Deep learning inference efficiency can be verified by measuring GPU utilization, KV cache size per token, floating-point operations per forward pass, and latency