transformers 5.10.2__tar.gz → 5.10.4__tar.gz

{transformers-5.10.2/src/transformers.egg-info → transformers-5.10.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: transformers
-Version: 5.10.2
+Version: 5.10.4
 Summary: Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.
 Home-page: https://github.com/huggingface/transformers
 Author: The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)

{transformers-5.10.2 → transformers-5.10.4}/setup.py

@@ -104,7 +104,7 @@ _deps = [
     "pandas<2.3.0",  # `datasets` requires `pandas` while `pandas==2.3.0` has issues with CircleCI on 2025/06/05
     "packaging>=20.0",
     "parameterized>=0.9",  # older version of parameterized cause pytest collection to fail on .expand
-    "peft>=0.18.0",
+    "peft>=0.19.0",
     "phonemizer",
     "protobuf",
     "psutil",
@@ -325,7 +325,7 @@ if __name__ == "__main__":
 
     setup(
         name="transformers",
-        version="5.10.2",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+        version="5.10.4",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
         author="The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)",
         author_email="transformers@huggingface.co",
         description="Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.",

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/__init__.py

@@ -18,7 +18,7 @@
 # to defer the actual importing for when the objects are requested. This way `import transformers` provides the names
 # in the namespace without actually importing anything (and especially none of the backends).
 
-__version__ = "5.10.2"
+__version__ = "5.10.4"
 
 import importlib
 import sys

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/dependency_versions_table.py

@@ -36,7 +36,7 @@ deps = {
     "pandas": "pandas<2.3.0",
     "packaging": "packaging>=20.0",
     "parameterized": "parameterized>=0.9",
-    "peft": "peft>=0.18.0",
+    "peft": "peft>=0.19.0",
     "phonemizer": "phonemizer",
     "protobuf": "protobuf",
     "psutil": "psutil",

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/integrations/peft.py

@@ -58,7 +58,7 @@ if is_accelerate_available():
     from accelerate.utils import get_balanced_memory, infer_auto_device_map
 
 # Minimum PEFT version supported for the integration
-MIN_PEFT_VERSION = "0.18.2"
+MIN_PEFT_VERSION = "0.19.0"
 
 
 logger = logging.get_logger(__name__)
@@ -415,7 +415,7 @@ class PeftAdapterMixin:
     prompt tuning, prompt learning are out of scope as these adapters are not "injectable" into a torch module. For
     using these methods, please refer to the usage guide of PEFT library.
 
-    With this mixin, if the correct PEFT version is installed (>= 0.18.0), it is possible to:
+    With this mixin, if the correct PEFT version is installed (>= 0.19.0), it is possible to:
 
     - Load an adapter stored on a local path or in a remote Hub repository, and inject it in the model
     - Attach new adapters in the model and train them with Trainer or by your own.

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/audioflamingo3/configuration_audioflamingo3.py

@@ -100,7 +100,6 @@ class AudioFlamingo3Config(PreTrainedConfig):
     audio_token_id: int = 151669
     projector_hidden_act: str = "gelu"
     projector_bias: bool = True
-    tie_word_embeddings: bool = True
 
     def __post_init__(self, **kwargs):
         if isinstance(self.audio_config, dict):

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/audioflamingo3/modeling_audioflamingo3.py

@@ -567,7 +567,7 @@ class AudioFlamingo3Model(AudioFlamingo3PreTrainedModel):
 )
 class AudioFlamingo3ForConditionalGeneration(AudioFlamingo3PreTrainedModel, GenerationMixin):
     _keep_in_fp32_modules_strict = ["embed_positions"]
-    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
+    _tied_weights_keys = None
 
     def __init__(self, config):
         super().__init__(config)

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/audioflamingo3/modular_audioflamingo3.py

@@ -273,6 +273,8 @@ class AudioFlamingo3Model(VoxtralModel):
     """
 )
 class AudioFlamingo3ForConditionalGeneration(VoxtralForConditionalGeneration):
+    _tied_weights_keys = None
+
     def __init__(self, config):
         super().__init__(config)
         self.model = AudioFlamingo3Model(config)

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/auto/tokenization_auto.py

@@ -33,7 +33,7 @@ from ...utils import (
     is_tokenizers_available,
     logging,
 )
-from ...utils.hub import cached_file
+from ...utils.hub import cached_file, has_file
 from ..encoder_decoder import EncoderDecoderConfig
 from .auto_factory import _LazyAutoMapping
 from .configuration_auto import (
@@ -363,7 +363,6 @@ MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS: set[str] = {
     "h2ovl_chat",
     "hyperclovax_vlm",
     "internlm2",
-    "internvl_chat",
     "jamba",
     "janus",
     "llava",
@@ -413,6 +412,25 @@ def load_merges(merges_file):
     return merges
 
 
+def _has_tekken_tokenizer_file(
+    pretrained_model_name_or_path: str | os.PathLike[str],
+    **kwargs,
+) -> bool:
+    subfolder = kwargs.get("subfolder", "")
+    tekken_filename = os.path.join(subfolder, "tekken.json") if subfolder else "tekken.json"
+    try:
+        return has_file(
+            pretrained_model_name_or_path,
+            tekken_filename,
+            revision=kwargs.get("revision"),
+            token=kwargs.get("token"),
+            cache_dir=kwargs.get("cache_dir"),
+            local_files_only=kwargs.get("local_files_only", False),
+        )
+    except OSError:
+        return False
+
+
 def tokenizer_class_from_name(class_name: str) -> type[Any] | None:
     # Bloom tokenizer classes were removed but should map to the fast backend for BC
     if class_name in {"BloomTokenizer", "BloomTokenizerFast"}:
@@ -698,6 +716,7 @@ class AutoTokenizer:
                 config = PreTrainedConfig.from_pretrained(pretrained_model_name_or_path, **kwargs)
 
         config_model_type = config.model_type
+        config_model_name = config.model_name if hasattr(config, "model_name") else None
 
         # Next, let's try to use the tokenizer_config file to get the tokenizer class.
         tokenizer_config = get_tokenizer_config(pretrained_model_name_or_path, **kwargs)
@@ -714,22 +733,32 @@ class AutoTokenizer:
 
         # if there is a config, we can check that the tokenizer class != than model class.
         # Use the config class if it's a specialized tokenizer, otherwise fall back to TokenizersBackend.
+        # Hub class should prioritize tokenizer_config.json or fallback to config.tokenizer_class.
+        _hub_class = tokenizer_config_class or getattr(config, "tokenizer_class", None)
         if (
             tokenizer_auto_map is None
-            and tokenizer_config_class is not None
+            and _hub_class is not None
             and config_model_type is not None
             and config_model_type != ""
             and TOKENIZER_MAPPING_NAMES.get(config_model_type) is not None
             and (TOKENIZER_MAPPING_NAMES.get(config_model_type).removesuffix("Fast"))
-            != (tokenizer_config_class.removesuffix("Fast"))
+            != (_hub_class.removesuffix("Fast"))
         ):
             registered_class_name = TOKENIZER_MAPPING_NAMES.get(config_model_type).removesuffix("Fast")
-            if registered_class_name not in ("TokenizersBackend", "PythonBackend", "PreTrainedTokenizerFast"):
+            if registered_class_name not in (
+                "TokenizersBackend",
+                "PythonBackend",
+                "PreTrainedTokenizerFast",
+                "MistralCommonBackend",
+            ):
                 # If the hub class is known incorrect for this model type, use the registered class; otherwise trust the hub.
                 class_name = (
                     registered_class_name
-                    if config_model_type in MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS
-                    else tokenizer_config_class
+                    if (
+                        config_model_type in MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS
+                        or config_model_name in MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS
+                    )
+                    else _hub_class
                 )
                 tokenizer_class = tokenizer_class_from_name(class_name)
                 if tokenizer_class is not None and tokenizer_class.__name__ not in (
@@ -739,11 +768,21 @@ class AutoTokenizer:
                 ):
                     return tokenizer_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
 
+            if (
+                registered_class_name == "MistralCommonBackend"
+                and is_mistral_common_available()
+                and "fix_mistral_regex" not in kwargs
+                and _has_tekken_tokenizer_file(pretrained_model_name_or_path, **kwargs)
+            ):
+                tokenizer_class = tokenizer_class_from_name("MistralCommonBackend")
+                if tokenizer_class is not None:
+                    return tokenizer_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
+
             if TokenizersBackend is not None:
                 return TokenizersBackend.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
 
             raise ValueError(
-                f"Tokenizer class '{tokenizer_config_class}' specified in the tokenizer config was not found. "
+                f"Tokenizer class '{_hub_class}' specified in the tokenizer config was not found. "
                 f"The tokenizer may need to be converted or re-saved."
             )
 
@@ -773,7 +812,11 @@ class AutoTokenizer:
             )
         )
         # V5: Skip remote tokenizer for custom models with incorrect hub tokenizer class
-        if has_remote_code and config_model_type in MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS:
+        if (
+            has_remote_code
+            and config_model_type in MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS
+            and trust_remote_code is not True
+        ):
             has_remote_code = False
             tokenizer_auto_map = None
 
@@ -836,6 +879,11 @@ class AutoTokenizer:
         if model_type is not None:
             tokenizer_class = TOKENIZER_MAPPING.get(type(config), TokenizersBackend)
             if tokenizer_class is not None:
+                if getattr(tokenizer_class, "__name__", None) == "MistralCommonBackend" and (
+                    "fix_mistral_regex" in kwargs
+                    or not _has_tekken_tokenizer_file(pretrained_model_name_or_path, **kwargs)
+                ):
+                    tokenizer_class = TokenizersBackend
                 return tokenizer_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
 
         # Fallback: try tokenizer_class from tokenizer_config.json

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/glmasr/modular_glmasr.py

@@ -379,6 +379,8 @@ class GlmAsrModel(AudioFlamingo3Model):
     """
 )
 class GlmAsrForConditionalGeneration(AudioFlamingo3ForConditionalGeneration):
+    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
+
     def __init__(self, config):
         super().__init__(config)
         self.model = GlmAsrModel(config)

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/musicflamingo/configuration_musicflamingo.py

@@ -66,7 +66,6 @@ class MusicFlamingoConfig(PreTrainedConfig):
     audio_token_id: int = 151669
     projector_hidden_act: str = "gelu"
     projector_bias: bool = True
-    tie_word_embeddings: bool = False
 
     audio_bos_token_id: int = 151670
     audio_eos_token_id: int = 151671

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/musicflamingo/modeling_musicflamingo.py

@@ -409,7 +409,7 @@ class MusicFlamingoCausalLMOutputWithPast(ModelOutput):
 )
 class MusicFlamingoForConditionalGeneration(MusicFlamingoPreTrainedModel, GenerationMixin):
     _keep_in_fp32_modules_strict = ["embed_positions"]
-    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
+    _tied_weights_keys = None
 
     def __init__(self, config: MusicFlamingoConfig):
         super().__init__(config)

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/musicflamingo/modular_musicflamingo.py

@@ -78,7 +78,6 @@ class MusicFlamingoConfig(AudioFlamingo3Config):
     audio_eos_token_id: int = 151671
     audio_frame_step: float = 0.01
     rope_parameters: dict | None = None
-    tie_word_embeddings: bool = False
 
     def __post_init__(self, **kwargs):
         if self.rope_parameters is None:

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/qwen2_audio/configuration_qwen2_audio.py

@@ -98,7 +98,6 @@ class Qwen2AudioConfig(PreTrainedConfig):
     audio_config: dict | PreTrainedConfig | None = None
     text_config: dict | PreTrainedConfig | None = None
     audio_token_index: int = 151646
-    tie_word_embeddings: bool = True
 
     def __post_init__(self, **kwargs):
         if isinstance(self.audio_config, dict):

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/qwen2_audio/modeling_qwen2_audio.py

@@ -766,8 +766,6 @@ class Qwen2AudioModel(Qwen2AudioPreTrainedModel):
     """
 )
 class Qwen2AudioForConditionalGeneration(Qwen2AudioPreTrainedModel, GenerationMixin):
-    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
-
     def __init__(self, config: Qwen2AudioConfig):
         super().__init__(config)
         self.model = Qwen2AudioModel(config)

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/vibevoice_asr/configuration_vibevoice_asr.py

@@ -75,7 +75,6 @@ class VibeVoiceAsrConfig(PreTrainedConfig):
     audio_bos_token_id: int = 151646
     audio_eos_token_id: int = 151647
     acoustic_tokenizer_chunk_size: int = 1440000
-    tie_word_embeddings: bool = True
 
     def __post_init__(self, **kwargs):
         if isinstance(self.acoustic_tokenizer_encoder_config, dict):

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/vibevoice_asr/modeling_vibevoice_asr.py

@@ -444,8 +444,6 @@ class VibeVoiceAsrModel(VibeVoiceAsrPreTrainedModel):
     """
 )
 class VibeVoiceAsrForConditionalGeneration(VibeVoiceAsrPreTrainedModel, GenerationMixin):
-    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
-
     def __init__(self, config: VibeVoiceAsrConfig):
         super().__init__(config)
         self.model = VibeVoiceAsrModel(config)

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/vibevoice_asr/modular_vibevoice_asr.py

@@ -88,7 +88,6 @@ class VibeVoiceAsrConfig(PreTrainedConfig):
     audio_bos_token_id: int = 151646
     audio_eos_token_id: int = 151647
     acoustic_tokenizer_chunk_size: int = 1440000
-    tie_word_embeddings: bool = True
 
     def __post_init__(self, **kwargs):
         if isinstance(self.acoustic_tokenizer_encoder_config, dict):
@@ -361,8 +360,6 @@ class VibeVoiceAsrModel(VibeVoiceAsrPreTrainedModel):
     """
 )
 class VibeVoiceAsrForConditionalGeneration(VibeVoiceAsrPreTrainedModel, GenerationMixin):
-    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
-
     def __init__(self, config: VibeVoiceAsrConfig):
         super().__init__(config)
         self.model = VibeVoiceAsrModel(config)

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/voxtral/configuration_voxtral.py

@@ -110,7 +110,6 @@ class VoxtralConfig(PreTrainedConfig):
     text_config: dict | PreTrainedConfig | None = None
     audio_token_id: int | None = None
     projector_hidden_act: str = "gelu"
-    tie_word_embeddings: bool = True
 
     def __post_init__(self, **kwargs):
         if isinstance(self.audio_config, dict):

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/voxtral/modeling_voxtral.py

@@ -487,7 +487,6 @@ class VoxtralModel(VoxtralPreTrainedModel):
 )
 class VoxtralForConditionalGeneration(VoxtralPreTrainedModel, GenerationMixin):
     _keep_in_fp32_modules_strict = ["embed_positions"]
-    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
 
     def __init__(self, config):
         super().__init__(config)

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/voxtral/modular_voxtral.py

@@ -257,7 +257,6 @@ class VoxtralModel(VoxtralPreTrainedModel):
 )
 class VoxtralForConditionalGeneration(VoxtralPreTrainedModel, GenerationMixin):
     _keep_in_fp32_modules_strict = ["embed_positions"]
-    _tied_weights_keys = {"lm_head.weight": "model.language_model.embed_tokens.weight"}
 
     def __init__(self, config):
         super().__init__(config)

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/models/voxtral_realtime/configuration_voxtral_realtime.py

@@ -56,7 +56,6 @@ class VoxtralRealtimeTextConfig(PreTrainedConfig):
     pad_token_id: int | None = None
     bos_token_id: int | None = 1
     eos_token_id: int | list[int] | None = 2
-    tie_word_embeddings: bool = False
     rope_parameters: RopeParameters | dict | None = None
     sliding_window: int | None = 4096
     attention_dropout: float | int = 0.0

{transformers-5.10.2 → transformers-5.10.4}/src/transformers/processing_utils.py

@@ -873,24 +873,30 @@ class ProcessorMixin(PushToHubMixin):
         batch_replacement_offsets = []
         for batch_idx in range(len(text)):
             last = 0
+            offset = 0
             replacement_offsets = []
             expanded_sample = []
             for m in re.finditer(regex_special_mm_tokens, text[batch_idx]):
                 start, end = m.span()
                 expanded_sample.append(text[batch_idx][last:start])
 
+                # adjust spans using running offset if one sample has several MM data associated
+                start_with_offset = start + offset
+
                 mm_type = m.lastgroup
                 replacement_text = next(replacements_iters[mm_type])
                 replacement_offsets.append(
                     {
                         "type": mm_type,
                         "span": (start, end),
-                        "new_span": (start, start + len(replacement_text)),
+                        "new_span": (start_with_offset, start_with_offset + len(replacement_text)),
                         "text": m.group(),
                         "replacement": replacement_text,
                     }
                 )
                 expanded_sample.append(replacement_text)
+                # update the offsets and the last position
+                offset += len(replacement_text) - (end - start)
                 last = end
 
             expanded_sample.append(text[batch_idx][last:])
@@ -941,16 +947,34 @@ class ProcessorMixin(PushToHubMixin):
     # These values are used to build `mm_token_type_ids`
     @property
     def image_token_ids(self) -> list[int | None]:
+        if _image_token_ids := getattr(self, "_image_token_ids", None):
+            return _image_token_ids
         return [getattr(self, "image_token_id", None)]
 
+    @image_token_ids.setter
+    def image_token_ids(self, value: list[int | None]):
+        setattr(self, "_image_token_ids", value)
+
     @property
     def video_token_ids(self) -> list[int | None]:
+        if _video_token_ids := getattr(self, "_video_token_ids", None):
+            return _video_token_ids
         return [getattr(self, "video_token_id", None)]
 
+    @video_token_ids.setter
+    def video_token_ids(self, value: list[int | None]):
+        setattr(self, "_video_token_ids", value)
+
     @property
     def audio_token_ids(self) -> list[int | None]:
+        if _audio_token_ids := getattr(self, "_audio_token_ids", None):
+            return _audio_token_ids
         return [getattr(self, "audio_token_id", None)]
 
+    @audio_token_ids.setter
+    def audio_token_ids(self, value: list[int | None]):
+        setattr(self, "_audio_token_ids", value)
+
     def check_argument_for_proper_class(self, argument_name, argument):
         """
         Checks the passed argument's class against the expected transformers class. In case of an unexpected

{transformers-5.10.2 → transformers-5.10.4/src/transformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: transformers
-Version: 5.10.2
+Version: 5.10.4
 Summary: Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.
 Home-page: https://github.com/huggingface/transformers
 Author: The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)

{transformers-5.10.2 → transformers-5.10.4}/tests/test_processing_common.py

@@ -2076,7 +2076,8 @@ class ProcessorTesterMixin:
         for h, w in image_sizes:
             image_inputs.append(np.random.randint(255, size=(h, w, 3), dtype=np.uint8))
 
-        text = [f"This is an image {getattr(self, 'image_token', '')}"] * len(image_inputs)
+        image_token = getattr(self, "image_token", "")
+        text = [f"This is an image {image_token}"] * len(image_inputs)
         inputs = processor(
             text=text, images=image_inputs, padding=True, return_mm_token_type_ids=True, return_tensors="pt"
         )
@@ -2087,3 +2088,17 @@ class ProcessorTesterMixin:
         num_image_tokens_from_call = inputs.mm_token_type_ids.sum(-1).tolist()
         num_image_tokens_from_helper = processor._get_num_multimodal_tokens(image_sizes=image_sizes)
         self.assertListEqual(num_image_tokens_from_call, num_image_tokens_from_helper["num_image_tokens"])
+
+        # Test with two images per single text
+        text = [f"These are two images {image_token}{image_token}"] * len(image_inputs)
+        inputs = processor(
+            text=text,
+            images=image_inputs * 2,
+            padding=True,
+            return_mm_token_type_ids=True,
+            return_tensors="pt",
+        )
+
+        num_image_tokens_from_call = inputs.mm_token_type_ids.sum(-1).tolist()
+        num_image_tokens_from_helper = processor._get_num_multimodal_tokens(image_sizes=image_sizes * 2)
+        self.assertEqual(sum(num_image_tokens_from_call), sum(num_image_tokens_from_helper["num_image_tokens"]))