-
Notifications
You must be signed in to change notification settings - Fork 3.6k
[CI] Small fixes for DeepSpeed. #7823
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
Conversation
Codecov Report
@@ Coverage Diff @@
## master #7823 +/- ##
=======================================
- Coverage 93% 88% -5%
=======================================
Files 199 199
Lines 13016 13026 +10
=======================================
- Hits 12051 11438 -613
- Misses 965 1588 +623 |
| model_parallel_context = super().model_sharded_context() | ||
|
|
||
| with model_parallel_context: | ||
| with torch.cuda.amp.autocast(), model_parallel_context: |
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.
I think for ZeRO 2/3, autocast isn't used as we rely on a custom DeepSpeed FP16Optimizer to handle this, which converts the entire model to half iirc.
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.
With newer DeepSpeed version, they check the weights are created in half precision and currently fails in CI.
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.
oh very interesting! i'll check myself when I get time, do they throw an error if the weights are in half precision or not in half precision?
|
If you go through the special tests (for some reason they pass even though a test fails) there still is a failing test. I've removed a line in #7841 and will debug there, as I'm unsure why the changes in this PR are necessary and need more information since the version of DeepSpeed shouldn't have changed |
@awaelchli you looked at this already somewhere else, right? |
|
Yes here #7790 |
|
Hey @SeanNaren, Closing this PR for now. Best, |
What does this PR do?
Fixes #<issue_number>
Before submitting
PR review
Anyone in the community is free to review the PR once the tests have passed.
Before you start reviewing make sure you have read Review guidelines. In short, see the following bullet-list:
Did you have fun?
Make sure you had fun coding 🙃