-
Notifications
You must be signed in to change notification settings - Fork 31.3k
Fix EncoderDecoder cache #41612
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Fix EncoderDecoder cache #41612
Conversation
|
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. |
Cyrilvallez
left a comment
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Whisper and a few others used to rely on the same __init__ as dp/ddp, so they are not happy with the change unfortunately 🥲 Should not be too much trouble to fix it by adding a None though as the other tests on the other PR!
|
@remi-or I tested this branch on PEFT ( |
|
Nice! I think the failing tests have nothing to do with the PR, so this is ready to be merged imo :) |
|
run-slow: vit |
|
This comment contains run-slow, running the specified jobs: models: ['models/vit'] |
|
run-slow: vit |
|
This comment contains run-slow, running the specified jobs: models: ['models/vit'] |
Cyrilvallez
left a comment
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, LGTM! Thanks for the fix!
|
[For maintainers] Suggested jobs to run (before merge) run-slow: rag, whisper |
* Fix EncoderDecoder cache * Add the option for the ddp data tuples to have 2 elems * Modifiy the order of the KV and sliding * Adapted RAG and Whisper to new EncoderDecoderCache * A single comma * Remove kwargs in map * Fixed order in manual injection cache test * Slight changes to support legacy format * Removed Nonnes
In #41569 we restored thr
__iter__method toDynamicCachebut I missed the fact that it was also removed fromEncoderDecoderCache. This PR fixes that and modifies the__init__ofEncoderDecoderCachein the case of DDP, so it is compatible with the new system.