Improve execution through changes such as launch reduction, graph capture, stream and event coordination, compute-communication overlap, memory reuse, batching, and scheduling Build automated benchmarks and regression detection that preserve performance across models, tensor shapes, data types, software releases, and hardware targets. Isolate bottlenecks across compute, memory bandwidth, cache behavior, kernel launch, synchronization, host-device interaction, communication, and runtime scheduling Use profiling, hardware counters, controlled experiments, and performance models to explain why a workload behaves as it does and identify the highest-leverage optimization.