It’d be worth doing a save_warmup=1 run and plotting traceplots.
Plot a few parameters and see what’s happening. If it doesn’t look like (from a far off view) that the chains are getting to the same region of parameter space by 150-200 samples, then it might either be worth increasing the initial adaptation window (the init_buffer argument to cmdstan). Pay attention to the treedepths as well and see if there are differences in different sections of warmup. There’s a graphic of how things work in the Cmdstan manual (Figure 9.3 in the 2.18.1 manual).