Auto classes

AutoConfig, AutoModel, AutoModelFor<Task>, AutoTokenizer, AutoImageProcessor, AutoVideoProcessor, AutoFeatureExtractor, and AutoProcessor are the late-binding factories most users interact with. They take a checkpoint identifier, look at its model_type, and instantiate the right concrete class.

Why they exist

The library has 462 architectures. Forcing every user to import the exact class (from transformers import LlamaForCausalLM) would be brittle: a checkpoint might switch architectures on the Hub, or a notebook might try several models. The auto factories decouple the user code from the architecture.

Where they live

src/transformers/models/auto/
├── __init__.py
├── auto_factory.py             # the magic: from_pretrained → resolve class → instantiate
├── auto_mappings.py            # OrderedDicts mapping model_type → class
├── configuration_auto.py       # AutoConfig + CONFIG_MAPPING
├── feature_extraction_auto.py  # AutoFeatureExtractor + FEATURE_EXTRACTOR_MAPPING
├── image_processing_auto.py    # AutoImageProcessor + IMAGE_PROCESSOR_MAPPING
├── modeling_auto.py            # Auto*Model* classes + MODEL_*_MAPPING
├── processing_auto.py          # AutoProcessor + PROCESSOR_MAPPING
├── tokenization_auto.py        # AutoTokenizer + TOKENIZER_MAPPING
└── video_processing_auto.py    # AutoVideoProcessor + VIDEO_PROCESSOR_MAPPING

auto_factory.py defines the meta-class machinery; auto_mappings.py is the registry of model_type strings to classes.

How AutoModelForCausalLM.from_pretrained resolves

graph LR
    User --> AM[AutoModelForCausalLM.from_pretrained repo_id]
    AM --> CFG[AutoConfig.from_pretrained]
    CFG --> JSON[Read config.json]
    JSON --> MT[config.model_type = llama]
    MT --> Map[MODEL_FOR_CAUSAL_LM_MAPPING llama → LlamaForCausalLM]
    Map --> Inst[LlamaForCausalLM.from_pretrained]

The mapping MODEL_FOR_CAUSAL_LM_MAPPING (and dozens of siblings: MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING, MODEL_FOR_VISION_2_SEQ_MAPPING, MODEL_FOR_OBJECT_DETECTION_MAPPING, MODEL_FOR_IMAGE_TEXT_TO_TEXT_MAPPING, …) is hand-edited in auto_mappings.py but make fix-repo enforces alphabetical order via utils/sort_auto_mappings.py.

All auto-class flavours

Auto class	Mapping	Use
AutoConfig	CONFIG_MAPPING	Load any config
AutoModel	MODEL_MAPPING	Backbone (no head)
AutoModelForCausalLM	MODEL_FOR_CAUSAL_LM_MAPPING	Decoder LMs
AutoModelForMaskedLM	MODEL_FOR_MASKED_LM_MAPPING	BERT-style
AutoModelForSeq2SeqLM	MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING	T5, BART, …
AutoModelForSequenceClassification	…	Classification
AutoModelForTokenClassification	…	NER
AutoModelForQuestionAnswering	…	Extractive QA
AutoModelForMultipleChoice	…	Multi-choice
AutoModelForVision2Seq	…	Image captioning, OCR
AutoModelForImageTextToText	…	VLMs
AutoModelForImageClassification	…	ViT-style
AutoModelForObjectDetection	…	DETR-style
AutoModelForSemanticSegmentation	…	Segmenter
AutoModelForUniversalSegmentation	…	OneFormer / Mask2Former
AutoModelForDepthEstimation	…	Depth nets
AutoModelForVideoClassification	…	TimeSformer, VideoMAE
AutoModelForAudioClassification	…	AST
AutoModelForCTC, AutoModelForSpeechSeq2Seq, AutoModelForAudioFrameClassification, AutoModelForXVector, AutoModelForAudioXVector	…	ASR / audio
AutoModelForTextToWaveform, AutoModelForTextToSpectrogram	…	TTS
AutoModelForMaskGeneration	…	SAM
AutoModelForKeypointDetection, AutoModelForKeypointMatching	…	Keypoints
AutoModelForZeroShotImageClassification, AutoModelForZeroShotObjectDetection	…	Zero-shot vision
AutoModelForAnyToAny	…	"Any modality in, any modality out" models
AutoTokenizer	TOKENIZER_MAPPING	Tokenizer
AutoImageProcessor	IMAGE_PROCESSOR_MAPPING	Image preprocessor
AutoVideoProcessor	VIDEO_PROCESSOR_MAPPING	Video preprocessor
AutoFeatureExtractor	FEATURE_EXTRACTOR_MAPPING	Audio preprocessor
AutoProcessor	PROCESSOR_MAPPING	Multimodal bundler
AutoBackbone	BACKBONE_MAPPING	Backbone for detection/segmentation

The auto_mappings.py file is the single source of truth — auto_factory.py reads it at import time.

Registering a new model

When you add a new architecture, edit auto_mappings.py to add an entry:

("my_arch", "MyArchConfig"),                     # to CONFIG_MAPPING_NAMES
("my_arch", "MyArchForCausalLM"),                # to MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
("my_arch", ("MyArchTokenizer", "MyArchTokenizerFast")),  # to TOKENIZER_MAPPING_NAMES

make fix-repo then sorts the lists. The matching import is generated from the names; do not add manual imports.

Trust remote code

Auto classes also handle the trust_remote_code=True codepath. When a checkpoint declares auto_map: {AutoModelForCausalLM: <module.MyClass>} in its config, AutoModelForCausalLM.from_pretrained(..., trust_remote_code=True) will fetch that module from the Hub and instantiate the class. This is how community models that aren't merged into transformers still work.

The dynamic-loader is in src/transformers/dynamic_module_utils.py (37K LOC). resolve_trust_remote_code enforces an interactive confirmation when running outside a known-safe environment.

Integration points

• Configuration — AutoConfig is the entry point.
• Modeling — AutoModelFor<Task> instantiates concrete classes.
• Pipelines — pipeline() uses the auto classes internally.
• from_pretrained — auto classes layer on top of the universal loader.

Entry points for modification

• New auto class for a new task → add to modeling_auto.py, register the mapping, add tests in tests/models/auto/.
• New model_type → edit auto_mappings.py and run make fix-repo.
• Trust-remote-code policy → src/transformers/dynamic_module_utils.py.