You are correct — this is related to the number of sequences, but not due to memory constraints.
@BenKaehler reported this bug to scikit-learn — did you figure out a way to address this in q2-feature-classifier @BenKaehler?
If you have not done so already, I would recommend dereplicating your sequences to reduce the number of unique sequences, rather than training on the full sequence set.
Otherwise, let’s see what Ben has to say!