Usually, this is related to the start value of the chain. I dont know your model so I cant say why excatly, but it might be that your model has some local maximum, and some of them have a very different geometry than the others. Say your second chain started near a local maximum with geometry that very different than the rest (i.e., high curvature), NUTS would adapt to the local curvature resulting a small step size and large leapfrog steps.
One way you can try is to set init to pm.sample(..., init='adapt_diag') so all chain starting from the same point, or supply starting value by hand pm.sample(..., start=[...])