How huggingface/peft works
PEFT is a Python library for adapting pretrained models by training small adapter parameters rather than the full model. The main workflow is: a user supplies a pretrained model and PEFT configuration, the library wraps the model with a selected tuner, an external training loop updates adapter parameters, and the resulting adapter can be saved and later loaded for inference. The README also highlights quantization and integration with Transformers, Diffusers, and Accelerate. The graph captures those documented capabilities; sampled source is partial, so uncertain internal call wiring is omitted.
Main entry points: peft_model.py PEFT model · mapping.py Tuner mapping · mapping_func.py Model factory · __init__.py PEFT exports
Components 21
Public API
- PEFT exports
src/peft/__init__.py - Adapter config
src/peft/config.py - Tuner mapping
src/peft/mapping.py - Model factory
src/peft/mapping_func.py
Adapter Methods
- LoRA tuner
src/peft/tuners/lora/model.py - Prompt tuners
src/peft/tuners/prompt_tuning/model.py - Other tuners
src/peft/tuners - Mixed adapters
src/peft/mixed_model.py - PEFT optimizers
src/peft/optimizers
Model Runtime
- PEFT model
src/peft/peft_model.py - Auto loading
src/peft/auto.py - Tuner base utilities
src/peft/tuners/tuners_utils.py
Persistence And Integrations
- Adapter save/load
src/peft/utils/save_and_load.py - Ecosystem integrations
src/peft/utils/integrations.py - Quantization helpers
src/peft/utils/quantization_utils.py
Other components
- Library user
- Pretrained model
- Training loop
- Transformers
- Diffusers
- Accelerate
Connections 25
- Library user → PEFT exports: uses
- PEFT exports → Adapter config: configures
- Library user → Pretrained model: provides
- Library user → Model factory: calls factory
- Model factory → Tuner mapping: selects tuner
- Model factory → PEFT model: wraps model
- Tuner mapping → LoRA tuner: maps method
- Tuner mapping → Prompt tuners: maps method
- Tuner mapping → Other tuners: maps methods
- PEFT model → Pretrained model: wraps
- PEFT model → Tuner base utilities: uses
- Training loop → PEFT model: trains adapters
- Library user → Training loop: runs
- Library user → Adapter save/load: saves adapter
- Adapter save/load → PEFT model: loads weights
- Library user → Auto loading: loads for inference
- Auto loading → PEFT model: constructs
- PEFT model → Pretrained model: runs inference
- PEFT model → Mixed adapters: supports mixing
- Library user → PEFT optimizers: may use
- PEFT model → Quantization helpers: supports quantized models
- Transformers → Pretrained model: provides models
- Diffusers → PEFT model: integrates adapters
- Accelerate → Training loop: distributes training
- PEFT model → Ecosystem integrations: uses integrations