Developed a pluggable collective backend framework for the facebookresearch/param repository, focusing on distributed training with PyTorch. Leveraging Python and the mixin pattern, the work introduced a plugin-friendly architecture that allows users to integrate custom collective backends and kernels with minimal friction. The implementation included a new pytorch_ncclx_backend supporting dynamic all-to-all operations, enhancing performance and flexibility for distributed systems. Refactoring efforts improved synchronization correctness by ensuring collective operations are cleared after device synchronization, leading to better resource management. The approach emphasized extensibility and maintainability, enabling easier experimentation and collaboration for backend development in distributed machine learning workflows.