Developed and integrated the Mamba3 selective state-space layer for the fla-org/flash-linear-attention repository, focusing on scalable attention mechanisms for deep learning workloads. The work replaced causal convolutions with MIMO projections and per-head biases, optimizing both performance and maintainability. Leveraging PyTorch and Python, the implementation included comprehensive unit testing and stability improvements, such as refined weight initialization and enhanced error handling. Adjustments to inference caching and test workflows aligned the project with production standards, increasing test coverage and reliability. These contributions laid the foundation for broader adoption of the Mamba3 pathway, supporting faster and more robust deployment of attention-based models.