Worked on the flash-linear-attention repository to deliver Parallax, a parameterized local linear attention mechanism featuring a secondary query stream for improved long-sequence performance. Developed end-to-end integration using Python, PyTorch, and GPU-accelerated Triton kernels, supporting variable-length sequences and optimized decoding. The engineering effort included implementing a dedicated Parallax layer and model, integrating rotary embeddings, and enabling robust KV-cache decoding with sliding-window support. Comprehensive test coverage ensured parity between training and decoding, with support for multiple attention configurations and numeric formats. Refactoring and documentation streamlined naming conventions and release readiness, while style improvements and reproducibility measures enhanced code quality and maintainability.