It seems that the Truncated logp switches all values to negative, if any of them is invalid:
import pymc as pm
with pm.Model() as m:
x = pm.TruncatedNormal("x", mu=0, sigma=1, lower=0, upper=None, transform=None, shape=2)
# -1 is invalid, but 1 is fine
print(m.compile_logp(sum=False)({"x": [-1, 1]})) # [array([-inf, -inf])]
It should return -inf for the first, and a finite value for the second