Adapter Activation and Composition

One of the great advantages of using adapters is the possibility to combine multiple adapters trained on different tasks in various ways. To enable such adapter compositions, adapter-transformers comes with a modular and flexible concept to define how the input to the model should flow through the available adapters. This not only allows stacking (MAD-X) and fusing (AdapterFusion) adapters, but also even more complex adapter setups.

The single location where all the adapter composition magic happens is the active_adapters property of the model class. In the simplest case, you can set the name of a single adapter here to activate it:

model.active_adapters = "adapter_name"

Note that we also could have used model.set_active_adapters("adapter_name") which does the same.


active_adapters defines which of the available adapters are used in each forward and backward pass through the model. This means:

  • You cannot activate an adapter not previously added to the model using either add_adapter() or load_adapter().

  • All adapters not mentioned anywhere in the active_adapters setup are ignored although they might be loaded into the model. Thus, after adding an adapter, make sure to activate it.

The basic building blocks of the more advanced setups are simple objects derived from AdapterCompositionBlock, each representing a different possibility to combine single adapters. They are presented in more detail in the following.


Illustration of stacking adapters.

Stacking adapters using the ‘Stack’ block.

The Stack block can be used to stack multiple adapters on top of each other. This kind of adapter composition is used e.g. in the MAD-X framework for cross-lingual transfer (Pfeiffer et al., 2020), where language and task adapters are stacked on top of each other. For more, check out this Colab notebook on cross-lingual transfer.

In the following example, we stack the adapters a, b and c so that in each layer, the input is first passed through a, the output of a is then inputted to b and the output of b is finally inputted to c.

import transformers.adapters.composition as ac

// ...


model.active_adapters = ac.Stack("a", "b", "c")

In v1.x of adapter-transformers, stacking adapters was done using a list of adapter names, i.e. the example from above would be defined as ["a", "b", "c"]. For backwards compatibility, you can still do this, although it is recommended to use the new syntax.


Illustration of AdapterFusion.

Fusing adapters with AdapterFusion.

The Fuse block can be used to activate a fusion layer of adapters. AdapterFusion is a non-destructive way to combine the knowledge of multiple pre-trained adapters on a new downstream task, proposed by Pfeiffer et al., 2021. In the following example, we activate the adapters d, e and f as well as the fusion layer that combines the outputs of all three. The fusion layer is added beforehand using model.add_fusion() where we specify the names of the adapters which should be fused.

import transformers.adapters.composition as ac

// ...

model.add_fusion(["d", "e", "f"])

model.active_adapters = ac.Fuse("d", "e", "f")


Fusing adapters with the Fuse block only works successfully if an adapter fusion layer combining all of the adapters listed in the Fuse has been added to the model. This can be done either using add_fusion() or load_adapter_fusion().

To learn how training an AdapterFusion layer works, check out this Colab notebook from the adapter-transformers repo.

In v1.x of adapter-transformers, fusing adapters was done using a nested list of adapter names, i.e. the example from above would be defined as [["d", "e", "f"]]. For backwards compatibility, you can still do this, although it is recommended to use the new syntax.


Illustration of splitting adapters.

Splitting the input between two adapters using the ‘Stack’ block.

The Split block can be used to split an input sequence between two adapters. This is done by specifying a split index, at which the sequences should be divided. In the following example, we split each input sequence between adapters g and h. For each sequence, all tokens from 0 up to 63 are forwarded through g while all tokens beginning at index 64 are forwarded through h:

import transformers.adapters.composition as ac

// ...


model.active_adapters = ac.Split("g", "h", split_index=64)


Illustration of parallel adapter inference.

Parallel adapter inference as implemented by the ‘Parallel’ block. The input is replicated at the first layer with parallel adapters.

The Parallel can be used to enable parallel multi-task inference on different adapters, each with their own prediction head. Parallel adapter inference was first used in AdapterDrop: On the Efficiency of Adapters in Transformers (Rücklé et al., 2020).

In the following example, we load two adapters for semantic textual similarity (sts) from the Hub, one trained on the STS benchmark, the other trained on the MRPC dataset. We activate a parallel setup where the input is passed through both adapters and their respective prediction heads.

model = AutoModelWithHeads.from_pretrained("distilbert-base-uncased")
tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")

adapter1 = model.load_adapter("sts/sts-b@ukp")
adapter2 = model.load_adapter("sts/mrpc@ukp")

model.active_adapters = ac.Parallel(adapter1, adapter2)

input_ids = tokenizer("Adapters are great!", "Adapters are awesome!", return_tensors="pt")

output1, output2 = model(**input_ids)

print("STS-B adapter output:", output1[0].item())
print("MRPC adapter output:", bool(torch.argmax(output2[0]).item()))

Note that the Parallel block is only intended for inference, not for training adapters.

Nesting composition blocks

Of course, it is also possible to combine different composition blocks in one adapter setup. E.g., we can nest a Split block within a Stack of adapters:

import transformers.adapters.composition as ac

model.active_adapters = ac.Stack("a", ac.Split("b", "c", split_index=60))

However, combinations of adapter composition blocks cannot be arbitrarily deep. All currently supported possibilities are visualized in the figure below.

Adapter composition block combinations

Allowed nestings of adapter composition blocks.