Mingxu developed a lightweight DeepSeek-671B model to streamline host-offloading workflow validation across the Intel-tensorflow/xla and ROCm/tensorflow-upstream repositories. By reducing the model to fewer layers, Mingxu enabled fast, repeatable performance testing and established benchmarking scaffolding using TensorFlow and Python. The work included targeted HLO adjustments and the integration of a dedicated benchmark artifact, supporting reproducible evaluation of host-offload scenarios. Mingxu’s contributions aligned changes across both forks, closed related issues, and improved testing coverage. This focused engineering effort provided a scalable foundation for future performance assessments, demonstrating depth in data processing, deep learning, and machine learning within a one-month period.