[Torch] Support aten::lstm

[masahi]

It seems the commonly used torch.nn.LSTM module generates aten::lstm op, which we currently don't support. It is reported in https://discuss.tvm.apache.org/t/import-rnn-t-pytorch-model-into-tvm/7874 and it is likely we'll get more requests for this op in the future.

The corresponding PyTorch implementation is here https://github.com/pytorch/pytorch/blob/70c8daf43946b53af6493d058899ef952d27d339/aten/src/ATen/native/RNN.cpp#L1396

[jroesch]

I believe the MXNet frontend lowers LSTM to Relay directly.

[zhangzhen507]

But, the pre-trained rnn-t model https://zenodo.org/record/3662521 is a pytorch model. Can you support the torch.nn.LSTM by aten::lstm in TVM ?

[masahi]

yes, torch.nn.LSTM is translated to aten::lstm in Torchscript.

[zhangzhen507]

yes, torch.nn.LSTM is translated to aten::lstm in Torchscript.

Hi, do you have a plan when you will support the aten::lstm in tvm/relay/frontend/pytorch.py ?

[masahi]

No from my end. I'm not expert on LSTM and don't have time for this.

[masahi]

#8447