Worked on the deepseek-ai/FlashMLA repository to deliver a targeted feature upgrade in DeviceAllocation memory management, focusing on improving PyTorch integration. Refactored the backward pass to replace manual CUDA memory management with PyTorch tensor allocation, thereby reducing reliance on cudaMalloc and cudaFree. This approach aligned memory handling with PyTorch semantics, enhancing maintainability and safety while laying the groundwork for broader integration and potential performance improvements. The work was implemented using C++, CUDA, and PyTorch, and addressed the need for more robust and future-proof memory management within the FlashMLA project, reflecting a thoughtful and technically sound engineering solution.