Skip to content

[hotfix] fix entropy calculate for pp; using old logp to calculate loss #6364

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 1 commit into from
Jul 22, 2025
Merged

[hotfix] fix entropy calculate for pp; using old logp to calculate loss #6364

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
32 changes: 13 additions & 19 deletions applications/ColossalChat/coati/distributed/grpo_consumer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -250,6 +250,9 @@ def step(self, step_idx: int, pbar: Any, **kwargs) -> Optional[float]:
input_ids_forward_micro_batch = data["input_ids"][
forward_micro_batch_start : forward_micro_batch_start + train_microbatch_size
]
old_action_log_probs_micro_batch = old_action_log_probs[
forward_micro_batch_start : forward_micro_batch_start + train_microbatch_size
]
attention_mask_forward_micro_batch = data["attention_mask"][
forward_micro_batch_start : forward_micro_batch_start + train_microbatch_size
]
Expand Down Expand Up @@ -306,17 +309,22 @@ def step(self, step_idx: int, pbar: Any, **kwargs) -> Optional[float]:
"action_mask": action_mask_forward_micro_batch,
"advantages": advantages_forward_micro_batch,
"loss_mask": loss_mask_forward_micro_batch,
"old_action_log_probs": old_action_log_probs_micro_batch,
"source": self.rank,
}
if reference_action_log_probs is not None:
data_policy_forward["reference_action_log_probs"] = reference_action_log_probs

kl = []
policy_model_logits = torch.empty_like(input_ids_forward_micro_batch, device=self.device)

def _criterion(outputs, inputs):
action_logits = outputs.logits
policy_model_logits.copy_(action_logits)
mini_batch_entropies.append(
(
((entropy_from_logits(action_logits[:, -num_action:]) * inputs["action_mask"]).sum(-1))
/ inputs["action_mask"].sum(-1)
).detach()
)
action_log_probs = memory_efficient_logprob(
action_logits / self.generate_config["temperature"],
inputs["input_ids"],
Expand All @@ -339,7 +347,7 @@ def _criterion(outputs, inputs):

loss, _ = self.policy_loss_fn(
action_log_probs,
action_log_probs,
inputs["old_action_log_probs"],
inputs["advantages"].repeat_interleave(action_log_probs.size(-1), dim=-1),
per_token_kl,
inputs["action_mask"],
Expand All @@ -363,20 +371,6 @@ def _criterion(outputs, inputs):
kl = all_reduce_mean(torch.mean(torch.stack(kl)).to(loss.device), self.plugin).data
mean_kl.append(kl)
mean_loss.append(all_reduce_mean(loss, self.plugin).data)
mini_batch_entropies.append(
all_reduce_mean(
(
(
(
entropy_from_logits(policy_model_logits[:, -num_action:])
* action_mask_forward_micro_batch
).sum(-1)
)
/ action_mask_forward_micro_batch.sum(-1)
).detach(),
self.plugin,
)
)
else:
policy_model_logits = self.policy_model(
input_ids=input_ids_forward_micro_batch,
Expand Down Expand Up @@ -415,7 +409,7 @@ def _criterion(outputs, inputs):

loss, _ = self.policy_loss_fn(
action_log_probs,
old_action_log_probs,
old_action_log_probs_micro_batch,
advantages_forward_micro_batch.repeat_interleave(action_log_probs.size(-1), dim=-1),
per_token_kl,
action_mask_forward_micro_batch,
Expand Down Expand Up @@ -455,7 +449,7 @@ def _criterion(outputs, inputs):
ans_acc = all_reduce_mean(ans_acc.mean(), self.plugin)
advantages = all_reduce_mean(advantages.mean(), self.plugin)
response_length = all_reduce_mean(response_length.mean(), self.plugin)
entropy = torch.cat(mini_batch_entropies, dim=0).mean()
entropy = all_reduce_mean(torch.cat(mini_batch_entropies, dim=0).mean(), self.plugin)
self.accum_loss.add_(sum(mean_loss) / len(mean_loss))
self.accum_entropy.add_(entropy.data)
if self.policy_loss_fn.beta > 0:
Expand Down