Whisper: fix asr pipeline with seq2seq assistant model #30726
Closed
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 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. |
kamilakesbi
approved these changes
May 10, 2024
There was a problem hiding this comment.
Hi @gante,
We pass the encoder output directly to generate when doing short form generation with Whisper, but it doesn't seem necessary indeed! @sanchit-gandhi WDYT ?
I can add the code modifications to #30637.
|
(diff included in #30637 ) |
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Outdated suggestions cannot be applied.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 #29869
Fixes #30407
Fixes #30611 (related PR: #30637)
The ASR pipeline was preparing the encoder outputs before
generate(and not passinginput_features), but that's not needed: the exact same preparation is done insidegenerate, as it is a hard requirement to generate with encoder-decoder models.However, by not passing
input_features, it was blocking the proper use of encoder-decoder assistants that used a differentencoderoutput shape (= decoder input shape), and thus needed the inputs to run their own encoding step.Contrarily to #30637, this fix relies on lowering the complexity of our codebase 馃懠