Worked on enhancing the ROCm/onnxruntime repository by delivering support for dynamic input shapes in the QNN preprocess workflow. Developed and integrated a new --dynamic_input_shapes argument within the qnn.preprocess function, enabling models to accept dynamic input shapes and allowing static shapes to be defined after preprocessing. This approach addressed existing limitations in the QNN Execution Provider regarding dynamic shapes, thereby expanding model compatibility and improving inference performance across diverse workloads. The work leveraged Python programming and applied machine learning and model optimization techniques, focusing on practical solutions to increase flexibility and efficiency in model deployment scenarios without introducing new bugs.