Skip to content

chore: patch KL loss to prevent nans #876

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
parthchadha merged 8 commits into from
Sep 25, 2025
Merged

chore: patch KL loss to prevent nans #876

Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
b7e7896
kl logprob clamp
rohitrango Aug 8, 2025
8677e4b
Merge remote-tracking branch 'origin/main' into rohit/kl_fix
parthchadha Sep 17, 2025
465e6ff
chore: align Megatron-LM submodule pointer with origin/main
parthchadha Sep 17, 2025
97ff7ea
Fix merge
parthchadha Sep 17, 2025
4a4e5de
revert: restore third-party files to match origin/main (undo unintend…
parthchadha Sep 17, 2025
29f7072
Merge branch 'main' into rohit/kl_fix
parthchadha Sep 17, 2025
bcd7638
Merge remote-tracking branch 'origin/main' into rohit/kl_fix
parthchadha Sep 24, 2025
aedac78
Merge branch 'main' into rohit/kl_fix
parthchadha Sep 25, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 5 additions & 1 deletion nemo_rl/algorithms/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,7 +31,9 @@


def calculate_kl_penalty_joschu2020(
logprobs_policy: torch.Tensor, logprobs_reference: torch.Tensor
logprobs_policy: torch.Tensor,
logprobs_reference: torch.Tensor,
clamp_value: Optional[float] = 20.0,
) -> torch.Tensor:
"""Calculates a per-token estimate of the KL Divergence between two log_probs.

Expand All @@ -41,6 +43,8 @@ def calculate_kl_penalty_joschu2020(
logprobs_reference: torch.Tensor (b, s)
"""
r = logprobs_reference - logprobs_policy
if clamp_value is not None:
r = r.clamp(min=-clamp_value, max=clamp_value)
return torch.exp(r) - r - 1


Expand Down
Loading