Expose DeepSpeed FP16 parameters due to loss instability #6115
Conversation
This file contains hidden or 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
SeanNaren
requested review from
Borda,
awaelchli,
carmocca,
justusschock,
tchaton and
williamFalcon
as code owners
SeanNaren
added
3rd party
SeanNaren
commented
Feb 21, 2021
Codecov Report
@@ Coverage Diff @@
## master #6115 +/- ##
======================================
- Coverage 93% 93% -0%
======================================
Files 160 160
Lines 11405 11405
======================================
- Hits 10661 10629 -32
- Misses 744 776 +32 |
carmocca
approved these changes
Feb 21, 2021
Co-authored-by: Carlos Mocholí <[email protected]>
kaushikb11
approved these changes
Feb 21, 2021
awaelchli
approved these changes
Feb 21, 2021
Borda
approved these changes
Feb 21, 2021
SeanNaren
added a commit
that referenced
this pull request
Mar 16, 2021
* Expose deepspeed config parameters to init function due to instability in parameters * See if tests can run on normal CI, without special tests * Add changelog * Update pytorch_lightning/plugins/training_type/deepspeed.py Co-authored-by: Carlos Mocholí <[email protected]> Co-authored-by: Carlos Mocholí <[email protected]> (cherry picked from commit 432e563)
Borda
pushed a commit
that referenced
this pull request
Mar 16, 2021
* Expose deepspeed config parameters to init function due to instability in parameters * See if tests can run on normal CI, without special tests * Add changelog * Update pytorch_lightning/plugins/training_type/deepspeed.py Co-authored-by: Carlos Mocholí <[email protected]> Co-authored-by: Carlos Mocholí <[email protected]> (cherry picked from commit 432e563) Add missing config
lexierule
pushed a commit
that referenced
this pull request
Mar 16, 2021
* Expose deepspeed config parameters to init function due to instability in parameters * See if tests can run on normal CI, without special tests * Add changelog * Update pytorch_lightning/plugins/training_type/deepspeed.py Co-authored-by: Carlos Mocholí <[email protected]> Co-authored-by: Carlos Mocholí <[email protected]> (cherry picked from commit 432e563) Add missing config
This was referenced Mar 17, 2021
Closed
Merged
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
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?
Received a few reports that DeepSpeed quickly gives off NaNs when using ZeRO. This seems to be related to default Loss scaling values, which are not exposed currently. As a result the user needs to override everything to set these values.
Expose these values so that the user can access them. I thought about putting the values into the Precision plugin, but then I'd need the training_type_plugin to be aware of the precision plugin and I'm not a fan of that, given a longer term plan of the Training Type plugin handling precision.
Also I've included a few tests of edge cases that needed to be checked + seeing if we need to run the single GPU tests as special tests (might need to revert this, we'll see).
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 🙃