MLCommons Inference reference repo, language/llama2-70b/README.md - the Interactive scenario latency rule

The authoritative statement of what the Interactive scenario actually enforces, quoted verbatim: 'For official, Llama2-70b submissions it is also possible to submit in the interactive category. This sets a more strict latency requirements for Time to First Token (ttft) and Time per Output Token (tpot). Specifically, the interactive category requires loadgen to enforce ttft <= 450ms and tpot <= 40ms'. This is the concrete reason an Interactive throughput number is not comparable to a Server number on the same node: Server enforces only the target_qps=0.5 arrival rate plus the per-query 99th-percentile bounds, while Interactive additionally hard-caps inter-token latency at 40ms, which caps how many concurrent streams a node can hold. Note also that both llama2-70b and llama3.1-8b user.conf files in the same repo contain only '*.Offline.min_duration = 600000', '*.Offline.min_query_count = 2000' (llama2 only), '*.Server.target_qps = 0.5' and '*.Server.min_duration = 120000' - the per-scenario latency bounds are NOT in user.conf, they are compiled into the harness for each model and scenario, so a submitter's user.conf cannot be read to learn the latency target that applied.