Adapters

AdapterHub earns its place by covering ground most PEFT libraries skip. The method list runs from classical bottleneck adapters, whose config strings were renamed for clarity (houlsby is now double_seq_bn, pfeiffer is now seq_bn, parallel is par_seq_bn, with +inv variants), through Language Adapters with invertible adapters, Prefix Tuning, Compacter, LoRA, (IA)^3, Vera, DoRA, Prompt Tuning and ReFT. Where it pulls clearly ahead is composition and multi-task work: Adapter Activation and Composition gives you Stack, Fuse, Split, BatchSplit, MultiTask, Parallel output averaging and Nesting composition blocks, Method Combinations offers Mix-and-Match Adapters and UniPELT, and Multi Task Methods adds MTL-LoRA and AdapterFusion. That is why researchers reach for it when they need one harness to compare methods rather than one method in isolation. The engineering utilities are practical rather than decorative: Merging LoRA Adapters (plus other LoRA variants) folds adapters back into base weights for deployment, Quantized Model Training and Gradient Checkpointing cut memory, Prediction Heads supports AdapterModel classes, static heads from Hugging Face Transformers, automatic conversion and custom heads, and Embeddings covers adding, deleting, training and saving embeddings. Loading and Sharing connects to Hugging Face's Model Hub for downloading and uploading, so published adapters are discoverable rather than trapped locally. The honest costs: it is PyTorch-only for the models on the Model Overview page, it is a stand-alone package rather than a drop-in replacement for adapter-transformers, so imports move from transformers to adapters and Hugging Face model classes need an explicit adapters.init(model) call. Migration also drops three capabilities — transformers pipelines with adapters, invertible adapters in Hugging Face model classes (use AdapterModel instead), and loading save_pretrained model-and-adapter checkpoints through Hugging Face classes. Adapter weights themselves stay backward compatible, so previously trained adapters load without degradation. The documentation is split between this library and the legacy adapter-transformers docs at docs-legacy.adapterhub.ml, and the legacy codebase now lives in an archival repository — a real source of confusion when following older tutorials. Treat AdapterHub as research and experimental infrastructure: excellent for benchmarking PEFT methods and building modular multi-task NLP systems on limited compute, poor as a managed service, since nothing here supplies auto-scaling, monitoring or a no-code interface. Teams without PyTorch and Transformers fluency will spend their first week on setup rather than on modeling.