Dhruva Kaushal delivered stabilization and performance optimization for the Flex Attention Benchmark in the pytorch-labs/tritonbench repository. Focusing on benchmarking and CUDA, Dhruva addressed runtime compatibility by disabling Alibi mode for Flash Attention v3 and improved benchmark fidelity by changing the default mask type to ‘all’ and increasing the sliding window size from 128 to 4096. These C++ and Python code changes ensure the benchmark more accurately reflects real-world attention workloads, providing more reliable performance data for future planning. The work, implemented through two well-documented commits, demonstrates a focused approach to code configuration and performance tuning within a complex benchmarking suite.