Inference efficiency can be measured using metrics including GPU utilization, key-value cache size per token, floating-point operations per forward pass, and latency