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

transformers/models/audioflamingo3/modeling_audioflamingo3.py

@@ -19,7 +19,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-
 import math
 from collections.abc import Callable
 
@@ -32,10 +31,11 @@ from ...generation import GenerationMixin
 from ...masking_utils import create_bidirectional_mask
 from ...modeling_flash_attention_utils import FlashAttentionKwargs
 from ...modeling_layers import GradientCheckpointingLayer
-from ...modeling_outputs import BaseModelOutput, CausalLMOutputWithPast
+from ...modeling_outputs import BaseModelOutputWithPooling, CausalLMOutputWithPast
 from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ...processing_utils import Unpack
 from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
+from ...utils.generic import check_model_inputs
 from ..auto import AutoModel, AutoModelForCausalLM
 from .configuration_audioflamingo3 import AudioFlamingo3Config, AudioFlamingo3EncoderConfig
 
@@ -171,9 +171,9 @@ class AudioFlamingo3Attention(nn.Module):
                     key_states, value_states, self.layer_idx, {"cache_position": cache_position}
                 )
 
-        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,
@@ -280,6 +280,11 @@ class AudioFlamingo3Encoder(AudioFlamingo3PreTrainedModel):
     input_modalities = "audio"
     _no_split_modules = ["AudioFlamingo3EncoderLayer"]
 
+    _can_record_outputs = {
+        "hidden_states": AudioFlamingo3EncoderLayer,
+        "attentions": AudioFlamingo3Attention,
+    }
+
     def __init__(self, config: AudioFlamingo3EncoderConfig):
         super().__init__(config)
         self.dropout = config.dropout
@@ -287,7 +292,6 @@ class AudioFlamingo3Encoder(AudioFlamingo3PreTrainedModel):
 
         embed_dim = config.d_model
         self.num_mel_bins = config.num_mel_bins
-        self.padding_idx = config.pad_token_id
         self.max_source_positions = config.max_source_positions
         self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0
 
@@ -317,13 +321,13 @@ class AudioFlamingo3Encoder(AudioFlamingo3PreTrainedModel):
     def set_input_embeddings(self, value: nn.Module):
         self.conv1 = value
 
-    @can_return_tuple
+    @check_model_inputs
     def forward(
         self,
         input_features: torch.Tensor,
         input_features_mask: torch.Tensor | None = None,
         **kwargs,
-    ):
+    ) -> tuple | BaseModelOutputWithPooling:
         r"""
         Args:
             input_features (`torch.FloatTensor` of shape `(batch_size, feature_size, sequence_length)`):
@@ -367,7 +371,7 @@ class AudioFlamingo3Encoder(AudioFlamingo3PreTrainedModel):
         hidden_states = self.avg_pooler(hidden_states).permute(0, 2, 1)
         hidden_states = self.layer_norm(hidden_states)
 
-        return BaseModelOutput(
+        return BaseModelOutputWithPooling(
             last_hidden_state=hidden_states,
         )
 
@@ -442,35 +446,40 @@ class AudioFlamingo3ForConditionalGeneration(AudioFlamingo3PreTrainedModel, Gene
     def get_decoder(self):
         return self.language_model.get_decoder()
 
+    @can_return_tuple
+    @auto_docstring(
+        custom_intro="This method is used to get the audio embeddings from input features (a log mel spectrogram), meaning inferring the audio encoder and the multi-modal projector."
+    )
     def get_audio_features(
-        self, input_features: torch.FloatTensor, input_features_mask: torch.Tensor
-    ) -> torch.FloatTensor:
-        """
-        This method is used to get the audio embeddings from input features (a log mel spectrogram), meaning inferring the audio encoder and the multi-modal projector.
-        Args:
-            input_features (`torch.FloatTensor`):
-                Float values of mel features extracted from the raw speech waveform. Raw speech waveform can be
-                obtained by loading a `.flac` or `.wav` audio file into an array of type `list[float]` or a
-                `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
-                `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding
-                and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
-            input_features_mask (`torch.Tensor` of shape `(batch_size, feature_sequence_length)`):
-                Mask to avoid performing attention on padded feature indices.
-
-        Returns:
-            `torch.FloatTensor`:
-                The audio embeddings.
+        self,
+        input_features: torch.FloatTensor,
+        input_features_mask: torch.Tensor,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        input_features (`torch.FloatTensor`):
+            Float values of mel features extracted from the raw speech waveform. Raw speech waveform can be
+            obtained by loading a `.flac` or `.wav` audio file into an array of type `list[float]` or a
+            `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
+            `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding
+            and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
+        input_features_mask (`torch.Tensor` of shape `(batch_size, feature_sequence_length)`):
+            Mask to avoid performing attention on padded feature indices.
         """
 
         # Encode audio
-        encoder_output = self.audio_tower(input_features, input_features_mask=input_features_mask)
-        audio_embeds = self.multi_modal_projector(encoder_output.last_hidden_state)
+        audio_output = self.audio_tower(
+            input_features, input_features_mask=input_features_mask, return_dict=True, **kwargs
+        )
+        audio_embeds = self.multi_modal_projector(audio_output.last_hidden_state)
 
         # Mask according to avg pooling (which is after attention blocks)
         post_lengths = (input_features_mask.sum(-1) - 2) // 2 + 1
         valid_mask = torch.arange(audio_embeds.shape[1], device=post_lengths.device)[None, :] < post_lengths[:, None]
         audio_embeds = audio_embeds[valid_mask.to(audio_embeds.device)]
-        return audio_embeds
+        audio_output.pooler_output = audio_embeds
+
+        return audio_output
 
     @can_return_tuple
     @auto_docstring
@@ -556,7 +565,7 @@ class AudioFlamingo3ForConditionalGeneration(AudioFlamingo3PreTrainedModel, Gene
             inputs_embeds = self.get_input_embeddings()(input_ids)
 
         if input_features is not None and input_ids is not None:
-            audio_embeds = self.get_audio_features(input_features, input_features_mask)
+            audio_embeds = self.get_audio_features(input_features, input_features_mask, return_dict=True).pooler_output
 
             # replace text-audio token placeholders with audio embeddings
             audio_token_mask = (input_ids == self.config.audio_token_id).unsqueeze(-1)

transformers/models/audioflamingo3/modular_audioflamingo3.py

@@ -13,28 +13,32 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-
 import torch
 from torch import nn
 
 from ...activations import ACT2FN
 from ...cache_utils import Cache
 from ...masking_utils import create_bidirectional_mask
-from ...modeling_outputs import BaseModelOutput, CausalLMOutputWithPast
+from ...modeling_outputs import BaseModelOutputWithPooling, CausalLMOutputWithPast
 from ...processing_utils import Unpack
 from ...utils import TransformersKwargs, auto_docstring, can_return_tuple, logging
+from ...utils.generic import check_model_inputs
 from ..qwen2_audio.modeling_qwen2_audio import (
     Qwen2AudioEncoder,
     Qwen2AudioPreTrainedModel,
 )
 from ..voxtral.modeling_voxtral import VoxtralForConditionalGeneration, VoxtralMultiModalProjector
-from ..whisper.modeling_whisper import WhisperEncoderLayer
+from ..whisper.modeling_whisper import WhisperAttention, WhisperEncoderLayer
 from .configuration_audioflamingo3 import AudioFlamingo3Config
 
 
 logger = logging.get_logger(__name__)
 
 
+class AudioFlamingo3Attention(WhisperAttention):
+    pass
+
+
 class AudioFlamingo3EncoderLayer(WhisperEncoderLayer):
     pass
 
@@ -53,13 +57,18 @@ class AudioFlamingo3Encoder(Qwen2AudioEncoder):
     AudioFlamingo3 encoder: Whisper encoder, average pool (time/2), then LayerNorm.
     """
 
-    @can_return_tuple
+    _can_record_outputs = {
+        "hidden_states": AudioFlamingo3EncoderLayer,
+        "attentions": AudioFlamingo3Attention,
+    }
+
+    @check_model_inputs
     def forward(
         self,
         input_features: torch.Tensor,
         input_features_mask: torch.Tensor | None = None,
         **kwargs,
-    ):
+    ) -> tuple | BaseModelOutputWithPooling:
         r"""
         Args:
             input_features (`torch.FloatTensor` of shape `(batch_size, feature_size, sequence_length)`):
@@ -103,7 +112,7 @@ class AudioFlamingo3Encoder(Qwen2AudioEncoder):
         hidden_states = self.avg_pooler(hidden_states).permute(0, 2, 1)
         hidden_states = self.layer_norm(hidden_states)
 
-        return BaseModelOutput(
+        return BaseModelOutputWithPooling(
             last_hidden_state=hidden_states,
         )
 
@@ -138,35 +147,40 @@ class AudioFlamingo3ForConditionalGeneration(VoxtralForConditionalGeneration):
     def __init__(self, config):
         super().__init__(config)
 
+    @can_return_tuple
+    @auto_docstring(
+        custom_intro="This method is used to get the audio embeddings from input features (a log mel spectrogram), meaning inferring the audio encoder and the multi-modal projector."
+    )
     def get_audio_features(
-        self, input_features: torch.FloatTensor, input_features_mask: torch.Tensor
-    ) -> torch.FloatTensor:
-        """
-        This method is used to get the audio embeddings from input features (a log mel spectrogram), meaning inferring the audio encoder and the multi-modal projector.
-        Args:
-            input_features (`torch.FloatTensor`):
-                Float values of mel features extracted from the raw speech waveform. Raw speech waveform can be
-                obtained by loading a `.flac` or `.wav` audio file into an array of type `list[float]` or a
-                `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
-                `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding
-                and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
-            input_features_mask (`torch.Tensor` of shape `(batch_size, feature_sequence_length)`):
-                Mask to avoid performing attention on padded feature indices.
-
-        Returns:
-            `torch.FloatTensor`:
-                The audio embeddings.
+        self,
+        input_features: torch.FloatTensor,
+        input_features_mask: torch.Tensor,
+        **kwargs: Unpack[TransformersKwargs],
+    ) -> tuple | BaseModelOutputWithPooling:
+        r"""
+        input_features (`torch.FloatTensor`):
+            Float values of mel features extracted from the raw speech waveform. Raw speech waveform can be
+            obtained by loading a `.flac` or `.wav` audio file into an array of type `list[float]` or a
+            `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
+            `input_features`, the [`AutoFeatureExtractor`] should be used for extracting the mel features, padding
+            and conversion into a tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`]
+        input_features_mask (`torch.Tensor` of shape `(batch_size, feature_sequence_length)`):
+            Mask to avoid performing attention on padded feature indices.
         """
 
         # Encode audio
-        encoder_output = self.audio_tower(input_features, input_features_mask=input_features_mask)
-        audio_embeds = self.multi_modal_projector(encoder_output.last_hidden_state)
+        audio_output = self.audio_tower(
+            input_features, input_features_mask=input_features_mask, return_dict=True, **kwargs
+        )
+        audio_embeds = self.multi_modal_projector(audio_output.last_hidden_state)
 
         # Mask according to avg pooling (which is after attention blocks)
         post_lengths = (input_features_mask.sum(-1) - 2) // 2 + 1
         valid_mask = torch.arange(audio_embeds.shape[1], device=post_lengths.device)[None, :] < post_lengths[:, None]
         audio_embeds = audio_embeds[valid_mask.to(audio_embeds.device)]
-        return audio_embeds
+        audio_output.pooler_output = audio_embeds
+
+        return audio_output
 
     @can_return_tuple
     @auto_docstring
@@ -252,7 +266,7 @@ class AudioFlamingo3ForConditionalGeneration(VoxtralForConditionalGeneration):
             inputs_embeds = self.get_input_embeddings()(input_ids)
 
         if input_features is not None and input_ids is not None:
-            audio_embeds = self.get_audio_features(input_features, input_features_mask)
+            audio_embeds = self.get_audio_features(input_features, input_features_mask, return_dict=True).pooler_output
 
             # replace text-audio token placeholders with audio embeddings
             audio_token_mask = (input_ids == self.config.audio_token_id).unsqueeze(-1)

transformers/models/auto/auto_factory.py

@@ -21,6 +21,8 @@ from collections import OrderedDict
 from collections.abc import Iterator
 from typing import Any, TypeVar
 
+from huggingface_hub import repo_exists
+
 from ...configuration_utils import PreTrainedConfig
 from ...dynamic_module_utils import get_class_from_dynamic_module, resolve_trust_remote_code
 from ...utils import (
@@ -68,7 +70,7 @@ FROM_CONFIG_DOCSTRING = """
 
                 List options
             attn_implementation (`str`, *optional*):
-                The attention implementation to use in the model (if relevant). Can be any of `"eager"` (manual implementation of the attention), `"sdpa"` (using [`F.scaled_dot_product_attention`](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention.html)), or `"flash_attention_2"` (using [Dao-AILab/flash-attention](https://github.com/Dao-AILab/flash-attention)). By default, if available, SDPA will be used for torch>=2.1.1. The default is otherwise the manual `"eager"` implementation.
+                The attention implementation to use in the model (if relevant). Can be any of `"eager"` (manual implementation of the attention), `"sdpa"` (using [`F.scaled_dot_product_attention`](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention.html)), `"flash_attention_2"` (using [Dao-AILab/flash-attention](https://github.com/Dao-AILab/flash-attention)), or `"flash_attention_3"` (using [Dao-AILab/flash-attention/hopper](https://github.com/Dao-AILab/flash-attention/tree/main/hopper)). By default, if available, SDPA will be used for torch>=2.1.1. The default is otherwise the manual `"eager"` implementation.
 
         Examples:
 
@@ -416,21 +418,21 @@ class _BaseAutoBackboneClass(_BaseAutoModelClass):
 
         num_channels = kwargs.pop("num_channels", config.num_channels)
         features_only = kwargs.pop("features_only", config.features_only)
-        use_pretrained_backbone = kwargs.pop("use_pretrained_backbone", config.use_pretrained_backbone)
         out_indices = kwargs.pop("out_indices", config.out_indices)
         config = TimmBackboneConfig(
             backbone=pretrained_model_name_or_path,
             num_channels=num_channels,
             features_only=features_only,
-            use_pretrained_backbone=use_pretrained_backbone,
             out_indices=out_indices,
         )
-        return super().from_config(config, **kwargs)
+        # Always load a pretrained model when `from_pretrained` is called
+        kwargs.pop("use_pretrained_backbone", None)
+        return super().from_config(config, pretrained=True, **kwargs)
 
     @classmethod
     def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
-        use_timm_backbone = kwargs.pop("use_timm_backbone", False)
-        if use_timm_backbone:
+        kwargs.pop("use_timm_backbone", None)
+        if not repo_exists(pretrained_model_name_or_path):
             return cls._load_timm_backbone_from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
 
         return super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)

transformers/models/auto/configuration_auto.py

@@ -138,6 +138,7 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("encodec", "EncodecConfig"),
         ("encoder-decoder", "EncoderDecoderConfig"),
         ("eomt", "EomtConfig"),
+        ("eomt_dinov3", "EomtDinov3Config"),
         ("ernie", "ErnieConfig"),
         ("ernie4_5", "Ernie4_5Config"),
         ("ernie4_5_moe", "Ernie4_5_MoeConfig"),
@@ -145,6 +146,7 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("esm", "EsmConfig"),
         ("evolla", "EvollaConfig"),
         ("exaone4", "Exaone4Config"),
+        ("exaone_moe", "ExaoneMoeConfig"),
         ("falcon", "FalconConfig"),
         ("falcon_h1", "FalconH1Config"),
         ("falcon_mamba", "FalconMambaConfig"),
@@ -184,6 +186,9 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("glm_image_text", "GlmImageTextConfig"),
         ("glm_image_vision", "GlmImageVisionConfig"),
         ("glm_image_vqmodel", "GlmImageVQVAEConfig"),
+        ("glm_ocr", "GlmOcrConfig"),
+        ("glm_ocr_text", "GlmOcrTextConfig"),
+        ("glm_ocr_vision", "GlmOcrVisionConfig"),
         ("glmasr", "GlmAsrConfig"),
         ("glmasr_encoder", "GlmAsrEncoderConfig"),
         ("glpn", "GLPNConfig"),
@@ -339,6 +344,7 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("plbart", "PLBartConfig"),
         ("poolformer", "PoolFormerConfig"),
         ("pop2piano", "Pop2PianoConfig"),
+        ("pp_doclayout_v3", "PPDocLayoutV3Config"),
         ("prompt_depth_anything", "PromptDepthAnythingConfig"),
         ("prophetnet", "ProphetNetConfig"),
         ("pvt", "PvtConfig"),
@@ -403,6 +409,7 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("smollm3", "SmolLM3Config"),
         ("smolvlm", "SmolVLMConfig"),
         ("smolvlm_vision", "SmolVLMVisionConfig"),
+        ("solar_open", "SolarOpenConfig"),
         ("speech-encoder-decoder", "SpeechEncoderDecoderConfig"),
         ("speech_to_text", "Speech2TextConfig"),
         ("speecht5", "SpeechT5Config"),
@@ -420,6 +427,7 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("t5", "T5Config"),
         ("t5gemma", "T5GemmaConfig"),
         ("t5gemma2", "T5Gemma2Config"),
+        ("t5gemma2_encoder", "T5Gemma2EncoderConfig"),
         ("table-transformer", "TableTransformerConfig"),
         ("tapas", "TapasConfig"),
         ("textnet", "TextNetConfig"),
@@ -474,6 +482,7 @@ CONFIG_MAPPING_NAMES = OrderedDict[str, str](
         ("xmod", "XmodConfig"),
         ("yolos", "YolosConfig"),
         ("yoso", "YosoConfig"),
+        ("youtu", "YoutuConfig"),
         ("zamba", "ZambaConfig"),
         ("zamba2", "Zamba2Config"),
         ("zoedepth", "ZoeDepthConfig"),
@@ -598,6 +607,7 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("encodec", "EnCodec"),
         ("encoder-decoder", "Encoder decoder"),
         ("eomt", "EoMT"),
+        ("eomt_dinov3", "EoMT-DINOv3"),
         ("ernie", "ERNIE"),
         ("ernie4_5", "Ernie4_5"),
         ("ernie4_5_moe", "Ernie4_5_MoE"),
@@ -605,6 +615,7 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("esm", "ESM"),
         ("evolla", "Evolla"),
         ("exaone4", "EXAONE-4.0"),
+        ("exaone_moe", "EXAONE-MoE"),
         ("falcon", "Falcon"),
         ("falcon3", "Falcon3"),
         ("falcon_h1", "FalconH1"),
@@ -647,6 +658,9 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("glm_image_text", "GlmImageText"),
         ("glm_image_vision", "GlmImageVisionModel"),
         ("glm_image_vqmodel", "GlmImageVQVAE"),
+        ("glm_ocr", "Glmocr"),
+        ("glm_ocr_text", "GlmOcrText"),
+        ("glm_ocr_vision", "GlmOcrVisionModel"),
         ("glmasr", "GLM-ASR"),
         ("glmasr_encoder", "GLM-ASR Encoder"),
         ("glpn", "GLPN"),
@@ -816,6 +830,7 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("plbart", "PLBart"),
         ("poolformer", "PoolFormer"),
         ("pop2piano", "Pop2Piano"),
+        ("pp_doclayout_v3", "PPDocLayoutV3"),
         ("prompt_depth_anything", "PromptDepthAnything"),
         ("prophetnet", "ProphetNet"),
         ("pvt", "PVT"),
@@ -880,6 +895,7 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("smollm3", "SmolLM3"),
         ("smolvlm", "SmolVLM"),
         ("smolvlm_vision", "SmolVLMVisionTransformer"),
+        ("solar_open", "SolarOpen"),
         ("speech-encoder-decoder", "Speech Encoder decoder"),
         ("speech_to_text", "Speech2Text"),
         ("speecht5", "SpeechT5"),
@@ -897,6 +913,7 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("t5", "T5"),
         ("t5gemma", "T5Gemma"),
         ("t5gemma2", "T5Gemma2"),
+        ("t5gemma2_encoder", "T5Gemma2Encoder"),
         ("t5v1.1", "T5v1.1"),
         ("table-transformer", "Table Transformer"),
         ("tapas", "TAPAS"),
@@ -957,6 +974,7 @@ MODEL_NAMES_MAPPING = OrderedDict[str, str](
         ("xmod", "X-MOD"),
         ("yolos", "YOLOS"),
         ("yoso", "YOSO"),
+        ("youtu", "Youtu"),
         ("zamba", "Zamba"),
         ("zamba2", "Zamba2"),
         ("zoedepth", "ZoeDepth"),
@@ -997,6 +1015,9 @@ SPECIAL_MODEL_TYPE_TO_MODULE_NAME = OrderedDict[str, str](
         ("glm_image_vision", "glm_image"),
         ("glm_image_vqmodel", "glm_image"),
         ("glm_image_text", "glm_image"),
+        ("glm_ocr_vision", "glm_ocr"),
+        ("glm_ocr_vqmodel", "glm_ocr"),
+        ("glm_ocr_text", "glm_ocr"),
         ("glmasr_encoder", "glmasr"),
         ("grounding-dino", "grounding_dino"),
         ("mm-grounding-dino", "mm_grounding_dino"),
@@ -1021,6 +1042,7 @@ SPECIAL_MODEL_TYPE_TO_MODULE_NAME = OrderedDict[str, str](
         ("sam3_vision_model", "sam3"),
         ("edgetam_vision_model", "edgetam"),
         ("sam_hq_vision_model", "sam_hq"),
+        ("t5gemma2_encoder", "t5gemma2"),
         ("llama4_text", "llama4"),
         ("blip_2_qformer", "blip_2"),
         ("fastspeech2_conformer_with_hifigan", "fastspeech2_conformer"),

transformers/models/auto/image_processing_auto.py

@@ -47,9 +47,14 @@ from .configuration_auto import (
 
 logger = logging.get_logger(__name__)
 
-
-FORCE_FAST_IMAGE_PROCESSOR = ["Qwen2VLImageProcessor"]
-
+# These image processors use Lanczos interpolation, which is not supported by fast image processors.
+# To avoid important differences in outputs, we default to using the slow image processors for these processors.
+DEFAULT_TO_SLOW_IMAGE_PROCESSORS = [
+    "ChameleonImageProcessor",
+    "FlavaImageProcessor",
+    "Idefics3ImageProcessor",
+    "SmolVLMImageProcessor",
+]
 
 if TYPE_CHECKING:
     # This significantly improves completion suggestion performance when
@@ -98,6 +103,7 @@ else:
             ("efficientnet", ("EfficientNetImageProcessor", "EfficientNetImageProcessorFast")),
             ("emu3", ("Emu3ImageProcessor", None)),
             ("eomt", ("EomtImageProcessor", "EomtImageProcessorFast")),
+            ("eomt_dinov3", ("EomtImageProcessor", "EomtImageProcessorFast")),
             ("ernie4_5_vl_moe", ("Ernie4_5_VL_MoeImageProcessor", "Ernie4_5_VL_MoeImageProcessorFast")),
             ("flava", ("FlavaImageProcessor", "FlavaImageProcessorFast")),
             ("florence2", ("CLIPImageProcessor", "CLIPImageProcessorFast")),
@@ -164,6 +170,7 @@ else:
             ("pixio", ("BitImageProcessor", "BitImageProcessorFast")),
             ("pixtral", ("PixtralImageProcessor", "PixtralImageProcessorFast")),
             ("poolformer", ("PoolFormerImageProcessor", "PoolFormerImageProcessorFast")),
+            ("pp_doclayout_v3", (None, "PPDocLayoutV3ImageProcessorFast")),
             ("prompt_depth_anything", ("PromptDepthAnythingImageProcessor", "PromptDepthAnythingImageProcessorFast")),
             ("pvt", ("PvtImageProcessor", "PvtImageProcessorFast")),
             ("pvt_v2", ("PvtImageProcessor", "PvtImageProcessorFast")),
@@ -196,6 +203,7 @@ else:
             ("swin2sr", ("Swin2SRImageProcessor", "Swin2SRImageProcessorFast")),
             ("swinv2", ("ViTImageProcessor", "ViTImageProcessorFast")),
             ("t5gemma2", ("Gemma3ImageProcessor", "Gemma3ImageProcessorFast")),
+            ("t5gemma2_encoder", ("Gemma3ImageProcessor", "Gemma3ImageProcessorFast")),
             ("table-transformer", ("DetrImageProcessor", "DetrImageProcessorFast")),
             ("textnet", ("TextNetImageProcessor", "TextNetImageProcessorFast")),
             ("timesformer", ("VideoMAEImageProcessor", None)),
@@ -535,24 +543,20 @@ class AutoImageProcessor:
                 image_processor_auto_map = config.auto_map["AutoImageProcessor"]
 
         image_processor_class = None
-        # TODO: @yoni, change logic in v4.52 (when use_fast set to True by default)
         if image_processor_type is not None:
             # if use_fast is not set and the processor was saved with a fast processor, we use it, otherwise we use the slow processor.
             if use_fast is None:
                 use_fast = image_processor_type.endswith("Fast")
-                if not use_fast and image_processor_type in FORCE_FAST_IMAGE_PROCESSOR and is_torchvision_available():
-                    use_fast = True
+                if (
+                    not use_fast
+                    and is_torchvision_available()
+                    and image_processor_type not in DEFAULT_TO_SLOW_IMAGE_PROCESSORS
+                ):
                     logger.warning_once(
                         f"The image processor of type `{image_processor_type}` is now loaded as a fast processor by default, even if the model checkpoint was saved with a slow processor. "
                         "This is a breaking change and may produce slightly different outputs. To continue using the slow processor, instantiate this class with `use_fast=False`. "
-                        "Note that this behavior will be extended to all models in a future release."
-                    )
-                if not use_fast:
-                    logger.warning_once(
-                        "Using a slow image processor as `use_fast` is unset and a slow processor was saved with this model. "
-                        "`use_fast=True` will be the default behavior in v4.52, even if the model was saved with a slow processor. "
-                        "This will result in minor differences in outputs. You'll still be able to use a slow processor with `use_fast=False`."
                     )
+                    use_fast = True
             if use_fast and not image_processor_type.endswith("Fast"):
                 image_processor_type += "Fast"
             if use_fast and not is_torchvision_available():

transformers/models/auto/modeling_auto.py

@@ -147,6 +147,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("esm", "EsmModel"),
         ("evolla", "EvollaModel"),
         ("exaone4", "Exaone4Model"),
+        ("exaone_moe", "ExaoneMoeModel"),
         ("falcon", "FalconModel"),
         ("falcon_h1", "FalconH1Model"),
         ("falcon_mamba", "FalconMambaModel"),
@@ -186,6 +187,9 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("glm_image_text", "GlmImageTextModel"),
         ("glm_image_vision", "GlmImageVisionModel"),
         ("glm_image_vqmodel", "GlmImageVQVAE"),
+        ("glm_ocr", "GlmOcrModel"),
+        ("glm_ocr_text", "GlmOcrTextModel"),
+        ("glm_ocr_vision", "GlmOcrVisionModel"),
         ("glmasr", "GlmAsrForConditionalGeneration"),
         ("glmasr_encoder", "GlmAsrEncoder"),
         ("glpn", "GLPNModel"),
@@ -333,6 +337,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("pixtral", "PixtralVisionModel"),
         ("plbart", "PLBartModel"),
         ("poolformer", "PoolFormerModel"),
+        ("pp_doclayout_v3", "PPDocLayoutV3Model"),
         ("prophetnet", "ProphetNetModel"),
         ("pvt", "PvtModel"),
         ("pvt_v2", "PvtV2Model"),
@@ -391,6 +396,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("smollm3", "SmolLM3Model"),
         ("smolvlm", "SmolVLMModel"),
         ("smolvlm_vision", "SmolVLMVisionTransformer"),
+        ("solar_open", "SolarOpenModel"),
         ("speech_to_text", "Speech2TextModel"),
         ("speecht5", "SpeechT5Model"),
         ("splinter", "SplinterModel"),
@@ -405,6 +411,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("t5", "T5Model"),
         ("t5gemma", "T5GemmaModel"),
         ("t5gemma2", "T5Gemma2Model"),
+        ("t5gemma2_encoder", "T5Gemma2Encoder"),
         ("table-transformer", "TableTransformerModel"),
         ("tapas", "TapasModel"),
         ("textnet", "TextNetModel"),
@@ -453,6 +460,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("xmod", "XmodModel"),
         ("yolos", "YolosModel"),
         ("yoso", "YosoModel"),
+        ("youtu", "YoutuModel"),
         ("zamba", "ZambaModel"),
         ("zamba2", "Zamba2Model"),
     ]
@@ -479,6 +487,7 @@ MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict(
         ("ernie", "ErnieForPreTraining"),
         ("evolla", "EvollaForProteinText2Text"),
         ("exaone4", "Exaone4ForCausalLM"),
+        ("exaone_moe", "ExaoneMoeForCausalLM"),
         ("falcon_mamba", "FalconMambaForCausalLM"),
         ("flaubert", "FlaubertWithLMHeadModel"),
         ("flava", "FlavaForPreTraining"),
@@ -590,6 +599,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("ernie4_5", "Ernie4_5ForCausalLM"),
         ("ernie4_5_moe", "Ernie4_5_MoeForCausalLM"),
         ("exaone4", "Exaone4ForCausalLM"),
+        ("exaone_moe", "ExaoneMoeForCausalLM"),
         ("falcon", "FalconForCausalLM"),
         ("falcon_h1", "FalconH1ForCausalLM"),
         ("falcon_mamba", "FalconMambaForCausalLM"),
@@ -680,6 +690,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("rwkv", "RwkvForCausalLM"),
         ("seed_oss", "SeedOssForCausalLM"),
         ("smollm3", "SmolLM3ForCausalLM"),
+        ("solar_open", "SolarOpenForCausalLM"),
         ("stablelm", "StableLmForCausalLM"),
         ("starcoder2", "Starcoder2ForCausalLM"),
         ("trocr", "TrOCRForCausalLM"),
@@ -692,6 +703,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("xlnet", "XLNetLMHeadModel"),
         ("xlstm", "xLSTMForCausalLM"),
         ("xmod", "XmodForCausalLM"),
+        ("youtu", "YoutuForCausalLM"),
         ("zamba", "ZambaForCausalLM"),
         ("zamba2", "Zamba2ForCausalLM"),
     ]
@@ -871,6 +883,7 @@ MODEL_FOR_UNIVERSAL_SEGMENTATION_MAPPING_NAMES = OrderedDict(
         # Model for Universal Segmentation mapping
         ("detr", "DetrForSegmentation"),
         ("eomt", "EomtForUniversalSegmentation"),
+        ("eomt_dinov3", "EomtDinov3ForUniversalSegmentation"),
         ("mask2former", "Mask2FormerForUniversalSegmentation"),
         ("maskformer", "MaskFormerForInstanceSegmentation"),
         ("oneformer", "OneFormerForUniversalSegmentation"),
@@ -914,6 +927,7 @@ MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING_NAMES = OrderedDict(
         ("glm46v", "Glm46VForConditionalGeneration"),
         ("glm4v", "Glm4vForConditionalGeneration"),
         ("glm4v_moe", "Glm4vMoeForConditionalGeneration"),
+        ("glm_ocr", "GlmOcrForConditionalGeneration"),
         ("got_ocr2", "GotOcr2ForConditionalGeneration"),
         ("idefics", "IdeficsForVisionText2Text"),
         ("idefics2", "Idefics2ForConditionalGeneration"),
@@ -1028,6 +1042,7 @@ MODEL_FOR_OBJECT_DETECTION_MAPPING_NAMES = OrderedDict(
         ("deformable_detr", "DeformableDetrForObjectDetection"),
         ("detr", "DetrForObjectDetection"),
         ("lw_detr", "LwDetrForObjectDetection"),
+        ("pp_doclayout_v3", "PPDocLayoutV3ForObjectDetection"),
         ("rt_detr", "RTDetrForObjectDetection"),
         ("rt_detr_v2", "RTDetrV2ForObjectDetection"),
         ("table-transformer", "TableTransformerForObjectDetection"),