Skip to content

Fix: skip KL backward graph when coefficient is zero #2258

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

Open
Dodojordi wants to merge 1 commit into
base: main
Choose a base branch
from
Open

Fix: skip KL backward graph when coefficient is zero #2258

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
18 changes: 14 additions & 4 deletions slime/backends/megatron_utils/loss.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1053,18 +1053,28 @@ def policy_loss_function(
if args.use_kl_loss:
ref_log_probs = batch["ref_log_probs"]
ref_log_probs = torch.cat(ref_log_probs, dim=0)
kl_loss_coef = float(args.kl_loss_coef)
with_kl_grad = kl_loss_coef != 0.0

# Keep reference KL metrics when the coefficient is zero, but detach
# their inputs so the metric-only path does not retain an actor
# backward graph or contaminate policy gradients with non-finite KL.
kl_log_probs = log_probs if with_kl_grad else log_probs.detach()
kl_old_log_probs = old_log_probs if with_kl_grad else old_log_probs.detach()
kl_ref_log_probs = ref_log_probs if with_kl_grad else ref_log_probs.detach()
importance_ratio = None
if args.use_unbiased_kl:
importance_ratio = torch.exp(log_probs - old_log_probs)
importance_ratio = torch.exp(kl_log_probs - kl_old_log_probs)
kl = compute_approx_kl(
log_probs,
ref_log_probs,
kl_log_probs,
kl_ref_log_probs,
kl_loss_type=args.kl_loss_type,
importance_ratio=importance_ratio,
)
kl_loss = sum_of_sample_mean(kl)

loss = loss + args.kl_loss_coef * kl_loss
if with_kl_grad:
loss = loss + kl_loss_coef * kl_loss

# make sure the gradient could backprop correctly.
if log_probs.numel() == 0:
Expand Down
Loading