Jon,
That's a different error from the first one you posted (which you in fact fixed by adding -num_glt 21), and it has a different cause. In this case, I would bet that your regressors are linearly dependent. I can't tell what your stim files contain from the command above, but this can happen if you are modelling the baseline as a separate stimulus or have in some other way accounted in your model for every single time point during your scan. Other possible reasons for the dependence could be two stimuli that are identical or a stimulus that's all zeroes (perhaps it represents the subject's errors at a task when in fact the subject made none); in this last case, since the linear dependence is intentional and not an error, you could add the option -regress_opts_3dD -allzero_OK to your command to tell 3dDeconvolve to go through with the analysis anyway.
Your problem could also be the TENT(0,14,9) basis. If your TR is longer than 1.75s, that basis function alone will be enough to cause problems unless 3dDeconvolve is smart enough to check for such things. When using a TENT(b,c,n) basis you'll need to have (c-b)/(n-1) >= TR (since otherwise you're trying to estimate the IRF at a finer resolution than the data you actually have).
Isaac Schwabacher