FIX: New bloom changes breaking prompt learning #1969
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
Bloom had two dimensions of the attention layer transposed (compared to all other transformers models), which was fixed by: huggingface/transformers#31445 Therefore, for future transformers versions, skip the special handling in PEFT. There is also an issue that prompt injection did not take place when past_key_values was a Cache object that is empty. This should now hopefully work as expected. Note: This needs a review from a transformers expert.
|
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 one of you @ArthurZucker or @gante be so kind to review this PR? Note that the code here is just unindented, so the actual diff is smaller than it seems. |
I don't think it will, for the foreseeable future. We're out of bandwidth to convert all models 💔 A note: the cache class has a |
BenjaminBossan
added a commit
to BenjaminBossan/peft
that referenced
this pull request
Aug 6, 2024
See huggingface#1969 The fix to the bloom architecture was not actually released in transformers 4.43.3, which makes the version check invalid. Instead, now checking an attribute on the BloomPreTrainedModel.
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Outdated suggestions cannot be applied.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.
Bloom had two dimensions of the attention layer transposed (compared to all other transformers models), which was fixed by:
huggingface/transformers#31445
Therefore, for future transformers versions, skip the special handling in PEFT.
There is also an issue that prompt injection did not take place when
past_key_valueswas aCacheobject that is empty. This should now hopefully work as expected.For the time being,
past_key_valuesis still being passed as a tuple of tensors, not sure if eventually this must be aCacheobject. If yes, this will need fixing in a future PR. Same goes for the usage ofget_seq_length()onCacheobjects.Note:
I tested this locally with transformers installed form main and bloom tests passed. PEFT CI only tests this nightly, not on PR.