Worked on the ROCm/AMDMIGraphX repository to deliver ONNX MultiHeadAttention operator support within the ONNX parser, enabling direct deployment of complex attention-based models on ROCm platforms. Developed parsing logic in C++ to handle multiple input formats and implemented graph transformations to support advanced attention mechanisms. This addition expanded the expressiveness and portability of models by reducing reliance on custom operators. The work involved deep learning operations, ONNX parsing, and thorough testing to ensure compatibility and reliability. No major bugs were addressed during this period, with the primary focus on feature development and cross-team collaboration to enhance ONNX model support.