Worked on the microsoft/DeepSpeed repository to improve the stability of distributed training workflows by addressing a critical resource management issue. Focused on the asynchronous I/O path, the developer identified and fixed a file descriptor leak in deepspeed_io_handle_t::wait(), adding a close() call after asynchronous operations to ensure proper release of system resources. This C++ solution prevented file descriptor accumulation during repeated checkpoint I/O, particularly benefiting long-running jobs using ZeRO offload NVMe. By isolating the fix within the I/O handling code, the work enhanced overall code maintainability and reduced the risk of resource exhaustion in production-scale system programming environments.