. This development is significant for the AI/ML community as it streamlines the labor-intensive process of high-performance GPU kernel optimization, a critical factor in enhancing the efficiency of machine learning applications. AutoKernel has demonstrated impressive results, outperforming existing PyTorch implementations and competing optimizers on various configurations, achieving speedups of up to 5.29 times. By covering nine kernel types prevalent in modern transformer architectures and integrating with the KernelBench benchmark suite, AutoKernel stands to revolutionize how machine learning practitioners can optimize their models, potentially leading to faster training and inference times across a wide range of applications.