🚨🚨 Generate: change order of ops in beam sample to avoid nans #26843
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
|
The documentation is not available anymore as the PR was closed or merged. |
gante
changed the title
ArthurZucker
approved these changes
Oct 17, 2023
Co-authored-by: Arthur <[email protected]>
4 tasks
EduardoPach
pushed a commit
to EduardoPach/transformers
that referenced
this pull request
Nov 19, 2023
…gface#26843) Co-authored-by: Arthur <[email protected]>
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Outdated suggestions cannot be applied.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?
Fixes #26332 -- see this comment for a full explanation.
TL;DR: in
beam_sample,logits_warperare now applied BEFORE adding the beam scores. We have been postponing this change to avoid introducing output differences, but the truth is that the order of operations causes issues (e.g. 1 2)This is technically a bug fix (
beam_sampleis unusable for long generations withtemperature < 1.0before this change), hence the lack of a deprecation cycle. However, it may alter some generated outputs (short generations with low temperature), hence the 🚨🚨.Please note that other common operators, like
top_pandtop_k, are unaltered by this change.