Worked on the fla-org/flash-linear-attention repository, developing a new autograd function for normalized linear attention using Python and PyTorch. This feature improved the accuracy and efficiency of the backward pass by addressing gradient and stability issues, optimizing tensor operations, and reducing redundant calculations. The implementation included targeted code refactors to enhance maintainability and test coverage, as well as updates to unit tests for better numerical stability. By refining the handling of key variables and clarifying function naming, the work contributed to more scalable attention mechanisms for long sequences, while also supporting collaborative development through comprehensive documentation and co-authorship.