StitchCUDA: An Automated Multi-Agents End-to-End GPU Programing Framework with Rubric-based Agentic Reinforcement Learning
StitchCUDA automates creating entire GPU programs by using three specialized agents—Planner, Coder, and Verifier—in an iterative loop to generate CUDA code. It uses rubric-based reinforcement learning to train the Coder to handle complex system requirements and optimize performance across multiple kernels simultaneously. This framework achieves near 100% success on difficult programming tasks.