Initial hidden state h_0 for multi-layer GRU/LSTM

I am trying to use a 2-layer GRU network, and I want to use two different initial hidden states [[h_0_1],
[h_0_2]]
for the two GRU layers. I am not sure if h_0_1 will be connected to the GRU that takes my direct input, or will it be connected to the GRU that generates the output?

Thanks for the help!