Worked on the openvinotoolkit/training_extensions repository to deliver XPU device support and upgrade PyTorch, broadening hardware compatibility and improving training performance. The engineering approach involved refactoring Python utilities for more robust device handling and addressing mixed-precision reliability to ensure stable training workflows. Installation processes and documentation were updated to reflect new dependencies and setup steps, simplifying adoption for users. By integrating deep learning techniques with system integration skills, the work enabled more reliable and efficient training across a wider range of accelerators. The changes focused on enhancing usability, stability, and performance for developers working with PyTorch and XPU hardware.