Large-Scale Agentic RL for CUDA Kernel Generation (cuda-agent.github.io)

The announcement of CUDA Agent marks a significant advancement in GPU kernel optimization through the application of large-scale agentic reinforcement learning (RL) methods. Traditionally, generating and optimizing CUDA kernels demands specialized hardware knowledge, but CUDA Agent enhances this process by integrating scalable data synthesis, a skill-augmented development environment, and stable long-context training techniques. This novel approach stands out for achieving state-of-the-art performance on KernelBench, significantly outperforming existing methods by up to 100% in speed on various levels of complexity. Key technical innovations include a meticulously crafted pipeline for data synthesis and training, involving a combination of seed problem crawling, LLM-based synthesis, and execution-driven filtering to ensure high-quality outputs. By releasing a curated dataset (CUDA-Agent-Ops-6K) of 6,000 diverse training samples, the project aims to foster reproducible research in RL-based kernel optimization. Moreover, the implementation of a structured reward system promotes the development of genuinely effective kernels, mitigating the tendencies toward shortcut behaviors during policy learning. This initiative not only accelerates the generation of efficient CUDA kernels but also enhances the capabilities of AI systems in performance-critical domains like deep learning.