Garrett enhanced benchmarking reliability for BLAS backends in the ROCm/flash-attention repository by developing backend-aware benchmarking functionality. He implemented automatic detection of CUDA and HIP support within the benchmark_gemm.py script, enabling dynamic selection of the appropriate backendBLAS for each run. Outputs and descriptions were updated to clearly indicate the active backend, improving the clarity and accuracy of performance comparisons between hipBLAS and cuBLAS. This work, completed in Python and leveraging skills in CUDA, HIP, and performance benchmarking, addressed the need for precise, cross-platform analysis. The depth of the solution reflects careful attention to benchmarking accuracy and maintainability.