Fix contiguous isse in bilstm export (#1270)

Summary:
Pull Request resolved: #1270

Call contiguous function on the input embedding in case it's not contiguous, which will cause issue in the Caffe2 operator

Reviewed By: shivanipoddariiith

Differential Revision: D20277324

fbshipit-source-id: 1cfe1e595f8a1014b08c7dba33f8e58a7cc73c34

pytext/models/representations/bilstm.py

@@ -123,7 +123,7 @@ def forward(
             )
             states = (state, state)
         if torch.onnx.is_in_onnx_export():
-            lstm_in = [embedded_tokens, states[0], states[1]] + [
+            lstm_in = [embedded_tokens.contiguous(), states[0], states[1]] + [
                 param.detach() for param in self.lstm._flat_weights
             ]
             rep, new_state_0, new_state_1 = torch.ops._caffe2.InferenceLSTM(