During September 2025, Wang Ziyue focused on stabilizing the VLM engine in the jd-opensource/xllm repository by addressing a critical runtime error encountered when tensor parallelism was enabled. By analyzing and correcting the handling of forward input data, Wang ensured accurate data propagation throughout the engine, directly improving reliability and reducing production risk for parallel inference workloads. This work required deep understanding of C++ and runtime error handling, as well as expertise in tensor parallelism. Although the contribution centered on a single bug fix, it demonstrated careful attention to system stability and enabled safer scaling of VLM workloads in production environments.