During May 2025, contributed to the InfiniTensor/InfiniCore repository by addressing a critical issue in floating-point data conversion. Focused on the numerical correctness and stability of FP16 workflows, the work involved refining the _f32_to_f16() function to properly handle edge cases involving infinity and NaN during FP32 to FP16 conversion. This low-level programming effort in C++ and floating-point arithmetic ensured that extreme values were accurately represented, preventing downstream errors in model inference and training pipelines. The fix improved the reliability of FP16 data handling across platforms, reducing the risk of corrupted representations and enhancing overall model stability.