Worked on the ROCm/Megatron-LM repository to refactor the parallel linear components, focusing on improving code organization and maintainability for distributed deep learning workflows. The approach involved abstracting the forward implementation logic into a shared member method, enabling both ColumnParallelLinear and RowParallelLinear modules to centralize their forward path selection based on gradient requirements. This reduced code duplication and streamlined future enhancements, particularly for gradient-aware optimizations and testing. The work was implemented using Python and PyTorch, leveraging model parallelism techniques to support scalable training. The changes laid a foundation for more maintainable distributed linear modules and improved long-term team velocity.