Worked on the swiss-ai/Megatron-LM repository to enhance the robustness of the MambaStack model architecture by addressing error handling in layer allocation. Refactored the codebase in Python to explicitly handle invalid layer types, ensuring that misconfigurations now raise a ValueError during model construction. Added targeted unit tests to validate this behavior, reducing the risk of runtime errors and improving deploy-time reliability. Focused on deep learning model stability, the changes streamline error detection and future maintenance. The work demonstrates a methodical approach to testing and error handling, contributing to more predictable and maintainable model builds within the Megatron-LM framework.