During January 2025, Sibofeng focused on enhancing the stability and reliability of mixed-precision training in the InternLM/InternEvo repository. Addressing a critical bug in the hybrid optimizer, Sibofeng corrected the handling of fp32 gradients during CPU offloading, ensuring gradients were properly scaled and transferred to the correct device before partitioning. This Python-based solution leveraged deep learning and distributed systems expertise to reduce numerical instability and improve reproducibility in multi-device training environments. Although no new features were introduced, the work demonstrated a deep understanding of optimization challenges and contributed to more robust and reliable training pipelines for complex configurations.