Worked on the rjg-lyh/vllm-ascend repository to deliver a configurable inference optimization feature focused on backend development and performance tuning. Developed a new configuration option that enables frozen parameters, allowing the memory addresses of model weights to remain fixed during inference, which can help reduce input address refresh time during graph execution. The implementation involved updates in Python and Markdown, including comprehensive documentation and test modifications to ensure accurate usage and robust test coverage. Emphasized configuration management best practices and maintained code quality through CI-friendly changes, resulting in a well-documented, maintainable feature that enhances inference stability and performance.