fix(mtp logging): Correctly accumulate MTP loss for logging when log_interval > 1 #1684
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
…interval > 1
In `multi_token_prediction.py`, the MTP loss is not correctly accumulated for logging when the
log_interval is greater than 1.
The current implementation overwrites the
`total_loss_dict[f"mtp_{i+1} loss"]` at each step. This means only the loss from the logging step itself (e.g., the 10th step
if `log_interval=10`) is stored. This single value is then incorrectly
divided by `log_interval`, resulting in a reported MTP loss that is
artificially low and does not reflect the true average.
This commit modifies the logic to correctly accumulate the `f"mtp_{i+1} loss"`
across all steps within a logging interval by checking for
the key's existence and using addition (`+=`) instead of assignment.
This fix ensures the reported MTP loss is accurate.
|
Thank you for the PR. We'll help merge it. |
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Outdated suggestions cannot be applied.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.
In
multi_token_prediction.py, the MTP loss is not correctly accumulated for logging when the log_interval is greater than 1.The current implementation overwrites the
total_loss_dict[f"mtp_{i+1} loss"]at each step. This means only the loss from the logging step itself (e.g., the 10th step iflog_interval=10) is stored. This single value is then incorrectly divided bylog_interval, resulting in a reported MTP loss that is artificially low and does not reflect the true average.This commit modifies the logic to correctly accumulate the
f"mtp_{i+1} loss"across all steps within a logging interval by checking for the key's existence and using addition (+=) instead of assignment.This fix ensures the reported MTP loss is accurate.