Tanima Dey developed a Unified Device Management API for DistributedDataParallel within the ROCm/pytorch repository, focusing on simplifying device initialization and usage across multi-GPU and accelerator environments. Using Python and leveraging expertise in PyTorch, distributed computing, and parallel processing, Tanima integrated XCCL changes to enhance cross-node communication and resilience in distributed training. This work reduced setup complexity and boilerplate, improving usability for training workflows and enabling faster onboarding with fewer configuration errors. The project laid a foundation for supporting additional device backends and future scalability, demonstrating depth in distributed systems engineering and a clear focus on practical workflow improvements.