Keeping the engine loaded saves one or two seconds
Rollout seconds on the steps that had no heat pause in both runs of a same-day pair, seed 1, engine memory fraction 0.25. Default offload against --no-offload-rollout: step 0, 69.4 against 68.0 s; step 1, 58.5 against 56.7 s; step 8, 55.8 against 55.2 s.
0
80
rollout seconds, clean steps only
69.4 s
step 0
68.0 s
step 0
58.5 s
step 1
56.7 s
step 1
55.8 s
step 8
55.2 s
step 8
default offload
engine kept loaded
Same step, same answer length: the difference is 0.6 to 1.8 s.