Developed a block-sparse paged attention kernel for sliding window attention on TPU within the apple/axlearn repository, focusing on optimizing long-context machine learning workloads. The work involved enhancing logit bias handling and refining mask functions to improve both accuracy and robustness of the attention mechanism. Leveraging JAX and Python, the developer prioritized memory efficiency and compute throughput, addressing the challenges of scaling attention for large input sequences. Comprehensive unit tests and benchmarks were added to validate kernel correctness and performance on TPU hardware. This contribution demonstrates a strong focus on performance optimization and deep understanding of TPU programming for advanced machine learning applications.