fix memory issue of exporter for bi-transformer

Summary: Before word_feat for bi-transformer is a required input feature, which makes a lot of stuff inefficient for bi-transformer fine-tuning model. So that we met memory issues when exporting hate speech models, eg f105928516. This diff is to fix the inefficiencies for bi-transformer fine-tuning model.

Differential Revision: D14694641

fbshipit-source-id: 85d183033b0490720bfb248756c4a3ae8395bc79

pytext/models/representations/pure_doc_attention.py

@@ -57,7 +57,7 @@ def __init__(self, config: Config, embed_dim: int) -> None:
             self.representation_dim = self.dense.out_dim
 
     def forward(
-        self, embedded_tokens: torch.Tensor, seq_lengths: torch.Tensor, *args
+        self, embedded_tokens: torch.Tensor, seq_lengths: torch.Tensor = None, *args
     ) -> Any:
         rep = self.dropout(embedded_tokens)