Worked on the fla-org/flash-linear-attention repository to address numerical stability issues in deep learning attention modules. Focused on improving the reliability of RMSNorm by enforcing float32 data types, which mitigated precision underflow during attention calculations. Enhanced the LayerNorm implementation by adding factory_kwargs support, making the API more flexible for integration with external models. Applied comprehensive lint improvements across the codebase to boost maintainability and continuous integration reliability. Utilized Python and PyTorch throughout the process, emphasizing robust numerical computation and code quality. The work reduced training instability risks and streamlined downstream model integration, reflecting a methodical approach to engineering challenges.