Adding Adapters to a Model

This document gives an overview of how new model architectures of Hugging Face Transformers can be supported by adapters. Before delving into implementation details, you should familiarize yourself with the main design philosophies of adapters:

  • Adapters should integrate seamlessly with existing model classes: If a model architecture supports adapters, it should be possible to use them with all model classes of this architecture.

  • Copied code should be minimal: adapters extensively uses Python mixins to add adapter support to HF models. Functions that cannot be sufficiently modified by mixins are copied and then modified. Try to avoid copying functions as much as possible.

Relevant Classes

Adding adapter support to an existing model architecture requires modifying some parts of the model forward pass logic. These modifications are realized by the four files in the src/adapters/models/<model_type>/ directory. Let’s examine the purpose of these files in the example of BERT. It’s important to note that we are adapting the original Hugging Face model, implemented in transformers/models/bert/modeling_bert.py. The files in src/adapters/models/bert/ are:

  1. src/adapters/models/bert/mixin_bert.py: This file contains mixins for each class we want to change. For example, in the BertSelfAttention class, we need to make changes for LoRA and Prefix Tuning. For this, we create a BertSelfAttentionAdaptersMixin to implement these changes. We will discuss how this works in detail below.

  2. src/adapters/models/bert/modeling_bert.py: For some classes of the BERT implementation (e.g. BertModel or BertLayer) the code can be sufficiently customized via mixins. For other classes (like BertSelfAttention), we need to edit the original code directly. These classes are copied into src/adapters/models/bert/modeling_bert.py and modified.

  3. src/adapters/models/bert/adapter_model.py: In this file, the adapter model class is defined. This class allows flexible adding of and switching between multiple prediction heads of different types. This looks about the same for each model, except that each model has different heads and thus different add_..._head() functions.

  4. src/adapters/models/bert/__init__.py: Defines Python’s import structure.

Implementation Steps 📝

Now that we have discussed the purpose of every file in src/adapters/models/<model_type>/, we go through the integration of adapters into an existing model architecture step by step. The following steps might not be applicable to every model architecture.

  1. Files:

    • Create the src/adapters/models/<model_type>/ directory and in it the 4 files: mixin_<model_type>.py, modeling_<model_type>.py adapter_model.py and __init__.py

  2. Mixins:

    • In src/adapters/models/<model_type>/mixin_<model_type>.py, create mixins for any class you want to change and where you can’t reuse an existing mixin from another class.

      • To figure out which classes to change, think about where to insert LoRA, Prefix Tuning, and bottleneck adapters.

      • You can use similar model implementations for guidance.

      • Often, existing mixins of another class can be reused. E.g. BertLayer, RobertaLayer, XLMRobertaLayer, DebertaLayer, DebertaV2Layer and BertGenerationLayer (all models derived from BERT) use the BertLayerAdaptersMixin.

    • To additionally support Prefix Tuning, it’s necessary to apply the forward call to the PrefixTuningLayer module in the respective attention layer (see step 3 for how to modify the code of an Hugging Face class).

    • Make sure the calls to bottleneck_layer_forward() are added in the right places.

    • The mixin for the whole base model class (e.g., BertModel) should derive from ModelBaseAdaptersMixin and (if possible) EmbeddingAdaptersMixin and/or InvertibleAdaptersMixin. This mixin should at least implement the iter_layers() method but might require additional modifications depending on the architecture.

      • If the model is a combination of different models, such as the EncoderDecoderModel, use ModelUsingSubmodelsAdaptersMixin instead of ModelBaseAdaptersMixin.

  3. Copied functions:

    • For those classes where the mixin is not enough to realize the wanted behavior, you must:

    • Create a new class in src/adapters/models/<model_type>/modeling_<model_type>.py with the name <class>WithAdapters. This class should derive from the corresponding mixin and HF class.

    • Copy the function you want to change into this class and modify it.

      • e.g., the forward method of the BertSelfAttention class must be adapted to support prefix tuning. We therefore create a class BertSelfAttentionWithAdapters(BertSelfAttentionAdaptersMixin, BertSelfAttention), copy the forward method into it and modify it.

      • if the forward method of a module is copied and modified, make sure to call adapters.utils.patch_forward() in the module’s init_adapters() method. This ensures adapters work correctly with the accelerate package.

  4. Modify MODEL_MIXIN_MAPPING

    • For each mixin whose class was not copied into modeling_<model_type>.py, add the mixin/class combination into MODEL_MIXIN_MAPPING in the file src/adapters/models/__init__.py.

  5. Create the adapter model:

    • Adapter-supporting architectures should provide a new model class <model_type>AdapterModel. This class allows flexible adding of and switching between multiple prediction heads of different types.

    • This is done in the adapter_model.py file:

      • This module should implement the <model_type>AdapterModel class, deriving from ModelWithFlexibleHeadsAdaptersMixin and <model_type>PreTrainedModel.

      • In the model class, add methods for those prediction heads that make sense for the new model architecture.

      • Again, have a look at existing implementations.

    • Add <model_type>AdapterModel to the ADAPTER_MODEL_MAPPING_NAMES mapping in src/adapters/models/auto/adapter_model.py and to src/adapters/__init__.py.

    • Define the classes to be added to Python’s import structure in src/adapters/models/<model_type>/__init__.py. This will likely only be the <model_type>AdapterModel.

  6. Adapt the config classes:

    • Adapt the config class to the requirements of adapters in src/adapters/wrappers/configuration.py.

    • There are some naming differences in the config attributes of different model architectures. The adapter implementation requires some additional attributes with a specific name to be available. These currently are num_attention_heads, hidden_size, hidden_dropout_prob and attention_probs_dropout_prob as in the BertConfig class. If your model config does not provide these, add corresponding mappings to CONFIG_CLASS_KEYS_MAPPING.

Additional (optional) implementation steps 📝

  • Parallel adapter inference via Parallel composition block (cf. documentation, PR#150).

  • Provide mappings for an architecture’s existing (static) prediction heads into adapters flex heads (cf. implementation).

Testing

❓ In addition to the general Hugging Face model tests, there are adapter-specific test cases. All tests are executed from the tests folder. You need to add two different test classes.

📝 Steps

  1. Add a new test_<model_type>.py module in tests/

    • This file is used to test that everything related to the usage of adapters (adding, removing, activating, …) works.

    • This module typically holds 2 test classes and a test base class:

      • <model_type>AdapterTestBase: This class contains the tokenizer_name, config_class and config.

      • <model_type>AdapterTest derives from a collection of test mixins that hold various adapter tests (depending on the implementation).

      • (optionally) <model_type>ClassConversionTest runs tests for correct class conversion if conversion of prediction heads is implemented.

  2. Add a new test_<model_type>.py module in tests/models/

    • This file is used to test the AdapterModel class.

    • This module typically holds 1 test class with the name <model_type>AdapterModelTest

      • <model_type>AdapterModelTest derives directly from Hugging Face’s existing model test class <model_type>ModelTest and adds <model_type>AdapterModel as a class to test.

Documentation

❓ The documentation for adapters lives in the docs folder.

📝 Steps

  • Add docs/classes/models/<model_type>.rst (oriented at the doc file in the HF docs). Make sure to include <model_type>AdapterModel autodoc. Finally, list the file in index.rst.

  • Add a new row for the model in the model table of the overview page at docs/model_overview.md, listing all the methods implemented by the new model.

Training Example Adapters

❓ To make sure the new adapter implementation works properly, it is useful to train some example adapters and compare the training results to full model fine-tuning. Ideally, this would include training adapters on one (or more) tasks that are good for demonstrating the new model architecture (e.g. GLUE benchmark for BERT, summarization for BART) and uploading them to AdapterHub.

We provide training scripts for many tasks here: https://github.com/Adapter-Hub/adapters/tree/main/examples/pytorch/