During November 2025, Ziqi contributed to the IBM/vllm repository by optimizing the AITER MLA backend, focusing on attention processing within the VLLM framework. He implemented a decoupling of the kernel block size, explicitly setting it to one, which streamlined attention operations and improved compatibility across various configurations. This backend development work, carried out in Python and leveraging deep learning and machine learning principles, targeted the AMD pathway to enhance stability and deployment readiness for broader AITER MLA workloads. The contribution addressed a nuanced backend challenge, demonstrating a solid understanding of system-level optimization in a complex machine learning infrastructure.