The program fails because it runs out of memory. Your time series is 6680 points long, and you have used the '012' penalty function set. The result least-squares matrix is 26717 X 6680, and the program runs out of memory when trying to set up this matrix's pseudo-inverse.
There is no easy fix for this. Since the deconvolution+penalty matrix is sparse, I could re-write 3dTfitter to try sparse matrix regression. But that would take some time. On your end, you could reduce the matrix size by using '01' as the penalty set instead of '012' -- this might reduce the matrix size enough for the program to run (it would be slow).