NVIDIA AI agent optimizes Kimi Delta Attention kernel, achieving 3x speedup

NVIDIA's research team utilized an AI agent to rewrite the GPU kernel for Kimi Delta Attention, a core mechanism in Moonshot AI's Kimi-Linear model. The AI-generated code achieved 2.96 times the speed of the official FlashKDA version on NVIDIA B300 hardware, based on the geometric mean of six test tasks. The kernel has been open-sourced after rigorous validation against real-world Kimi data and edge cases to ensure accuracy.