Fix FSDP mixed precision setting and loss w/ accelerate #465
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
Signed-off-by: Mustafa Eyceoz <[email protected]>
RobotSail
approved these changes
Apr 16, 2025
RobotSail
approved these changes
Apr 16, 2025
Member
RobotSail
left a comment
RobotSail
left a comment
There was a problem hiding this comment.
Tested locally and confirmed that this PR produces the same loss as DeepSpeed
Signed-off-by: Mustafa Eyceoz <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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Outdated suggestions cannot be applied.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.
Accelerate requires mixed_precision to be passed directly in config, else the fp32 upcast gets skipped and training loses overall precision, resulting in notably worse performance. Passing this also makes our manual mixed precision policy redundant, and since we are now actually using mixed precision, hybrid shard finally becomes the new default sharding strategy to keep memory expectations in parity.