diff --git a/src/transformers/generation/utils.py b/src/transformers/generation/utils.py
index 0d2baea6d85f..9cb5cc1558ac 100644
--- a/src/transformers/generation/utils.py
+++ b/src/transformers/generation/utils.py
@@ -40,13 +40,6 @@
 )
 from ..integrations.deepspeed import is_deepspeed_zero3_enabled
 from ..modeling_outputs import CausalLMOutputWithPast, Seq2SeqLMOutput
-from ..models.auto import (
-    MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING,
-    MODEL_FOR_CAUSAL_LM_MAPPING,
-    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
-    MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING,
-    MODEL_FOR_VISION_2_SEQ_MAPPING,
-)
 from ..pytorch_utils import isin_mps_friendly
 from ..tokenization_utils import ExtensionsTrie
 from ..utils import (
@@ -1113,32 +1106,6 @@ def compute_transition_scores(
 
         return transition_scores
 
-    def _validate_model_class(self):
-        """
-        Confirms that the model class is compatible with generation. If not, raises an exception that points to the
-        right class to use.
-        """
-        if not is_torchdynamo_compiling() and not self.can_generate():
-            generate_compatible_mappings = [
-                MODEL_FOR_CAUSAL_LM_MAPPING,
-                MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING,
-                MODEL_FOR_VISION_2_SEQ_MAPPING,
-                MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
-                MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING,
-            ]
-            generate_compatible_classes = set()
-            for model_mapping in generate_compatible_mappings:
-                supported_models = model_mapping.get(type(self.config), default=None)
-                if supported_models is not None:
-                    generate_compatible_classes.add(supported_models.__name__)
-            exception_message = (
-                f"The current model class ({self.__class__.__name__}) is not compatible with `.generate()`, as "
-                "it doesn't have a language model head."
-            )
-            if generate_compatible_classes:
-                exception_message += f" Please use one of the following classes instead: {generate_compatible_classes}"
-            raise TypeError(exception_message)
-
     def _validate_assistant(self, assistant_model):
         if assistant_model is None:
             return
@@ -1777,7 +1744,6 @@ def generate(
                     - [`~generation.GenerateBeamEncoderDecoderOutput`]
         """
         # 1. Handle `generation_config` and kwargs that might update it, and validate the `.generate()` call
-        self._validate_model_class()
         tokenizer = kwargs.pop("tokenizer", None)  # Pull this out first, we only use it for stopping criteria
         generation_config, model_kwargs = self._prepare_generation_config(generation_config, **kwargs)
         self._validate_model_kwargs(model_kwargs.copy())
diff --git a/src/transformers/modeling_utils.py b/src/transformers/modeling_utils.py
index b943b5e7989f..b5353995c017 100755
--- a/src/transformers/modeling_utils.py
+++ b/src/transformers/modeling_utils.py
@@ -212,7 +212,7 @@ def _skip_init(*args, **kwargs):
                 setattr(torch.nn.init, name, init_func)
 
 
-def get_parameter_device(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
+def get_parameter_device(parameter: Union[nn.Module, "ModuleUtilsMixin"]):
     try:
         return next(parameter.parameters()).device
     except StopIteration:
@@ -227,7 +227,7 @@ def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]:
         return first_tuple[1].device
 
 
-def get_first_parameter_dtype(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
+def get_first_parameter_dtype(parameter: Union[nn.Module, "ModuleUtilsMixin"]):
     """
     Returns the first parameter dtype (can be non-floating) or asserts if none were found.
     """
@@ -245,7 +245,7 @@ def find_tensor_attributes(module: nn.Module) -> List[Tuple[str, Tensor]]:
         return first_tuple[1].dtype
 
 
-def get_parameter_dtype(parameter: Union[nn.Module, GenerationMixin, "ModuleUtilsMixin"]):
+def get_parameter_dtype(parameter: Union[nn.Module, "ModuleUtilsMixin"]):
     """
     Returns the first found floating dtype in parameters if there is one, otherwise returns the last dtype it found.
     """
@@ -1295,7 +1295,7 @@ def floating_point_ops(
         return 6 * self.estimate_tokens(input_dict) * self.num_parameters(exclude_embeddings=exclude_embeddings)
 
 
-class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMixin, PeftAdapterMixin):
+class PreTrainedModel(nn.Module, ModuleUtilsMixin, PushToHubMixin, PeftAdapterMixin):
     r"""
     Base class for all models.
 
@@ -1624,11 +1624,7 @@ def can_generate(cls) -> bool:
         Returns:
             `bool`: Whether this model can generate sequences with `.generate()`.
         """
-        # Detects whether `prepare_inputs_for_generation` has been overwritten, which is a requirement for generation.
-        # Alternativelly, the model can also have a custom `generate` function.
-        if "GenerationMixin" in str(cls.prepare_inputs_for_generation) and "GenerationMixin" in str(cls.generate):
-            return False
-        return True
+        return issubclass(cls, GenerationMixin)
 
     @classmethod
     def _check_and_enable_flash_attn_2(
@@ -4716,6 +4712,36 @@ def _is_quantized_training_enabled(self):
 
         return self.hf_quantizer.is_trainable
 
+    def generate(self, *args, **kwargs):
+        raise NotImplementedError(
+            f"{self.__class__.__name__} doesn't have a `generate` method. If you were not "
+            "expecting this exception, here are some things to check:"
+            "\n  1. If you are using a custom model (or `trust_remote_code=True`): Load your model with an "
+            "`AutoModel...` class. If that is not possible, make sure your model class inherits from "
+            "`GenerationMixin` as its first inherited class. `PreTrainedModel` no longer inherits "
+            "`GenerationMixin` as of v4.45"
+            "\n  2. If you are using a transformers model: You might be using a class that is not meant for "
+            "generation. Your model must be compatible with one of the following auto classes: "
+            "`AutoModelForCausalLM`, `ForConditionalGeneration`, `AutoModelForSpeechSeq2Seq`, and "
+            "`AutoModelForVision2Seq`"
+            "\n  3. If none of the cases above apply, please open an issue on GitHub 🤗"
+        )
+
+    def prepare_inputs_for_generation(self, *args, **kwargs):
+        raise NotImplementedError(
+            f"{self.__class__.__name__} doesn't have a `prepare_inputs_for_generation` method. If you were not "
+            "expecting this exception, here are some things to check:"
+            "\n  1. If you are using a custom model (or `trust_remote_code=True`): Load your model with an "
+            "`AutoModel...` class. If that is not possible, make sure your model class inherits from "
+            "`GenerationMixin` as its first inherited class. `PreTrainedModel` no longer inherits "
+            "`GenerationMixin` as of v4.45"
+            "\n  2. If you are using a transformers model: You might be using a class that is not meant for "
+            "generation. Your model must be compatible with one of the following auto classes: "
+            "`AutoModelForCausalLM`, `ForConditionalGeneration`, `AutoModelForSpeechSeq2Seq`, and "
+            "`AutoModelForVision2Seq`"
+            "\n  3. If none of the cases above apply, please open an issue on GitHub 🤗"
+        )
+
 
 PreTrainedModel.push_to_hub = copy_func(PreTrainedModel.push_to_hub)
 if PreTrainedModel.push_to_hub.__doc__ is not None:
diff --git a/src/transformers/models/auto/auto_factory.py b/src/transformers/models/auto/auto_factory.py
index 6b572b252779..5e2b15bd434c 100644
--- a/src/transformers/models/auto/auto_factory.py
+++ b/src/transformers/models/auto/auto_factory.py
@@ -30,12 +30,17 @@
     extract_commit_hash,
     find_adapter_config_file,
     is_peft_available,
+    is_torch_available,
     logging,
     requires_backends,
 )
 from .configuration_auto import AutoConfig, model_type_to_module_name, replace_list_option_in_docstrings
 
 
+if is_torch_available():
+    from ...generation import GenerationMixin
+
+
 logger = logging.get_logger(__name__)
 
 
@@ -432,6 +437,7 @@ def from_config(cls, config, **kwargs):
             else:
                 cls.register(config.__class__, model_class, exist_ok=True)
             _ = kwargs.pop("code_revision", None)
+            model_class = add_generation_mixin_to_remote_model(model_class)
             return model_class._from_config(config, **kwargs)
         elif type(config) in cls._model_mapping.keys():
             model_class = _get_model_class(config, cls._model_mapping)
@@ -556,6 +562,7 @@ def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
                 model_class.register_for_auto_class(cls.__name__)
             else:
                 cls.register(config.__class__, model_class, exist_ok=True)
+            model_class = add_generation_mixin_to_remote_model(model_class)
             return model_class.from_pretrained(
                 pretrained_model_name_or_path, *model_args, config=config, **hub_kwargs, **kwargs
             )
@@ -705,6 +712,36 @@ def getattribute_from_module(module, attr):
         raise ValueError(f"Could not find {attr} in {transformers_module}!")
 
 
+def add_generation_mixin_to_remote_model(model_class):
+    """
+    Adds `GenerationMixin` to the inheritance of `model_class`, if `model_class` is a PyTorch model. This function is
+    used for backwards compatibility purposes: prior to v4.45, `PreTrainedModel` inherited from `GenerationMixin`.
+    Without this function, older models dynamically loaded from the Hub may not have the `generate` method.
+    """
+    # 1. If it is not a PT model (i.e. doesn't inherit Module), do nothing
+    if "torch.nn.modules.module.Module" not in str(model_class.__mro__):
+        return model_class
+
+    # 2. If it already inherits from GenerationMixin, do nothing
+    if issubclass(model_class, GenerationMixin):
+        return model_class
+
+    # 3. If the class name has a suffix that indicates that it should be able to generate, add `GenerationMixin` to
+    # the class inheritance. Otherwise, do nothing.
+    terminations_with_generation_support = [
+        "ForCausalLM",
+        "ForConditionalGeneration",
+        "ForSpeechSeq2Seq",
+        "ForVision2Seq",
+    ]
+    if any(model_class.__name__.endswith(termination) for termination in terminations_with_generation_support):
+        model_class_with_generation_mixin = type(
+            model_class.__name__, (GenerationMixin, model_class), {**model_class.__dict__}
+        )
+        return model_class_with_generation_mixin
+    return model_class
+
+
 class _LazyAutoMapping(OrderedDict):
     """
     " A mapping config to object (model or tokenizer for instance) that will load keys and values when it is accessed.
diff --git a/src/transformers/models/bark/modeling_bark.py b/src/transformers/models/bark/modeling_bark.py
index ac67ef4b37e4..658272048494 100644
--- a/src/transformers/models/bark/modeling_bark.py
+++ b/src/transformers/models/bark/modeling_bark.py
@@ -22,6 +22,7 @@
 from torch import nn
 from torch.nn import functional as F
 
+from ...generation import GenerationMixin
 from ...generation.logits_process import (
     AlternatingCodebooksLogitsProcessor,
     BarkEosPrioritizerLogitsProcessor,
@@ -546,7 +547,7 @@ def device(self) -> torch.device:
 
 
 # GPT2-like autoregressive model
-class BarkCausalModel(BarkPreTrainedModel):
+class BarkCausalModel(GenerationMixin, BarkPreTrainedModel):
     config_class = BarkSubModelConfig
 
     def __init__(self, config):
@@ -1136,7 +1137,7 @@ def generate(
     language modeling heads, one for each codebook.""",
     BARK_MODEL_START_DOCSTRING.format(config="BarkFineConfig"),
 )
-class BarkFineModel(BarkPreTrainedModel):
+class BarkFineModel(GenerationMixin, BarkPreTrainedModel):
     base_model_prefix = "fine_acoustics"
     config_class = BarkFineConfig
     main_input_name = "codebook_idx"
@@ -1536,7 +1537,7 @@ def generate(
     """,
     BARK_START_DOCSTRING,
 )
-class BarkModel(BarkPreTrainedModel):
+class BarkModel(GenerationMixin, BarkPreTrainedModel):
     config_class = BarkConfig
 
     def __init__(self, config):
diff --git a/src/transformers/models/bart/modeling_bart.py b/src/transformers/models/bart/modeling_bart.py
index fa928d05caa8..806e5bb4c522 100755
--- a/src/transformers/models/bart/modeling_bart.py
+++ b/src/transformers/models/bart/modeling_bart.py
@@ -25,6 +25,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import (
     _prepare_4d_attention_mask,
     _prepare_4d_attention_mask_for_sdpa,
@@ -1557,7 +1558,7 @@ def forward(
 @add_start_docstrings(
     "The BART Model with a language modeling head. Can be used for summarization.", BART_START_DOCSTRING
 )
-class BartForConditionalGeneration(BartPreTrainedModel):
+class BartForConditionalGeneration(GenerationMixin, BartPreTrainedModel):
     base_model_prefix = "model"
     _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
     _keys_to_ignore_on_load_missing = ["final_logits_bias"]
@@ -2010,7 +2011,7 @@ def forward(self, *args, **kwargs):
     """,
     BART_START_DOCSTRING,
 )
-class BartForCausalLM(BartPreTrainedModel):
+class BartForCausalLM(GenerationMixin, BartPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/bert/modeling_bert.py b/src/transformers/models/bert/modeling_bert.py
index 850e93ca59fb..ddbc9beb2cd7 100755
--- a/src/transformers/models/bert/modeling_bert.py
+++ b/src/transformers/models/bert/modeling_bert.py
@@ -28,6 +28,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import (
     _prepare_4d_attention_mask_for_sdpa,
     _prepare_4d_causal_attention_mask_for_sdpa,
@@ -1280,7 +1281,7 @@ def forward(
 @add_start_docstrings(
     """Bert Model with a `language modeling` head on top for CLM fine-tuning.""", BERT_START_DOCSTRING
 )
-class BertLMHeadModel(BertPreTrainedModel):
+class BertLMHeadModel(GenerationMixin, BertPreTrainedModel):
     _tied_weights_keys = ["cls.predictions.decoder.bias", "cls.predictions.decoder.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/bert_generation/modeling_bert_generation.py b/src/transformers/models/bert_generation/modeling_bert_generation.py
index a5fb3d053115..3359a0de0f95 100755
--- a/src/transformers/models/bert_generation/modeling_bert_generation.py
+++ b/src/transformers/models/bert_generation/modeling_bert_generation.py
@@ -23,6 +23,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import BaseModelOutputWithPastAndCrossAttentions, CausalLMOutputWithCrossAttentions
 from ...modeling_utils import PreTrainedModel
 from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
@@ -863,7 +864,7 @@ def _tie_weights(self):
     """BertGeneration Model with a `language modeling` head on top for CLM fine-tuning.""",
     BERT_GENERATION_START_DOCSTRING,
 )
-class BertGenerationDecoder(BertGenerationPreTrainedModel):
+class BertGenerationDecoder(GenerationMixin, BertGenerationPreTrainedModel):
     _tied_weights_keys = ["lm_head.decoder.weight", "lm_head.decoder.bias"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/big_bird/modeling_big_bird.py b/src/transformers/models/big_bird/modeling_big_bird.py
index a6b1660d5ae1..5852f96230cf 100755
--- a/src/transformers/models/big_bird/modeling_big_bird.py
+++ b/src/transformers/models/big_bird/modeling_big_bird.py
@@ -26,6 +26,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
     BaseModelOutputWithPoolingAndCrossAttentions,
@@ -2495,7 +2496,7 @@ def prepare_inputs_for_generation(self, input_ids, attention_mask=None, **model_
 @add_start_docstrings(
     """BigBird Model with a `language modeling` head on top for CLM fine-tuning.""", BIG_BIRD_START_DOCSTRING
 )
-class BigBirdForCausalLM(BigBirdPreTrainedModel):
+class BigBirdForCausalLM(GenerationMixin, BigBirdPreTrainedModel):
     _tied_weights_keys = ["cls.predictions.decoder.weight", "cls.predictions.decoder.bias"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py b/src/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py
index 9f8e3cd19cd8..a2c58d8b5a0e 100755
--- a/src/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py
+++ b/src/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py
@@ -24,6 +24,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -2436,7 +2437,7 @@ def forward(
     BIGBIRD_PEGASUS_START_DOCSTRING,
 )
 # Copied from transformers.models.bart.modeling_bart.BartForConditionalGeneration with Bart->BigBirdPegasus, BART->BIGBIRD_PEGASUS
-class BigBirdPegasusForConditionalGeneration(BigBirdPegasusPreTrainedModel):
+class BigBirdPegasusForConditionalGeneration(GenerationMixin, BigBirdPegasusPreTrainedModel):
     base_model_prefix = "model"
     _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
     _keys_to_ignore_on_load_missing = ["final_logits_bias"]
@@ -2882,7 +2883,7 @@ def forward(self, *args, **kwargs):
         return self.decoder(*args, **kwargs)
 
 
-class BigBirdPegasusForCausalLM(BigBirdPegasusPreTrainedModel):
+class BigBirdPegasusForCausalLM(GenerationMixin, BigBirdPegasusPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/biogpt/modeling_biogpt.py b/src/transformers/models/biogpt/modeling_biogpt.py
index 020f52833d5b..b09909337a68 100755
--- a/src/transformers/models/biogpt/modeling_biogpt.py
+++ b/src/transformers/models/biogpt/modeling_biogpt.py
@@ -23,6 +23,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
@@ -596,7 +597,7 @@ def forward(
 @add_start_docstrings(
     """BioGPT Model with a `language modeling` head on top for CLM fine-tuning.""", BIOGPT_START_DOCSTRING
 )
-class BioGptForCausalLM(BioGptPreTrainedModel):
+class BioGptForCausalLM(GenerationMixin, BioGptPreTrainedModel):
     _tied_weights_keys = ["output_projection.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/blenderbot/modeling_blenderbot.py b/src/transformers/models/blenderbot/modeling_blenderbot.py
index 12d259fde71e..7e42a88d4ddd 100755
--- a/src/transformers/models/blenderbot/modeling_blenderbot.py
+++ b/src/transformers/models/blenderbot/modeling_blenderbot.py
@@ -26,6 +26,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -1196,7 +1197,7 @@ def forward(
 @add_start_docstrings(
     "The Blenderbot Model with a language modeling head. Can be used for summarization.", BLENDERBOT_START_DOCSTRING
 )
-class BlenderbotForConditionalGeneration(BlenderbotPreTrainedModel):
+class BlenderbotForConditionalGeneration(GenerationMixin, BlenderbotPreTrainedModel):
     base_model_prefix = "model"
     _keys_to_ignore_on_load_missing = ["final_logits_bias"]
     _tied_weights_keys = ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight", "lm_head.weight"]
@@ -1397,7 +1398,7 @@ def forward(self, *args, **kwargs):
 
 
 # Copied from transformers.models.bart.modeling_bart.BartForCausalLM with Bart->Blenderbot, facebook/bart-base->facebook/blenderbot-400M-distill
-class BlenderbotForCausalLM(BlenderbotPreTrainedModel):
+class BlenderbotForCausalLM(GenerationMixin, BlenderbotPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/blenderbot_small/modeling_blenderbot_small.py b/src/transformers/models/blenderbot_small/modeling_blenderbot_small.py
index aa0e38bd8e91..94ed63dea616 100755
--- a/src/transformers/models/blenderbot_small/modeling_blenderbot_small.py
+++ b/src/transformers/models/blenderbot_small/modeling_blenderbot_small.py
@@ -24,6 +24,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -1163,7 +1164,7 @@ def forward(
     "The BlenderbotSmall Model with a language modeling head. Can be used for summarization.",
     BLENDERBOT_SMALL_START_DOCSTRING,
 )
-class BlenderbotSmallForConditionalGeneration(BlenderbotSmallPreTrainedModel):
+class BlenderbotSmallForConditionalGeneration(GenerationMixin, BlenderbotSmallPreTrainedModel):
     base_model_prefix = "model"
     _keys_to_ignore_on_load_missing = ["final_logits_bias"]
     _tied_weights_keys = ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight", "lm_head.weight"]
@@ -1349,7 +1350,7 @@ def forward(self, *args, **kwargs):
 
 
 # Copied from transformers.models.bart.modeling_bart.BartForCausalLM with Bart->BlenderbotSmall, facebook/bart-base->facebook/blenderbot_small-90M
-class BlenderbotSmallForCausalLM(BlenderbotSmallPreTrainedModel):
+class BlenderbotSmallForCausalLM(GenerationMixin, BlenderbotSmallPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/blip/modeling_blip.py b/src/transformers/models/blip/modeling_blip.py
index 46e3a6005b0a..38e40a5b8c4d 100644
--- a/src/transformers/models/blip/modeling_blip.py
+++ b/src/transformers/models/blip/modeling_blip.py
@@ -25,6 +25,7 @@
 from torch.nn.functional import normalize
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
 from ...modeling_utils import PreTrainedModel
 from ...utils import (
@@ -1027,7 +1028,7 @@ def forward(
     """,
     BLIP_START_DOCSTRING,
 )
-class BlipForConditionalGeneration(BlipPreTrainedModel):
+class BlipForConditionalGeneration(GenerationMixin, BlipPreTrainedModel):
     config_class = BlipConfig
     _tied_weights_keys = ["text_decoder.cls.predictions.decoder.bias"]
     main_input_name = "pixel_values"
diff --git a/src/transformers/models/blip/modeling_blip_text.py b/src/transformers/models/blip/modeling_blip_text.py
index a800ba89825d..2363f6f582eb 100644
--- a/src/transformers/models/blip/modeling_blip_text.py
+++ b/src/transformers/models/blip/modeling_blip_text.py
@@ -23,6 +23,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
     BaseModelOutputWithPoolingAndCrossAttentions,
@@ -808,7 +809,7 @@ def forward(
 
 
 # Adapted from https://github.com/salesforce/BLIP/blob/main/models/med.py#L811
-class BlipTextLMHeadModel(BlipTextPreTrainedModel):
+class BlipTextLMHeadModel(GenerationMixin, BlipTextPreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
 
diff --git a/src/transformers/models/blip_2/modeling_blip_2.py b/src/transformers/models/blip_2/modeling_blip_2.py
index fba4c98696a0..8ff06604cff9 100644
--- a/src/transformers/models/blip_2/modeling_blip_2.py
+++ b/src/transformers/models/blip_2/modeling_blip_2.py
@@ -24,6 +24,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutput,
     BaseModelOutputWithPastAndCrossAttentions,
@@ -1997,7 +1998,7 @@ def forward(
     """,
     BLIP_2_START_DOCSTRING,
 )
-class Blip2ForConditionalGeneration(Blip2PreTrainedModel):
+class Blip2ForConditionalGeneration(GenerationMixin, Blip2PreTrainedModel):
     config_class = Blip2Config
     main_input_name = "pixel_values"
 
diff --git a/src/transformers/models/bloom/modeling_bloom.py b/src/transformers/models/bloom/modeling_bloom.py
index 70e748343561..5f6e60c25bcf 100644
--- a/src/transformers/models/bloom/modeling_bloom.py
+++ b/src/transformers/models/bloom/modeling_bloom.py
@@ -26,6 +26,7 @@
 
 from ...cache_utils import Cache, DynamicCache, StaticCache
 from ...file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
@@ -856,7 +857,7 @@ def _update_causal_mask(
     """,
     BLOOM_START_DOCSTRING,
 )
-class BloomForCausalLM(BloomPreTrainedModel):
+class BloomForCausalLM(GenerationMixin, BloomPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config: BloomConfig):
diff --git a/src/transformers/models/camembert/modeling_camembert.py b/src/transformers/models/camembert/modeling_camembert.py
index 03a60a2856be..72f62ea0f9a4 100644
--- a/src/transformers/models/camembert/modeling_camembert.py
+++ b/src/transformers/models/camembert/modeling_camembert.py
@@ -25,6 +25,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN, gelu
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import (
     _prepare_4d_attention_mask_for_sdpa,
     _prepare_4d_causal_attention_mask_for_sdpa,
@@ -1544,7 +1545,7 @@ def forward(
     """CamemBERT Model with a `language modeling` head on top for CLM fine-tuning.""", CAMEMBERT_START_DOCSTRING
 )
 # Copied from transformers.models.roberta.modeling_roberta.RobertaForCausalLM with Roberta->Camembert, ROBERTA->CAMEMBERT, FacebookAI/roberta-base->almanach/camembert-base
-class CamembertForCausalLM(CamembertPreTrainedModel):
+class CamembertForCausalLM(GenerationMixin, CamembertPreTrainedModel):
     _tied_weights_keys = ["lm_head.decoder.weight", "lm_head.decoder.bias"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/chameleon/modeling_chameleon.py b/src/transformers/models/chameleon/modeling_chameleon.py
index 23334311ca95..43fe52bd94c0 100644
--- a/src/transformers/models/chameleon/modeling_chameleon.py
+++ b/src/transformers/models/chameleon/modeling_chameleon.py
@@ -26,6 +26,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_flash_attention_utils import _flash_attention_forward
 from ...modeling_outputs import (
@@ -1496,7 +1497,7 @@ def _update_causal_mask(
     "Chameleon Model with a head on top used for outputting logits for next token prediction.",
     CHAMELEON_START_DOCSTRING,
 )
-class ChameleonForConditionalGeneration(ChameleonPreTrainedModel):
+class ChameleonForConditionalGeneration(GenerationMixin, ChameleonPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/clvp/modeling_clvp.py b/src/transformers/models/clvp/modeling_clvp.py
index b6d025a0b8e2..81ac5d469397 100644
--- a/src/transformers/models/clvp/modeling_clvp.py
+++ b/src/transformers/models/clvp/modeling_clvp.py
@@ -26,7 +26,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
-from ...generation import GenerationConfig
+from ...generation import GenerationConfig, GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -1278,7 +1278,7 @@ def forward(
     "The CLVP decoder model with a language modelling head on top.",
     CLVP_START_DOCSTRING,
 )
-class ClvpForCausalLM(ClvpPreTrainedModel):
+class ClvpForCausalLM(GenerationMixin, ClvpPreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
 
@@ -1509,7 +1509,7 @@ def _reorder_cache(
     "together to filter out the best speech_ids.",
     CLVP_START_DOCSTRING,
 )
-class ClvpModelForConditionalGeneration(ClvpPreTrainedModel):
+class ClvpModelForConditionalGeneration(GenerationMixin, ClvpPreTrainedModel):
     config_class = ClvpConfig
 
     def __init__(self, config: ClvpConfig):
diff --git a/src/transformers/models/codegen/modeling_codegen.py b/src/transformers/models/codegen/modeling_codegen.py
index bfa591f7bdaf..7668ecda0083 100644
--- a/src/transformers/models/codegen/modeling_codegen.py
+++ b/src/transformers/models/codegen/modeling_codegen.py
@@ -23,6 +23,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 from ...modeling_utils import PreTrainedModel
@@ -697,7 +698,7 @@ def _update_causal_mask(
     """,
     CODEGEN_START_DOCSTRING,
 )
-class CodeGenForCausalLM(CodeGenPreTrainedModel):
+class CodeGenForCausalLM(GenerationMixin, CodeGenPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/cohere/modeling_cohere.py b/src/transformers/models/cohere/modeling_cohere.py
index 6912a4596370..c7aa0c421f49 100644
--- a/src/transformers/models/cohere/modeling_cohere.py
+++ b/src/transformers/models/cohere/modeling_cohere.py
@@ -32,6 +32,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
@@ -972,7 +973,7 @@ def _update_causal_mask(
 
 
 # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM with Llama->Cohere
-class CohereForCausalLM(CoherePreTrainedModel):
+class CohereForCausalLM(GenerationMixin, CoherePreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     # Ignore copy
diff --git a/src/transformers/models/cpmant/modeling_cpmant.py b/src/transformers/models/cpmant/modeling_cpmant.py
index c8a313505251..903a851d11fb 100755
--- a/src/transformers/models/cpmant/modeling_cpmant.py
+++ b/src/transformers/models/cpmant/modeling_cpmant.py
@@ -24,6 +24,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 from ...modeling_utils import PreTrainedModel
 from ...utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
@@ -736,7 +737,7 @@ def forward(
     """,
     CPMANT_START_DOCSTRING,
 )
-class CpmAntForCausalLM(CpmAntPreTrainedModel):
+class CpmAntForCausalLM(GenerationMixin, CpmAntPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config: CpmAntConfig):
diff --git a/src/transformers/models/ctrl/modeling_ctrl.py b/src/transformers/models/ctrl/modeling_ctrl.py
index bbf3b10a62ec..7aadf5819fa9 100644
--- a/src/transformers/models/ctrl/modeling_ctrl.py
+++ b/src/transformers/models/ctrl/modeling_ctrl.py
@@ -22,6 +22,7 @@
 from torch import nn
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
+from ...generation import GenerationMixin
 from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutput
 from ...modeling_utils import PreTrainedModel
 from ...pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_linear_layer
@@ -503,7 +504,7 @@ def forward(
     """,
     CTRL_START_DOCSTRING,
 )
-class CTRLLMHeadModel(CTRLPreTrainedModel):
+class CTRLLMHeadModel(GenerationMixin, CTRLPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/data2vec/modeling_data2vec_text.py b/src/transformers/models/data2vec/modeling_data2vec_text.py
index a41fdfb56ed1..19ec77d08c29 100644
--- a/src/transformers/models/data2vec/modeling_data2vec_text.py
+++ b/src/transformers/models/data2vec/modeling_data2vec_text.py
@@ -23,6 +23,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN, gelu
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
     BaseModelOutputWithPoolingAndCrossAttentions,
@@ -866,7 +867,7 @@ def forward(
 @add_start_docstrings(
     """Data2VecText Model with a `language modeling` head on top for CLM fine-tuning.""", DATA2VECTEXT_START_DOCSTRING
 )
-class Data2VecTextForCausalLM(Data2VecTextPreTrainedModel):
+class Data2VecTextForCausalLM(GenerationMixin, Data2VecTextPreTrainedModel):
     _tied_weights_keys = ["lm_head.decoder.weight", "lm_head.decoder.bias"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/dbrx/modeling_dbrx.py b/src/transformers/models/dbrx/modeling_dbrx.py
index 9684fd174733..482c0a0e90eb 100644
--- a/src/transformers/models/dbrx/modeling_dbrx.py
+++ b/src/transformers/models/dbrx/modeling_dbrx.py
@@ -23,6 +23,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast
 from ...modeling_utils import PreTrainedModel
@@ -1223,7 +1224,7 @@ def _update_causal_mask(
 
 
 @add_start_docstrings("The DBRX Model transformer for causal language modeling.", DBRX_START_DOCSTRING)
-class DbrxForCausalLM(DbrxPreTrainedModel):
+class DbrxForCausalLM(GenerationMixin, DbrxPreTrainedModel):
     def __init__(self, config: DbrxConfig):
         super().__init__(config)
         self.transformer = DbrxModel(config)
diff --git a/src/transformers/models/electra/modeling_electra.py b/src/transformers/models/electra/modeling_electra.py
index dd017170bef9..3682ad23ff36 100644
--- a/src/transformers/models/electra/modeling_electra.py
+++ b/src/transformers/models/electra/modeling_electra.py
@@ -25,6 +25,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN, get_activation
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutputWithCrossAttentions,
     BaseModelOutputWithPastAndCrossAttentions,
@@ -1524,7 +1525,7 @@ def forward(
 @add_start_docstrings(
     """ELECTRA Model with a `language modeling` head on top for CLM fine-tuning.""", ELECTRA_START_DOCSTRING
 )
-class ElectraForCausalLM(ElectraPreTrainedModel):
+class ElectraForCausalLM(GenerationMixin, ElectraPreTrainedModel):
     _tied_weights_keys = ["generator_lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/ernie/modeling_ernie.py b/src/transformers/models/ernie/modeling_ernie.py
index 6a0a26a5cbe5..11c0abd5fffe 100644
--- a/src/transformers/models/ernie/modeling_ernie.py
+++ b/src/transformers/models/ernie/modeling_ernie.py
@@ -25,6 +25,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
     BaseModelOutputWithPoolingAndCrossAttentions,
@@ -1081,7 +1082,7 @@ def forward(
 @add_start_docstrings(
     """Ernie Model with a `language modeling` head on top for CLM fine-tuning.""", ERNIE_START_DOCSTRING
 )
-class ErnieForCausalLM(ErniePreTrainedModel):
+class ErnieForCausalLM(GenerationMixin, ErniePreTrainedModel):
     _tied_weights_keys = ["cls.predictions.decoder.bias", "cls.predictions.decoder.weight"]
 
     # Copied from transformers.models.bert.modeling_bert.BertLMHeadModel.__init__ with BertLMHeadModel->ErnieForCausalLM,Bert->Ernie,bert->ernie
diff --git a/src/transformers/models/falcon/modeling_falcon.py b/src/transformers/models/falcon/modeling_falcon.py
index a9acd171c3ae..4c6f8df8c1b8 100644
--- a/src/transformers/models/falcon/modeling_falcon.py
+++ b/src/transformers/models/falcon/modeling_falcon.py
@@ -25,6 +25,7 @@
 
 from ...activations import get_activation
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import (
     AttentionMaskConverter,
 )
@@ -1220,7 +1221,7 @@ def _update_causal_mask(
     "The Falcon Model transformer with a language modeling head on top (linear layer with weights tied to the input embeddings).",
     FALCON_START_DOCSTRING,
 )
-class FalconForCausalLM(FalconPreTrainedModel):
+class FalconForCausalLM(GenerationMixin, FalconPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config: FalconConfig):
diff --git a/src/transformers/models/falcon_mamba/modeling_falcon_mamba.py b/src/transformers/models/falcon_mamba/modeling_falcon_mamba.py
index 07374fe1dfd7..05b8b6ff6117 100644
--- a/src/transformers/models/falcon_mamba/modeling_falcon_mamba.py
+++ b/src/transformers/models/falcon_mamba/modeling_falcon_mamba.py
@@ -25,6 +25,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import MambaCache
+from ...generation import GenerationMixin
 from ...modeling_utils import PreTrainedModel
 from ...utils import (
     ModelOutput,
@@ -702,7 +703,7 @@ def forward(
     FALCONMAMBA_START_DOCSTRING,
 )
 # Copied from transformers.models.mamba.modeling_mamba.MambaForCausalLM with MAMBA->FALCONMAMBA,Mamba->FalconMamba,mamba->falcon_mamba,FalconMambaCache->MambaCache
-class FalconMambaForCausalLM(FalconMambaPreTrainedModel):
+class FalconMambaForCausalLM(GenerationMixin, FalconMambaPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/flaubert/modeling_flaubert.py b/src/transformers/models/flaubert/modeling_flaubert.py
index 50c6f7ede222..1be2394cb3d7 100644
--- a/src/transformers/models/flaubert/modeling_flaubert.py
+++ b/src/transformers/models/flaubert/modeling_flaubert.py
@@ -25,6 +25,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import gelu
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutput,
     MaskedLMOutput,
@@ -644,7 +645,7 @@ def forward(
     FLAUBERT_START_DOCSTRING,
 )
 # Copied transformers.models.xlm.modeling_xlm.XLMWithLMHeadModel with XLM_INPUTS->FLAUBERT_INPUTS,XLM->Flaubert
-class FlaubertWithLMHeadModel(FlaubertPreTrainedModel):
+class FlaubertWithLMHeadModel(GenerationMixin, FlaubertPreTrainedModel):
     _tied_weights_keys = ["pred_layer.proj.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/fsmt/modeling_fsmt.py b/src/transformers/models/fsmt/modeling_fsmt.py
index 179408aba38e..6cf1ee9a842e 100644
--- a/src/transformers/models/fsmt/modeling_fsmt.py
+++ b/src/transformers/models/fsmt/modeling_fsmt.py
@@ -35,6 +35,7 @@
 from torch.nn import CrossEntropyLoss, LayerNorm
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...integrations.deepspeed import is_deepspeed_zero3_enabled
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -1173,7 +1174,7 @@ def set_output_embeddings(self, value):
 @add_start_docstrings(
     "The FSMT Model with a language modeling head. Can be used for summarization.", FSMT_START_DOCSTRING
 )
-class FSMTForConditionalGeneration(PretrainedFSMTModel):
+class FSMTForConditionalGeneration(GenerationMixin, PretrainedFSMTModel):
     base_model_prefix = "model"
     _tied_weights_keys = ["decoder.embed_tokens.weight", "decoder.output_projection.weight"]
 
diff --git a/src/transformers/models/fuyu/modeling_fuyu.py b/src/transformers/models/fuyu/modeling_fuyu.py
index b4b6330d0d86..82b7e6554f6a 100644
--- a/src/transformers/models/fuyu/modeling_fuyu.py
+++ b/src/transformers/models/fuyu/modeling_fuyu.py
@@ -20,6 +20,7 @@
 import torch.utils.checkpoint
 from torch import nn
 
+from ...generation import GenerationMixin
 from ...modeling_outputs import CausalLMOutputWithPast
 from ...modeling_utils import PreTrainedModel
 from ...models.auto.modeling_auto import AutoModelForCausalLM
@@ -145,7 +146,7 @@ def _init_weights(self, module):
     "Fuyu Model with a language modeling head on top for causal language model conditioned on image patches and text.",
     FUYU_START_DOCSTRING,
 )
-class FuyuForCausalLM(FuyuPreTrainedModel):
+class FuyuForCausalLM(GenerationMixin, FuyuPreTrainedModel):
     def __init__(self, config: FuyuConfig):
         super().__init__(config)
         self.padding_idx = config.pad_token_id
diff --git a/src/transformers/models/gemma/diff_gemma.py b/src/transformers/models/gemma/diff_gemma.py
index fcdb0f0b3d7d..6ec1aad286c2 100644
--- a/src/transformers/models/gemma/diff_gemma.py
+++ b/src/transformers/models/gemma/diff_gemma.py
@@ -34,6 +34,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_flash_attention_utils import _flash_attention_forward
 from ...modeling_outputs import CausalLMOutputWithPast
 from ...pytorch_utils import ALL_LAYERNORM_LAYERS
@@ -520,7 +521,7 @@ def forward(
 
 
 # Example where we ony modify the docstring and call super
-class GemmaForCausalLM(LlamaForCausalLM):
+class GemmaForCausalLM(GenerationMixin, LlamaForCausalLM):
     def forward(
         self,
         input_ids: torch.LongTensor = None,
diff --git a/src/transformers/models/gemma/modeling_gemma.py b/src/transformers/models/gemma/modeling_gemma.py
index 62917c73f332..1549bcf64f3d 100644
--- a/src/transformers/models/gemma/modeling_gemma.py
+++ b/src/transformers/models/gemma/modeling_gemma.py
@@ -29,6 +29,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_flash_attention_utils import _flash_attention_forward
 from ...modeling_outputs import (
@@ -989,7 +990,7 @@ def _update_causal_mask(
         return causal_mask
 
 
-class GemmaForCausalLM(GemmaPreTrainedModel):
+class GemmaForCausalLM(GenerationMixin, GemmaPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/gemma2/diff_gemma2.py b/src/transformers/models/gemma2/diff_gemma2.py
index 0e300c6337e2..c7c18415fed4 100644
--- a/src/transformers/models/gemma2/diff_gemma2.py
+++ b/src/transformers/models/gemma2/diff_gemma2.py
@@ -33,6 +33,7 @@
 )
 
 from ...cache_utils import Cache
+from ...generation import GenerationMixin
 from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 from ...utils import is_flash_attn_2_available, is_flash_attn_greater_or_equal_2_10, logging
 
@@ -476,7 +477,7 @@ def _update_causal_mask(
         return causal_mask
 
 
-class Gemma2ForCausalLM(GemmaForCausalLM):
+class Gemma2ForCausalLM(GenerationMixin, GemmaForCausalLM):
     def forward(
         self,
         input_ids: torch.LongTensor = None,
diff --git a/src/transformers/models/gemma2/modeling_gemma2.py b/src/transformers/models/gemma2/modeling_gemma2.py
index bf6ff76189d4..17b5133461d0 100644
--- a/src/transformers/models/gemma2/modeling_gemma2.py
+++ b/src/transformers/models/gemma2/modeling_gemma2.py
@@ -28,6 +28,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, HybridCache
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
     CausalLMOutputWithPast,
@@ -943,7 +944,7 @@ def _update_causal_mask(
         return causal_mask
 
 
-class Gemma2ForCausalLM(Gemma2PreTrainedModel):
+class Gemma2ForCausalLM(GenerationMixin, Gemma2PreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/git/modeling_git.py b/src/transformers/models/git/modeling_git.py
index 4807289c927c..e1510ffd513e 100644
--- a/src/transformers/models/git/modeling_git.py
+++ b/src/transformers/models/git/modeling_git.py
@@ -27,6 +27,7 @@
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...file_utils import ModelOutput
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -1318,7 +1319,7 @@ def forward(
 @add_start_docstrings(
     """GIT Model with a `language modeling` head on top for autoregressive language modeling.""", GIT_START_DOCSTRING
 )
-class GitForCausalLM(GitPreTrainedModel):
+class GitForCausalLM(GenerationMixin, GitPreTrainedModel):
     _tied_weights_keys = ["output.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/gpt2/modeling_gpt2.py b/src/transformers/models/gpt2/modeling_gpt2.py
index 8dfbfb906444..02ea192a8aa6 100644
--- a/src/transformers/models/gpt2/modeling_gpt2.py
+++ b/src/transformers/models/gpt2/modeling_gpt2.py
@@ -28,6 +28,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask_for_sdpa, _prepare_4d_causal_attention_mask_for_sdpa
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
@@ -1182,7 +1183,7 @@ def forward(
     """,
     GPT2_START_DOCSTRING,
 )
-class GPT2LMHeadModel(GPT2PreTrainedModel):
+class GPT2LMHeadModel(GenerationMixin, GPT2PreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
@@ -1384,7 +1385,7 @@ def _reorder_cache(
 """,
     GPT2_START_DOCSTRING,
 )
-class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
+class GPT2DoubleHeadsModel(GenerationMixin, GPT2PreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py b/src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
index 0f927a72469d..9caab1f3c47c 100644
--- a/src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
+++ b/src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
@@ -22,6 +22,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
@@ -1040,7 +1041,7 @@ def forward(
     """,
     GPT_BIGCODE_START_DOCSTRING,
 )
-class GPTBigCodeForCausalLM(GPTBigCodePreTrainedModel):
+class GPTBigCodeForCausalLM(GenerationMixin, GPTBigCodePreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/gpt_neo/modeling_gpt_neo.py b/src/transformers/models/gpt_neo/modeling_gpt_neo.py
index e59853677f83..ade982415958 100755
--- a/src/transformers/models/gpt_neo/modeling_gpt_neo.py
+++ b/src/transformers/models/gpt_neo/modeling_gpt_neo.py
@@ -24,6 +24,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
@@ -912,7 +913,7 @@ def _update_causal_mask(
     """,
     GPT_NEO_START_DOCSTRING,
 )
-class GPTNeoForCausalLM(GPTNeoPreTrainedModel):
+class GPTNeoForCausalLM(GenerationMixin, GPTNeoPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/gpt_neox/modeling_gpt_neox.py b/src/transformers/models/gpt_neox/modeling_gpt_neox.py
index 7be35c0d137d..13908a1a14fb 100755
--- a/src/transformers/models/gpt_neox/modeling_gpt_neox.py
+++ b/src/transformers/models/gpt_neox/modeling_gpt_neox.py
@@ -30,6 +30,7 @@
     add_start_docstrings_to_model_forward,
     replace_return_docstrings,
 )
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
@@ -1079,7 +1080,7 @@ def _update_causal_mask(
 @add_start_docstrings(
     """GPTNeoX Model with a `language modeling` head on top for CLM fine-tuning.""", GPT_NEOX_START_DOCSTRING
 )
-class GPTNeoXForCausalLM(GPTNeoXPreTrainedModel):
+class GPTNeoXForCausalLM(GenerationMixin, GPTNeoXPreTrainedModel):
     _tied_weights_keys = ["embed_out.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/gpt_neox_japanese/modeling_gpt_neox_japanese.py b/src/transformers/models/gpt_neox_japanese/modeling_gpt_neox_japanese.py
index b9c4cad0fdc5..137afe616d06 100755
--- a/src/transformers/models/gpt_neox_japanese/modeling_gpt_neox_japanese.py
+++ b/src/transformers/models/gpt_neox_japanese/modeling_gpt_neox_japanese.py
@@ -23,6 +23,7 @@
 
 from ...activations import ACT2FN
 from ...file_utils import add_start_docstrings, add_start_docstrings_to_model_forward, replace_return_docstrings
+from ...generation import GenerationMixin
 from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
 from ...modeling_utils import PreTrainedModel
 from ...utils import logging
@@ -589,7 +590,7 @@ def forward(
     """GPTNeoXJapanese Model with a `language modeling` head on top for Classifier Model fine-tuning.""",
     GPT_NEOX_JAPANESE_START_DOCSTRING,
 )
-class GPTNeoXJapaneseForCausalLM(GPTNeoXJapanesePreTrainedModel):
+class GPTNeoXJapaneseForCausalLM(GenerationMixin, GPTNeoXJapanesePreTrainedModel):
     _tied_weights_keys = ["embed_out.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/gptj/modeling_gptj.py b/src/transformers/models/gptj/modeling_gptj.py
index 1408bfe8a61d..aae8d0ea615e 100644
--- a/src/transformers/models/gptj/modeling_gptj.py
+++ b/src/transformers/models/gptj/modeling_gptj.py
@@ -25,6 +25,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
@@ -1006,7 +1007,7 @@ def _update_causal_mask(
     """,
     GPTJ_START_DOCSTRING,
 )
-class GPTJForCausalLM(GPTJPreTrainedModel):
+class GPTJForCausalLM(GenerationMixin, GPTJPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/idefics2/modeling_idefics2.py b/src/transformers/models/idefics2/modeling_idefics2.py
index 5d4f8e408eb0..c7c8e9f56d3f 100644
--- a/src/transformers/models/idefics2/modeling_idefics2.py
+++ b/src/transformers/models/idefics2/modeling_idefics2.py
@@ -23,11 +23,12 @@
 from torch import nn
 from torch.nn import CrossEntropyLoss
 
-from ... import PreTrainedModel
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask
 from ...modeling_outputs import BaseModelOutput, ModelOutput
+from ...modeling_utils import PreTrainedModel
 from ...utils import (
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
@@ -1438,7 +1439,7 @@ def forward(
     """The Idefics2 Model with a language modeling head. It is made up a SigLIP vision encoder, with a language modeling head on top. """,
     IDEFICS2_START_DOCSTRING,
 )
-class Idefics2ForConditionalGeneration(Idefics2PreTrainedModel):
+class Idefics2ForConditionalGeneration(GenerationMixin, Idefics2PreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/imagegpt/modeling_imagegpt.py b/src/transformers/models/imagegpt/modeling_imagegpt.py
index 5d59a4ed90e4..a202f2dbc9bf 100755
--- a/src/transformers/models/imagegpt/modeling_imagegpt.py
+++ b/src/transformers/models/imagegpt/modeling_imagegpt.py
@@ -26,6 +26,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
     CausalLMOutputWithCrossAttentions,
@@ -880,7 +881,7 @@ def forward(
     """,
     IMAGEGPT_START_DOCSTRING,
 )
-class ImageGPTForCausalImageModeling(ImageGPTPreTrainedModel):
+class ImageGPTForCausalImageModeling(GenerationMixin, ImageGPTPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config: ImageGPTConfig):
diff --git a/src/transformers/models/instructblip/modeling_instructblip.py b/src/transformers/models/instructblip/modeling_instructblip.py
index f59f72a6699c..beb65d130f8d 100644
--- a/src/transformers/models/instructblip/modeling_instructblip.py
+++ b/src/transformers/models/instructblip/modeling_instructblip.py
@@ -24,6 +24,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutput,
     BaseModelOutputWithPastAndCrossAttentions,
@@ -1274,7 +1275,7 @@ def forward(
     """,
     INSTRUCTBLIP_START_DOCSTRING,
 )
-class InstructBlipForConditionalGeneration(InstructBlipPreTrainedModel):
+class InstructBlipForConditionalGeneration(GenerationMixin, InstructBlipPreTrainedModel):
     config_class = InstructBlipConfig
     main_input_name = "pixel_values"
 
diff --git a/src/transformers/models/instructblipvideo/diff_instructblipvideo.py b/src/transformers/models/instructblipvideo/diff_instructblipvideo.py
index 506da83c5322..b6ccc67b1e83 100644
--- a/src/transformers/models/instructblipvideo/diff_instructblipvideo.py
+++ b/src/transformers/models/instructblipvideo/diff_instructblipvideo.py
@@ -45,6 +45,7 @@
     InstructBlipVisionModel,
 )
 
+from ...generation import GenerationMixin
 from ...utils import logging
 
 
@@ -128,7 +129,7 @@ class InstructBlipVideoQFormerModel(InstructBlipQFormerModel):
     pass
 
 
-class InstructBlipVideoForConditionalGeneration(InstructBlipForConditionalGeneration):
+class InstructBlipVideoForConditionalGeneration(GenerationMixin, InstructBlipForConditionalGeneration):
     def forward(
         self,
         pixel_values: torch.FloatTensor,
diff --git a/src/transformers/models/instructblipvideo/modeling_instructblipvideo.py b/src/transformers/models/instructblipvideo/modeling_instructblipvideo.py
index 701402241d4a..57c38bee9ffc 100644
--- a/src/transformers/models/instructblipvideo/modeling_instructblipvideo.py
+++ b/src/transformers/models/instructblipvideo/modeling_instructblipvideo.py
@@ -30,6 +30,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutput,
     BaseModelOutputWithPastAndCrossAttentions,
@@ -1283,7 +1284,7 @@ def forward(
     """,
     INSTRUCTBLIPVIDEO_START_DOCSTRING,
 )
-class InstructBlipVideoForConditionalGeneration(InstructBlipVideoPreTrainedModel):
+class InstructBlipVideoForConditionalGeneration(GenerationMixin, InstructBlipVideoPreTrainedModel):
     config_class = InstructBlipVideoConfig
     main_input_name = "pixel_values"
 
diff --git a/src/transformers/models/jamba/modeling_jamba.py b/src/transformers/models/jamba/modeling_jamba.py
index 60e1670a3c27..f55e3af49cb7 100755
--- a/src/transformers/models/jamba/modeling_jamba.py
+++ b/src/transformers/models/jamba/modeling_jamba.py
@@ -30,6 +30,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache  # we need __iter__ and __len__ of pkv
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import (
     AttentionMaskConverter,
 )
@@ -1424,7 +1425,7 @@ def _update_mamba_mask(self, attention_mask, cache_position):
 
 
 # Adapted from transformers.models.mixtral.modeling_mixtral.MixtralForCausalLM with MIXTRAL->JAMBA, Mixtral->Jamba
-class JambaForCausalLM(JambaPreTrainedModel):
+class JambaForCausalLM(GenerationMixin, JambaPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config: JambaConfig):
diff --git a/src/transformers/models/jetmoe/modeling_jetmoe.py b/src/transformers/models/jetmoe/modeling_jetmoe.py
index 162478a7258c..53853bb9d845 100644
--- a/src/transformers/models/jetmoe/modeling_jetmoe.py
+++ b/src/transformers/models/jetmoe/modeling_jetmoe.py
@@ -25,6 +25,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     MoeCausalLMOutputWithPast,
@@ -1195,7 +1196,7 @@ def _update_causal_mask(
         return causal_mask
 
 
-class JetMoeForCausalLM(JetMoePreTrainedModel):
+class JetMoeForCausalLM(GenerationMixin, JetMoePreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/kosmos2/modeling_kosmos2.py b/src/transformers/models/kosmos2/modeling_kosmos2.py
index 69641790b2db..625fb777d6e3 100644
--- a/src/transformers/models/kosmos2/modeling_kosmos2.py
+++ b/src/transformers/models/kosmos2/modeling_kosmos2.py
@@ -24,6 +24,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutput,
     BaseModelOutputWithPastAndCrossAttentions,
@@ -1521,7 +1522,7 @@ def forward(
     """,
     KOSMOS2_START_DOCSTRING,
 )
-class Kosmos2TextForCausalLM(Kosmos2PreTrainedModel):
+class Kosmos2TextForCausalLM(GenerationMixin, Kosmos2PreTrainedModel):
     config_class = Kosmos2TextConfig
     _tied_weights_keys = ["lm_head.weight"]
 
@@ -1864,7 +1865,7 @@ def forward(
     """,
     KOSMOS2_START_DOCSTRING,
 )
-class Kosmos2ForConditionalGeneration(Kosmos2PreTrainedModel):
+class Kosmos2ForConditionalGeneration(GenerationMixin, Kosmos2PreTrainedModel):
     config_class = Kosmos2Config
     main_input_name = "pixel_values"
     _tied_weights_keys = ["text_model.lm_head.weight"]
diff --git a/src/transformers/models/led/modeling_led.py b/src/transformers/models/led/modeling_led.py
index 41b6c0a2bea2..30081e636291 100755
--- a/src/transformers/models/led/modeling_led.py
+++ b/src/transformers/models/led/modeling_led.py
@@ -25,6 +25,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _create_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
@@ -2298,7 +2299,7 @@ def forward(
 @add_start_docstrings(
     "The LED Model with a language modeling head. Can be used for summarization.", LED_START_DOCSTRING
 )
-class LEDForConditionalGeneration(LEDPreTrainedModel):
+class LEDForConditionalGeneration(GenerationMixin, LEDPreTrainedModel):
     base_model_prefix = "led"
     _keys_to_ignore_on_load_missing = ["final_logits_bias"]
     _tied_weights_keys = ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight", "lm_head.weight"]
diff --git a/src/transformers/models/llama/modeling_llama.py b/src/transformers/models/llama/modeling_llama.py
index 022ae5ce74c4..aa8eb91689c3 100644
--- a/src/transformers/models/llama/modeling_llama.py
+++ b/src/transformers/models/llama/modeling_llama.py
@@ -28,6 +28,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_flash_attention_utils import _flash_attention_forward
 from ...modeling_outputs import (
@@ -1097,7 +1098,7 @@ def _update_causal_mask(
         return causal_mask
 
 
-class LlamaForCausalLM(LlamaPreTrainedModel):
+class LlamaForCausalLM(GenerationMixin, LlamaPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/llava/modeling_llava.py b/src/transformers/models/llava/modeling_llava.py
index 394c80edb540..290fc75517e3 100644
--- a/src/transformers/models/llava/modeling_llava.py
+++ b/src/transformers/models/llava/modeling_llava.py
@@ -23,6 +23,7 @@
 
 from ... import PreTrainedModel
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import ModelOutput
 from ...utils import (
     add_start_docstrings,
@@ -240,7 +241,7 @@ def _supports_sdpa(self):
     """The LLAVA model which consists of a vision backbone and a language model.""",
     LLAVA_START_DOCSTRING,
 )
-class LlavaForConditionalGeneration(LlavaPreTrainedModel):
+class LlavaForConditionalGeneration(GenerationMixin, LlavaPreTrainedModel):
     def __init__(self, config: LlavaConfig):
         super().__init__(config)
         self.vision_tower = AutoModel.from_config(config.vision_config)
diff --git a/src/transformers/models/llava_next/modeling_llava_next.py b/src/transformers/models/llava_next/modeling_llava_next.py
index 723d54c92dd9..460754fa7e91 100644
--- a/src/transformers/models/llava_next/modeling_llava_next.py
+++ b/src/transformers/models/llava_next/modeling_llava_next.py
@@ -23,10 +23,11 @@
 import torch.utils.checkpoint
 from torch import nn
 
-from ... import PreTrainedModel
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...image_processing_utils import select_best_resolution
 from ...modeling_outputs import ModelOutput
+from ...modeling_utils import PreTrainedModel
 from ...utils import (
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
@@ -346,7 +347,7 @@ def _supports_sdpa(self):
     """The LLAVA-NeXT model which consists of a vision backbone and a language model.""",
     LLAVA_NEXT_START_DOCSTRING,
 )
-class LlavaNextForConditionalGeneration(LlavaNextPreTrainedModel):
+class LlavaNextForConditionalGeneration(GenerationMixin, LlavaNextPreTrainedModel):
     def __init__(self, config: LlavaNextConfig):
         super().__init__(config)
         self.vision_tower = AutoModel.from_config(config.vision_config)
diff --git a/src/transformers/models/llava_next_video/diff_llava_next_video.py b/src/transformers/models/llava_next_video/diff_llava_next_video.py
index b4018db586e7..5242eb1ed3a2 100644
--- a/src/transformers/models/llava_next_video/diff_llava_next_video.py
+++ b/src/transformers/models/llava_next_video/diff_llava_next_video.py
@@ -30,6 +30,7 @@
 )
 
 from ...cache_utils import Cache
+from ...generation import GenerationMixin
 from ...utils import (
     logging,
     replace_return_docstrings,
@@ -219,7 +220,7 @@ class LlavaNextVideoMultiModalProjector(LlavaNextMultiModalProjector):
     pass
 
 
-class LlavaNextVideoForConditionalGeneration(LlavaNextForConditionalGeneration):
+class LlavaNextVideoForConditionalGeneration(GenerationMixin, LlavaNextForConditionalGeneration):
     def __init__(self, config: LlavaNextVideoConfig, **super_kwargs):
         super().__init__(config, **super_kwargs)
         self.vision_resampler = LlavaNextVideoPooler(config)
diff --git a/src/transformers/models/llava_next_video/modeling_llava_next_video.py b/src/transformers/models/llava_next_video/modeling_llava_next_video.py
index 3430fbe590aa..8fecf89f5965 100644
--- a/src/transformers/models/llava_next_video/modeling_llava_next_video.py
+++ b/src/transformers/models/llava_next_video/modeling_llava_next_video.py
@@ -32,6 +32,7 @@
 from ... import PreTrainedModel
 from ...activations import ACT2FN
 from ...cache_utils import Cache
+from ...generation import GenerationMixin
 from ...image_processing_utils import select_best_resolution
 from ...modeling_outputs import ModelOutput
 from ...utils import (
@@ -387,7 +388,7 @@ def _supports_sdpa(self):
     """The LLAVA-NeXT model which consists of a vision backbone and a language model.""",
     LLAVA_NEXT_VIDEO_START_DOCSTRING,
 )
-class LlavaNextVideoForConditionalGeneration(LlavaNextVideoPreTrainedModel):
+class LlavaNextVideoForConditionalGeneration(GenerationMixin, LlavaNextVideoPreTrainedModel):
     def __init__(
         self,
         config: LlavaNextVideoConfig,
diff --git a/src/transformers/models/longt5/modeling_longt5.py b/src/transformers/models/longt5/modeling_longt5.py
index b2a6ed11ca57..b6f3a5a5d33a 100644
--- a/src/transformers/models/longt5/modeling_longt5.py
+++ b/src/transformers/models/longt5/modeling_longt5.py
@@ -24,6 +24,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutput,
     BaseModelOutputWithPastAndCrossAttentions,
@@ -1900,7 +1901,7 @@ def forward(
 
 
 @add_start_docstrings("""LONGT5 Model with a `language modeling` head on top.""", LONGT5_START_DOCSTRING)
-class LongT5ForConditionalGeneration(LongT5PreTrainedModel):
+class LongT5ForConditionalGeneration(GenerationMixin, LongT5PreTrainedModel):
     _keys_to_ignore_on_load_unexpected = [
         r"decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight",
     ]
diff --git a/src/transformers/models/m2m_100/modeling_m2m_100.py b/src/transformers/models/m2m_100/modeling_m2m_100.py
index 23a855fff256..93294e3592c3 100755
--- a/src/transformers/models/m2m_100/modeling_m2m_100.py
+++ b/src/transformers/models/m2m_100/modeling_m2m_100.py
@@ -22,6 +22,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...integrations.deepspeed import is_deepspeed_zero3_enabled
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
@@ -1342,7 +1343,7 @@ def forward(
 @add_start_docstrings(
     "The M2M100 Model with a language modeling head. Can be used for summarization.", M2M_100_START_DOCSTRING
 )
-class M2M100ForConditionalGeneration(M2M100PreTrainedModel):
+class M2M100ForConditionalGeneration(GenerationMixin, M2M100PreTrainedModel):
     base_model_prefix = "model"
     _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
 
diff --git a/src/transformers/models/mamba/modeling_mamba.py b/src/transformers/models/mamba/modeling_mamba.py
index 14a3dea1d1cc..ad96469a9063 100644
--- a/src/transformers/models/mamba/modeling_mamba.py
+++ b/src/transformers/models/mamba/modeling_mamba.py
@@ -25,6 +25,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import MambaCache
+from ...generation import GenerationMixin
 from ...modeling_utils import PreTrainedModel
 from ...utils import (
     ModelOutput,
@@ -657,7 +658,7 @@ def forward(
     """,
     MAMBA_START_DOCSTRING,
 )
-class MambaForCausalLM(MambaPreTrainedModel):
+class MambaForCausalLM(GenerationMixin, MambaPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/mamba2/modeling_mamba2.py b/src/transformers/models/mamba2/modeling_mamba2.py
index 69390ea9ad2b..fd49f6cbf6d2 100644
--- a/src/transformers/models/mamba2/modeling_mamba2.py
+++ b/src/transformers/models/mamba2/modeling_mamba2.py
@@ -24,6 +24,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_utils import PreTrainedModel
 from ...utils import (
     ModelOutput,
@@ -931,7 +932,7 @@ def forward(
     """,
     MAMBA2_START_DOCSTRING,
 )
-class Mamba2ForCausalLM(Mamba2PreTrainedModel):
+class Mamba2ForCausalLM(GenerationMixin, Mamba2PreTrainedModel):
     _tied_weights_keys = []
 
     def __init__(self, config):
diff --git a/src/transformers/models/marian/modeling_marian.py b/src/transformers/models/marian/modeling_marian.py
index 2045f673540f..7aa1fd60c52a 100755
--- a/src/transformers/models/marian/modeling_marian.py
+++ b/src/transformers/models/marian/modeling_marian.py
@@ -25,6 +25,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -1224,7 +1225,7 @@ def forward(
 @add_start_docstrings(
     "The Marian Model with a language modeling head. Can be used for summarization.", MARIAN_START_DOCSTRING
 )
-class MarianMTModel(MarianPreTrainedModel):
+class MarianMTModel(GenerationMixin, MarianPreTrainedModel):
     base_model_prefix = "model"
     _keys_to_ignore_on_load_missing = [
         "final_logits_bias",
@@ -1504,7 +1505,7 @@ def forward(self, *args, **kwargs):
 
 
 # Copied from transformers.models.bart.modeling_bart.BartForCausalLM with Bart->Marian, facebook/bart-base->Helsinki-NLP/opus-mt-fr-en
-class MarianForCausalLM(MarianPreTrainedModel):
+class MarianForCausalLM(GenerationMixin, MarianPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/mbart/modeling_mbart.py b/src/transformers/models/mbart/modeling_mbart.py
index 6cad7b08f994..c9a72e30cc37 100755
--- a/src/transformers/models/mbart/modeling_mbart.py
+++ b/src/transformers/models/mbart/modeling_mbart.py
@@ -24,6 +24,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -1389,7 +1390,7 @@ def forward(
     "The MBART Model with a language modeling head. Can be used for summarization, after fine-tuning the pretrained models.",
     MBART_START_DOCSTRING,
 )
-class MBartForConditionalGeneration(MBartPreTrainedModel):
+class MBartForConditionalGeneration(GenerationMixin, MBartPreTrainedModel):
     base_model_prefix = "model"
     _keys_to_ignore_on_load_missing = ["final_logits_bias"]
     _tied_weights_keys = ["model.encoder.embed_tokens.weight", "model.decoder.embed_tokens.weight", "lm_head.weight"]
@@ -1830,7 +1831,7 @@ def forward(self, *args, **kwargs):
 
 
 # Copied from transformers.models.bart.modeling_bart.BartForCausalLM with Bart->MBart, facebook/bart-base->facebook/mbart-large-cc25
-class MBartForCausalLM(MBartPreTrainedModel):
+class MBartForCausalLM(GenerationMixin, MBartPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/megatron_bert/modeling_megatron_bert.py b/src/transformers/models/megatron_bert/modeling_megatron_bert.py
index 16641655e203..5c0f920b02fd 100755
--- a/src/transformers/models/megatron_bert/modeling_megatron_bert.py
+++ b/src/transformers/models/megatron_bert/modeling_megatron_bert.py
@@ -27,6 +27,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
     BaseModelOutputWithPoolingAndCrossAttentions,
@@ -1110,7 +1111,7 @@ def forward(
     """MegatronBert Model with a `language modeling` head on top for CLM fine-tuning.""",
     MEGATRON_BERT_START_DOCSTRING,
 )
-class MegatronBertForCausalLM(MegatronBertPreTrainedModel):
+class MegatronBertForCausalLM(GenerationMixin, MegatronBertPreTrainedModel):
     _tied_weights_keys = ["cls.predictions.decoder"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/mistral/modeling_mistral.py b/src/transformers/models/mistral/modeling_mistral.py
index 240e229e0bb0..ce85cfa25a9a 100644
--- a/src/transformers/models/mistral/modeling_mistral.py
+++ b/src/transformers/models/mistral/modeling_mistral.py
@@ -29,6 +29,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
@@ -947,7 +948,7 @@ def _update_causal_mask(
         return causal_mask
 
 
-class MistralForCausalLM(MistralPreTrainedModel):
+class MistralForCausalLM(GenerationMixin, MistralPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/mixtral/modeling_mixtral.py b/src/transformers/models/mixtral/modeling_mixtral.py
index 919f32abc7fc..fa01764c9148 100644
--- a/src/transformers/models/mixtral/modeling_mixtral.py
+++ b/src/transformers/models/mixtral/modeling_mixtral.py
@@ -30,6 +30,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     MoeCausalLMOutputWithPast,
@@ -1180,7 +1181,7 @@ def _update_causal_mask(
         return causal_mask
 
 
-class MixtralForCausalLM(MixtralPreTrainedModel):
+class MixtralForCausalLM(GenerationMixin, MixtralPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/mpt/modeling_mpt.py b/src/transformers/models/mpt/modeling_mpt.py
index 85579636dcc4..fa9c2362324a 100644
--- a/src/transformers/models/mpt/modeling_mpt.py
+++ b/src/transformers/models/mpt/modeling_mpt.py
@@ -24,6 +24,7 @@
 from torch.nn import functional as F
 
 from ...file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
@@ -500,7 +501,7 @@ def forward(
     """,
     MPT_START_DOCSTRING,
 )
-class MptForCausalLM(MptPreTrainedModel):
+class MptForCausalLM(GenerationMixin, MptPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config: MptConfig):
diff --git a/src/transformers/models/mt5/modeling_mt5.py b/src/transformers/models/mt5/modeling_mt5.py
index 54943cf982dd..6d2d99df290b 100644
--- a/src/transformers/models/mt5/modeling_mt5.py
+++ b/src/transformers/models/mt5/modeling_mt5.py
@@ -25,6 +25,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutput,
     BaseModelOutputWithPastAndCrossAttentions,
@@ -1550,7 +1551,7 @@ def forward(
 
 
 @add_start_docstrings("""MT5 Model with a `language modeling` head on top.""", MT5_START_DOCSTRING)
-class MT5ForConditionalGeneration(MT5PreTrainedModel):
+class MT5ForConditionalGeneration(GenerationMixin, MT5PreTrainedModel):
     r"""
     Examples:
 
diff --git a/src/transformers/models/musicgen/modeling_musicgen.py b/src/transformers/models/musicgen/modeling_musicgen.py
index f720faac038e..0e6bc8c90a35 100644
--- a/src/transformers/models/musicgen/modeling_musicgen.py
+++ b/src/transformers/models/musicgen/modeling_musicgen.py
@@ -26,9 +26,14 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
-from ...generation.configuration_utils import GenerationConfig, GenerationMode
-from ...generation.logits_process import ClassifierFreeGuidanceLogitsProcessor, LogitsProcessorList
-from ...generation.stopping_criteria import StoppingCriteriaList
+from ...generation import (
+    ClassifierFreeGuidanceLogitsProcessor,
+    GenerationConfig,
+    GenerationMixin,
+    GenerationMode,
+    LogitsProcessorList,
+    StoppingCriteriaList,
+)
 from ...modeling_attn_mask_utils import (
     _prepare_4d_attention_mask,
     _prepare_4d_attention_mask_for_sdpa,
@@ -1206,7 +1211,7 @@ def forward(
     "The MusicGen decoder model with a language modelling head on top.",
     MUSICGEN_START_DOCSTRING,
 )
-class MusicgenForCausalLM(MusicgenPreTrainedModel):
+class MusicgenForCausalLM(GenerationMixin, MusicgenPreTrainedModel):
     def __init__(self, config: MusicgenDecoderConfig):
         super().__init__(config)
 
@@ -1658,7 +1663,7 @@ def generate(
     "for music generation tasks with one or both of text and audio prompts.",
     MUSICGEN_START_DOCSTRING,
 )
-class MusicgenForConditionalGeneration(PreTrainedModel):
+class MusicgenForConditionalGeneration(GenerationMixin, PreTrainedModel):
     config_class = MusicgenConfig
     base_model_prefix = "encoder_decoder"
     main_input_name = "input_ids"
diff --git a/src/transformers/models/musicgen_melody/modeling_musicgen_melody.py b/src/transformers/models/musicgen_melody/modeling_musicgen_melody.py
index a8a8fe960989..8584594b825c 100644
--- a/src/transformers/models/musicgen_melody/modeling_musicgen_melody.py
+++ b/src/transformers/models/musicgen_melody/modeling_musicgen_melody.py
@@ -26,9 +26,14 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
-from ...generation.configuration_utils import GenerationConfig, GenerationMode
-from ...generation.logits_process import ClassifierFreeGuidanceLogitsProcessor, LogitsProcessorList
-from ...generation.stopping_criteria import StoppingCriteriaList
+from ...generation import (
+    ClassifierFreeGuidanceLogitsProcessor,
+    GenerationConfig,
+    GenerationMixin,
+    GenerationMode,
+    LogitsProcessorList,
+    StoppingCriteriaList,
+)
 from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
@@ -1117,7 +1122,7 @@ def forward(
     MUSICGEN_MELODY_START_DOCSTRING,
 )
 # Copied from transformers.models.musicgen.modeling_musicgen.MusicgenForCausalLM with MUSICGEN->MUSICGEN_MELODY,Musicgen->MusicgenMelody,MusicGen->Musicgen Melody
-class MusicgenMelodyForCausalLM(MusicgenMelodyPreTrainedModel):
+class MusicgenMelodyForCausalLM(GenerationMixin, MusicgenMelodyPreTrainedModel):
     def __init__(self, config: MusicgenMelodyDecoderConfig):
         super().__init__(config)
 
@@ -1585,7 +1590,7 @@ def generate(
         decoder (`Optional[MusicgenMelodyForCausalLM]`, *optional*): MusicGen Melody decoder used to generate audio codes.
     """,
 )
-class MusicgenMelodyForConditionalGeneration(PreTrainedModel):
+class MusicgenMelodyForConditionalGeneration(GenerationMixin, PreTrainedModel):
     config_class = MusicgenMelodyConfig
     main_input_name = "input_ids"
     supports_gradient_checkpointing = True
diff --git a/src/transformers/models/mvp/modeling_mvp.py b/src/transformers/models/mvp/modeling_mvp.py
index 319f1760cef9..0868af9af081 100644
--- a/src/transformers/models/mvp/modeling_mvp.py
+++ b/src/transformers/models/mvp/modeling_mvp.py
@@ -24,6 +24,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -1351,7 +1352,7 @@ def forward(
 @add_start_docstrings(
     "The MVP Model with a language modeling head. Can be used for various text generation tasks.", MVP_START_DOCSTRING
 )
-class MvpForConditionalGeneration(MvpPreTrainedModel):
+class MvpForConditionalGeneration(GenerationMixin, MvpPreTrainedModel):
     _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
 
     def __init__(self, config: MvpConfig):
@@ -1791,7 +1792,7 @@ def forward(self, *args, **kwargs):
         return self.decoder(*args, **kwargs)
 
 
-class MvpForCausalLM(MvpPreTrainedModel):
+class MvpForCausalLM(GenerationMixin, MvpPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/nemotron/modeling_nemotron.py b/src/transformers/models/nemotron/modeling_nemotron.py
index 548732b371a5..8c986e075785 100644
--- a/src/transformers/models/nemotron/modeling_nemotron.py
+++ b/src/transformers/models/nemotron/modeling_nemotron.py
@@ -26,6 +26,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_flash_attention_utils import _flash_attention_forward
 from ...modeling_outputs import (
@@ -979,7 +980,7 @@ def _update_causal_mask(
 
 
 # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM with LLAMA->NEMOTRON,Llama->Nemotron,llama->nemotron
-class NemotronForCausalLM(NemotronPreTrainedModel):
+class NemotronForCausalLM(GenerationMixin, NemotronPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/nllb_moe/modeling_nllb_moe.py b/src/transformers/models/nllb_moe/modeling_nllb_moe.py
index 2bec0fb84dce..a419b8eccff3 100644
--- a/src/transformers/models/nllb_moe/modeling_nllb_moe.py
+++ b/src/transformers/models/nllb_moe/modeling_nllb_moe.py
@@ -22,6 +22,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...integrations.deepspeed import is_deepspeed_zero3_enabled
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
@@ -1604,7 +1605,7 @@ def forward(
 @add_start_docstrings(
     "The NllbMoe Model with a language modeling head. Can be used for summarization.", NLLB_MOE_START_DOCSTRING
 )
-class NllbMoeForConditionalGeneration(NllbMoePreTrainedModel):
+class NllbMoeForConditionalGeneration(GenerationMixin, NllbMoePreTrainedModel):
     base_model_prefix = "model"
     _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
 
diff --git a/src/transformers/models/olmo/modeling_olmo.py b/src/transformers/models/olmo/modeling_olmo.py
index 587ef92e4585..61006c1cc43f 100644
--- a/src/transformers/models/olmo/modeling_olmo.py
+++ b/src/transformers/models/olmo/modeling_olmo.py
@@ -30,6 +30,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
@@ -1018,7 +1019,7 @@ def _update_causal_mask(
 
 
 # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM with LLAMA->OLMO,Llama->Olmo
-class OlmoForCausalLM(OlmoPreTrainedModel):
+class OlmoForCausalLM(GenerationMixin, OlmoPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/openai/modeling_openai.py b/src/transformers/models/openai/modeling_openai.py
index 2b24850f3f0c..2b252cfc76d2 100644
--- a/src/transformers/models/openai/modeling_openai.py
+++ b/src/transformers/models/openai/modeling_openai.py
@@ -26,6 +26,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import gelu_new, silu
+from ...generation import GenerationMixin
 from ...modeling_outputs import BaseModelOutput, CausalLMOutput, SequenceClassifierOutput
 from ...modeling_utils import PreTrainedModel, SequenceSummary
 from ...pytorch_utils import Conv1D, find_pruneable_heads_and_indices, prune_conv1d_layer
@@ -524,7 +525,7 @@ def forward(
     """,
     OPENAI_GPT_START_DOCSTRING,
 )
-class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
+class OpenAIGPTLMHeadModel(GenerationMixin, OpenAIGPTPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/opt/modeling_opt.py b/src/transformers/models/opt/modeling_opt.py
index 8f058171778e..858f32073454 100644
--- a/src/transformers/models/opt/modeling_opt.py
+++ b/src/transformers/models/opt/modeling_opt.py
@@ -22,6 +22,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
@@ -882,7 +883,7 @@ def forward(
         )
 
 
-class OPTForCausalLM(OPTPreTrainedModel):
+class OPTForCausalLM(GenerationMixin, OPTPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/paligemma/modeling_paligemma.py b/src/transformers/models/paligemma/modeling_paligemma.py
index 8eff8cce50cc..31ea0e93fcc7 100644
--- a/src/transformers/models/paligemma/modeling_paligemma.py
+++ b/src/transformers/models/paligemma/modeling_paligemma.py
@@ -22,6 +22,7 @@
 from torch import nn
 
 from ...cache_utils import Cache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_utils import PreTrainedModel
 from ...utils import (
     ModelOutput,
@@ -236,7 +237,7 @@ def _supports_sdpa(self):
     """The PALIGEMMA model which consists of a vision backbone and a language model.""",
     PALIGEMMA_START_DOCSTRING,
 )
-class PaliGemmaForConditionalGeneration(PaliGemmaPreTrainedModel):
+class PaliGemmaForConditionalGeneration(GenerationMixin, PaliGemmaPreTrainedModel):
     def __init__(self, config: PaliGemmaConfig):
         super().__init__(config)
         self.vision_tower = AutoModel.from_config(config=config.vision_config)
diff --git a/src/transformers/models/pegasus/modeling_pegasus.py b/src/transformers/models/pegasus/modeling_pegasus.py
index 42cef3a63558..69890f162c14 100755
--- a/src/transformers/models/pegasus/modeling_pegasus.py
+++ b/src/transformers/models/pegasus/modeling_pegasus.py
@@ -25,6 +25,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -1244,7 +1245,7 @@ def forward(
 @add_start_docstrings(
     "The PEGASUS Model with a language modeling head. Can be used for summarization.", PEGASUS_START_DOCSTRING
 )
-class PegasusForConditionalGeneration(PegasusPreTrainedModel):
+class PegasusForConditionalGeneration(GenerationMixin, PegasusPreTrainedModel):
     base_model_prefix = "model"
     _keys_to_ignore_on_load_missing = ["final_logits_bias"]
     _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
@@ -1456,7 +1457,7 @@ def forward(self, *args, **kwargs):
         return self.decoder(*args, **kwargs)
 
 
-class PegasusForCausalLM(PegasusPreTrainedModel):
+class PegasusForCausalLM(GenerationMixin, PegasusPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/pegasus_x/modeling_pegasus_x.py b/src/transformers/models/pegasus_x/modeling_pegasus_x.py
index 6d9072777bf6..1cc5d44bc2a3 100755
--- a/src/transformers/models/pegasus_x/modeling_pegasus_x.py
+++ b/src/transformers/models/pegasus_x/modeling_pegasus_x.py
@@ -25,6 +25,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -1464,7 +1465,7 @@ def forward(
 
 
 @add_start_docstrings("The PEGASUS-X for conditional generation (e.g. summarization).", PEGASUS_X_START_DOCSTRING)
-class PegasusXForConditionalGeneration(PegasusXPreTrainedModel):
+class PegasusXForConditionalGeneration(GenerationMixin, PegasusXPreTrainedModel):
     base_model_prefix = "model"
     _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
 
diff --git a/src/transformers/models/persimmon/modeling_persimmon.py b/src/transformers/models/persimmon/modeling_persimmon.py
index 90a7f355992e..695e64fe973c 100644
--- a/src/transformers/models/persimmon/modeling_persimmon.py
+++ b/src/transformers/models/persimmon/modeling_persimmon.py
@@ -29,6 +29,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
@@ -828,7 +829,7 @@ def _update_causal_mask(
         return causal_mask
 
 
-class PersimmonForCausalLM(PersimmonPreTrainedModel):
+class PersimmonForCausalLM(GenerationMixin, PersimmonPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.__init__ with LLAMA->PERSIMMON,Llama->Persimmon
diff --git a/src/transformers/models/phi/modeling_phi.py b/src/transformers/models/phi/modeling_phi.py
index 3c647a9d8d81..08d11d9b96f1 100644
--- a/src/transformers/models/phi/modeling_phi.py
+++ b/src/transformers/models/phi/modeling_phi.py
@@ -26,6 +26,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
@@ -1113,7 +1114,7 @@ def _update_causal_mask(
         return causal_mask
 
 
-class PhiForCausalLM(PhiPreTrainedModel):
+class PhiForCausalLM(GenerationMixin, PhiPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.__init__ with Llama->Phi,bias=False->bias=True
diff --git a/src/transformers/models/phi3/modeling_phi3.py b/src/transformers/models/phi3/modeling_phi3.py
index 4652294980fd..ec47e6fa3e69 100644
--- a/src/transformers/models/phi3/modeling_phi3.py
+++ b/src/transformers/models/phi3/modeling_phi3.py
@@ -26,6 +26,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
@@ -1153,7 +1154,7 @@ def _update_causal_mask(
         return causal_mask
 
 
-class Phi3ForCausalLM(Phi3PreTrainedModel):
+class Phi3ForCausalLM(GenerationMixin, Phi3PreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.__init__ with Llama->Phi3
diff --git a/src/transformers/models/pix2struct/modeling_pix2struct.py b/src/transformers/models/pix2struct/modeling_pix2struct.py
index 94d882c80566..df7f84a19f83 100644
--- a/src/transformers/models/pix2struct/modeling_pix2struct.py
+++ b/src/transformers/models/pix2struct/modeling_pix2struct.py
@@ -22,6 +22,7 @@
 from torch import nn
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutput,
     BaseModelOutputWithPooling,
@@ -1553,7 +1554,7 @@ def forward(
     "A conditional generation model with a language modeling head. Can be used for sequence generation tasks.",
     PIX2STRUCT_START_DOCSTRING,
 )
-class Pix2StructForConditionalGeneration(Pix2StructPreTrainedModel):
+class Pix2StructForConditionalGeneration(GenerationMixin, Pix2StructPreTrainedModel):
     config_class = Pix2StructConfig
     main_input_name = "flattened_patches"
     _tied_weights_keys = ["decoder.lm_head.weight"]
diff --git a/src/transformers/models/plbart/modeling_plbart.py b/src/transformers/models/plbart/modeling_plbart.py
index 93d91e160089..520a994e9afa 100644
--- a/src/transformers/models/plbart/modeling_plbart.py
+++ b/src/transformers/models/plbart/modeling_plbart.py
@@ -24,6 +24,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import (
     _prepare_4d_attention_mask,
     _prepare_4d_attention_mask_for_sdpa,
@@ -1254,7 +1255,7 @@ def forward(
     "The PLBART Model with a language modeling head. Can be used for code-to-text, text-to-code and code-to-code.",
     PLBART_START_DOCSTRING,
 )
-class PLBartForConditionalGeneration(PLBartPreTrainedModel):
+class PLBartForConditionalGeneration(GenerationMixin, PLBartPreTrainedModel):
     base_model_prefix = "model"
     _keys_to_ignore_on_load_missing = ["final_logits_bias"]
     _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
@@ -1568,7 +1569,7 @@ def forward(self, *args, **kwargs):
 
 
 # Copied from transformers.models.bart.modeling_bart.BartForCausalLM with Bart->PLBart, facebook/bart-base->uclanlp/plbart-base
-class PLBartForCausalLM(PLBartPreTrainedModel):
+class PLBartForCausalLM(GenerationMixin, PLBartPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/pop2piano/modeling_pop2piano.py b/src/transformers/models/pop2piano/modeling_pop2piano.py
index c769cff3c454..0ecd4af73edb 100644
--- a/src/transformers/models/pop2piano/modeling_pop2piano.py
+++ b/src/transformers/models/pop2piano/modeling_pop2piano.py
@@ -25,6 +25,7 @@
 from transformers.generation import GenerationConfig
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutput,
     BaseModelOutputWithPastAndCrossAttentions,
@@ -1001,7 +1002,7 @@ def forward(self, feature, index_value, embedding_offset):
 
 
 @add_start_docstrings("""Pop2Piano Model with a `language modeling` head on top.""", Pop2Piano_START_DOCSTRING)
-class Pop2PianoForConditionalGeneration(Pop2PianoPreTrainedModel):
+class Pop2PianoForConditionalGeneration(GenerationMixin, Pop2PianoPreTrainedModel):
     _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
 
     def __init__(self, config: Pop2PianoConfig):
diff --git a/src/transformers/models/prophetnet/modeling_prophetnet.py b/src/transformers/models/prophetnet/modeling_prophetnet.py
index 96fa2e2c12e5..1acc3e66ec2a 100644
--- a/src/transformers/models/prophetnet/modeling_prophetnet.py
+++ b/src/transformers/models/prophetnet/modeling_prophetnet.py
@@ -26,6 +26,7 @@
 from torch.nn import LayerNorm
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import BaseModelOutput
 from ...modeling_utils import PreTrainedModel
 from ...utils import (
@@ -1856,7 +1857,7 @@ def forward(
     "The ProphetNet Model with a language modeling head. Can be used for sequence generation tasks.",
     PROPHETNET_START_DOCSTRING,
 )
-class ProphetNetForConditionalGeneration(ProphetNetPreTrainedModel):
+class ProphetNetForConditionalGeneration(GenerationMixin, ProphetNetPreTrainedModel):
     _tied_weights_keys = ["encoder.word_embeddings.weight", "decoder.word_embeddings.weight", "lm_head.weight"]
 
     def __init__(self, config: ProphetNetConfig):
@@ -2073,7 +2074,7 @@ def get_decoder(self):
     " language modeling.",
     PROPHETNET_START_DOCSTRING,
 )
-class ProphetNetForCausalLM(ProphetNetPreTrainedModel):
+class ProphetNetForCausalLM(GenerationMixin, ProphetNetPreTrainedModel):
     _tied_weights_keys = [
         "prophetnet.word_embeddings.weight",
         "prophetnet.decoder.word_embeddings.weight",
diff --git a/src/transformers/models/qwen2/modeling_qwen2.py b/src/transformers/models/qwen2/modeling_qwen2.py
index 59413730ad4a..0772fe6bd626 100644
--- a/src/transformers/models/qwen2/modeling_qwen2.py
+++ b/src/transformers/models/qwen2/modeling_qwen2.py
@@ -29,6 +29,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
@@ -1018,7 +1019,7 @@ def _update_causal_mask(
         return causal_mask
 
 
-class Qwen2ForCausalLM(Qwen2PreTrainedModel):
+class Qwen2ForCausalLM(GenerationMixin, Qwen2PreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/qwen2_audio/modeling_qwen2_audio.py b/src/transformers/models/qwen2_audio/modeling_qwen2_audio.py
index 14235bf0aaf6..489434350111 100644
--- a/src/transformers/models/qwen2_audio/modeling_qwen2_audio.py
+++ b/src/transformers/models/qwen2_audio/modeling_qwen2_audio.py
@@ -22,10 +22,11 @@
 import torch.utils.checkpoint
 from torch import nn
 
-from ... import PreTrainedModel
 from ...activations import ACT2FN
 from ...cache_utils import Cache, EncoderDecoderCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_outputs import BaseModelOutput, ModelOutput
+from ...modeling_utils import PreTrainedModel
 from ...utils import (
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
@@ -855,7 +856,7 @@ def forward(self, audio_features):
     """The QWEN2AUDIO model which consists of a audio backbone and a language model.""",
     QWEN2AUDIO_START_DOCSTRING,
 )
-class Qwen2AudioForConditionalGeneration(Qwen2AudioPreTrainedModel):
+class Qwen2AudioForConditionalGeneration(GenerationMixin, Qwen2AudioPreTrainedModel):
     def __init__(self, config: Qwen2AudioConfig):
         super().__init__(config)
         self.audio_tower = AutoModel.from_config(config.audio_config, attn_implementation=config._attn_implementation)
diff --git a/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py b/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py
index c08735f45345..c3558a5b2851 100644
--- a/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py
+++ b/src/transformers/models/qwen2_moe/modeling_qwen2_moe.py
@@ -30,6 +30,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     MoeCausalLMOutputWithPast,
@@ -1191,7 +1192,7 @@ def _update_causal_mask(
         return causal_mask
 
 
-class Qwen2MoeForCausalLM(Qwen2MoePreTrainedModel):
+class Qwen2MoeForCausalLM(GenerationMixin, Qwen2MoePreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py b/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py
index 6ab813ad9ade..c6c2cedc7813 100644
--- a/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py
+++ b/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py
@@ -31,6 +31,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import (
     AttentionMaskConverter,
 )
@@ -1306,7 +1307,7 @@ def _update_causal_mask(
 """
 
 
-class Qwen2VLForConditionalGeneration(Qwen2VLPreTrainedModel):
+class Qwen2VLForConditionalGeneration(GenerationMixin, Qwen2VLPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/recurrent_gemma/modeling_recurrent_gemma.py b/src/transformers/models/recurrent_gemma/modeling_recurrent_gemma.py
index a8f076fad79c..28055de872ab 100644
--- a/src/transformers/models/recurrent_gemma/modeling_recurrent_gemma.py
+++ b/src/transformers/models/recurrent_gemma/modeling_recurrent_gemma.py
@@ -24,6 +24,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import BaseModelOutputWithNoAttention, CausalLMOutput
 from ...modeling_utils import PreTrainedModel
@@ -777,7 +778,7 @@ def _update_causal_mask(self, attention_mask, input_tensor, cache_position):
 
 
 # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM with LLAMA->RECURRENTGEMMA,Llama->RecurrentGemma,llama->gemma
-class RecurrentGemmaForCausalLM(RecurrentGemmaPreTrainedModel):
+class RecurrentGemmaForCausalLM(GenerationMixin, RecurrentGemmaPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/reformer/modeling_reformer.py b/src/transformers/models/reformer/modeling_reformer.py
index 2e98a07217e6..c1b074d86ca3 100755
--- a/src/transformers/models/reformer/modeling_reformer.py
+++ b/src/transformers/models/reformer/modeling_reformer.py
@@ -29,6 +29,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import CausalLMOutput, MaskedLMOutput, QuestionAnsweringModelOutput, SequenceClassifierOutput
 from ...modeling_utils import PreTrainedModel
 from ...pytorch_utils import apply_chunking_to_forward
@@ -2183,7 +2184,7 @@ def _pad_to_mult_of_chunk_length(
 
 
 @add_start_docstrings("""Reformer Model with a `language modeling` head on top.""", REFORMER_START_DOCSTRING)
-class ReformerModelWithLMHead(ReformerPreTrainedModel):
+class ReformerModelWithLMHead(GenerationMixin, ReformerPreTrainedModel):
     _tied_weights_keys = ["lm_head.decoder.weight", "lm_head.decoder.bias"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/rembert/modeling_rembert.py b/src/transformers/models/rembert/modeling_rembert.py
index 31f7e3dce454..0604bdb1d864 100755
--- a/src/transformers/models/rembert/modeling_rembert.py
+++ b/src/transformers/models/rembert/modeling_rembert.py
@@ -24,6 +24,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
     BaseModelOutputWithPoolingAndCrossAttentions,
@@ -1002,7 +1003,7 @@ def prepare_inputs_for_generation(self, input_ids, attention_mask=None, **model_
 @add_start_docstrings(
     """RemBERT Model with a `language modeling` head on top for CLM fine-tuning.""", REMBERT_START_DOCSTRING
 )
-class RemBertForCausalLM(RemBertPreTrainedModel):
+class RemBertForCausalLM(GenerationMixin, RemBertPreTrainedModel):
     _tied_weights_keys = ["cls.predictions.decoder.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/roberta/modeling_roberta.py b/src/transformers/models/roberta/modeling_roberta.py
index f1f83147527d..9ccff336cd6e 100644
--- a/src/transformers/models/roberta/modeling_roberta.py
+++ b/src/transformers/models/roberta/modeling_roberta.py
@@ -25,6 +25,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN, gelu
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import (
     _prepare_4d_attention_mask_for_sdpa,
     _prepare_4d_causal_attention_mask_for_sdpa,
@@ -1003,7 +1004,7 @@ def forward(
 @add_start_docstrings(
     """RoBERTa Model with a `language modeling` head on top for CLM fine-tuning.""", ROBERTA_START_DOCSTRING
 )
-class RobertaForCausalLM(RobertaPreTrainedModel):
+class RobertaForCausalLM(GenerationMixin, RobertaPreTrainedModel):
     _tied_weights_keys = ["lm_head.decoder.weight", "lm_head.decoder.bias"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/roberta_prelayernorm/modeling_roberta_prelayernorm.py b/src/transformers/models/roberta_prelayernorm/modeling_roberta_prelayernorm.py
index 95657c260dc7..a06eae4e4a54 100644
--- a/src/transformers/models/roberta_prelayernorm/modeling_roberta_prelayernorm.py
+++ b/src/transformers/models/roberta_prelayernorm/modeling_roberta_prelayernorm.py
@@ -24,6 +24,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN, gelu
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
     BaseModelOutputWithPoolingAndCrossAttentions,
@@ -855,7 +856,7 @@ def forward(
     ROBERTA_PRELAYERNORM_START_DOCSTRING,
 )
 # Copied from transformers.models.roberta.modeling_roberta.RobertaForCausalLM with FacebookAI/roberta-base->andreasmadsen/efficient_mlm_m0.40,ROBERTA->ROBERTA_PRELAYERNORM,Roberta->RobertaPreLayerNorm,roberta->roberta_prelayernorm, RobertaPreLayerNormTokenizer->RobertaTokenizer
-class RobertaPreLayerNormForCausalLM(RobertaPreLayerNormPreTrainedModel):
+class RobertaPreLayerNormForCausalLM(GenerationMixin, RobertaPreLayerNormPreTrainedModel):
     _tied_weights_keys = ["lm_head.decoder.weight", "lm_head.decoder.bias"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/roc_bert/modeling_roc_bert.py b/src/transformers/models/roc_bert/modeling_roc_bert.py
index c4efbf16323e..1c641fb4a510 100644
--- a/src/transformers/models/roc_bert/modeling_roc_bert.py
+++ b/src/transformers/models/roc_bert/modeling_roc_bert.py
@@ -24,6 +24,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
     BaseModelOutputWithPoolingAndCrossAttentions,
@@ -1403,7 +1404,7 @@ def prepare_inputs_for_generation(
 @add_start_docstrings(
     """RoCBert Model with a `language modeling` head on top for CLM fine-tuning.""", ROC_BERT_START_DOCSTRING
 )
-class RoCBertForCausalLM(RoCBertPreTrainedModel):
+class RoCBertForCausalLM(GenerationMixin, RoCBertPreTrainedModel):
     _tied_weights_keys = ["cls.predictions.decoder.weight", "cls.predictions.decoder.bias"]
 
     # Copied from transformers.models.bert.modeling_bert.BertLMHeadModel.__init__ with BertLMHeadModel->RoCBertForCausalLM,Bert->RoCBert,bert->roc_bert
diff --git a/src/transformers/models/roformer/modeling_roformer.py b/src/transformers/models/roformer/modeling_roformer.py
index 69588ff743a0..fe461fc60475 100644
--- a/src/transformers/models/roformer/modeling_roformer.py
+++ b/src/transformers/models/roformer/modeling_roformer.py
@@ -25,6 +25,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
     CausalLMOutputWithCrossAttentions,
@@ -1033,7 +1034,7 @@ def prepare_inputs_for_generation(self, input_ids, attention_mask=None, **model_
 @add_start_docstrings(
     """RoFormer Model with a `language modeling` head on top for CLM fine-tuning.""", ROFORMER_START_DOCSTRING
 )
-class RoFormerForCausalLM(RoFormerPreTrainedModel):
+class RoFormerForCausalLM(GenerationMixin, RoFormerPreTrainedModel):
     _tied_weights_keys = ["cls.predictions.decoder.bias", "cls.predictions.decoder.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/rwkv/modeling_rwkv.py b/src/transformers/models/rwkv/modeling_rwkv.py
index 7dec1f26e1a3..8acc8bc2f496 100644
--- a/src/transformers/models/rwkv/modeling_rwkv.py
+++ b/src/transformers/models/rwkv/modeling_rwkv.py
@@ -25,6 +25,7 @@
 from torch import nn
 from torch.nn import CrossEntropyLoss
 
+from ...generation import GenerationMixin
 from ...modeling_utils import PreTrainedModel
 from ...utils import (
     ModelOutput,
@@ -751,7 +752,7 @@ def _bnb_4bit_dequantize_and_rescale(self, target_layer, block_id):
     """,
     RWKV_START_DOCSTRING,
 )
-class RwkvForCausalLM(RwkvPreTrainedModel):
+class RwkvForCausalLM(GenerationMixin, RwkvPreTrainedModel):
     _tied_weights_keys = ["head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/seamless_m4t/modeling_seamless_m4t.py b/src/transformers/models/seamless_m4t/modeling_seamless_m4t.py
index a79d1d4cf2b9..c1e4ad9fedd0 100755
--- a/src/transformers/models/seamless_m4t/modeling_seamless_m4t.py
+++ b/src/transformers/models/seamless_m4t/modeling_seamless_m4t.py
@@ -26,6 +26,7 @@
 
 from ...activations import ACT2FN
 from ...deepspeed import is_deepspeed_zero3_enabled
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -2150,7 +2151,7 @@ def forward(
         embed_tokens_decoder (`nn.Embedding`, *optional*): input embedding of the decoder.
     """,
 )
-class SeamlessM4TTextToUnitForConditionalGeneration(SeamlessM4TPreTrainedModel):
+class SeamlessM4TTextToUnitForConditionalGeneration(GenerationMixin, SeamlessM4TPreTrainedModel):
     _keys_to_ignore_on_load_missing = [
         "vocoder",
         "speech_encoder",
@@ -2656,7 +2657,7 @@ def remove_weight_norm(self):
     "The text-to-text SeamlessM4T Model transformer which can be used for T2TT.",
     SEAMLESS_M4T_START_DOCSTRING,
 )
-class SeamlessM4TForTextToText(SeamlessM4TPreTrainedModel):
+class SeamlessM4TForTextToText(GenerationMixin, SeamlessM4TPreTrainedModel):
     _keys_to_ignore_on_load_missing = ["speech_encoder", "t2u_model", "vocoder"]
     main_input_name = "input_ids"
 
diff --git a/src/transformers/models/seamless_m4t_v2/modeling_seamless_m4t_v2.py b/src/transformers/models/seamless_m4t_v2/modeling_seamless_m4t_v2.py
index a53f544bb34f..fae1eb61ba53 100644
--- a/src/transformers/models/seamless_m4t_v2/modeling_seamless_m4t_v2.py
+++ b/src/transformers/models/seamless_m4t_v2/modeling_seamless_m4t_v2.py
@@ -26,6 +26,7 @@
 
 from ...activations import ACT2FN
 from ...deepspeed import is_deepspeed_zero3_enabled
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -2439,7 +2440,7 @@ def forward(
         embed_tokens_decoder (`nn.Embedding`, *optional*): input embedding of the decoder.
     """,
 )
-class SeamlessM4Tv2TextToUnitForConditionalGeneration(SeamlessM4Tv2PreTrainedModel):
+class SeamlessM4Tv2TextToUnitForConditionalGeneration(GenerationMixin, SeamlessM4Tv2PreTrainedModel):
     _keys_to_ignore_on_load_missing = [
         "vocoder",
         "speech_encoder",
@@ -2914,7 +2915,7 @@ def remove_weight_norm(self):
     SEAMLESS_M4T_V2_START_DOCSTRING,
 )
 # Copied from transformers.models.seamless_m4t.modeling_seamless_m4t.SeamlessM4TForTextToText with SeamlessM4T->SeamlessM4Tv2,SeamlessM4Tv2Tokenizer->SeamlessM4TTokenizer, SeamlessM4Tv2Processor->SeamlessM4TProcessor
-class SeamlessM4Tv2ForTextToText(SeamlessM4Tv2PreTrainedModel):
+class SeamlessM4Tv2ForTextToText(GenerationMixin, SeamlessM4Tv2PreTrainedModel):
     _keys_to_ignore_on_load_missing = ["speech_encoder", "t2u_model", "vocoder"]
     main_input_name = "input_ids"
 
diff --git a/src/transformers/models/speech_to_text/modeling_speech_to_text.py b/src/transformers/models/speech_to_text/modeling_speech_to_text.py
index 8353a172b212..f014a00bc387 100755
--- a/src/transformers/models/speech_to_text/modeling_speech_to_text.py
+++ b/src/transformers/models/speech_to_text/modeling_speech_to_text.py
@@ -22,6 +22,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -1207,7 +1208,7 @@ def forward(
     "The Speech2Text Model with a language modeling head. Can be used for summarization.",
     SPEECH_TO_TEXT_START_DOCSTRING,
 )
-class Speech2TextForConditionalGeneration(Speech2TextPreTrainedModel):
+class Speech2TextForConditionalGeneration(GenerationMixin, Speech2TextPreTrainedModel):
     base_model_prefix = "model"
     _tied_weights_keys = ["lm_head.weight"]
 
diff --git a/src/transformers/models/stablelm/modeling_stablelm.py b/src/transformers/models/stablelm/modeling_stablelm.py
index 1ec4665fcfb7..489f751a27df 100755
--- a/src/transformers/models/stablelm/modeling_stablelm.py
+++ b/src/transformers/models/stablelm/modeling_stablelm.py
@@ -29,6 +29,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
@@ -1106,7 +1107,7 @@ def _update_causal_mask(
 
 
 # Copied from transformers.models.persimmon.modeling_persimmon.PersimmonForCausalLM with PERSIMMON->STABLELM,Persimmon->StableLm
-class StableLmForCausalLM(StableLmPreTrainedModel):
+class StableLmForCausalLM(GenerationMixin, StableLmPreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.__init__ with LLAMA->STABLELM,Llama->StableLm
diff --git a/src/transformers/models/starcoder2/modeling_starcoder2.py b/src/transformers/models/starcoder2/modeling_starcoder2.py
index 90603fd4e51e..f1f322c5d52f 100644
--- a/src/transformers/models/starcoder2/modeling_starcoder2.py
+++ b/src/transformers/models/starcoder2/modeling_starcoder2.py
@@ -29,6 +29,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutputWithPast,
@@ -993,7 +994,7 @@ def _update_causal_mask(
 
 
 # Copied from transformers.models.qwen2.modeling_qwen2.Qwen2ForCausalLM with QWEN2->STARCODER2,Qwen2->Starcoder2
-class Starcoder2ForCausalLM(Starcoder2PreTrainedModel):
+class Starcoder2ForCausalLM(GenerationMixin, Starcoder2PreTrainedModel):
     _tied_weights_keys = ["lm_head.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/switch_transformers/modeling_switch_transformers.py b/src/transformers/models/switch_transformers/modeling_switch_transformers.py
index c5797d4573b7..dcd86d8884af 100644
--- a/src/transformers/models/switch_transformers/modeling_switch_transformers.py
+++ b/src/transformers/models/switch_transformers/modeling_switch_transformers.py
@@ -24,6 +24,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     MoEModelOutput,
     MoEModelOutputWithPastAndCrossAttentions,
@@ -1456,7 +1457,7 @@ def forward(
 @add_start_docstrings(
     """SWITCH_TRANSFORMERS Model with a `language modeling` head on top.""", SWITCH_TRANSFORMERS_START_DOCSTRING
 )
-class SwitchTransformersForConditionalGeneration(SwitchTransformersPreTrainedModel):
+class SwitchTransformersForConditionalGeneration(GenerationMixin, SwitchTransformersPreTrainedModel):
     _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]
 
     def __init__(self, config: SwitchTransformersConfig):
diff --git a/src/transformers/models/t5/modeling_t5.py b/src/transformers/models/t5/modeling_t5.py
index a90101924c5b..3dba75ad8467 100644
--- a/src/transformers/models/t5/modeling_t5.py
+++ b/src/transformers/models/t5/modeling_t5.py
@@ -25,6 +25,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutput,
     BaseModelOutputWithPastAndCrossAttentions,
@@ -1542,7 +1543,7 @@ def forward(
 
 
 @add_start_docstrings("""T5 Model with a `language modeling` head on top.""", T5_START_DOCSTRING)
-class T5ForConditionalGeneration(T5PreTrainedModel):
+class T5ForConditionalGeneration(GenerationMixin, T5PreTrainedModel):
     _keys_to_ignore_on_load_unexpected = [
         "decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight",
     ]
diff --git a/src/transformers/models/trocr/modeling_trocr.py b/src/transformers/models/trocr/modeling_trocr.py
index 04eb40ab2a2f..ad79e3c0ff5e 100644
--- a/src/transformers/models/trocr/modeling_trocr.py
+++ b/src/transformers/models/trocr/modeling_trocr.py
@@ -23,6 +23,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import BaseModelOutputWithPastAndCrossAttentions, CausalLMOutputWithCrossAttentions
 from ...modeling_utils import PreTrainedModel
@@ -736,7 +737,7 @@ def forward(self, *args, **kwargs):
     " [`VisionEncoderDecoder`].",
     TROCR_START_DOCSTRING,
 )
-class TrOCRForCausalLM(TrOCRPreTrainedModel):
+class TrOCRForCausalLM(GenerationMixin, TrOCRPreTrainedModel):
     _tied_weights_keys = ["output_projection.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/udop/modeling_udop.py b/src/transformers/models/udop/modeling_udop.py
index 972248daaae5..f2ea13ffa4fe 100644
--- a/src/transformers/models/udop/modeling_udop.py
+++ b/src/transformers/models/udop/modeling_udop.py
@@ -34,6 +34,7 @@
 )
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_utils import PreTrainedModel
 from ...pytorch_utils import find_pruneable_heads_and_indices, prune_linear_layer
 from ...utils import (
@@ -1679,7 +1680,7 @@ def forward(
     This class is based on [`T5ForConditionalGeneration`], extended to deal with images and layout (2D) data.""",
     UDOP_START_DOCSTRING,
 )
-class UdopForConditionalGeneration(UdopPreTrainedModel):
+class UdopForConditionalGeneration(GenerationMixin, UdopPreTrainedModel):
     _tied_weights_keys = [
         "encoder.embed_tokens.weight",
         "decoder.embed_tokens.weight",
diff --git a/src/transformers/models/umt5/modeling_umt5.py b/src/transformers/models/umt5/modeling_umt5.py
index 3271689540b9..26ed40a5a17a 100644
--- a/src/transformers/models/umt5/modeling_umt5.py
+++ b/src/transformers/models/umt5/modeling_umt5.py
@@ -23,6 +23,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutput,
     BaseModelOutputWithPastAndCrossAttentions,
@@ -1101,7 +1102,7 @@ def forward(
 
 
 @add_start_docstrings("""UMT5 Model with a `language modeling` head on top.""", UMT5_START_DOCSTRING)
-class UMT5ForConditionalGeneration(UMT5PreTrainedModel):
+class UMT5ForConditionalGeneration(GenerationMixin, UMT5PreTrainedModel):
     r"""
     Examples:
 
diff --git a/src/transformers/models/video_llava/modeling_video_llava.py b/src/transformers/models/video_llava/modeling_video_llava.py
index 425d46bd7741..0cc4ec27d4db 100644
--- a/src/transformers/models/video_llava/modeling_video_llava.py
+++ b/src/transformers/models/video_llava/modeling_video_llava.py
@@ -21,9 +21,10 @@
 import torch.utils.checkpoint
 from torch import nn
 
-from ... import PreTrainedModel
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import BaseModelOutputWithPooling, ModelOutput
+from ...modeling_utils import PreTrainedModel
 from ...utils import (
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
@@ -238,7 +239,7 @@ def _supports_sdpa(self):
     """The VideoLlava model which consists of a vision backbone and a language model.""",
     VIDEO_LLAVA_START_DOCSTRING,
 )
-class VideoLlavaForConditionalGeneration(VideoLlavaPreTrainedModel):
+class VideoLlavaForConditionalGeneration(GenerationMixin, VideoLlavaPreTrainedModel):
     def __init__(self, config: VideoLlavaConfig):
         super().__init__(config)
         self.video_tower = AutoModel.from_config(config.vision_config)
diff --git a/src/transformers/models/vipllava/modeling_vipllava.py b/src/transformers/models/vipllava/modeling_vipllava.py
index b1df10fdb3dc..008afd6de7f6 100644
--- a/src/transformers/models/vipllava/modeling_vipllava.py
+++ b/src/transformers/models/vipllava/modeling_vipllava.py
@@ -21,9 +21,10 @@
 import torch.utils.checkpoint
 from torch import nn
 
-from ... import PreTrainedModel
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_outputs import ModelOutput
+from ...modeling_utils import PreTrainedModel
 from ...utils import (
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
@@ -242,7 +243,7 @@ def _supports_sdpa(self):
     VIPLLAVA_START_DOCSTRING,
 )
 # Copied from transformers.models.llava.modeling_llava.LlavaForConditionalGeneration with LLAVA->VIPLLAVA,Llava->VipLlava
-class VipLlavaForConditionalGeneration(VipLlavaPreTrainedModel):
+class VipLlavaForConditionalGeneration(GenerationMixin, VipLlavaPreTrainedModel):
     def __init__(self, config: VipLlavaConfig):
         super().__init__(config)
         self.vision_tower = AutoModel.from_config(config.vision_config)
diff --git a/src/transformers/models/whisper/generation_whisper.py b/src/transformers/models/whisper/generation_whisper.py
index 7da4e49de98c..1073bb040cab 100644
--- a/src/transformers/models/whisper/generation_whisper.py
+++ b/src/transformers/models/whisper/generation_whisper.py
@@ -25,7 +25,7 @@
 
 from transformers.cache_utils import EncoderDecoderCache
 
-from ...generation.configuration_utils import GenerationConfig
+from ...generation import GenerationConfig, GenerationMixin
 from ...generation.logits_process import (
     LogitsProcessorList,
     SuppressTokensAtBeginLogitsProcessor,
@@ -172,7 +172,7 @@ def _pad_to_max_length(
     return sequences
 
 
-class WhisperGenerationMixin:
+class WhisperGenerationMixin(GenerationMixin):
     def _extract_token_timestamps(self, generate_outputs, alignment_heads, time_precision=0.02, num_frames=None):
         """
         Calculates token-level timestamps using the encoder-decoder cross-attentions and dynamic time-warping (DTW) to
diff --git a/src/transformers/models/whisper/modeling_whisper.py b/src/transformers/models/whisper/modeling_whisper.py
index 81f60edbfa98..bb24b68df620 100644
--- a/src/transformers/models/whisper/modeling_whisper.py
+++ b/src/transformers/models/whisper/modeling_whisper.py
@@ -25,6 +25,7 @@
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache, EncoderDecoderCache, StaticCache
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import AttentionMaskConverter
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -1910,7 +1911,7 @@ def forward(self, *args, **kwargs):
     """,
     WHISPER_START_DOCSTRING,
 )
-class WhisperForCausalLM(WhisperPreTrainedModel):
+class WhisperForCausalLM(GenerationMixin, WhisperPreTrainedModel):
     _tied_weights_keys = ["proj_out.weight"]
     main_input_name = "input_ids"
 
diff --git a/src/transformers/models/xglm/modeling_xglm.py b/src/transformers/models/xglm/modeling_xglm.py
index 4f1693583494..191834a9eac5 100755
--- a/src/transformers/models/xglm/modeling_xglm.py
+++ b/src/transformers/models/xglm/modeling_xglm.py
@@ -23,6 +23,7 @@
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask, _prepare_4d_causal_attention_mask
 from ...modeling_outputs import BaseModelOutputWithPastAndCrossAttentions, CausalLMOutputWithCrossAttentions
 from ...modeling_utils import PreTrainedModel
@@ -696,7 +697,7 @@ def forward(
     """,
     XGLM_START_DOCSTRING,
 )
-class XGLMForCausalLM(XGLMPreTrainedModel):
+class XGLMForCausalLM(GenerationMixin, XGLMPreTrainedModel):
     base_model_prefix = "model"
     _tied_weights_keys = ["lm_head.weight"]
 
diff --git a/src/transformers/models/xlm/modeling_xlm.py b/src/transformers/models/xlm/modeling_xlm.py
index 280383630987..ee36c1a6008b 100755
--- a/src/transformers/models/xlm/modeling_xlm.py
+++ b/src/transformers/models/xlm/modeling_xlm.py
@@ -27,6 +27,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import gelu
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutput,
     MaskedLMOutput,
@@ -657,7 +658,7 @@ def forward(self, x, y=None):
     """,
     XLM_START_DOCSTRING,
 )
-class XLMWithLMHeadModel(XLMPreTrainedModel):
+class XLMWithLMHeadModel(GenerationMixin, XLMPreTrainedModel):
     _tied_weights_keys = ["pred_layer.proj.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/xlm_roberta/modeling_xlm_roberta.py b/src/transformers/models/xlm_roberta/modeling_xlm_roberta.py
index 3ac94e75f92f..ee5036dbedd1 100644
--- a/src/transformers/models/xlm_roberta/modeling_xlm_roberta.py
+++ b/src/transformers/models/xlm_roberta/modeling_xlm_roberta.py
@@ -25,6 +25,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN, gelu
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import (
     _prepare_4d_attention_mask_for_sdpa,
     _prepare_4d_causal_attention_mask_for_sdpa,
@@ -1006,7 +1007,7 @@ def forward(
     XLM_ROBERTA_START_DOCSTRING,
 )
 # Copied from transformers.models.roberta.modeling_roberta.RobertaForCausalLM with Roberta->XLMRoberta, ROBERTA->XLM_ROBERTA
-class XLMRobertaForCausalLM(XLMRobertaPreTrainedModel):
+class XLMRobertaForCausalLM(GenerationMixin, XLMRobertaPreTrainedModel):
     _tied_weights_keys = ["lm_head.decoder.weight", "lm_head.decoder.bias"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/xlm_roberta_xl/modeling_xlm_roberta_xl.py b/src/transformers/models/xlm_roberta_xl/modeling_xlm_roberta_xl.py
index f66a32291794..a677c61cba1c 100644
--- a/src/transformers/models/xlm_roberta_xl/modeling_xlm_roberta_xl.py
+++ b/src/transformers/models/xlm_roberta_xl/modeling_xlm_roberta_xl.py
@@ -24,6 +24,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN, gelu
+from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import (
     _prepare_4d_attention_mask_for_sdpa,
     _prepare_4d_causal_attention_mask_for_sdpa,
@@ -986,7 +987,7 @@ def forward(
     """XLM-RoBERTa-XL Model with a `language modeling` head on top for CLM fine-tuning.""",
     XLM_ROBERTA_XL_START_DOCSTRING,
 )
-class XLMRobertaXLForCausalLM(XLMRobertaXLPreTrainedModel):
+class XLMRobertaXLForCausalLM(GenerationMixin, XLMRobertaXLPreTrainedModel):
     _tied_weights_keys = ["lm_head.decoder.weight", "lm_head.decoder.bias"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/xlnet/modeling_xlnet.py b/src/transformers/models/xlnet/modeling_xlnet.py
index 5d424ebe12dd..5d7e59c215f9 100755
--- a/src/transformers/models/xlnet/modeling_xlnet.py
+++ b/src/transformers/models/xlnet/modeling_xlnet.py
@@ -26,6 +26,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
+from ...generation import GenerationMixin
 from ...modeling_utils import PoolerAnswerClass, PoolerEndLogits, PoolerStartLogits, PreTrainedModel, SequenceSummary
 from ...pytorch_utils import apply_chunking_to_forward
 from ...utils import (
@@ -1286,7 +1287,7 @@ def forward(
     """,
     XLNET_START_DOCSTRING,
 )
-class XLNetLMHeadModel(XLNetPreTrainedModel):
+class XLNetLMHeadModel(GenerationMixin, XLNetPreTrainedModel):
     _tied_weights_keys = ["lm_loss.weight"]
 
     def __init__(self, config):
diff --git a/src/transformers/models/xmod/modeling_xmod.py b/src/transformers/models/xmod/modeling_xmod.py
index b1ca8116a72a..0586b725f7d4 100644
--- a/src/transformers/models/xmod/modeling_xmod.py
+++ b/src/transformers/models/xmod/modeling_xmod.py
@@ -23,6 +23,7 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN, gelu
+from ...generation import GenerationMixin
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
     BaseModelOutputWithPoolingAndCrossAttentions,
@@ -956,7 +957,7 @@ def forward(
     "X-MOD Model with a `language modeling` head on top for CLM fine-tuning.",
     XMOD_START_DOCSTRING,
 )
-class XmodForCausalLM(XmodPreTrainedModel):
+class XmodForCausalLM(GenerationMixin, XmodPreTrainedModel):
     _tied_weights_keys = ["lm_head.decoder.weight", "lm_head.decoder.bias"]
 
     # Copied from transformers.models.roberta.modeling_roberta.RobertaForCausalLM.__init__ with Roberta->Xmod