Worked on the tenstorrent/tt-metal repository to enhance the correctness and reliability of the scaled dot product attention (SDPA) decode path. Focused on fixing tensor dimension handling by aligning buffer allocations with actual tensor shapes, which improved the accuracy of attention operations. Updated the sdpa_decode program factory to ensure tensor shapes were correctly reflected throughout the workflow. Applied targeted memory allocation tuning and refined the test harness to reduce decode-path errors and test flakiness. Utilized C++ and Python, leveraging deep learning and machine learning expertise to stabilize attention computations and enable more robust downstream workloads within the TT-Metal framework.