Suppose I loop the fit call manually on a stateful network with batch_input_shape = (1, 1, N)

Will backpropagation still backprop through all the past states to get the full gradient? Or will it just treat the state as another input which it can't backprop through?