Worked on the flash-linear-attention repository to enhance autotuning correctness and efficiency by implementing autotuning parameter deduplication. Addressed a bug where duplicate autotune keys led to redundant parameter evaluations, resulting in unnecessary compute usage and unpredictable performance. Refactored Python code, particularly in chunk_o_bwd.py, to ensure only unique parameters are considered during the autotuning process. This change improved the reliability and predictability of autotuning across various workloads. Demonstrated skills in code refactoring, performance optimization, and effective use of version control. The work focused on backend improvements rather than feature development, emphasizing robust engineering and maintainability within the autotuning workflow.