Why is Pytorch built-in LSTM much faster than custom written ones?

Question

I have tried several custom LSTM cells, including this but they're all very slow compared to the built-in torch.nn.LSTM. Why is this the case? How can I write an optimum built-in LSTM?