Implementation:Huggingface Transformers Create Dummy Models

Knowledge Sources	Huggingface_Transformers
Domains	Testing_Infrastructure, Model_Architecture
Last Updated	2026-02-13 20:00 GMT

Overview

Concrete tool for creating tiny (minimal-size) versions of all model architectures in the Transformers library for fast CI testing.

Description

The create_dummy_models.py utility (1477 lines) generates tiny-random-* models used across the CI test suite. For each model config class, it retrieves processor classes from auto mappings, builds processors from Hub checkpoints (with recursive fallback strategies), obtains tiny configs from model tester classes, converts tokenizers to reduced vocabulary (~1024 tokens) using train_new_from_iterator, shrinks image processor sizes to match tiny configs, then instantiates and saves each model architecture with the tiny config. Handles composite models (encoder-decoder, vision-text) specially. Supports multiprocessing, uploading to HuggingFace Hub, and generates detailed reports.

Usage

Run periodically to regenerate tiny models when new model architectures are added, or when model test infrastructure changes require updated tiny models.

Code Reference

Source Location

• Repository: Huggingface_Transformers
• File: utils/create_dummy_models.py
• Lines: 1-1477

Signature

def get_tiny_config(config_class: type) -> PretrainedConfig:
    """Retrieve a minimal configuration from the model tester class."""

def convert_tokenizer(tokenizer, tiny_config) -> PreTrainedTokenizer:
    """Reduce tokenizer vocabulary to ~1024 tokens."""

def build_processor(config_class: type) -> Optional[ProcessorMixin]:
    """Build the appropriate processor for a model config class."""

def create_tiny_models(
    output_dir: str,
    all_model_classes: List[type] = None,
    upload: bool = False,
    organization: str = "hf-internal-testing",
    num_workers: int = 1,
) -> Dict:
    """Create tiny random models for all architectures."""

Import

python utils/create_dummy_models.py --output_dir tiny_models/ --upload

I/O Contract

Inputs

Name	Type	Required	Description
--output_dir	str	Yes	Directory to save tiny models
--upload	flag	No	Upload to HuggingFace Hub
--organization	str	No	Hub organization (default: hf-internal-testing)
--num_workers	int	No	Number of parallel workers

Outputs

Name	Type	Description
Model directories	Directories	Tiny model files for each architecture
Report	JSON	Summary of created/failed models
Hub uploads	Hub repos	Uploaded tiny-random-* repos (if --upload)

Usage Examples

Creating Tiny Models

# Create tiny models locally
python utils/create_dummy_models.py --output_dir ./tiny_models/

# Create and upload to Hub
python utils/create_dummy_models.py --output_dir ./tiny_models/ \
    --upload --organization hf-internal-testing --num_workers 4

Related Pages

• Principle:Huggingface_Transformers_Tiny_Model_Generation