Yeah I did everything via CmdstanPy, was kind of awkward (but great in the end, thank you dear developers, @mitzimorris and @ahartikainen I guess?), but certainly 100% better than if I had tried to abuse the C++ code.
That’s the window for the step size adaptation before we start to sample, but we do not want to start to sample until the very last run, which includes all measurements. Once we decide to sample, we do use the regular adapt_step_size before.
Edit: I’m actually not 100% sure whether it may be better to include the last window, and then use that step size for the next iteration. Preliminary tests looked like this was worse, but they were really not exhaustive.
Edit2: I believe the point of the last warmup window is to adapt the stepsize such that we get the aimed for acceptance rate? But this doesn’t appear to be relevant for the warmup phase? It looks like for the second warmup phase, the stepsize is always just one?