adapters

What it is and what it does

Adapters is an extension library for HuggingFace Transformers that integrates parameter-efficient fine-tuning methods into transformer models. Instead of updating all model weights during training, adapters add small trainable modules (like LoRA, prefix tuning, or bottleneck adapters) that achieve competitive performance with far fewer parameters. The library provides a unified interface for training these adapters, loading pre-trained ones from AdapterHub, and composing multiple adapters together for multi-task or ensemble scenarios.

The package depends on transformers and packaging, and requires Python 3.9+ and PyTorch 2.0+. It's designed for researchers and practitioners working on NLP tasks who want efficient fine-tuning without the memory and compute costs of full model updates. You can use it to adapt existing model setups, load community-shared adapters, or combine multiple adapters in a single model using composition blocks.

Use it for:

• Fine-tune large language models on domain-specific tasks with minimal GPU memory by training only adapter weights instead of all parameters.
• Load and apply pre-trained adapters from AdapterHub to perform sentiment analysis, question classification, or other NLP tasks on models like RoBERTa or T5.
• Compose multiple task-specific adapters in parallel or sequence to handle multi-task learning or ensemble predictions in a single model.
• Experiment with different parameter-efficient methods (LoRA, prefix tuning, bottleneck adapters) on the same base model using a unified configuration interface.
• Train quantized adapters (Q-LoRA, Q-Bottleneck) for even smaller memory footprints on resource-constrained hardware.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Adapters adds parameter-efficient fine-tuning methods to HuggingFace Transformers models, letting you train and compose multiple adapter modules on top of pre-trained models with minimal parameter overhead.

Yes. The package is actively maintained, has no known vulnerabilities, low install friction, and a permissive license. It solves a real problem—efficient fine-tuning of large models—with a well-documented, unified interface. Install it if you're doing transfer learning on transformer models and want to reduce training costs and memory usage.

Install

adapters on PyPI

pip install adapters

uv add adapters

poetry add adapters

Installing adapters

Before you install

Low install friction with only 2 runtime dependencies (transformers and packaging). The package is actively maintained with a recent release and no known vulnerabilities.

License in practice

Licensed under Apache (permissive), allowing commercial and private use with minimal restrictions.

Quickstart

pip install -U adapters

from adapters import AutoAdapterModel
from transformers import AutoTokenizer

model = AutoAdapterModel.from_pretrained("roberta-base")
tokenizer = AutoTokenizer.from_pretrained("roberta-base")
model.load_adapter("AdapterHub/roberta-base-pf-imdb", source="hf", set_active=True)
print(model(**tokenizer("This works great!", return_tensors="pt")).logits)

Requires Python 3.9+ and PyTorch 2.0+

Verify before relying

• Whether all 10+ adapter methods mentioned are fully functional in version 1.3.0
• Compatibility scope with the full range of 20+ Transformer models claimed
• Performance characteristics and memory savings compared to full model fine-tuning

Package facts

Evidence: adapters-1.3.0-py3-none-any.whl

Keywords: NLP, deep, learning, transformer, pytorch, BERT, adapters, PEFT, LoRA

Tags

More Artificial Intelligence packages

Further reading