Developed and delivered a performance optimization feature for the kvcache-ai/sglang repository, targeting Mixture of Experts (MOE) workloads on SM90 GPUs. The work focused on implementing the SwapAB optimization in the Triton fused MOE kernel, which conditionally swaps the dimensions of accumulator and input tensors to better utilize device capabilities and configuration settings. Using Python and leveraging deep learning and GPU programming expertise, this approach reduced kernel latency and increased throughput, enabling higher-concurrency inference and more efficient GPU utilization. The solution incorporated device- and configuration-aware logic, enhancing robustness and adaptability across varying hardware environments without introducing new bugs.