Worked on the ROCm/onnxruntime repository to address a critical issue in the QNN Model Wrapper related to tensor memory type handling. Using C++ and leveraging expertise in machine learning and tensor operations, implemented a targeted fix that restricts the use of MemHandle to Graph IO tensors while assigning RAW memory type to all other tensors. This adjustment resolved persistent model composition failures by ensuring correct memory-type configuration during static graph assembly. The solution aligned the codebase with ROCm/onnxruntime memory handling guidelines, improved maintainability through explicit tensor memory typing, and enhanced the reliability of QNN model execution on ROCm platforms.