Refactor adapter composition implementation #591
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Merged
lenglaender
approved these changes
Oct 18, 2023
There was a problem hiding this comment.
Looks great to me 👍
Can be merged once the backwards compatibility tests have been run (and once the part of the TODOs we want to do in this PR has been implemented).
And maybe rename tests_adapters/methods/test_adapter_common.py to sth. like tests_adapters/methods/test_bottleneck_adapter.py
Co-authored-by: Leon Engländer <[email protected]>
hSterz
reviewed
Oct 25, 2023
|
@hSterz @lenglaender also updated the docs accordingly now. Feel free to have another look. |
|
The documentation looks good to me |
2 tasks
calpt
added a commit
that referenced
this pull request
Nov 18, 2023
Follow-up to #591. This PR provides initial support for adapter composition in LoRA & (IA)³ modules. Currently LoRA & (IA)³ don't support composition. With this PR, the following blocks will be supported: **Stack, BatchSplit, Average, Parallel** Additionally, the LoRA implementation is refactored a bit in an effort to make it cleaner. ### Limitations - Split & Fuse compositions are **not** supported - LoRA/ (IA)³ composition is **not** supported for models using the `LoRAMergedLinear` implementation. These currently are: **GPT-2, DeBERTa (v1)**
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Refactors the implementation of composition blocks in the model forward pass such that more of the logic is shared between all adapter methods.
Changes
methodsfolderComposableAdapterLayerBaseas subclass ofAdapterLayerbaseas shared base class of all adapter methods that support composition.Stack,Parallel,BatchSplit,Average) which can be used by all subclasses.ComposableAdapterLayerBasemust be implemented. See https://github.com/calpt/adapter-transformers/blob/55fdc0cbe2f695914108a9c0e208127b13bc617e/src/adapters/methods/adapter_layer_base.py#L132-L222.NamedTuplein the base class. Deriving classes should define concreteNamedTuple-derived state classes. E.g., see https://github.com/calpt/adapter-transformers/blob/55fdc0cbe2f695914108a9c0e208127b13bc617e/src/adapters/methods/bottleneck.py#L22Splitcomposition block to support more than two child blocks. Splits are defined as a list of split indices, ie.Split("a", "b", "c", splits=[64, 64, 64]). Breaking changeAdapterLayer->BottleneckLayer;PrefixTuningShim->PrefixTuningLayerTodo
cc @TimoImhof would be nice to run some backwards compability tests here :)