Skip to content

[FEAT][kernels]: implement fused GRPO loss with in-place group reward normalization #93

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
Flink-ddd merged 6 commits into from
Jun 10, 2026
+1,293 −0
Merged

[FEAT][kernels]: implement fused GRPO loss with in-place group reward normalization #93

Changes from 1 commit
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
13195f5
Add fused GRPO loss op (PyTorch reference + Triton kernel)
KJLdefeated Jun 9, 2026
25e19bf
Add GRPO loss tests and native-vs-Triton benchmark
KJLdefeated Jun 9, 2026
f434e31
Document the GRPO loss operator
KJLdefeated Jun 9, 2026
e851c48
(doc) add grpo-loss kernel doc
KJLdefeated Jun 9, 2026
a4ca453
(bug) sanitize masked position
KJLdefeated Jun 9, 2026
f1bc93a
(bug) sqrt(0) numerical issue, max group size try catch, block sum ou…
KJLdefeated Jun 10, 2026
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
Prev Previous commit
Next Next commit
(bug) sanitize masked position
  • Loading branch information
@KJLdefeated
KJLdefeated committed Jun 9, 2026
commit a4ca4535cf7868dafe3e8a5f6de50642618c47ce
5 changes: 3 additions & 2 deletions rl_engine/kernels/ops/pytorch/loss/grpo_loss.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -98,14 +98,15 @@ def apply(
"""Compute ``(loss, policy_loss, kl)`` from per-token log-probabilities."""
bool_mask = completion_mask.bool()

ratio = torch.exp(current_logps.float() - old_logps.float())
delta = (current_logps.float() - old_logps.float()).masked_fill(~bool_mask, 0.0)
diff = (ref_logps.float() - current_logps.float()).masked_fill(~bool_mask, 0.0)
ratio = torch.exp(delta)
adv = advantages.float()
unclipped = ratio * adv
clipped = torch.clamp(ratio, 1.0 - clip_eps, 1.0 + clip_eps) * adv
policy_loss_terms = -torch.minimum(unclipped, clipped)

# k3 reference-KL estimator: exp(d) - d - 1, with d = ref - current.
diff = ref_logps.float() - current_logps.float()
kl_terms = torch.exp(diff) - diff - 1.0

policy_loss = self._masked_mean(policy_loss_terms, bool_mask)
Expand Down
Loading