cuPilot: A Strategy-Coordinated Multi-agent Framework for CUDA Kernel Evolution

Optimizing CUDA kernels is complex and labor-intensive. This paper introduces cuPilot, a multi-agent framework that uses strategy as an intermediate semantic representation for kernel evolution, addressing suboptimal agent designs and mismatched evolution representations in current LLM-based optimization. Experimental results show cuPilot generated kernels achieve an average speedup of 3.09x over PyTorch.