Inference systems
One cache flag, two attention backends: benchmarking Qwen3.8 on GB200
A prefix-cache benchmark reached 2.29× throughput, then a configuration check revealed different attention backends. Here is what the measurements actually establish.