During December 2024, Dakersnar enhanced the NVPTX backend in the espressif/llvm-project repository by generalizing the upsizing of 8- and 16-bit vector loads and stores. This work centralized the logic for vector upsizing, streamlining code generation and simplifying future maintenance. By focusing on low-level optimization and leveraging expertise in C++ and LLVM IR, Dakersnar enabled more efficient lowering of vector types to PTX, which can improve codegen efficiency and runtime performance. The depth of the contribution lies in its targeted approach to maintainability and extensibility, addressing a specific need in GPU architecture without introducing unnecessary complexity.