transformers 5.0.0rc3__py3-none-any.whl → 5.1.0__py3-none-any.whl

transformers/models/paddleocr_vl/modeling_paddleocr_vl.py

@@ -42,8 +42,8 @@ from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPast, BaseMo
 from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
-from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging, torch_int
-from ...utils.generic import check_model_inputs, maybe_autocast
+from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging, torch_compilable_check, torch_int
+from ...utils.generic import check_model_inputs, is_flash_attention_requested, maybe_autocast
 from .configuration_paddleocr_vl import PaddleOCRTextConfig, PaddleOCRVisionConfig, PaddleOCRVLConfig
 
 
@@ -340,9 +340,9 @@ class PaddleOCRAttention(nn.Module):
             cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
             key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx, cache_kwargs)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         attn_output, attn_weights = attention_interface(
             self,
@@ -544,42 +544,6 @@ class PaddleOCRTextModel(PaddleOCRVLPreTrainedModel):
         )
 
 
-class PaddleOCRVisionModel(PaddleOCRVLPreTrainedModel):
-    config: PaddleOCRVisionConfig
-    main_input_name = "pixel_values"
-    input_modalities = "image"
-
-    def __init__(self, config: PaddleOCRVisionConfig):
-        super().__init__(config)
-
-        self.vision_model = PaddleOCRVisionTransformer(config)
-
-        # Initialize weights and apply final processing
-        self.post_init()
-
-    def forward(
-        self,
-        pixel_values: torch.FloatTensor,
-        cu_seqlens: torch.Tensor,
-        image_grid_thw: list[tuple[int, int, int] | list[tuple[int, int, int]]] | None = None,
-        **kwargs,
-    ) -> BaseModelOutputWithPooling:
-        """
-        Args:
-            pixel_values (`torch.FloatTensor` of shape `(batch_size, sequence_length, image_channels, patch_size, patch_size)`):
-                The tensors corresponding to the input images.
-            cu_seqlens (`torch.Tensor` of shape `(num_images + 1,)`):
-                The cumulative sequence lengths of each image or video feature.
-            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
-                The temporal, height and width of feature shape of each image in LLM.
-        """
-        return self.vision_model(
-            pixel_values=pixel_values,
-            cu_seqlens=cu_seqlens,
-            image_grid_thw=image_grid_thw,
-        )
-
-
 class PaddleOCRVisionEmbeddings(nn.Module):
     def __init__(self, config: PaddleOCRVisionConfig):
         super().__init__()
@@ -731,11 +695,11 @@ class PaddleOCRVisionAttention(nn.Module):
         key_states = key_states.transpose(0, 1).unsqueeze(0)
         value_states = value_states.transpose(0, 1).unsqueeze(0)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
-        if self.config._attn_implementation == "flash_attention_2":
+        if is_flash_attention_requested(self.config):
             # Flash Attention 2: Use cu_seqlens for variable length attention
             max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
             attn_output, attn_weights = attention_interface(
@@ -970,6 +934,47 @@ class PaddleOCRVisionTransformer(PaddleOCRVLPreTrainedModel):
         )
 
 
+class PaddleOCRVisionModel(PaddleOCRVLPreTrainedModel):
+    config: PaddleOCRVisionConfig
+    main_input_name = "pixel_values"
+    input_modalities = "image"
+    _can_record_outputs = {
+        "hidden_states": PaddleOCRVisionEncoderLayer,
+        "attentions": PaddleOCRVisionAttention,
+    }
+
+    def __init__(self, config: PaddleOCRVisionConfig):
+        super().__init__(config)
+
+        self.vision_model = PaddleOCRVisionTransformer(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @check_model_inputs(tie_last_hidden_states=False)
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        cu_seqlens: torch.Tensor,
+        image_grid_thw: list[tuple[int, int, int] | list[tuple[int, int, int]]] | None = None,
+        **kwargs,
+    ) -> tuple | BaseModelOutputWithPooling:
+        """
+        Args:
+            pixel_values (`torch.FloatTensor` of shape `(batch_size, sequence_length, image_channels, patch_size, patch_size)`):
+                The tensors corresponding to the input images.
+            cu_seqlens (`torch.Tensor` of shape `(num_images + 1,)`):
+                The cumulative sequence lengths of each image or video feature.
+            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+                The temporal, height and width of feature shape of each image in LLM.
+        """
+        return self.vision_model(
+            pixel_values=pixel_values,
+            cu_seqlens=cu_seqlens,
+            image_grid_thw=image_grid_thw,
+        )
+
+
 @dataclass
 @auto_docstring(
     custom_intro="""
@@ -1196,33 +1201,19 @@ class PaddleOCRVLModel(PaddleOCRVLPreTrainedModel):
 
             return position_ids, mrope_position_deltas
 
-    def get_video_features(
-        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: torch.LongTensor | None = None
-    ):
-        """
-        Encodes videos into continuous embeddings that can be forwarded to the language model.
-
-        Args:
-            pixel_values_videos (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
-                The tensors corresponding to the input videos.
-            video_grid_thw (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
-                The temporal, height and width of feature shape of each video in LLM.
-        """
-        pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
-        video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
-        split_sizes = (video_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
-        video_embeds = torch.split(video_embeds, split_sizes)
-        return video_embeds
-
-    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: torch.LongTensor | None = None):
-        """
-        Encodes images into continuous embeddings that can be forwarded to the language model.
-
-        Args:
-            pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
-                The tensors corresponding to the input images.
-            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
-                The temporal, height and width of feature shape of each image in LLM.
+    @can_return_tuple
+    @auto_docstring
+    def get_image_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        image_grid_thw: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+            The tensors corresponding to the input images.
+        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+            The temporal, height and width of feature shape of each image in LLM.
         """
         pixel_values = pixel_values.type(self.visual.dtype).unsqueeze(0)
         cu_seqlens = torch.repeat_interleave(image_grid_thw[:, 1] * image_grid_thw[:, 2], image_grid_thw[:, 0]).cumsum(
@@ -1238,10 +1229,14 @@ class PaddleOCRVLModel(PaddleOCRVLPreTrainedModel):
             pixel_values=pixel_values,
             image_grid_thw=image_grid_thw,
             cu_seqlens=cu_seqlens,
+            return_dict=True,
+            **kwargs,
         )
         image_embeds = vision_outputs.last_hidden_state
         image_embeds = self.projector(image_embeds, image_grid_thw)
-        return image_embeds
+        vision_outputs.pooler_output = image_embeds
+
+        return vision_outputs
 
     def get_placeholder_mask(
         self, input_ids: torch.LongTensor, inputs_embeds: torch.FloatTensor, image_features: torch.FloatTensor
@@ -1261,10 +1256,10 @@ class PaddleOCRVLModel(PaddleOCRVLPreTrainedModel):
         n_image_tokens = special_image_mask.sum()
         special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
         n_image_features = image_features.shape[0] * image_features.shape[1]
-        if inputs_embeds[special_image_mask].numel() != image_features.numel():
-            raise ValueError(
-                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
-            )
+        torch_compilable_check(
+            inputs_embeds[special_image_mask].numel() == image_features.numel(),
+            f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}",
+        )
         return special_image_mask
 
     @can_return_tuple
@@ -1292,9 +1287,8 @@ class PaddleOCRVLModel(PaddleOCRVLPreTrainedModel):
             inputs_embeds = self.language_model.embed_tokens(input_ids)
 
         if pixel_values is not None:
-            image_embeds = self.get_image_features(pixel_values, image_grid_thw).to(
-                inputs_embeds.device, inputs_embeds.dtype
-            )
+            image_embeds = self.get_image_features(pixel_values, image_grid_thw, return_dict=True).pooler_output
+            image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
             image_mask = self.get_placeholder_mask(input_ids, inputs_embeds=inputs_embeds, image_features=image_embeds)
             inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
 
@@ -1360,13 +1354,20 @@ class PaddleOCRVLForConditionalGeneration(PaddleOCRVLPreTrainedModel, Generation
     def set_input_embeddings(self, value):
         self.model.set_input_embeddings(value)
 
-    def get_video_features(
-        self, pixel_values_videos: torch.FloatTensor, video_grid_thw: torch.LongTensor | None = None
-    ):
-        return self.model.get_video_features(pixel_values_videos, video_grid_thw)
-
-    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: torch.LongTensor | None = None):
-        return self.model.get_image_features(pixel_values, image_grid_thw)
+    @auto_docstring
+    def get_image_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        image_grid_thw: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+            The tensors corresponding to the input images.
+        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+            The temporal, height and width of feature shape of each image in LLM.
+        """
+        return self.model.get_image_features(pixel_values=pixel_values, image_grid_thw=image_grid_thw, **kwargs)
 
     @can_return_tuple
     @auto_docstring

transformers/models/paddleocr_vl/modular_paddleocr_vl.py

@@ -54,7 +54,15 @@ from ...processing_utils import (
     Unpack,
 )
 from ...tokenization_utils_base import PreTokenizedInput, TextInput
-from ...utils import TensorType, TransformersKwargs, auto_docstring, can_return_tuple, logging, torch_int
+from ...utils import (
+    TensorType,
+    TransformersKwargs,
+    auto_docstring,
+    can_return_tuple,
+    logging,
+    torch_compilable_check,
+    torch_int,
+)
 from ...utils.generic import check_model_inputs
 from ..ernie4_5.configuration_ernie4_5 import Ernie4_5Config
 from ..ernie4_5.modeling_ernie4_5 import (
@@ -660,6 +668,8 @@ class PaddleOCRVLConfig(Qwen2VLConfig):
             The token index to denote start of vision input.
         vision_end_token_id (`int`, *optional*, defaults to 101306):
             The token index to denote end of vision input.
+        tie_word_embeddings (`bool`, *optional*, defaults to `True`):
+            Whether the model's input and output word embeddings should be tied.
 
     ```python
     >>> from transformers import PaddleOCRVLForConditionalGeneration, PaddleOCRVLConfig
@@ -684,6 +694,7 @@ class PaddleOCRVLConfig(Qwen2VLConfig):
         video_token_id=100296,
         vision_start_token_id=101305,
         vision_end_token_id=101306,
+        tie_word_embeddings=True,
         **kwargs,
     ):
         super().__init__()
@@ -860,42 +871,6 @@ class PaddleOCRTextModel(PaddleOCRVLPreTrainedModel, Ernie4_5Model):
         )
 
 
-class PaddleOCRVisionModel(PaddleOCRVLPreTrainedModel):
-    config: PaddleOCRVisionConfig
-    main_input_name = "pixel_values"
-    input_modalities = "image"
-
-    def __init__(self, config: PaddleOCRVisionConfig):
-        super().__init__(config)
-
-        self.vision_model = PaddleOCRVisionTransformer(config)
-
-        # Initialize weights and apply final processing
-        self.post_init()
-
-    def forward(
-        self,
-        pixel_values: torch.FloatTensor,
-        cu_seqlens: torch.Tensor,
-        image_grid_thw: list[tuple[int, int, int] | list[tuple[int, int, int]]] | None = None,
-        **kwargs,
-    ) -> BaseModelOutputWithPooling:
-        """
-        Args:
-            pixel_values (`torch.FloatTensor` of shape `(batch_size, sequence_length, image_channels, patch_size, patch_size)`):
-                The tensors corresponding to the input images.
-            cu_seqlens (`torch.Tensor` of shape `(num_images + 1,)`):
-                The cumulative sequence lengths of each image or video feature.
-            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
-                The temporal, height and width of feature shape of each image in LLM.
-        """
-        return self.vision_model(
-            pixel_values=pixel_values,
-            cu_seqlens=cu_seqlens,
-            image_grid_thw=image_grid_thw,
-        )
-
-
 class PaddleOCRVisionEmbeddings(SiglipVisionEmbeddings):
     def __init__(self, config: PaddleOCRVisionConfig):
         super().__init__()
@@ -1086,6 +1061,47 @@ class PaddleOCRVisionTransformer(PaddleOCRVLPreTrainedModel):
         )
 
 
+class PaddleOCRVisionModel(PaddleOCRVLPreTrainedModel):
+    config: PaddleOCRVisionConfig
+    main_input_name = "pixel_values"
+    input_modalities = "image"
+    _can_record_outputs = {
+        "hidden_states": PaddleOCRVisionEncoderLayer,
+        "attentions": PaddleOCRVisionAttention,
+    }
+
+    def __init__(self, config: PaddleOCRVisionConfig):
+        super().__init__(config)
+
+        self.vision_model = PaddleOCRVisionTransformer(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @check_model_inputs(tie_last_hidden_states=False)
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        cu_seqlens: torch.Tensor,
+        image_grid_thw: list[tuple[int, int, int] | list[tuple[int, int, int]]] | None = None,
+        **kwargs,
+    ) -> tuple | BaseModelOutputWithPooling:
+        """
+        Args:
+            pixel_values (`torch.FloatTensor` of shape `(batch_size, sequence_length, image_channels, patch_size, patch_size)`):
+                The tensors corresponding to the input images.
+            cu_seqlens (`torch.Tensor` of shape `(num_images + 1,)`):
+                The cumulative sequence lengths of each image or video feature.
+            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+                The temporal, height and width of feature shape of each image in LLM.
+        """
+        return self.vision_model(
+            pixel_values=pixel_values,
+            cu_seqlens=cu_seqlens,
+            image_grid_thw=image_grid_thw,
+        )
+
+
 class PaddleOCRVLModelOutputWithPast(Qwen2VLModelOutputWithPast):
     pass
 
@@ -1113,15 +1129,22 @@ class PaddleOCRVLModel(Qwen2VLModel):
     def set_input_embeddings(self, value):
         self.language_model.embed_tokens = value
 
-    def get_image_features(self, pixel_values: torch.FloatTensor, image_grid_thw: torch.LongTensor | None = None):
-        """
-        Encodes images into continuous embeddings that can be forwarded to the language model.
+    def get_video_features(self):
+        raise AttributeError("PaddleOCRVLModel does not support video.")
 
-        Args:
-            pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
-                The tensors corresponding to the input images.
-            image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
-                The temporal, height and width of feature shape of each image in LLM.
+    @can_return_tuple
+    @auto_docstring
+    def get_image_features(
+        self,
+        pixel_values: torch.FloatTensor,
+        image_grid_thw: torch.LongTensor | None = None,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, image_size, image_size)`):
+            The tensors corresponding to the input images.
+        image_grid_thw (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+            The temporal, height and width of feature shape of each image in LLM.
         """
         pixel_values = pixel_values.type(self.visual.dtype).unsqueeze(0)
         cu_seqlens = torch.repeat_interleave(image_grid_thw[:, 1] * image_grid_thw[:, 2], image_grid_thw[:, 0]).cumsum(
@@ -1137,10 +1160,14 @@ class PaddleOCRVLModel(Qwen2VLModel):
             pixel_values=pixel_values,
             image_grid_thw=image_grid_thw,
             cu_seqlens=cu_seqlens,
+            return_dict=True,
+            **kwargs,
         )
         image_embeds = vision_outputs.last_hidden_state
         image_embeds = self.projector(image_embeds, image_grid_thw)
-        return image_embeds
+        vision_outputs.pooler_output = image_embeds
+
+        return vision_outputs
 
     def get_placeholder_mask(
         self, input_ids: torch.LongTensor, inputs_embeds: torch.FloatTensor, image_features: torch.FloatTensor
@@ -1160,10 +1187,10 @@ class PaddleOCRVLModel(Qwen2VLModel):
         n_image_tokens = special_image_mask.sum()
         special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
         n_image_features = image_features.shape[0] * image_features.shape[1]
-        if inputs_embeds[special_image_mask].numel() != image_features.numel():
-            raise ValueError(
-                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
-            )
+        torch_compilable_check(
+            inputs_embeds[special_image_mask].numel() == image_features.numel(),
+            f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}",
+        )
         return special_image_mask
 
     @can_return_tuple
@@ -1191,9 +1218,8 @@ class PaddleOCRVLModel(Qwen2VLModel):
             inputs_embeds = self.language_model.embed_tokens(input_ids)
 
         if pixel_values is not None:
-            image_embeds = self.get_image_features(pixel_values, image_grid_thw).to(
-                inputs_embeds.device, inputs_embeds.dtype
-            )
+            image_embeds = self.get_image_features(pixel_values, image_grid_thw, return_dict=True).pooler_output
+            image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
             image_mask = self.get_placeholder_mask(input_ids, inputs_embeds=inputs_embeds, image_features=image_embeds)
             inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
 
@@ -1245,6 +1271,9 @@ class PaddleOCRVLForConditionalGeneration(Qwen2VLForConditionalGeneration):
     }
     _keys_to_ignore_on_load_unexpected = ["packing_position_embedding", "vision_model.head"]
 
+    def get_video_features(self):
+        raise AttributeError("PaddleOCRVLForConditionalGeneration does not support video.")
+
     @can_return_tuple
     @auto_docstring
     def forward(

transformers/models/paligemma/configuration_paligemma.py

@@ -45,6 +45,8 @@ class PaliGemmaConfig(PreTrainedConfig):
             Dimension of the multimodal projection space.
         hidden_size (`int`, *optional*, defaults to 2048):
             Dimension of the hidden layer of the Language model.
+        tie_word_embeddings (`bool`, *optional*, defaults to `True`):
+            Whether to tie weight embeddings
 
     Example:
 
@@ -82,12 +84,14 @@ class PaliGemmaConfig(PreTrainedConfig):
         vocab_size=257152,
         projection_dim=2048,
         hidden_size=2048,
+        tie_word_embeddings: bool | None = True,
         **kwargs,
     ):
         self.image_token_index = image_token_index
         self.projection_dim = projection_dim
         self.hidden_size = hidden_size
         self.vision_config = vision_config
+        self.tie_word_embeddings = tie_word_embeddings
         self.is_encoder_decoder = False
 
         if isinstance(self.vision_config, dict):

transformers/models/paligemma/modeling_paligemma.py

@@ -24,7 +24,7 @@ from ...configuration_utils import PreTrainedConfig
 from ...generation import GenerationMixin
 from ...masking_utils import create_masks_for_generate
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
-from ...modeling_outputs import BaseModelOutputWithPast
+from ...modeling_outputs import BaseModelOutputWithPast, BaseModelOutputWithPooling
 from ...modeling_utils import PreTrainedModel
 from ...processing_utils import Unpack
 from ...utils import (
@@ -33,6 +33,7 @@ from ...utils import (
     auto_docstring,
     can_return_tuple,
     logging,
+    torch_compilable_check,
 )
 from ..auto import AutoModel
 from .configuration_paligemma import PaliGemmaConfig
@@ -246,7 +247,6 @@ class PaliGemmaModel(PaliGemmaPreTrainedModel):
         language_model = AutoModel.from_config(config=config.text_config)
         self.language_model = language_model
 
-        self.pad_token_id = self.config.pad_token_id if self.config.pad_token_id is not None else -1
         self.text_config_dtype = self.config.get_text_config().dtype or self.dtype
         self.post_init()
 
@@ -258,21 +258,20 @@ class PaliGemmaModel(PaliGemmaPreTrainedModel):
     def set_input_embeddings(self, value):
         self.language_model.set_input_embeddings(value)
 
-    def get_image_features(self, pixel_values: torch.FloatTensor):
-        """
-        Obtains image last hidden states from the vision tower and apply multimodal projection.
-
-        Args:
-            pixel_values (`torch.FloatTensor]` of shape `(batch_size, channels, height, width)`)
-               The tensors corresponding to the input images.
-        Returns:
-            image_features (`torch.Tensor`): Image feature tensor of shape `(num_images, image_length, embed_dim)`).
-        """
-        image_outputs = self.vision_tower(pixel_values)
+    @can_return_tuple
+    @auto_docstring(
+        custom_intro="Obtains image last hidden states from the vision tower and apply multimodal projection."
+    )
+    def get_image_features(
+        self, pixel_values: torch.FloatTensor, **kwargs: Unpack[TransformersKwargs]
+    ) -> tuple | BaseModelOutputWithPooling:
+        image_outputs = self.vision_tower(pixel_values, return_dict=True, **kwargs)
         selected_image_feature = image_outputs.last_hidden_state
         image_features = self.multi_modal_projector(selected_image_feature)
         image_features = image_features / (self.config.text_config.hidden_size**0.5)
-        return image_features
+        image_outputs.pooler_output = image_features
+
+        return image_outputs
 
     def get_placeholder_mask(
         self, input_ids: torch.LongTensor, inputs_embeds: torch.FloatTensor, image_features: torch.FloatTensor
@@ -290,12 +289,12 @@ class PaliGemmaModel(PaliGemmaPreTrainedModel):
             special_image_mask = input_ids == self.config.image_token_id
 
         n_image_tokens = special_image_mask.sum()
-        special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
         n_image_features = image_features.shape[0] * image_features.shape[1]
-        if inputs_embeds[special_image_mask].numel() != image_features.numel():
-            raise ValueError(
-                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
-            )
+        special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
+        torch_compilable_check(
+            inputs_embeds[special_image_mask].numel() == image_features.numel(),
+            f"Image features and image tokens do not match, tokens: {n_image_tokens}, features: {n_image_features}",
+        )
         return special_image_mask
 
     @can_return_tuple
@@ -327,7 +326,8 @@ class PaliGemmaModel(PaliGemmaPreTrainedModel):
 
         ```python
         >>> from PIL import Image
-        >>> import requests
+        >>> import httpx
+        >>> from io import BytesIO
         >>> from transformers import AutoProcessor, PaliGemmaForConditionalGeneration
 
         >>> model = PaliGemmaForConditionalGeneration.from_pretrained("google/paligemma2-3b-mix-224")
@@ -335,7 +335,8 @@ class PaliGemmaModel(PaliGemmaPreTrainedModel):
 
         >>> prompt = "Where is the cat standing?"
         >>> url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
-        >>> image = Image.open(requests.get(url, stream=True).raw)
+        >>> with httpx.stream("GET", url) as response:
+        ...     image = Image.open(BytesIO(response.read()))
 
         >>> inputs = processor(images=image, text=prompt,  return_tensors="pt")
 
@@ -376,7 +377,7 @@ class PaliGemmaModel(PaliGemmaPreTrainedModel):
 
         # Merge text and images
         if pixel_values is not None:
-            image_features = self.get_image_features(pixel_values)
+            image_features = self.get_image_features(pixel_values, return_dict=True).pooler_output
             image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
             special_image_mask = self.get_placeholder_mask(
                 input_ids, inputs_embeds=inputs_embeds, image_features=image_features
@@ -445,8 +446,9 @@ class PaliGemmaForConditionalGeneration(PaliGemmaPreTrainedModel, GenerationMixi
     def set_input_embeddings(self, value):
         self.model.set_input_embeddings(value)
 
-    def get_image_features(self, pixel_values):
-        return self.model.get_image_features(pixel_values)
+    @auto_docstring
+    def get_image_features(self, pixel_values: torch.FloatTensor, **kwargs: Unpack[TransformersKwargs]):
+        return self.model.get_image_features(pixel_values, **kwargs)
 
     @can_return_tuple
     @auto_docstring
@@ -478,7 +480,8 @@ class PaliGemmaForConditionalGeneration(PaliGemmaPreTrainedModel, GenerationMixi
 
         ```python
         >>> from PIL import Image
-        >>> import requests
+        >>> import httpx
+        >>> from io import BytesIO
         >>> from transformers import AutoProcessor, PaliGemmaForConditionalGeneration
 
         >>> model = PaliGemmaForConditionalGeneration.from_pretrained("google/paligemma2-3b-mix-224")
@@ -486,7 +489,8 @@ class PaliGemmaForConditionalGeneration(PaliGemmaPreTrainedModel, GenerationMixi
 
         >>> prompt = "Where is the cat standing?"
         >>> url = "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg"
-        >>> image = Image.open(requests.get(url, stream=True).raw)
+        >>> with httpx.stream("GET", url) as response:
+        ...     image = Image.open(BytesIO(response.read()))
 
         >>> inputs = processor(images=image, text=prompt,  return_tensors="pt")
 

transformers/models/parakeet/configuration_parakeet.py

@@ -213,11 +213,9 @@ class ParakeetCTCConfig(PreTrainedConfig):
 
         self.encoder_config = self.encoder_config
         self.initializer_range = self.encoder_config.initializer_range
+        self.pad_token_id = pad_token_id
 
-        super().__init__(
-            pad_token_id=pad_token_id,
-            **kwargs,
-        )
+        super().__init__(**kwargs)
 
     @classmethod
     def from_encoder_config(cls, encoder_config: ParakeetEncoderConfig, **kwargs):

transformers/models/parakeet/modeling_parakeet.py

@@ -303,9 +303,9 @@ class ParakeetEncoderAttention(nn.Module):
         key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
         value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         query_states_with_bias_u = query_states + self.bias_u.view(
             1, self.config.num_attention_heads, 1, self.head_dim

transformers/models/parakeet/modular_parakeet.py

@@ -141,9 +141,9 @@ class ParakeetEncoderAttention(LlamaAttention):
         key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
         value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
 
-        attention_interface: Callable = eager_attention_forward
-        if self.config._attn_implementation != "eager":
-            attention_interface = ALL_ATTENTION_FUNCTIONS[self.config._attn_implementation]
+        attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface(
+            self.config._attn_implementation, eager_attention_forward
+        )
 
         query_states_with_bias_u = query_states + self.bias_u.view(
             1, self.config.num_attention_heads, 1, self.head_dim