Skip to content

[trainer, worker] fix: setting old log probs equal to log probs for on policy training #3119

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
merged 3 commits into from
Aug 19, 2025
Merged

[trainer, worker] fix: setting old log probs equal to log probs for on policy training #3119

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f0babc4
Update dp_actor.py
sahilpatelsp Aug 18, 2025
b2ef7fe
fix with the calling of old log probs
sahilpatelsp Aug 18, 2025
74d810d
fix lint
zhaochen20 Aug 18, 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
8 changes: 7 additions & 1 deletion verl/workers/actor/dp_actor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -386,6 +386,8 @@ def update_policy(self, data: DataProto):
# See PPO paper for details. https://arxiv.org/abs/1707.06347
mini_batches = data.split(self.config.ppo_mini_batch_size)

on_policy = len(mini_batches) == 1 and self.config.ppo_epochs == 1

metrics = {}
for _ in range(self.config.ppo_epochs):
for batch_idx, mini_batch in enumerate(mini_batches):
Expand All @@ -405,7 +407,6 @@ def update_policy(self, data: DataProto):
micro_batch_metrics = {}
model_inputs = {**micro_batch.batch, **micro_batch.non_tensor_batch}
response_mask = model_inputs["response_mask"]
old_log_prob = model_inputs["old_log_probs"]
advantages = model_inputs["advantages"]

entropy_coeff = self.config.entropy_coeff
Expand All @@ -424,6 +425,11 @@ def update_policy(self, data: DataProto):
model_inputs, temperature=temperature, calculate_entropy=calculate_entropy
)

if on_policy:
old_log_prob = log_prob.detach()
else:
old_log_prob = model_inputs["old_log_probs"]

loss_mode = self.config.policy_loss.get("loss_mode", "vanilla")
# vanilla -> verl.trainer.ppo.core_algos.compute_policy_loss_vanilla
# gpg -> verl.trainer.ppo.core_algos.compute_policy_loss_gpg
Expand Down
Loading