When you say 8 lines per stim type, I assume that means you are doing deconvolution with something like 'TENT(0,14,8)'. One problem could be if you have a run with only one stimulus in a class and that stimulus occurs near the end of the run. In that case, the later TENT amplitudes could be pushed off the end of the run, and so those parameters become unestimatable. But this would only occur if that stimulus type ONLY occured in this end-of-run scenario.
If you don't really care about these very rare stimuli, then you could model them with something like 'BLOCK(2)' (assuming 2 s=activation duration), and then use '-stim_base' to put those parameters into the baseline model.