python3 run_eval.py --arm autoregressive --upstream http://localhost:8000 --model $PLAIN --run 2026-09-23-l4-latency --limit 100 --concurrency 1
python3 run_eval.py --arm diffusion --upstream http://localhost:8000 --model $DIFF --run 2026-09-23-l4-latency --limit 100 --concurrency 1 --reads 4 --keep-top 5