When should one use `torch.nn.RNNCell` instead of `torch.nn.RNN`?

1

I saw some applications use torch.nn.RNNCell to build stacked RNN layers. Will this manually built block has performance issue compared to torch.nn.RNN?

2

RNN is usually more performant as it uses cuDNN’s implementation under the hood. RNNCell, on the other hand, is more flexible.