Conversation
Noticed our js_divergence_error metric builds the mixture distribution the same way as log(0.5*exp(a) + 0.5*exp(b)), which produces log(0) -> NaN once both logprobs underflow in fp32. Switching to logaddexp keeps it stable (same thing Automodel's parity_metrics.py already does). Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>
Adds a small CPU test that feeds ClippedPGLossFn logprobs low enough to underflow to zero in fp32, so it catches the NaN this same commit fixed. Signed-off-by: Kashif Rasul <kashif.rasul@gmail.com>
kashif
force-pushed
the
fix/jsd-metric-logaddexp-underflow
branch
from
September 6, 2026 11:25
ef7040a to
3cf6b2c
Compare
svcnvidia-nemo-ci
added
the
waiting-on-maintainers
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 ?
Fixes a NaN in the
js_divergence_errortraining metric when logprobs underflow to zero.Issues
Our JSD metric builds the mixture distribution as
log(0.5*exp(a) + 0.5*exp(b)). Once a logprob drops low enough thatexp()underflows to 0 in fp32, this becomeslog(0)and the KL terms downstream turn intoinf - inf, i.e. NaN. Switched totorch.logaddexp, which stays stable in that regime.Only affects the logged metric, not gradients/training - but a NaN there makes debugging confusing.
Usage
No usage change.
Before your PR is "Ready for review"
Pre checks:
Additional Information