update for transformers >= 4.29.1

modeling_lsg_bert.py

@@ -189,19 +189,25 @@ class CausalAttentionProduct(nn.Module):
         del key_layer
 
         if attention_mask is not None:
-            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
-            attention_scores = attention_scores + attention_mask
-
             # Add causal mask
             causal_shape = (self.block_size, self.block_size) if causal_shape is None else causal_shape
             causal_mask = torch.tril(
                 torch.ones(*causal_shape, device=attention_mask.device, dtype=attention_scores.dtype), 
                 diagonal=-1
                 ) 
-            causal_mask = causal_mask.T * torch.finfo(attention_scores.dtype).min
-            attention_scores[..., -causal_shape[0]:, -causal_shape[1] + 1:] = causal_mask[:, 1:]
+            
+            # Min value
+            dtype_min = torch.tensor(
+                        torch.finfo(attention_scores.dtype).min, device=attention_scores.device, dtype=attention_scores.dtype
+                    )
 
+            # Build causal + attention_mask
+            causal_mask = torch.nn.functional.pad(causal_mask.T * dtype_min, (attention_mask.size()[-1] - self.block_size, 0), value=0)
+            attention_mask = torch.max(attention_mask + causal_mask.unsqueeze(0).unsqueeze(0).unsqueeze(0), dtype_min)
+
+            attention_scores = attention_scores + attention_mask
             del attention_mask
+            del causal_mask
 
         # Normalize the attention scores to probabilities.
         attention_probs = nn.Softmax(dim=-1)(attention_scores)
@@ -991,8 +997,6 @@ class LSGBertModel(LSGBertPreTrainedModel, BertModel):
     documentation alongside usage examples.
     """
 
-    config_class = LSGBertConfig
-
     def __init__(self, config, add_pooling_layer=True):
         
         LSGBertPreTrainedModel.__init__(self, config)
@@ -1031,6 +1035,8 @@ class LSGBertModel(LSGBertPreTrainedModel, BertModel):
 
 class LSGBertForPreTraining(LSGBertPreTrainedModel, BertForPreTraining):
 
+    _tied_weights_keys = ["predictions.decoder.bias", "cls.predictions.decoder.weight"]
+
     def __init__(self, config):
         
         LSGBertPreTrainedModel.__init__(self, config)
@@ -1044,8 +1050,7 @@ class LSGBertForPreTraining(LSGBertPreTrainedModel, BertForPreTraining):
 
 class LSGBertLMHeadModel(LSGBertPreTrainedModel, BertLMHeadModel):
 
-    _keys_to_ignore_on_load_unexpected = [r"pooler"]
-    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+    _tied_weights_keys = ["predictions.decoder.bias", "cls.predictions.decoder.weight"]
 
     def __init__(self, config):
         
@@ -1067,9 +1072,7 @@ class LSGBertForMaskedLM(LSGBertPreTrainedModel, BertForMaskedLM):
     documentation alongside usage examples.
     """
 
-    config_class = LSGBertConfig
-    _keys_to_ignore_on_load_unexpected = [r"pooler"]
-    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+    _tied_weights_keys = ["predictions.decoder.bias", "cls.predictions.decoder.weight"]
 
     def __init__(self, config):
 
@@ -1107,8 +1110,6 @@ class LSGBertForSequenceClassification(LSGBertPreTrainedModel, BertForSequenceCl
     appropriate documentation alongside usage examples.
     """
 
-    config_class = LSGBertConfig
-
     def __init__(self, config):
         
         LSGBertPreTrainedModel.__init__(self, config)
@@ -1133,8 +1134,6 @@ class LSGBertForMultipleChoice(LSGBertPreTrainedModel, BertForMultipleChoice):
     appropriate documentation alongside usage examples.
     """
 
-    config_class = LSGBertConfig
-
     def __init__(self, config):
         
         LSGBertPreTrainedModel.__init__(self, config)
@@ -1156,9 +1155,6 @@ class LSGBertForTokenClassification(LSGBertPreTrainedModel, BertForTokenClassifi
     appropriate documentation alongside usage examples.
     """
 
-    config_class = LSGBertConfig
-    _keys_to_ignore_on_load_unexpected = [r"pooler"]
-
     def __init__(self, config):
         
         LSGBertPreTrainedModel.__init__(self, config)
@@ -1182,9 +1178,6 @@ class LSGBertForQuestionAnswering(LSGBertPreTrainedModel, BertForQuestionAnsweri
     appropriate documentation alongside usage examples.
     """
 
-    config_class = LSGBertConfig
-    _keys_to_ignore_on_load_unexpected = [r"pooler"]
-
     def __init__(self, config):
         
         LSGBertPreTrainedModel.__init__(self, config)