Merge pull request #1770 from huggingface/initi-encoder-mask

Only init encoder_attention_mask if stack is decoder

transformers/modeling_bert.py

@@ -660,8 +660,6 @@ class BertModel(BertPreTrainedModel):
 
         if attention_mask is None:
             attention_mask = torch.ones(input_shape, device=device)
-        if encoder_attention_mask is None:
-            encoder_attention_mask = torch.ones(input_shape, device=device)
         if token_type_ids is None:
             token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
 
@@ -692,6 +690,10 @@ class BertModel(BertPreTrainedModel):
 
         # If a 2D ou 3D attention mask is provided for the cross-attention
         # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
+        if self.config.is_decoder:
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(input_shape, device=device)
+
             if encoder_attention_mask.dim() == 3:
                 encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
             if encoder_attention_mask.dim() == 2:
@@ -699,6 +701,8 @@ class BertModel(BertPreTrainedModel):
 
             encoder_extended_attention_mask = encoder_extended_attention_mask.to(dtype=next(self.parameters()).dtype)  # fp16 compatibility
             encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -10000.0
+        else:
+            encoder_extended_attention_mask = None
 
         # Prepare head mask if needed
         # 1.0 in head_mask indicate we keep the head