Worked on expanding distributed capabilities in the facebookresearch/param repository by implementing object-based communication support within the PyTorch distributed backend. Developed and integrated object-based collective operations, including all_gather_object and broadcast_object_list, to enable more flexible distributed training patterns. The technical approach involved refining backend utilities, preparation functions, and updating tensor allocation and bandwidth logic to efficiently handle object-based operations. Leveraged skills in backend development, C++, and distributed systems, with Python as the primary language. This work improved memory and bandwidth efficiency for distributed training and established a foundation for future performance optimizations, with all changes tracked through traceable commits.