Worked on stabilizing and optimizing the Flex Attention Benchmark within the pytorch-labs/tritonbench repository, focusing on improving benchmark fidelity and runtime compatibility. The engineering approach involved adjusting default configurations, such as setting the mask type to 'all' for more comprehensive evaluation and increasing the sliding window size to better represent real-world attention workloads. Additionally, Alibi mode was disabled for Flash Attention v3 to address incompatibilities. The work was implemented through two traceable commits using C++ and Python, leveraging skills in benchmarking, CUDA, and performance optimization to deliver more reliable performance data for future planning and optimization of attention mechanisms.