Pytorch cudnn RNN backward can only be called in training mode

[ptrblck]

You could call either

• call .train() on the rnn module after using model.eval()
• call .eval() only on the necessary modules
• or disable cudnn via torch.backends.cudnn.enabled = False