Yuliang worked on the HabanaAI/optimum-habana-fork repository, focusing on improving the stability of model evaluation for the FP8 Baichuan-13B model. By introducing a max_graphs parameter to the wrap_in_hpu_graph function, Yuliang enabled controlled usage of HPU graphs during lm_eval, directly addressing out-of-memory failures that previously disrupted benchmarking. This solution, implemented in Python and leveraging deep learning frameworks, optimized memory management and enhanced the reliability of model deployment workflows. The work demonstrated a targeted approach to performance optimization, resolving a critical bug and contributing to more robust evaluation processes without introducing new features during the development period.