Worked on enhancing the robustness of model weight loading in the sglang repository, specifically addressing the NemotronH model’s handling of expert scale weights. Implemented a fix in Python using PyTorch to ensure that missing expert scale weights at runtime no longer cause errors, thereby improving the reliability of NemotronH deployments. Developed comprehensive unit tests to validate the model’s behavior when parameters are absent, increasing test coverage and safeguarding against regressions. Focused on maintaining clear traceability and review readiness by aligning the changes with the relevant issue. The work emphasized model loading reliability and thorough testing within the sglang codebase.