Worked on the microsoft/DeepSpeed repository to address a critical bug in the Adam optimizer’s subgroup handling, focusing on stability and reproducibility for large-scale distributed training. Using C++ and Python, applied deep learning and optimization algorithm expertise to implement idempotent logic for repeated same-step calls, preventing unnecessary recomputation and ensuring consistent bias-correction metadata across subgroup updates. Ported fixes from previous pull requests to the latest codebase, aligning step-transition logic between CPU and XPU implementations. Added comprehensive regression tests for CPUAdam to validate subgroup-style updates, ultimately enhancing training stability, reproducibility, and performance in multi-update scenarios without introducing new features.