Wang Li focused on enhancing the reliability and memory management of the unsloth-zoo repository, addressing a critical GPU memory handling bug that previously risked out-of-memory scenarios during inference. By enforcing a minimum value for max_num_seqs in GPU memory calculations, Wang Li improved the robustness of resource allocation and reduced the likelihood of invalid memory sizing. The work involved updating Python modules, particularly vllm_utils.py, to introduce input validation and clearer error messaging, thereby strengthening error handling. Collaborating across teams, Wang Li applied expertise in Python and GPU optimization to deliver more stable and predictable memory utilization for production workloads.