Worked on the huggingface/transformers repository to implement cross-version TF32 API support, focusing on improving compatibility and performance across different PyTorch releases. Developed a PyTorch-version-aware selector for TF32 APIs, introduced a global TF32 setting, and added MUSA compatibility to broaden GPU support. The refactored TF32 handling stabilized training and inference performance while simplifying configuration for users. All changes were backed by automated tests and code quality checks, including enhancements to packaging and export logic. Utilized Python, PyTorch, and deep learning expertise to deliver a feature that accelerates model workflows and reduces integration friction for machine learning practitioners and researchers.