Reuse tuning for next sampling call

Questions
7

I did a sanity check to see whether this method can be used to simply resume sampling:

with model:  # initial sample
    step = pm.NUTS()
    trace = pm.sample(draws=100, step=step, tune=3000, cores=n_chains)

from pymc3.step_methods.hmc import quadpotential
with model:
    cov = np.atleast_1d(pm.trace_cov(trace))
    start = list(np.random.choice(trace, n_chains))
    potential = quadpotential.QuadPotentialFull(cov)

with pm.Model() as model2:
    # Reset model here using the same observed data
    step = pm.NUTS(potential=potential)
    trace2 = pm.sample(draws=100, step=step, tune=0, cores=n_chains, start=start)

Which does not work at all - I get acceptance probabilities of 10-40% (for the second sampling). Which other tuning information do I need to carry over?