Developed performance-focused enhancements for the linkedin/Liger-Kernel repository by implementing Fused Neighborhood Attention (FNA) optimized for Atlas A2 NPUs. Refactored the attention grid to a 1D structure, improving thread mapping and preventing local memory overflow, while tuning NPU-affinity softmax tiling and grid sizing to maximize throughput under memory constraints. Leveraged deep learning expertise with PyTorch and Python to reduce synchronization overhead and increase memory efficiency for attention-heavy workloads. Conducted comprehensive end-to-end validation, including benchmark scripts and unit tests, ensuring code quality and adherence to style guidelines. The work enables higher throughput and efficiency for downstream models on NPU architectures.