Correctly drop tokens in SwitchTransformer #37123
Open
+16
−4
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
Previously, the identity function was used for dropped tokens with a weight from the expert that was not applied to the hidden states. This was misleading, because dropping means, the expert weight is zero. Instead of trying to fix the weight, we take an easier approach by initializing with zeros. Fixes issue huggingface#37017
|
Hi 👋, thank you for opening this pull request! The pull request is converted to draft by default. The CI will be paused while the PR is in draft mode. When it is ready for review, please click the |
4 tasks
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
Would love to be part of this project one day. |
|
LGTM but I'm not an expert on MoE routing! @zucchini-nlp @ArthurZucker if you're happy with it feel free to merge |
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Outdated suggestions cannot be applied.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.
What does this PR do?
Previously, the identity function was used for dropped tokens with a weight from the expert that was not applied to the hidden states. This was misleading, because dropping means, the expert weight is zero. Instead of trying to fix the weight, we take an easier approach by initializing with zeros.
Fixes #37017
Related work
https://github.com/tensorflow/mesh/blob/e6798a2610a2c2f4c4cd236d8214422cb1ecc00a/mesh_tensorflow/transformer/moe.py#L1144 mentions that it needs to be zeroed out.
https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/moe.py#L507C18-L507C31 combines the results without any clone initialization beforehand.
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
text models: @ArthurZucker
last major changing person: @zucchini-nlp
person who requested PR: @Rocketknight1