I think with timed based refresh, we want to show the estimated time to completion in the output. I put that in the other PR. It was basically just calculating iterations per second as the number of iterations completed since the last time-based refresh divided by the refresh interval and then multiplying by the number of iterations remaining.

That’s going to be a very wobbly time estimate given the discrete epochs of warmup and the fact that warmup usually takes about twice as long per iteration as sampling averaged over all iterations. I’m OK with trying it to see if it’s helpful.