Jingyi Jiang enhanced the nvidia-cosmos/cosmos-transfer1 repository by developing improved tooling for converting Tensor Parallel checkpoints to Fully Sharded Data Parallel (FSDP) format. Using Python and deep learning expertise, Jingyi delivered a robust conversion script with argument parsing and an option to include the base model, addressing key pain points in large-model training workflows. Comprehensive Markdown documentation and expanded usage instructions were added to streamline onboarding and reproducibility. The work included critical fixes to the TP to FSDP conversion logic, resulting in a more reliable and accessible process for deploying FSDP in production environments and supporting model conversion at scale.