It is done on the CPU instead of the GPU. GPU would seem like the natural choice for a convolution heavy model but was not used here for the mentioned reasons.