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

Add Pytorch Native AMP support in Trainer #6151

Merged
merged 7 commits into from
Jul 31, 2020

Conversation

prajjwal1
Copy link
Contributor

Pytorch 1.6 introduces Native AMP support. This eliminates the need to build and install Apex and provides improvements over problems highlighted in Apex #818 and flexibility. This is the recommended way to use AMP.

With this PR, Trainer will automatically use Pytorch's native AMP if 1.6 version is installed, otherwise, it will use Apex.

This PR will close #6115.

Copy link
Collaborator

@sgugger sgugger left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks a lot for the PR! It's very clean, I just have one suggestion.

src/transformers/trainer.py Outdated Show resolved Hide resolved
Copy link
Member

@LysandreJik LysandreJik left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That's really cool, thanks a lot @prajjwal1 !

@LysandreJik LysandreJik merged commit 0034a1d into huggingface:master Jul 31, 2020
@prajjwal1 prajjwal1 deleted the pytorch_native_amp branch July 31, 2020 08:38
Mehrad0711 pushed a commit to Mehrad0711/transformers that referenced this pull request Aug 3, 2020
* fixed type; add Pytorch Native CUDA AMP support

* reverted commit on modeling_utils

* confirming to HF black formatting rule

* changed bool value of _use_apex

* scaler support for gradient clipping

* fix inplace operation of clip_grad_norm

* removed not while version comparison
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.

Usage of Pytorch Native AMP in place of apex (Pytorch 1.6) in Trainer
4 participants