Fix BERT

transformers/modeling_bert.py

@@ -170,7 +170,7 @@ class BertEmbeddings(nn.Module):
             position_ids = torch.arange(seq_length, dtype=torch.long, device=device)
             position_ids = position_ids.unsqueeze(0).expand(input_shape)
         if token_type_ids is None:
-            token_type_ids = torch.zeros(input_shape, dtype=torch.long)
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
 
         if inputs_embeds is None:
             inputs_embeds = self.word_embeddings(input_ids)
@@ -655,11 +655,11 @@ class BertModel(BertPreTrainedModel):
         device = input_ids.device if input_ids is not None else inputs_embeds.device
 
         if attention_mask is None:
-            attention_mask = torch.ones(input_shape)
+            attention_mask = torch.ones(input_shape, device=device)
         if encoder_attention_mask is None:
-            encoder_attention_mask = torch.ones(input_shape)
+            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)
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
 
         # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
         # ourselves in which case we just need to make it broadcastable to all heads.