How huggingface/transformers works
Transformers is a model-definition library for inference and training across text, vision, audio, video, and multimodal tasks. A typical user workflow is to select a pretrained model, prepare inputs through a pipeline or modality processor, run model inference or generation, and receive task-specific outputs. The repository also provides model/configuration and tokenizer support, distributed execution, export backends, and a CLI serving interface. Arrows follow README-described behavior or references and calls visible in sampled files; unsampled subsystem links are omitted where their wiring is not established.
Main entry points: modeling_utils.py Pretrained Models · __init__.py Task Pipelines · processing_utils.py Multimodal Processors · transformers.py Transformers CLI
Components 20
User Interfaces
- Task Pipelines
src/transformers/pipelines/__init__.py - Transformers CLI
src/transformers/cli/transformers.py - Serving Server
src/transformers/cli/serving/server.py - Model Manager
src/transformers/cli/serving/model_manager.py
Input Preparation
- Multimodal Processors
src/transformers/processing_utils.py - Tokenizers
src/transformers/tokenization_utils_base.py - Feature Extraction
src/transformers/feature_extraction_utils.py
Models and Generation
- Pretrained Models
src/transformers/modeling_utils.py - Model Architectures
src/transformers/models - Auto Model Selection
src/transformers/models/auto/modeling_auto.py - Text Generation
src/transformers/generation/utils.py - Continuous Batching
src/transformers/generation/continuous_batching/continuous_api.py - Model Outputs
src/transformers/modeling_outputs.py
Execution and Export
- Distributed Execution
src/transformers/distributed - Framework Integrations
src/transformers/integrations - Model Export
src/transformers/exporters
Shared Model Assets
- Model Configuration
src/transformers/configuration_utils.py - Generation Cache
src/transformers/cache_utils.py - Hugging Face Hub
Other components
- User
Connections 22
- User → Task Pipelines: submits task
README.md - Task Pipelines → Multimodal Processors: preprocesses input
README.md - Task Pipelines → Pretrained Models: runs inference
README.md - Task Pipelines → Model Outputs: returns results
README.md - User → Transformers CLI: runs command
README.md - Transformers CLI → Serving Server: registers serve
src/transformers/cli/transformers.py - Serving Server → Model Manager: dispatches model requests
src/transformers/cli/serving/server.py - Model Manager → Auto Model Selection: selects model class
src/transformers/cli/serving/model_manager.py - Auto Model Selection → Model Architectures: resolves architecture
src/transformers/models/rag/modeling_rag.py - Pretrained Models → Model Configuration: loads configuration
src/transformers/modeling_utils.py - Model Architectures → Pretrained Models: implements model definitions
src/transformers/models/rag/modeling_rag.py - Pretrained Models → Model Outputs: produces outputs
README.md - Pretrained Models → Text Generation: supports generation
README.md - Text Generation → Generation Cache: uses decoding cache
- Multimodal Processors → Tokenizers: uses tokenization
src/transformers/processing_utils.py - Multimodal Processors → Feature Extraction: prepares modality features
src/transformers/processing_utils.py - Continuous Batching → Generation Cache: manages cache
- Pretrained Models → Distributed Execution: supports distributed loading
src/transformers/modeling_utils.py - Pretrained Models → Framework Integrations: uses optional backends
src/transformers/modeling_utils.py - User → Hugging Face Hub: selects checkpoint
README.md - Pretrained Models → Hugging Face Hub: loads pretrained checkpoint
README.md - Pretrained Models → Model Export: can export model