Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Revert low cpu mem tie weights #29135

Merged
merged 2 commits into from
Feb 20, 2024
Merged

Revert low cpu mem tie weights #29135

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
858a23f
Revert "Add tie_weights() to LM heads and set bias in set_output_embe…
amyeroberts Feb 20, 2024
75e514d
Revert "Patch to skip failing `test_save_load_low_cpu_mem_usage` test…
amyeroberts Feb 20, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 0 additions & 6 deletions src/transformers/models/bert/modeling_bert.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -692,9 +692,6 @@ 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)
Expand Down Expand Up @@ -1065,7 +1062,6 @@ 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)
Expand Down Expand Up @@ -1175,7 +1171,6 @@ 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(
Expand Down Expand Up @@ -1329,7 +1324,6 @@ 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(
Expand Down
6 changes: 0 additions & 6 deletions src/transformers/models/big_bird/modeling_big_bird.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1707,9 +1707,6 @@ 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)
Expand Down Expand Up @@ -2269,7 +2266,6 @@ 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)
Expand Down Expand Up @@ -2382,7 +2378,6 @@ 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)
Expand Down Expand Up @@ -2524,7 +2519,6 @@ 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(
Expand Down
4 changes: 0 additions & 4 deletions src/transformers/models/blip/modeling_blip_text.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -523,9 +523,6 @@ 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)
Expand Down Expand Up @@ -820,7 +817,6 @@ 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,
Expand Down
6 changes: 0 additions & 6 deletions src/transformers/models/ernie/modeling_ernie.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -608,9 +608,6 @@ 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)
Expand Down Expand Up @@ -998,7 +995,6 @@ 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)
Expand Down Expand Up @@ -1113,7 +1109,6 @@ 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(
Expand Down Expand Up @@ -1274,7 +1269,6 @@ 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(
Expand Down
4 changes: 0 additions & 4 deletions src/transformers/models/layoutlm/modeling_layoutlm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -589,9 +589,6 @@ 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)
Expand Down Expand Up @@ -872,7 +869,6 @@ 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)
Expand Down
3 changes: 0 additions & 3 deletions src/transformers/models/markuplm/modeling_markuplm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -318,9 +318,6 @@ 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)
Expand Down
6 changes: 0 additions & 6 deletions src/transformers/models/megatron_bert/modeling_megatron_bert.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -659,9 +659,6 @@ 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)
Expand Down Expand Up @@ -1026,7 +1023,6 @@ 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)
Expand Down Expand Up @@ -1136,7 +1132,6 @@ 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)
Expand Down Expand Up @@ -1295,7 +1290,6 @@ 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(
Expand Down
4 changes: 0 additions & 4 deletions src/transformers/models/mpnet/modeling_mpnet.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -587,7 +587,6 @@ 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(
Expand Down Expand Up @@ -660,9 +659,6 @@ 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)
Expand Down
4 changes: 0 additions & 4 deletions src/transformers/models/mra/modeling_mra.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -820,9 +820,6 @@ 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)
Expand Down Expand Up @@ -1056,7 +1053,6 @@ 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(
Expand Down
5 changes: 0 additions & 5 deletions src/transformers/models/nezha/modeling_nezha.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -679,9 +679,6 @@ 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)
Expand Down Expand Up @@ -1047,7 +1044,6 @@ 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)
Expand Down Expand Up @@ -1156,7 +1152,6 @@ 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(
Expand Down
4 changes: 0 additions & 4 deletions src/transformers/models/nystromformer/modeling_nystromformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -428,9 +428,6 @@ 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)
Expand Down Expand Up @@ -669,7 +666,6 @@ 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(
Expand Down
5 changes: 0 additions & 5 deletions src/transformers/models/qdqbert/modeling_qdqbert.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -683,9 +683,6 @@ 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)
Expand Down Expand Up @@ -1027,7 +1024,6 @@ 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)
Expand Down Expand Up @@ -1194,7 +1190,6 @@ 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(
Expand Down
6 changes: 0 additions & 6 deletions src/transformers/models/roc_bert/modeling_roc_bert.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -744,9 +744,6 @@ 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)
Expand Down Expand Up @@ -1093,7 +1090,6 @@ 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)
Expand Down Expand Up @@ -1286,7 +1282,6 @@ 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(
Expand Down Expand Up @@ -1424,7 +1419,6 @@ 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)
Expand Down
4 changes: 0 additions & 4 deletions src/transformers/models/tapas/modeling_tapas.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -729,9 +729,6 @@ 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)
Expand Down Expand Up @@ -1011,7 +1008,6 @@ 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)
Expand Down
4 changes: 0 additions & 4 deletions src/transformers/models/vilt/modeling_vilt.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -896,7 +896,6 @@ 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)
Expand Down Expand Up @@ -1043,9 +1042,6 @@ 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)
Expand Down
4 changes: 0 additions & 4 deletions src/transformers/models/visual_bert/modeling_visual_bert.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -499,9 +499,6 @@ 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)
Expand Down Expand Up @@ -882,7 +879,6 @@ 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)
Expand Down
4 changes: 0 additions & 4 deletions src/transformers/models/yoso/modeling_yoso.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -619,9 +619,6 @@ 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)
Expand Down Expand Up @@ -860,7 +857,6 @@ 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(
Expand Down
Loading
Loading