transformers 5.0.0rc0__py3-none-any.whl → 5.0.0rc1__py3-none-any.whl

transformers/models/qwen3_vl_moe/modeling_qwen3_vl_moe.py

@@ -31,7 +31,7 @@ from ... import initialization as init
 from ...activations import ACT2FN
 from ...cache_utils import Cache, DynamicCache
 from ...generation import GenerationMixin
-from ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub
+from ...integrations import use_kernel_forward_from_hub, use_kernel_func_from_hub, use_kernelized_func
 from ...masking_utils import create_causal_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
@@ -39,8 +39,8 @@ from ...modeling_outputs import BaseModelOutputWithPast, ModelOutput
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, auto_docstring, is_torchdynamo_compiling
-from ...utils.generic import OutputRecorder, check_model_inputs
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple
+from ...utils.generic import OutputRecorder, check_model_inputs, maybe_autocast
 from .configuration_qwen3_vl_moe import Qwen3VLMoeConfig, Qwen3VLMoeTextConfig, Qwen3VLMoeVisionConfig
 
 
@@ -226,6 +226,7 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
     return q_embed, k_embed
 
 
+@use_kernelized_func(apply_rotary_pos_emb)
 class Qwen3VLMoeTextAttention(nn.Module):
     """Multi-headed attention from 'Attention Is All You Need' paper"""
 
@@ -252,7 +253,6 @@ class Qwen3VLMoeTextAttention(nn.Module):
         self.o_proj = nn.Linear(
             config.num_attention_heads * self.head_dim, config.hidden_size, bias=config.attention_bias
         )
-        self.rotary_fn = apply_rotary_pos_emb
         self.q_norm = Qwen3VLMoeTextRMSNorm(
             self.head_dim, eps=config.rms_norm_eps
         )  # unlike olmo, only on the head dim!
@@ -860,7 +860,7 @@ class Qwen3VLMoeTextRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, :, None, :].float()  # shape (3, bs, 1, positions)
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(2, 3)
             freqs = self.apply_interleaved_mrope(freqs, self.mrope_section)
             emb = torch.cat((freqs, freqs), dim=-1)
@@ -1358,44 +1358,19 @@ class Qwen3VLMoeModel(Qwen3VLMoePreTrainedModel):
             deepstack_visual_embeds = deepstack_video_embeds
 
         if position_ids is None:
-            attention_mask_tensor = (
-                attention_mask if not isinstance(attention_mask, dict) else attention_mask["full_attention"]
-            )
-            if attention_mask_tensor is not None and attention_mask_tensor.ndim == 4:
-                attention_mask_tensor = torch.diagonal(attention_mask_tensor[:, 0], dim1=1, dim2=2)
-                # Only apply conversion for floating point tensors (inverted masks)
-                if attention_mask_tensor.dtype.is_floating_point:
-                    attention_mask_tensor = attention_mask_tensor / torch.finfo(attention_mask_tensor.dtype).min
-                    attention_mask_tensor = (1.0 - attention_mask_tensor).int()
-
-            # Calculate RoPE index once per generation in the pre-fill stage only.
-            # When compiling, we can't check tensor values thus we check only input length
-            # It is safe to assume that `length!=1` means we're in pre-fill because compiled
-            # models currently cannot do asssisted decoding
-            prefill_compiled_stage = is_torchdynamo_compiling() and (
-                (input_ids is not None and input_ids.shape[1] != 1)
-                or (inputs_embeds is not None and inputs_embeds.shape[1] != 1)
-            )
-            prefill_noncompiled_stage = not is_torchdynamo_compiling() and (
-                (cache_position is not None and cache_position[0] == 0)
-                or (past_key_values is None or past_key_values.get_seq_length() == 0)
-            )
-            if (prefill_compiled_stage or prefill_noncompiled_stage) or self.rope_deltas is None:
+            past_key_values_length = 0 if past_key_values is None else past_key_values.get_seq_length()
+            if self.rope_deltas is None or past_key_values_length == 0:
                 position_ids, rope_deltas = self.get_rope_index(
                     input_ids,
                     image_grid_thw,
                     video_grid_thw,
-                    attention_mask=attention_mask_tensor,
+                    attention_mask=attention_mask,
                 )
                 self.rope_deltas = rope_deltas
             # then use the prev pre-calculated rope-deltas to get the correct position ids
             else:
                 batch_size, seq_length, _ = inputs_embeds.shape
-                delta = (
-                    (cache_position[0] + self.rope_deltas).to(inputs_embeds.device)
-                    if cache_position is not None
-                    else 0
-                )
+                delta = (past_key_values_length + self.rope_deltas).to(inputs_embeds.device)
                 position_ids = torch.arange(seq_length, device=inputs_embeds.device)
                 position_ids = position_ids.view(1, -1).expand(batch_size, -1)
                 if cache_position is not None:  # otherwise `deltas` is an int `0`
@@ -1532,7 +1507,7 @@ class Qwen3VLMoeForConditionalGeneration(Qwen3VLMoePreTrainedModel, GenerationMi
     def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: Optional[torch.LongTensor] = None):
         return self.model.get_image_features(pixel_values, image_grid_thw)
 
-    @check_model_inputs
+    @can_return_tuple
     def forward(
         self,
         input_ids: torch.LongTensor = None,
@@ -1642,6 +1617,8 @@ class Qwen3VLMoeForConditionalGeneration(Qwen3VLMoePreTrainedModel, GenerationMi
             aux_loss=aux_loss,
             logits=logits,
             past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
             rope_deltas=outputs.rope_deltas,
         )
 
@@ -1677,8 +1654,33 @@ class Qwen3VLMoeForConditionalGeneration(Qwen3VLMoePreTrainedModel, GenerationMi
             **kwargs,
         )
 
-        # Qwen3VLMoe position_ids are prepareed with rope_deltas in forward
-        model_inputs["position_ids"] = None
+        # Qwen3VLMoe position_ids are prepared with rope_deltas
+        if position_ids is None:
+            # Calculate RoPE index once per generation in the pre-fill stage only.
+            # When compiling, we can't check tensor values thus we check only input length
+            # It is safe to assume that `length!=1` means we're in pre-fill because compiled
+            # models currently cannot do asssisted decoding
+            if model_inputs["cache_position"][0] == 0 or self.model.rope_deltas is None:
+                vision_positions, rope_deltas = self.model.get_rope_index(
+                    model_inputs.get("input_ids", None),
+                    image_grid_thw=image_grid_thw,
+                    video_grid_thw=video_grid_thw,
+                    attention_mask=attention_mask,
+                )
+                self.model.rope_deltas = rope_deltas
+            # then use the prev pre-calculated rope-deltas to get the correct position ids
+            elif "position_ids" in model_inputs:
+                batch_size, seq_length = model_inputs["position_ids"].shape
+                device = model_inputs["position_ids"].device
+                position_ids = torch.arange(seq_length, device=device)
+                position_ids = position_ids.view(1, 1, -1).expand(3, batch_size, -1)
+                delta = cache_position[0] + self.model.rope_deltas
+                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)
+                vision_positions = position_ids + delta.expand_as(position_ids)
+
+            # Concatenate "text + vision" positions into [4, bs, seq-len]
+            text_positions = model_inputs["position_ids"][None, ...]
+            model_inputs["position_ids"] = torch.cat([text_positions, vision_positions], dim=0)
 
         if cache_position[0] != 0:
             model_inputs["pixel_values"] = None

transformers/models/qwen3_vl_moe/modular_qwen3_vl_moe.py

@@ -26,7 +26,7 @@ from ...configuration_utils import PreTrainedConfig
 from ...modeling_rope_utils import RopeParameters
 from ...modeling_utils import PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, logging
+from ...utils import TransformersKwargs, can_return_tuple, logging
 from ..qwen3_moe.modeling_qwen3_moe import (
     Qwen3MoeDecoderLayer,
     Qwen3MoePreTrainedModel,
@@ -387,6 +387,7 @@ class Qwen3VLMoeModel(Qwen3VLModel):
 
 
 class Qwen3VLMoeForConditionalGeneration(Qwen3VLForConditionalGeneration):
+    @can_return_tuple
     def forward(
         self,
         input_ids: torch.LongTensor = None,
@@ -496,6 +497,8 @@ class Qwen3VLMoeForConditionalGeneration(Qwen3VLForConditionalGeneration):
             aux_loss=aux_loss,
             logits=logits,
             past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
             rope_deltas=outputs.rope_deltas,
         )
 

transformers/models/rag/modeling_rag.py

@@ -439,6 +439,7 @@ class RagModel(RagPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         output_retrieved: Optional[bool] = None,
         n_docs: Optional[int] = None,
+        **kwargs,
     ) -> Union[tuple[torch.Tensor], RetrievAugLMOutput]:
         r"""
         input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):

transformers/models/recurrent_gemma/modeling_recurrent_gemma.py

@@ -31,6 +31,7 @@ from ...modeling_outputs import BaseModelOutputWithNoAttention, CausalLMOutput
 from ...modeling_rope_utils import dynamic_rope_update
 from ...modeling_utils import PreTrainedModel
 from ...utils import auto_docstring, logging
+from ...utils.generic import maybe_autocast
 from ...utils.import_utils import is_tracing
 from .configuration_recurrent_gemma import RecurrentGemmaConfig
 
@@ -121,7 +122,7 @@ class RecurrentGemmaRotaryEmbedding(nn.Module):
         position_ids_expanded = position_ids[:, None, :].float()
 
         device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):  # Force float32
+        with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
             freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
             emb = torch.cat((freqs, freqs), dim=-1)
             cos = emb.cos() * self.attention_scaling
@@ -460,6 +461,7 @@ class RecurrentGemmaRecurrentBlock(nn.Module):
         use_cache: bool = True,
     ) -> tuple[torch.Tensor, dict[str, torch.Tensor]]:
         _, seq_len, _ = input_states.shape
+        batch_size = input_states.shape[0]
 
         y_branch = self.linear_y(input_states)
         y_branch = self.act_fn(y_branch)
@@ -468,6 +470,17 @@ class RecurrentGemmaRecurrentBlock(nn.Module):
         x_branch = x_branch.transpose(1, 2)
 
         if use_cache:
+            # Check if cache needs initialization (None or batch size mismatch)
+            if self.conv1d_state is None or self.conv1d_state.shape[0] != batch_size:
+                self.conv1d_state = torch.zeros(
+                    (batch_size, self.hidden_size, self.conv1d_width - 1),
+                    device=input_states.device,
+                    dtype=input_states.dtype,
+                )
+                self.rg_lru.recurrent_states = torch.zeros(
+                    (batch_size, self.lru_width), device=input_states.device, dtype=torch.float32
+                )
+
             if cache_position.shape[0] != 1:  # prefill
                 self.conv1d_state = nn.functional.pad(x_branch, (self.conv1d_width - x_branch.shape[-1] - 1, 0))
                 x_branch = self.conv_1d(x_branch)[..., :seq_len]
@@ -643,6 +656,7 @@ class RecurrentGemmaModel(RecurrentGemmaPreTrainedModel):
         use_cache: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[tuple, BaseModelOutputWithNoAttention]:
         output_hidden_states = (
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states

transformers/models/reformer/modeling_reformer.py

@@ -1946,6 +1946,7 @@ class ReformerModel(ReformerPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         output_attentions: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[tuple, ReformerModelOutput]:
         r"""
         input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
@@ -2297,6 +2298,7 @@ class ReformerForMaskedLM(ReformerPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         output_attentions: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[tuple, MaskedLMOutput]:
         r"""
         input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
@@ -2428,6 +2430,7 @@ class ReformerForSequenceClassification(ReformerPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         output_attentions: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[tuple, SequenceClassifierOutput]:
         r"""
         input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
@@ -2577,6 +2580,7 @@ class ReformerForQuestionAnswering(ReformerPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         output_attentions: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[tuple, QuestionAnsweringModelOutput]:
         r"""
         input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):

transformers/models/reformer/tokenization_reformer.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 """Tokenization class for model Reformer."""
 
-from typing import Optional
+from typing import Optional, Union
 
 from tokenizers import Regex, Tokenizer, decoders, normalizers, pre_tokenizers
 from tokenizers.models import BPE
@@ -60,38 +60,27 @@ class ReformerTokenizer(TokenizersBackend):
             The token used for padding, for example when batching sequences of different lengths.
         additional_special_tokens (`list[str]`, *optional*):
             Additional special tokens used by the tokenizer.
-        vocab (`dict`, *optional*):
-            Custom vocabulary dictionary. If not provided, vocabulary is loaded from vocab_file.
-        merges (`list`, *optional*):
-            Custom merges list. If not provided, merges are loaded from vocab_file.
+        vocab (`str` or `dict[str, int]`, *optional*):
+            Custom vocabulary dictionary. If not provided, vocabulary is loaded from `vocab_file`.
+        merges (`str` or `list[str]`, *optional*):
+            Custom merges list. If not provided, merges are loaded from `vocab_file`.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
     model_input_names = ["input_ids", "attention_mask"]
-    slow_tokenizer_class = None
+    model = BPE
 
     def __init__(
         self,
-        vocab_file: Optional[str] = None,
+        vocab: Optional[Union[str, dict[str, int]]] = None,
+        merges: Optional[Union[str, list[str]]] = None,
         eos_token: str = "</s>",
         unk_token: str = "<unk>",
         additional_special_tokens: Optional[list] = None,
-        vocab: Optional[dict] = None,
-        merges: Optional[list] = None,
         **kwargs,
     ):
-        self.vocab_file = vocab_file
-
-        if vocab is not None:
-            self._vocab = vocab
-        else:
-            self._vocab = {}
-
-        if merges is not None:
-            # Convert lists to tuples if necessary (happens when loading from JSON)
-            self._merges = [tuple(merge) if isinstance(merge, list) else merge for merge in merges]
-        else:
-            self._merges = []
+        self._vocab = vocab or {}
+        self._merges = merges or []
 
         self._tokenizer = Tokenizer(
             BPE(
@@ -106,10 +95,7 @@ class ReformerTokenizer(TokenizersBackend):
 
         self._tokenizer.normalizer = normalizers.Sequence(
             [
-                normalizers.Replace("\n", " "),
-                normalizers.Replace("\r", " "),
-                normalizers.Replace("\t", " "),
-                normalizers.Replace(Regex(r" {2,}"), " "),
+                normalizers.Replace(Regex(r"\s{2,}|[\n\r\t]"), " "),
                 normalizers.NFC(),
                 normalizers.Strip(left=False, right=True),
             ]
@@ -118,10 +104,7 @@ class ReformerTokenizer(TokenizersBackend):
         self._tokenizer.pre_tokenizer = pre_tokenizers.Metaspace(replacement="▁", prepend_scheme="always")
         self._tokenizer.decoder = decoders.Metaspace(replacement="▁", prepend_scheme="always")
 
-        tokenizer_object = self._tokenizer
-
         super().__init__(
-            tokenizer_object=tokenizer_object,
             eos_token=eos_token,
             unk_token=unk_token,
             additional_special_tokens=additional_special_tokens or [],

transformers/models/regnet/modeling_regnet.py

@@ -294,7 +294,11 @@ class RegNetModel(RegNetPreTrainedModel):
 
     @auto_docstring
     def forward(
-        self, pixel_values: Tensor, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None
+        self,
+        pixel_values: Tensor,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> BaseModelOutputWithPoolingAndNoAttention:
         output_hidden_states = (
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
@@ -348,6 +352,7 @@ class RegNetForImageClassification(RegNetPreTrainedModel):
         labels: Optional[torch.LongTensor] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> ImageClassifierOutputWithNoAttention:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):

transformers/models/rembert/modeling_rembert.py

@@ -540,6 +540,7 @@ class RemBertModel(RemBertPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.Tensor] = None,
+        **kwargs,
     ) -> Union[tuple, BaseModelOutputWithPoolingAndCrossAttentions]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -659,6 +660,7 @@ class RemBertForMaskedLM(RemBertPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[tuple, MaskedLMOutput]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -857,6 +859,7 @@ class RemBertForSequenceClassification(RemBertPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[tuple, SequenceClassifierOutput]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
@@ -940,6 +943,7 @@ class RemBertForMultipleChoice(RemBertPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[tuple, MultipleChoiceModelOutput]:
         r"""
         input_ids (`torch.LongTensor` of shape `(batch_size, num_choices, sequence_length)`):
@@ -1043,6 +1047,7 @@ class RemBertForTokenClassification(RemBertPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[tuple, TokenClassifierOutput]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1109,6 +1114,7 @@ class RemBertForQuestionAnswering(RemBertPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[tuple, QuestionAnsweringModelOutput]:
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 

transformers/models/rembert/tokenization_rembert.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 """Tokenization classes for RemBert model."""
 
-from typing import Optional
+from typing import Optional, Union
 
 from tokenizers import Regex, Tokenizer, decoders, normalizers, pre_tokenizers, processors
 from tokenizers.models import Unigram
@@ -74,11 +74,11 @@ class RemBertTokenizer(TokenizersBackend):
 
     vocab_files_names = VOCAB_FILES_NAMES
     model_input_names = ["input_ids", "attention_mask"]
-    slow_tokenizer_class = None
+    model = Unigram
 
     def __init__(
         self,
-        vocab_file: Optional[str] = None,
+        vocab: Optional[Union[str, list[tuple[str, float]]]] = None,
         do_lower_case: bool = False,
         keep_accents: bool = False,
         bos_token: str = "[CLS]",
@@ -90,11 +90,8 @@ class RemBertTokenizer(TokenizersBackend):
         mask_token: str = "[MASK]",
         add_prefix_space: bool = True,
         remove_space: bool = True,
-        vocab: Optional[dict] = None,
-        merges: Optional[list] = None,
         **kwargs,
     ):
-        self.vocab_file = vocab_file
         self.remove_space = remove_space
         self.do_lower_case = do_lower_case
         self.keep_accents = keep_accents
@@ -147,11 +144,7 @@ class RemBertTokenizer(TokenizersBackend):
         self._tokenizer.pre_tokenizer = pre_tokenizers.Metaspace(replacement="▁", prepend_scheme=prepend_scheme)
 
         self._tokenizer.decoder = decoders.Metaspace(replacement="▁", prepend_scheme=prepend_scheme)
-
-        tokenizer_object = self._tokenizer
-
         super().__init__(
-            tokenizer_object=tokenizer_object,
             add_prefix_space=add_prefix_space,
             do_lower_case=do_lower_case,
             keep_accents=keep_accents,

transformers/models/resnet/modeling_resnet.py

@@ -280,7 +280,11 @@ class ResNetModel(ResNetPreTrainedModel):
 
     @auto_docstring
     def forward(
-        self, pixel_values: Tensor, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None
+        self,
+        pixel_values: Tensor,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> BaseModelOutputWithPoolingAndNoAttention:
         output_hidden_states = (
             output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
@@ -333,6 +337,7 @@ class ResNetForImageClassification(ResNetPreTrainedModel):
         labels: Optional[torch.LongTensor] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> ImageClassifierOutputWithNoAttention:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
@@ -380,7 +385,11 @@ class ResNetBackbone(ResNetPreTrainedModel, BackboneMixin):
 
     @auto_docstring
     def forward(
-        self, pixel_values: Tensor, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None
+        self,
+        pixel_values: Tensor,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> BackboneOutput:
         r"""
         Examples:

transformers/models/roberta/tokenization_roberta.py

@@ -14,7 +14,7 @@
 # limitations under the License.
 """Tokenization classes for RoBERTa."""
 
-from typing import Optional
+from typing import Optional, Union
 
 from tokenizers import Tokenizer, decoders, pre_tokenizers, processors
 from tokenizers.models import BPE
@@ -59,6 +59,10 @@ class RobertaTokenizer(TokenizersBackend):
     this superclass for more information regarding those methods.
 
     Args:
+        vocab (`str`, `dict` or `list`, *optional*):
+            Custom vocabulary dictionary. If not provided, vocabulary is loaded from vocab_file.
+        merges (`str` or `list`, *optional*):
+            Custom merges list. If not provided, merges are loaded from merges_file.
         errors (`str`, *optional*, defaults to `"replace"`):
             Paradigm to follow when decoding bytes to UTF-8. See
             [bytes.decode](https://docs.python.org/3/library/stdtypes.html#bytes.decode) for more information.
@@ -102,18 +106,16 @@ class RobertaTokenizer(TokenizersBackend):
             other word. (RoBERTa tokenizer detect beginning of words by the preceding space).
         trim_offsets (`bool`, *optional*, defaults to `True`):
             Whether the post processing step should trim offsets to avoid including whitespaces.
-        vocab (`dict`, *optional*):
-            Custom vocabulary dictionary. If not provided, vocabulary is loaded from vocab_file.
-        merges (`list`, *optional*):
-            Custom merges list. If not provided, merges are loaded from merges_file.
     """
 
     vocab_files_names = VOCAB_FILES_NAMES
     model_input_names = ["input_ids", "attention_mask"]
-    slow_tokenizer_class = None
+    model = BPE
 
     def __init__(
         self,
+        vocab: Optional[Union[str, dict[str, int]]] = None,
+        merges: Optional[Union[str, list[str]]] = None,
         errors: str = "replace",
         bos_token: str = "<s>",
         eos_token: str = "</s>",
@@ -124,30 +126,22 @@ class RobertaTokenizer(TokenizersBackend):
         mask_token: str = "<mask>",
         add_prefix_space: bool = False,
         trim_offsets: bool = True,
-        vocab: Optional[dict] = None,
-        merges: Optional[list] = None,
         **kwargs,
     ):
         self.add_prefix_space = add_prefix_space
         self.trim_offsets = trim_offsets
 
-        if vocab is not None:
-            self._vocab = (
-                {token: idx for idx, (token, _score) in enumerate(vocab)} if isinstance(vocab, list) else vocab
-            )
-        else:
-            self._vocab = {
+        if vocab is None:
+            vocab = {
                 str(pad_token): 0,
                 str(unk_token): 1,
                 str(cls_token): 2,
                 str(sep_token): 3,
                 str(mask_token): 4,
             }
+        self._vocab = vocab
 
-        if merges is not None:
-            self._merges = [tuple(merge) if isinstance(merge, list) else merge for merge in merges]
-        else:
-            self._merges = []
+        self._merges = merges or []
 
         self._tokenizer = Tokenizer(
             BPE(
@@ -162,17 +156,8 @@ class RobertaTokenizer(TokenizersBackend):
 
         self._tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=add_prefix_space)
         self._tokenizer.decoder = decoders.ByteLevel()
-        self._tokenizer.post_processor = processors.RobertaProcessing(
-            sep=(str(sep_token), self._vocab.get(str(sep_token), 3)),
-            cls=(str(cls_token), self._vocab.get(str(cls_token), 2)),
-            add_prefix_space=add_prefix_space,
-            trim_offsets=trim_offsets,
-        )
-
-        tokenizer_object = self._tokenizer
 
         super().__init__(
-            tokenizer_object=tokenizer_object,
             errors=errors,
             bos_token=bos_token,
             eos_token=eos_token,
@@ -185,6 +170,12 @@ class RobertaTokenizer(TokenizersBackend):
             trim_offsets=trim_offsets,
             **kwargs,
         )
+        self._tokenizer.post_processor = processors.RobertaProcessing(
+            sep=(str(sep_token), self.sep_token_id),
+            cls=(str(cls_token), self.cls_token_id),
+            add_prefix_space=add_prefix_space,
+            trim_offsets=trim_offsets,
+        )
 
 
 __all__ = ["RobertaTokenizer"]

transformers/models/roformer/modeling_roformer.py

@@ -693,6 +693,7 @@ class RoFormerModel(RoFormerPreTrainedModel):
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
         cache_position: Optional[torch.Tensor] = None,
+        **kwargs,
     ) -> Union[BaseModelOutputWithPastAndCrossAttentions, tuple[torch.Tensor]]:
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -821,6 +822,7 @@ class RoFormerForMaskedLM(RoFormerPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[MaskedLMOutput, tuple[torch.Tensor]]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1035,6 +1037,7 @@ class RoFormerForSequenceClassification(RoFormerPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[SequenceClassifierOutput, tuple[torch.Tensor]]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
@@ -1114,6 +1117,7 @@ class RoFormerForMultipleChoice(RoFormerPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[MultipleChoiceModelOutput, tuple[torch.Tensor]]:
         r"""
         input_ids (`torch.LongTensor` of shape `(batch_size, num_choices, sequence_length)`):
@@ -1210,6 +1214,7 @@ class RoFormerForTokenClassification(RoFormerPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[TokenClassifierOutput, tuple[torch.Tensor]]:
         r"""
         labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
@@ -1275,6 +1280,7 @@ class RoFormerForQuestionAnswering(RoFormerPreTrainedModel):
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        **kwargs,
     ) -> Union[QuestionAnsweringModelOutput, tuple[torch.Tensor]]:
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict