Fix bug in forward_embedding (#931)

Summary:
See: https://twitter.com/nymwa/status/1200684169115734016
Pull Request resolved: https://github.com/fairinternal/fairseq-py/pull/931

Differential Revision: D18782345

Pulled By: myleott

fbshipit-source-id: 9e74287a8ce677237042c5fdbe0bdbd4774b5ce6

fairseq/models/transformer.py

@@ -353,7 +353,7 @@ class TransformerEncoder(FairseqEncoder):
 
     def forward_embedding(self, src_tokens):
         # embed tokens and positions
-        embed = self.embed_scale * self.embed_tokens(src_tokens)
+        x = embed = self.embed_scale * self.embed_tokens(src_tokens)
         if self.embed_positions is not None:
             x = embed + self.embed_positions(src_tokens)
         if self.layernorm_embedding: