Hi Francesco,
You can still use a queue system for running probtackx2_gpu.
By default, probtackx2_gpu will get the first free gpu (not any specific GPU is selected), exactly the same that in bedpostx_gpu.
Your queue system will need to assign different GPUs to different processes.
Before CUDA 7.5, one way to deal with this problem was to set the Compute mode of the GPUs to EXCLUSIVE_THREAD.
That way, when a CPU thread asked for a GPU, the GPU was blocked and it was not assigned to any other CPU thread until that thread finished. Then, the queue system did not have to deal with the problem.
Since CUDA 7.5, EXCLUSIVE_THREAD mode has been removed, but you can use EXCLUSIVE_PROCESS mode, which will have the same effect on bedpostx_gpu and probtackx2_gpu.
You can check and set the Compute mode of your GPUs using the command 'nvidia-smi'
Moises.