Worked on HabanaAI’s optimum-habana-fork repository to address out-of-memory errors during model evaluation with the FP8 Baichuan-13B model. Introduced a max_graphs parameter to the wrap_in_hpu_graph function, enabling control over the number of HPU graphs used during lm_eval and thereby optimizing memory usage. This change stabilized benchmarking and prevented OOM failures, improving the reliability of model deployment workflows. The solution was implemented in Python and leveraged deep learning frameworks with a focus on performance optimization. All modifications were linked to a specific commit for traceability, reflecting a methodical approach to debugging and maintaining high-quality evaluation pipelines.