Worked on deepseek-ai/DeepEP to deliver a major performance optimization for GPU-to-GPU data transfer using RDMA. Refactored the Internode Normal Kernel to utilize multiple Queue Pairs (QPs) with IBGAD/IBGDA, replacing the previous single-QP IBRC approach and enabling parallel data paths for improved throughput. Updated the project’s documentation in Markdown to include new performance metrics and bottleneck analysis, supporting scalability in dual-port NIC and RoCE environments. Leveraged C++, CUDA, and GPU computing expertise to enhance kernel efficiency, laying the groundwork for more scalable and cost-effective training workloads in data-center networking scenarios without introducing new bugs.