Fixes for gradient accumulation test #125
Merged
Conversation
This file contains 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
Earlier, we used deepcopy, which seems to copy the tensor python objects, but keeps a pointer to the same tensor, meaning we were not properly testing gradient accumulation. This commit mainly fixes that, using load_state_dict() instead to make a model copy. Unfortunately, the test fails once we do this, and a few changes are necessary to make it pass: 1. Switch AdamW -> SGD 2. Switch to float32 It's unclear why this was necessary, and solving this is left for a future commit.
|
Created #126 to keep track of the AdamW + grad accum issue. |
GeorgiosSmyrnis
approved these changes
Dec 4, 2023
There was a problem hiding this comment.
LGTM - let's keep track of the issue #126 and see what's going on.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
Earlier, we used deepcopy, which seems to copy the tensor python objects, but keeps a pointer to the same tensor, meaning we were not properly testing gradient accumulation. This commit mainly fixes that, using load_state_dict() instead to make a model copy.
Unfortunately, the test fails once we do this, and a few changes are necessary to make it pass:
It's unclear why this was necessary, and solving this is left for a future commit.