Tsinghua AIR and ByteDance Seed introduce CUDA Agent for LLM-generated CUDA kernels
AITsinghua AIR and ByteDance Seed introduce CUDA Agent, a large-scale agentic reinforcement learning system for generating high-performance CUDA kernels. The system uses a synthesized six-thousand-sample dataset, milestone-based rewards, and a staged warm-up training strategy, and the source reports that it outperforms torch.compile on most KernelBench Level-1 and Level-2 tasks.
AIWhy it matters.
The source explains how the CUDA Agent system uses agentic reinforcement learning and a staged training recipe to generate GPU kernels faster than torch.compile.
Source details
Tsinghua AIR · air.tsinghua.edu.cnPublished · added here
Original title: AIR科研｜突破大语言模型 CUDA 代码生成瓶颈：基于智能体强化学习的 CUDA Agent 系统