Worked on the ai-dynamo/nixl repository to address a critical issue in the UCX backend involving multi-GPU data transfer and latency accounting. Focused on backend development and GPU computing using C++, the work corrected calculations for total data transferred and average latency in complex multi-initiator, multi-GPU scenarios. The solution ensured that after a progress-thread restart, the main thread’s context was properly reapplied, restoring accurate operation and measurement. This fix improved the reliability and observability of performance metrics, supporting more effective performance optimization in demanding GPU deployments. The contribution reflects careful attention to detail in high-performance computing environments.