Worked on the intel/torch-xpu-ops repository to stabilize XPU profiling and normalize cross-version event traces. Focused on reducing profiling test flakiness by filtering XPU-specific runtime noise and unifying kernel launch event names across SYCL runtime versions. Used Python for both performance optimization and profiling, implementing functions to drop irrelevant runtime events while preserving accurate operator attribution through FX stack traces. Updated test baselines to reflect actual XPU operation traces, ensuring consistency across environments and operating systems. This work enabled more reliable performance comparisons and streamlined development of XPU features, supporting faster iteration and improved test suite stability without introducing new bugs.