Worked on NVIDIA-NeMo/Automodel to enhance distributed training reliability by addressing a PyTorch compatibility issue in the Distributed Parallelizer. The solution involved omitting the reshard_after_forward argument when its value is None, preventing runtime errors in PyTorch versions that do not accept explicit None values. This change was reinforced with comprehensive unit tests covering both single- and mixed-dtype sharding scenarios, ensuring robust validation across different configurations. Leveraging expertise in distributed systems, Python, and PyTorch, the work improved cross-version stability for users deploying Automodel at scale, focusing on practical compatibility and maintainability rather than introducing new features during the period.