Moning Chen delivered a major performance optimization for the deepseek-ai/DeepEP repository by refactoring the Internode Normal Kernel to use multiple Queue Pairs for RDMA data transfer between GPUs. Leveraging CUDA and C++, Moning replaced the previous single-QP IBRC approach with a multi-QP architecture using IBGAD and IBGDA, enabling parallel data paths and improving kernel throughput in dual-port NIC and RoCE environments. The work included updating documentation in Markdown to present new performance metrics and bottleneck analysis. This engineering effort enhanced GPU-to-GPU communication scalability, addressing network performance bottlenecks and supporting more efficient distributed training workloads.