Worked on optimizing large-scale recommender model training by delivering two core features across the pytorch/torchrec and pytorch/FBGEMM repositories. Focused on distributed systems and deep learning, the developer enhanced nested DistributedDataParallel workflows by propagating no_sync contexts, reducing unnecessary gradient synchronization and improving training efficiency for complex model structures. In FBGEMM, introduced support for preallocated host buffers in embedding codegen, enabling external tensor injection for weights and optimizer states, which improved memory efficiency and multi-process parameter sharing. All work was implemented in Python, leveraging GPU programming and PyTorch to address performance, scalability, and configuration safety in distributed machine learning environments.