Config kwargs

src/transformers/configuration_utils.py

@@ -19,7 +19,8 @@
 import math
 import os
 from collections.abc import Sequence
-from dataclasses import dataclass
+from dataclasses import MISSING, dataclass, fields
+from functools import wraps
 from typing import TYPE_CHECKING, Any, ClassVar, Literal, TypeVar, Union
 
 from huggingface_hub import create_repo
@@ -71,6 +72,46 @@
 )
 
 
+# copied from huggingface_hub.dataclasses.strict when `accept_kwargs=True`
+def wrap_init_to_accept_kwargs(cls: dataclass):
+    original_init = cls.__init__
+
+    @wraps(original_init)
+    def __init__(self, *args, **kwargs: Any) -> None:
+        # Extract only the fields that are part of the dataclass
+        dataclass_fields = {f.name for f in fields(cls)}
+        standard_kwargs = {k: v for k, v in kwargs.items() if k in dataclass_fields}
+
+        # We need to call bare `__init__` without `__post_init__` but the `original_init` of
+        # any dataclas contains a call to post-init at the end (without kwargs)
+        if len(args) > 0:
+            raise ValueError(
+                f"{cls.__name__} accepts only keyword arguments, but found `{len(args)}` positional args."
+            )
+
+        for f in fields(cls):  # type: ignore
+            if f.name in standard_kwargs:
+                setattr(self, f.name, standard_kwargs[f.name])
+            elif f.default is not MISSING:
+                setattr(self, f.name, f.default)
+            elif f.default_factory is not MISSING:
+                setattr(self, f.name, f.default_factory())
+            else:
+                raise TypeError(f"Missing required field - '{f.name}'")
+
+        # Pass any additional kwargs to `__post_init__` and let the object
+        # decide whether to set the attr or use for different purposes (e.g. BC checks)
+        additional_kwargs = {}
+        for name, value in kwargs.items():
+            if name not in dataclass_fields:
+                additional_kwargs[name] = value
+
+        self.__post_init__(**additional_kwargs)
+
+    cls.__init__ = __init__
+    return cls
+
+
 @strict(accept_kwargs=True)
 @dataclass(repr=False)
 class PreTrainedConfig(PushToHubMixin, RotaryEmbeddingConfigMixin):
@@ -252,8 +293,16 @@ def __post_init__(self, **kwargs):
 
     def __init_subclass__(cls, *args, **kwargs):
         super().__init_subclass__(*args, **kwargs)
+        cls_has_custom_init = "__init__" in cls.__dict__
         cls = dataclass(cls, repr=False)
 
+        if not cls_has_custom_init:
+            # Wrap all subclasses to accept arbitrary kwargs for BC
+            # only if the subclass has no custom `__init__`. Most
+            # remote code has an init defined, but some model are not
+            # See https://huggingface.co/hmellor/Ilama-3.2-1B/blob/main/configuration_ilama.py
+            cls = wrap_init_to_accept_kwargs(cls)
+
     @property
     def name_or_path(self) -> str | None:
         return getattr(self, "_name_or_path", None)

src/transformers/models/afmoe/configuration_afmoe.py

@@ -20,7 +20,7 @@
 from ...utils import auto_docstring
 
 
-@strict(accept_kwargs=True)
+@strict
 @auto_docstring(
     custom_intro="""
     AFMoE is an Adaptive Feedforward MoE (Mixture of Experts) model with token-choice routing, shared experts, and a

src/transformers/models/aimv2/configuration_aimv2.py

@@ -28,7 +28,7 @@
 
 
 @auto_docstring(checkpoint="apple/aimv2-large-patch14-224-lit")
-@strict(accept_kwargs=True)
+@strict
 class Aimv2VisionConfig(PreTrainedConfig):
     r"""
     use_head (`str`, *optional*, defaults to `True`):
@@ -72,7 +72,7 @@ class Aimv2VisionConfig(PreTrainedConfig):
 
 
 @auto_docstring(checkpoint="apple/aimv2-large-patch14-224-lit")
-@strict(accept_kwargs=True)
+@strict
 class Aimv2TextConfig(PreTrainedConfig):
     r"""
     Example:
@@ -111,7 +111,7 @@ def __post_init__(self, **kwargs):
 
 
 @auto_docstring(checkpoint="apple/aimv2-large-patch14-224-lit")
-@strict(accept_kwargs=True)
+@strict
 class Aimv2Config(PreTrainedConfig):
     r"""
     max_logit_scale (`float`, *optional*, defaults to `100.0`):

src/transformers/models/aimv2/modular_aimv2.py

@@ -38,7 +38,7 @@
 
 
 @auto_docstring(checkpoint="apple/aimv2-large-patch14-224-lit")
-@strict(accept_kwargs=True)
+@strict
 class Aimv2VisionConfig(SiglipVisionConfig):
     r"""
     use_head (`str`, *optional*, defaults to `True`):
@@ -79,7 +79,7 @@ class Aimv2VisionConfig(SiglipVisionConfig):
 
 
 @auto_docstring(checkpoint="apple/aimv2-large-patch14-224-lit")
-@strict(accept_kwargs=True)
+@strict
 class Aimv2TextConfig(SiglipTextConfig):
     vocab_size: int = 49408
     hidden_size: int = 768
@@ -102,7 +102,7 @@ def __post_init__(self, **kwargs):
 
 
 @auto_docstring(checkpoint="apple/aimv2-large-patch14-224-lit")
-@strict(accept_kwargs=True)
+@strict
 class Aimv2Config(SiglipConfig):
     r"""
     max_logit_scale (`float`, *optional*, defaults to `100.0`):

src/transformers/models/albert/configuration_albert.py

@@ -21,7 +21,7 @@
 
 
 @auto_docstring(checkpoint="albert/albert-xxlarge-v2")
-@strict(accept_kwargs=True)
+@strict
 class AlbertConfig(PreTrainedConfig):
     r"""
     inner_group_num (`int`, *optional*, defaults to 1):

src/transformers/models/align/configuration_align.py

@@ -23,7 +23,7 @@
 
 
 @auto_docstring(checkpoint="kakaobrain/align-base")
-@strict(accept_kwargs=True)
+@strict
 class AlignTextConfig(PreTrainedConfig):
     r"""
     Example:
@@ -62,7 +62,7 @@ class AlignTextConfig(PreTrainedConfig):
 
 
 @auto_docstring(checkpoint="kakaobrain/align-base")
-@strict(accept_kwargs=True)
+@strict
 class AlignVisionConfig(PreTrainedConfig):
     r"""
     width_coefficient (`float`, *optional*, defaults to 2.0):
@@ -155,7 +155,7 @@ def __post_init__(self, **kwargs):
 
 
 @auto_docstring(checkpoint="kakaobrain/align-base")
-@strict(accept_kwargs=True)
+@strict
 class AlignConfig(PreTrainedConfig):
     r"""
     temperature_init_value (`float`, *optional*, defaults to 1.0):

src/transformers/models/altclip/configuration_altclip.py

@@ -23,7 +23,7 @@
 
 
 @auto_docstring(checkpoint="BAAI/AltCLIP")
-@strict(accept_kwargs=True)
+@strict
 class AltCLIPTextConfig(PreTrainedConfig):
     r"""
     project_dim (`int`, *optional*, defaults to 768):
@@ -66,7 +66,7 @@ class AltCLIPTextConfig(PreTrainedConfig):
 
 
 @auto_docstring(checkpoint="BAAI/AltCLIP")
-@strict(accept_kwargs=True)
+@strict
 class AltCLIPVisionConfig(PreTrainedConfig):
     r"""
     Example:
@@ -103,7 +103,7 @@ class AltCLIPVisionConfig(PreTrainedConfig):
 
 
 @auto_docstring(checkpoint="BAAI/AltCLIP")
-@strict(accept_kwargs=True)
+@strict
 class AltCLIPConfig(PreTrainedConfig):
     r"""
     Example:

src/transformers/models/apertus/configuration_apertus.py

@@ -26,7 +26,7 @@
 
 
 @auto_docstring(checkpoint="swiss-ai/Apertus-8B-Instruct-2509")
-@strict(accept_kwargs=True)
+@strict
 class ApertusConfig(PreTrainedConfig):
     r"""
     ```python

src/transformers/models/apertus/modular_apertus.py

@@ -44,7 +44,7 @@
 
 
 @auto_docstring(checkpoint="swiss-ai/Apertus-8B-Instruct-2509")
-@strict(accept_kwargs=True)
+@strict
 class ApertusConfig(PreTrainedConfig):
     r"""
     ```python

src/transformers/models/arcee/configuration_arcee.py

@@ -27,7 +27,7 @@
 
 
 @auto_docstring(checkpoint="arcee-ai/AFM-4.5B")
-@strict(accept_kwargs=True)
+@strict
 class ArceeConfig(PreTrainedConfig):
     r"""
     ```python

src/transformers/models/arcee/modular_arcee.py

@@ -32,7 +32,7 @@
 
 
 @auto_docstring(checkpoint="arcee-ai/AFM-4.5B")
-@strict(accept_kwargs=True)
+@strict
 class ArceeConfig(LlamaConfig):
     r"""
     ```python

src/transformers/models/aria/configuration_aria.py

@@ -27,7 +27,7 @@
 
 
 @auto_docstring(checkpoint="rhymes-ai/Aria")
-@strict(accept_kwargs=True)
+@strict
 class AriaTextConfig(PreTrainedConfig):
     r"""
     moe_num_experts (`int`, *optional*, defaults to 8):
@@ -100,7 +100,7 @@ def validate_architecture(self):
 
 
 @auto_docstring(checkpoint="rhymes-ai/Aria")
-@strict(accept_kwargs=True)
+@strict
 class AriaConfig(PreTrainedConfig):
     r"""
     projector_patch_to_query_dict (`dict`, *optional*):

src/transformers/models/aria/modular_aria.py

@@ -98,7 +98,7 @@ def sequential_experts_gemm(token_states, expert_weights, tokens_per_expert):
 
 
 @auto_docstring(checkpoint="rhymes-ai/Aria")
-@strict(accept_kwargs=True)
+@strict
 class AriaTextConfig(LlamaConfig):
     r"""
     moe_num_experts (`int`, *optional*, defaults to 8):
@@ -129,7 +129,7 @@ class AriaTextConfig(LlamaConfig):
 
 
 @auto_docstring(checkpoint="rhymes-ai/Aria")
-@strict(accept_kwargs=True)
+@strict
 class AriaConfig(PreTrainedConfig):
     r"""
     projector_patch_to_query_dict (`dict`, *optional*):

src/transformers/models/audio_spectrogram_transformer/configuration_audio_spectrogram_transformer.py

@@ -20,7 +20,7 @@
 
 
 @auto_docstring(checkpoint="MIT/ast-finetuned-audioset-10-10-0.4593")
-@strict(accept_kwargs=True)
+@strict
 class ASTConfig(PreTrainedConfig):
     r"""
     frequency_stride (`int`, *optional*, defaults to 10):

src/transformers/models/audioflamingo3/configuration_audioflamingo3.py

@@ -22,7 +22,7 @@
 
 
 @auto_docstring(checkpoint="nvidia/audio-flamingo-3-hf")
-@strict(accept_kwargs=True)
+@strict
 class AudioFlamingo3EncoderConfig(PreTrainedConfig):
     r"""
     max_source_positions (`int`, *optional*, defaults to 1500):
@@ -69,7 +69,7 @@ class AudioFlamingo3EncoderConfig(PreTrainedConfig):
 
 
 @auto_docstring(checkpoint="nvidia/audio-flamingo-3-hf")
-@strict(accept_kwargs=True)
+@strict
 class AudioFlamingo3Config(PreTrainedConfig):
     r"""
     Example:

src/transformers/models/autoformer/configuration_autoformer.py

@@ -20,7 +20,7 @@
 
 
 @auto_docstring(checkpoint="huggingface/autoformer-tourism-monthly")
-@strict(accept_kwargs=True)
+@strict
 class AutoformerConfig(PreTrainedConfig):
     r"""
     prediction_length (`int`):

src/transformers/models/aya_vision/configuration_aya_vision.py

@@ -21,7 +21,7 @@
 
 
 @auto_docstring(checkpoint="CohereForAI/aya-vision-8b")
-@strict(accept_kwargs=True)
+@strict
 class AyaVisionConfig(PreTrainedConfig):
     r"""
     downsample_factor (`int`, *optional*, defaults to 2):

src/transformers/models/bamba/configuration_bamba.py

@@ -20,7 +20,7 @@
 from ...utils import auto_docstring
 
 
-@strict(accept_kwargs=True)
+@strict
 @auto_docstring(
     custom_intro="""
     The BambaModel is a hybrid [mamba2](https://github.com/state-spaces/mamba) architecture with SwiGLU.

src/transformers/models/bark/configuration_bark.py

@@ -24,7 +24,7 @@
 
 
 @auto_docstring(checkpoint="suno/bark")
-@strict(accept_kwargs=True)
+@strict
 class BarkSubModelConfig(PreTrainedConfig):
     r"""
     block_size (`int`, *optional*, defaults to 1024):
@@ -64,7 +64,7 @@ class BarkSubModelConfig(PreTrainedConfig):
 
 
 @auto_docstring(checkpoint="suno/bark")
-@strict(accept_kwargs=True)
+@strict
 class BarkSemanticConfig(BarkSubModelConfig):
     r"""
     block_size (`int`, *optional*, defaults to 1024):
@@ -101,7 +101,7 @@ class BarkSemanticConfig(BarkSubModelConfig):
 
 
 @auto_docstring(checkpoint="suno/bark")
-@strict(accept_kwargs=True)
+@strict
 class BarkCoarseConfig(BarkSubModelConfig):
     r"""
     block_size (`int`, *optional*, defaults to 1024):
@@ -138,7 +138,7 @@ class BarkCoarseConfig(BarkSubModelConfig):
 
 
 @auto_docstring(checkpoint="suno/bark")
-@strict(accept_kwargs=True)
+@strict
 class BarkFineConfig(BarkSubModelConfig):
     r"""
     block_size (`int`, *optional*, defaults to 1024):
@@ -184,7 +184,7 @@ class BarkFineConfig(BarkSubModelConfig):
 
 
 @auto_docstring(checkpoint="suno/bark")
-@strict(accept_kwargs=True)
+@strict
 class BarkConfig(PreTrainedConfig):
     r"""
     semantic_config ([`BarkSemanticConfig`], *optional*):

src/transformers/models/bart/configuration_bart.py

@@ -20,7 +20,7 @@
 
 
 @auto_docstring(checkpoint="facebook/bart-large")
-@strict(accept_kwargs=True)
+@strict
 class BartConfig(PreTrainedConfig):
     r"""
     Example:

src/transformers/models/beit/configuration_beit.py

@@ -21,7 +21,7 @@
 
 
 @auto_docstring(checkpoint="microsoft/beit-base-patch16-224-pt22k")
-@strict(accept_kwargs=True)
+@strict
 class BeitConfig(BackboneConfigMixin, PreTrainedConfig):
     r"""
     use_mask_token (`bool`, *optional*, defaults to `False`):

src/transformers/models/bert/configuration_bert.py

@@ -21,7 +21,7 @@
 
 
 @auto_docstring(checkpoint="google-bert/bert-base-uncased")
-@strict(accept_kwargs=True)
+@strict
 class BertConfig(PreTrainedConfig):
     r"""
     Examples: