Worked on enhancing the stability and reliability of mixed-precision training in the InternLM/InternEvo repository by addressing a critical bug in the hybrid optimizer’s CPU offloading pipeline. Focused on deep learning and distributed systems, the developer used Python to fix incorrect scaling and device movement of fp32 gradients, ensuring gradients are properly transferred to the target device before partitioning. This technical approach reduced numerical instability and improved reproducibility for CPU-offloaded multi-device training. The work did not introduce new user-facing features but delivered a targeted optimization that supports more robust and reliable training across diverse hardware configurations and optimization scenarios.