Fix model kwargs

src/transformers/modeling_utils.py

@@ -5198,6 +5198,9 @@ def tplize(mod: torch.nn.Module) -> None:
 
     @property
     def loss_function(self):
+        if hasattr(self, "_loss_function"):
+            return self._loss_function
+
         loss_type = getattr(self, "loss_type", None)
 
         if loss_type is None or loss_type not in LOSS_MAPPING:
@@ -5208,6 +5211,10 @@ def loss_function(self):
             loss_type = "ForCausalLM"
         return LOSS_MAPPING[loss_type]
 
+    @loss_function.setter
+    def loss_function(self, value):
+        self._loss_function = value
+
     def get_compiled_call(self, compile_config: CompileConfig):
         """Return a `torch.compile`'d version of `self.__call__`. This is useful to dynamically choose between
         non-compiled/compiled `forward` during inference, especially to switch between prefill (where we don't

src/transformers/models/bamba/modeling_bamba.py

@@ -1208,6 +1208,7 @@ def forward(
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,  # NOOP kwargs, for now
     ) -> Union[Tuple, BaseModelOutputWithPast]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (

src/transformers/models/bamba/modular_bamba.py

@@ -949,6 +949,7 @@ def forward(
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,  # NOOP kwargs, for now
     ) -> Union[Tuple, BaseModelOutputWithPast]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (

src/transformers/models/bert_generation/modeling_bert_generation.py

@@ -20,7 +20,6 @@
 import torch
 import torch.utils.checkpoint
 from torch import nn
-from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
 from ...generation import GenerationMixin
@@ -734,6 +733,7 @@ def forward(
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,  # NOOP kwargs, for now
     ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
         r"""
         encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
@@ -901,6 +901,7 @@ def forward(
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
         r"""
         encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
@@ -963,18 +964,20 @@ def forward(
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
+            **kwargs,
         )
 
         sequence_output = outputs[0]
         prediction_scores = self.lm_head(sequence_output)
 
         lm_loss = None
         if labels is not None:
-            # we are doing next-token prediction; shift prediction scores and input ids by one
-            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
-            labels = labels[:, 1:].contiguous()
-            loss_fct = CrossEntropyLoss()
-            lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            lm_loss = self.loss_function(
+                prediction_scores,
+                labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
 
         if not return_dict:
             output = (prediction_scores,) + outputs[1:]

src/transformers/models/big_bird/modeling_big_bird.py

@@ -1983,6 +1983,7 @@ def forward(
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,  # NOOP kwargs, for now
     ) -> Union[BaseModelOutputWithPoolingAndCrossAttentions, Tuple[torch.FloatTensor]]:
         r"""
         encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
@@ -2540,6 +2541,7 @@ def forward(
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[CausalLMOutputWithCrossAttentions, Tuple[torch.FloatTensor]]:
         r"""
         encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
@@ -2580,18 +2582,20 @@ def forward(
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
+            **kwargs,
         )
 
         sequence_output = outputs[0]
         prediction_scores = self.cls(sequence_output)
 
         lm_loss = None
         if labels is not None:
-            # we are doing next-token prediction; shift prediction scores and input ids by one
-            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
-            labels = labels[:, 1:].contiguous()
-            loss_fct = CrossEntropyLoss()
-            lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            lm_loss = self.loss_function(
+                prediction_scores,
+                labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
 
         if not return_dict:
             output = (prediction_scores,) + outputs[2:]

src/transformers/models/biogpt/modeling_biogpt.py

@@ -588,6 +588,7 @@ def forward(
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,  # NOOP kwargs, for now
     ) -> Union[Tuple, BaseModelOutputWithPastAndCrossAttentions]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -757,6 +758,7 @@ def forward(
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -783,11 +785,12 @@ def forward(
 
         lm_loss = None
         if labels is not None:
-            # we are doing next-token prediction; shift prediction scores and input ids by one
-            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
-            labels = labels[:, 1:].contiguous()
-            loss_fct = CrossEntropyLoss()
-            lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            lm_loss = self.loss_function(
+                prediction_scores,
+                labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
 
         if not return_dict:
             output = (prediction_scores,) + outputs[1:]

src/transformers/models/bloom/modeling_bloom.py

@@ -967,6 +967,8 @@ def forward(
             `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
             are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
         """
+        # Bloom has deprecated kwargs, so we need to pop num_items_in_batch explicitly
+        num_items_in_batch = deprecated_arguments.pop("num_items_in_batch", None)
         if deprecated_arguments.pop("position_ids", False) is not False:
             # `position_ids` could have been `torch.Tensor` or `None` so defaulting pop to `False` allows to detect if users were passing explicitly `None`
             warnings.warn(
@@ -999,14 +1001,12 @@ def forward(
         if labels is not None:
             # move labels to correct device to enable model parallelism
             labels = labels.to(lm_logits.device)
-            # Shift so that tokens < n predict n
-            shift_logits = lm_logits[..., :-1, :].contiguous()
-            shift_labels = labels[..., 1:].contiguous()
-            batch_size, seq_length, vocab_size = shift_logits.shape
             # Flatten the tokens
-            loss_fct = CrossEntropyLoss()
-            loss = loss_fct(
-                shift_logits.view(batch_size * seq_length, vocab_size), shift_labels.view(batch_size * seq_length)
+            loss = self.loss_function(
+                lm_logits,
+                labels,
+                vocab_size=self.config.vocab_size,
+                num_items_in_batch=num_items_in_batch,
             )
 
         if not return_dict:

src/transformers/models/camembert/modeling_camembert.py

@@ -1584,6 +1584,7 @@ def forward(
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
         r"""
         encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
@@ -1655,11 +1656,12 @@ def forward(
         if labels is not None:
             # move labels to correct device to enable model parallelism
             labels = labels.to(prediction_scores.device)
-            # we are doing next-token prediction; shift prediction scores and input ids by one
-            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
-            labels = labels[:, 1:].contiguous()
-            loss_fct = CrossEntropyLoss()
-            lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            lm_loss = self.loss_function(
+                prediction_scores,
+                labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
 
         if not return_dict:
             output = (prediction_scores,) + outputs[2:]

src/transformers/models/codegen/modeling_codegen.py

@@ -19,7 +19,6 @@
 import torch
 import torch.utils.checkpoint
 from torch import nn
-from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
@@ -450,6 +449,7 @@ def forward(
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,  # NOOP kwargs, for now
     ) -> Union[Tuple, BaseModelOutputWithPast]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -741,6 +741,7 @@ def forward(
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,
     ) -> Union[Tuple, CausalLMOutputWithPast]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -775,12 +776,13 @@ def forward(
         if labels is not None:
             # move labels to correct device to enable model parallelism
             labels = labels.to(lm_logits.device)
-            # Shift so that tokens < n predict n
-            shift_logits = lm_logits[..., :-1, :].contiguous()
-            shift_labels = labels[..., 1:].contiguous()
             # Flatten the tokens
-            loss_fct = CrossEntropyLoss()
-            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+            loss = self.loss_function(
+                lm_logits,
+                labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
 
             loss = loss.to(hidden_states.dtype)
 

src/transformers/models/ctrl/modeling_ctrl.py

@@ -360,6 +360,7 @@ def forward(
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,  # NOOP kwargs, for now
     ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPast]:
         r"""
         Returns:
@@ -537,6 +538,7 @@ def forward(
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithPast]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -593,12 +595,12 @@ def forward(
 
         loss = None
         if labels is not None:
-            # Shift so that tokens < n predict n
-            shift_logits = lm_logits[..., :-1, :].contiguous()
-            shift_labels = labels[..., 1:].contiguous()
-            # Flatten the tokens
-            loss_fct = CrossEntropyLoss()
-            loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+            loss = self.loss_function(
+                lm_logits,
+                labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
 
         if not return_dict:
             output = (lm_logits,) + transformer_outputs[1:]

src/transformers/models/data2vec/modeling_data2vec_text.py

@@ -906,6 +906,7 @@ def forward(
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
         r"""
         encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
@@ -975,13 +976,12 @@ def forward(
 
         lm_loss = None
         if labels is not None:
-            # we are doing next-token prediction; shift prediction scores and input ids by one
-            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
-            labels = labels[:, 1:].contiguous()
-            loss_fct = CrossEntropyLoss()
-
-            labels = labels.to(shifted_prediction_scores.device)
-            lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            lm_loss = self.loss_function(
+                prediction_scores,
+                labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
 
         if not return_dict:
             output = (prediction_scores,) + outputs[2:]

src/transformers/models/dbrx/modeling_dbrx.py

@@ -980,6 +980,7 @@ def forward(
         output_router_logits: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,  # NOOP kwargs, for now
     ) -> Union[Tuple, MoeModelOutputWithPast]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -1278,6 +1279,7 @@ def forward(
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.LongTensor] = None,
         logits_to_keep: Union[int, torch.Tensor] = 0,
+        **kwargs,
     ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
         r"""Forward function for causal language modeling.
 
@@ -1344,16 +1346,12 @@ def forward(
 
         loss = None
         if labels is not None:
-            # Shift so that tokens < n predict n
-            shift_logits = logits[..., :-1, :].contiguous()
-            shift_labels = labels[..., 1:].contiguous()
-            # Flatten the tokens
-            loss_fct = nn.CrossEntropyLoss()
-            shift_logits = shift_logits.view(-1, self.config.vocab_size)
-            shift_labels = shift_labels.view(-1)
-            # Enable model parallelism
-            shift_labels = shift_labels.to(shift_logits.device)
-            loss = loss_fct(shift_logits, shift_labels)
+            loss = self.loss_function(
+                logits,
+                labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
 
         aux_loss = None
         if output_router_logits:

src/transformers/models/electra/modeling_electra.py

@@ -1564,6 +1564,7 @@ def forward(
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
         r"""
         encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
@@ -1633,11 +1634,12 @@ def forward(
 
         lm_loss = None
         if labels is not None:
-            # we are doing next-token prediction; shift prediction scores and input ids by one
-            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
-            labels = labels[:, 1:].contiguous()
-            loss_fct = CrossEntropyLoss()
-            lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+            lm_loss = self.loss_function(
+                prediction_scores,
+                labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
 
         if not return_dict:
             output = (prediction_scores,) + outputs[1:]