In March 2025, Mingfei Zhang optimized the Notify Dispatch: Metadata Calculation path in the deepseek-ai/DeepEP repository by implementing dynamic warp sizing using CUDA and C++. He aligned GPU parallelism with the number of channels, enabling a single loop to process metadata for all channels and reducing unnecessary loop iterations. This approach improved throughput and GPU occupancy, enhancing scalability and lowering latency in metadata preparation. The work demonstrated a strong grasp of performance optimization and per-SM warp scheduling, resulting in maintainable code with clear documentation. Zhang’s contribution delivered more efficient resource utilization and laid groundwork for future extensions in DeepEP’s architecture.