Truncated BPTT through LSTM

I’m trying to implement truncated BPTT through LSTMs. However, I noticed that the following code produces c_t, h_t as volatile variables.
self.lstm1 = nn.LSTMCell(1, 20)
h_t, c_t = self.lstm1(input_t, (h_t, c_t))
Doesn’t this prevent backprop through them?


if the input is volatile, the output will be volatile. volatile is a viral flag and it is not back-propable.