The filter costs one or two extra batches of prompts
Rollout seconds per step with the heat-guard pause time taken out, for 12 steps with seed 1. Without the filter the median was 58.0 s. With the filter and 48 prompts per batch, steps that dropped 16 groups or fewer needed one batch and took a median of 88.7 s. Steps that dropped more needed a second batch and took 164.4 s.
0
100
200
0
3
6
9
11
step
rollout seconds, pause time taken out
one batch
one batch
no filter
DAPO filter, batches of 48 prompts
Median 58 s without the filter, 89 s with one batch, 164 s with two.