Is this a single task, or are there multiple tasks/conditions involved? The former case might be tractable, the latter sounds pretty hard.

In the single task case, you could setup your 3dDeconvolve analysis run. Then
analyze the same data repeatedly, adding various levels of the "-stim_times_subtract SS" option, where "SS" is the amount of time (in seconds) to subtract from each input time. You are allowed to use a negative value for "SS"! Then look at each result and see which one is the "best". It might be a little tedious, but it might also work and save the data. It would probably work better if you have a block design rather than a short event-related design.

How much to shift by? Steps of 1 second make sense to me. How far out to go? That depends on how bad you thinks things are! You could run with -9 -8 ... -1 0 1 ... 8 9 (19 cases) , for example.

Good luck!