Worked on stabilizing mixed-precision training in the timholy/boltz repository by addressing a critical issue in dropout mask generation. Focused on ensuring that dropout masks maintained consistent floating-point types and device placement, which prevented unintended dropout behavior during training. Improved the determinism and reproducibility of the dropout logic by refining how random intervals were handled and by applying a more precise probability comparison. All changes were delivered through targeted code fixes with minimal impact on the existing codebase. Utilized Python, PyTorch, and deep learning expertise to enhance the reliability of mixed-precision workflows without introducing new features.