Worked on the yhyang201/sglang repository to enhance the robustness of the AMX GQA extend attention mechanism, focusing on reducing edge-case failures and improving runtime reliability. Addressed a critical bug by refining the handling of the s_delta variable and introduced targeted unit tests to validate partial extension scenarios with a prefix. This approach strengthened regression coverage and lowered the risk of miscomputations in production environments. Collaborated closely with peers to ensure code quality and maintainability. The work leveraged deep learning and machine learning expertise, utilizing both Python and C++ to reinforce the reliability of core attention path components.