Skip to content
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

Keep FFN output layer in float32 for T5 models #1239

Draft
wants to merge 6 commits into
base: master
Choose a base branch
from

Conversation

guillaumekln
Copy link
Collaborator

No description provided.

@guillaumekln guillaumekln marked this pull request as draft June 6, 2023 08:59
@bradfox2
Copy link

bradfox2 commented Jul 24, 2023

Hey @guillaumekln, thanks for the great library. Any idea as to when you might merge this to main? I'm currently utilizing the binaries of this PR built during CI and would love to get the performance improvements of 3.16+.

@guillaumekln
Copy link
Collaborator Author

I updated this branch if you want to try new binaries: https://github.com/OpenNMT/CTranslate2/actions/runs/5644199537?pr=1239

I'm still not sure we will merge this branch. The change is specific to a single model (T5) and I feel like it will get in the way of other features in the future.

If you have enough memory you could now try to use "bfloat16" with T5 models.

@bradfox2
Copy link

bradfox2 commented Aug 3, 2023

Thank you for the update. It's working well so far.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants