Developed a foundational Call-Context Argument Propagation Framework for Keras within the ROCm/tensorflow-upstream repository, focusing on enabling flexible propagation of call-context data through Keras models. This work established the groundwork for future enhancements in control-flow arguments and performance optimizations on ROCm-enabled deployments. Leveraging deep learning expertise and technologies such as TensorFlow, Keras, and Python, the contribution improved the portability and maintainability of Keras models across diverse hardware. By preparing this upstream framework, the developer facilitated future downstream gains in performance and flexibility, aligning with broader goals to enhance machine learning workload efficiency on ROCm platforms without introducing new bugs.