During November 2025, Zhengjun Xing focused on stabilizing benchmarking workflows in the pytorch/FBGEMM repository by addressing file descriptor exhaustion during large-copy scenarios. He implemented a file_system multiprocessing sharing strategy in Python, allowing users to toggle the approach via the PYTORCH_SHARE_STRATEGY environment variable. This solution prevented OSError issues related to excessive open files, improving the reliability and reproducibility of performance measurements without requiring manual adjustment of OS file descriptor limits. Zhengjun’s work leveraged skills in benchmarking, multiprocessing, and performance optimization, enabling more robust large-scale benchmarking runs and reducing operational overhead for developers working with high-volume data copies.