Add tie_weights() to LM heads and set bias in set_output_embeddings() (…

…huggingface#28948)

* Add tie_weights() to LM heads and set bias in set_output_embeddings()

The bias were not tied correctly in some LM heads, and this change should fix that.

* Moving test_save_and_load_low_cpu_mem_usage to ModelTesterMixin

* Adding _tie_weights() to MPNet and Vilt

* Skip test for low cpu mem usage for Deta/DeformableDetr since they cannot init on meta device

* Rename to test name to save_load to match the convention

src/transformers/models/bert/modeling_bert.py

@@ -692,6 +692,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)
@@ -1062,6 +1065,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=BertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
@@ -1171,6 +1175,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
@@ -1324,6 +1329,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(

src/transformers/models/big_bird/modeling_big_bird.py

@@ -1707,6 +1707,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)
@@ -2266,6 +2269,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(BIG_BIRD_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=BigBirdForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
@@ -2378,6 +2382,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(BIG_BIRD_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=MaskedLMOutput, config_class=_CONFIG_FOR_DOC)
@@ -2519,6 +2524,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(BIG_BIRD_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(

src/transformers/models/blip/modeling_blip_text.py

@@ -523,6 +523,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)
@@ -816,6 +819,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     def forward(
         self,

src/transformers/models/ernie/modeling_ernie.py

@@ -608,6 +608,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)
@@ -995,6 +998,7 @@ def get_output_embeddings(self):
     # Copied from transformers.models.bert.modeling_bert.BertForPreTraining.set_output_embeddings
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(ERNIE_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=ErnieForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
@@ -1109,6 +1113,7 @@ def get_output_embeddings(self):
     # Copied from transformers.models.bert.modeling_bert.BertLMHeadModel.set_output_embeddings
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(ERNIE_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
@@ -1269,6 +1274,7 @@ def get_output_embeddings(self):
     # Copied from transformers.models.bert.modeling_bert.BertForMaskedLM.set_output_embeddings
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(ERNIE_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(

src/transformers/models/layoutlm/modeling_layoutlm.py

@@ -589,6 +589,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)
@@ -869,6 +872,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(LAYOUTLM_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=MaskedLMOutput, config_class=_CONFIG_FOR_DOC)

src/transformers/models/markuplm/modeling_markuplm.py

@@ -318,6 +318,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)

src/transformers/models/megatron_bert/modeling_megatron_bert.py

@@ -659,6 +659,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)
@@ -1023,6 +1026,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(MEGATRON_BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=MegatronBertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
@@ -1132,6 +1136,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(MEGATRON_BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)
@@ -1290,6 +1295,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(MEGATRON_BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(

src/transformers/models/mpnet/modeling_mpnet.py

@@ -587,6 +587,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.lm_head.decoder = new_embeddings
+        self.lm_head.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(MPNET_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
@@ -659,6 +660,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, features, **kwargs):
         x = self.dense(features)
         x = gelu(x)

src/transformers/models/mra/modeling_mra.py

@@ -820,6 +820,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)
@@ -1053,6 +1056,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(MRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(

src/transformers/models/nezha/modeling_nezha.py

@@ -679,6 +679,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)
@@ -1044,6 +1047,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(NEZHA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=NezhaForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)
@@ -1152,6 +1156,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(NEZHA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(

src/transformers/models/nystromformer/modeling_nystromformer.py

@@ -428,6 +428,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)
@@ -666,6 +669,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(NYSTROMFORMER_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(

src/transformers/models/qdqbert/modeling_qdqbert.py

@@ -683,6 +683,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)
@@ -1024,6 +1027,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(QDQBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)
@@ -1190,6 +1194,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(QDQBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(

src/transformers/models/roc_bert/modeling_roc_bert.py

@@ -744,6 +744,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)
@@ -1090,6 +1093,7 @@ def get_output_embeddings(self):
     # Copied from transformers.models.bert.modeling_bert.BertForPreTraining.set_output_embeddings
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(ROC_BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=MaskedLMOutput, config_class=_CONFIG_FOR_DOC)
@@ -1282,6 +1286,7 @@ def get_output_embeddings(self):
     # Copied from transformers.models.bert.modeling_bert.BertForMaskedLM.set_output_embeddings
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(ROC_BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     def forward(
@@ -1419,6 +1424,7 @@ def get_output_embeddings(self):
     # Copied from transformers.models.bert.modeling_bert.BertLMHeadModel.set_output_embeddings
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(ROC_BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)

src/transformers/models/tapas/modeling_tapas.py

@@ -729,6 +729,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)
@@ -1008,6 +1011,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(TAPAS_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=MaskedLMOutput, config_class=_CONFIG_FOR_DOC)

src/transformers/models/vilt/modeling_vilt.py

@@ -896,6 +896,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.mlm_score.decoder = new_embeddings
+        self.mlm_score.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(VILT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=MaskedLMOutput, config_class=_CONFIG_FOR_DOC)
@@ -1042,6 +1043,9 @@ def __init__(self, config, weight=None):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, x):
         x = self.transform(x)
         x = self.decoder(x)

src/transformers/models/visual_bert/modeling_visual_bert.py

@@ -499,6 +499,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)
@@ -879,6 +882,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(VISUAL_BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @replace_return_docstrings(output_type=VisualBertForPreTrainingOutput, config_class=_CONFIG_FOR_DOC)

src/transformers/models/yoso/modeling_yoso.py

@@ -619,6 +619,9 @@ def __init__(self, config):
         # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
         self.decoder.bias = self.bias
 
+    def _tie_weights(self):
+        self.decoder.bias = self.bias
+
     def forward(self, hidden_states):
         hidden_states = self.transform(hidden_states)
         hidden_states = self.decoder(hidden_states)
@@ -857,6 +860,7 @@ def get_output_embeddings(self):
 
     def set_output_embeddings(self, new_embeddings):
         self.cls.predictions.decoder = new_embeddings
+        self.cls.predictions.bias = new_embeddings.bias
 
     @add_start_docstrings_to_model_forward(YOSO_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(