During November 2025, Dezhantu focused on enhancing the stability and reliability of the IBM/vllm repository by addressing critical issues in weight loading and pipeline parallelism for large language models. Using Python and leveraging expertise in deep learning and model optimization, Dezhantu corrected parameter handling in the GptOssModel to ensure accurate weight loading and modified the Llama4 pipeline to gracefully skip missing layers, preventing assertion errors during parallel execution. These targeted bug fixes reduced runtime failures and improved deployment predictability. The work demonstrated careful attention to edge cases and maintainability, with clear documentation supporting future development and troubleshooting efforts.