Skip to content

fix grpo filter overlong #3844

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 5 commits into from
Apr 12, 2025
Merged

fix grpo filter overlong #3844

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
35aad7d
fix
hjh0119 Apr 11, 2025
40f2686
fix
hjh0119 Apr 11, 2025
4f13a87
Rever
hjh0119 Apr 12, 2025
0517d3e
fix
hjh0119 Apr 12, 2025
490874b
rm nanstd
hjh0119 Apr 12, 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
20 changes: 11 additions & 9 deletions swift/trainers/rlhf_trainer/grpo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1010,25 +1010,24 @@ def _log_metrics(self, inputs, messages, completions, rewards, rewards_per_func)
reward_func_name = reward_func.__class__.__name__

reward_func_names.append(reward_func_name)
metrics_mask = ~agg_truncated_mask if self.args.overlong_filter else torch.ones(
agg_completion_mask.shape[0], dtype=torch.bool)

for i, reward_func_name in enumerate(reward_func_names):
mean_rewards = (rewards_per_func[:, i][metrics_mask]).mean().item()
mean_rewards = rewards_per_func[:, i].mean().item()
self._metrics[mode][f'rewards/{reward_func_name}/mean'].append(mean_rewards)
std_rewards = (rewards_per_func[:, i][metrics_mask]).std().item()
std_rewards = rewards_per_func[:, i].std().item()
self._metrics[mode][f'rewards/{reward_func_name}/std'].append(std_rewards)

# Log overall reward stats
grouped_rewards = rewards[metrics_mask].view(-1, self.num_generations)
grouped_rewards = rewards.view(-1, self.num_generations)
self._metrics[mode]['reward'].append(grouped_rewards.mean().item())
self._metrics[mode]['reward_std'].append(grouped_rewards.std(dim=1).mean().item())

# Log prompt and completion texts
self._textual_logs['prompt'].extend(m for m, mask in zip(gather_object(messages), metrics_mask) if mask)
self._textual_logs['completion'].extend(c for c, mask in zip(gather_object(completions), metrics_mask) if mask)
self._textual_logs['prompt'].extend(gather_object(messages))
self._textual_logs['completion'].extend(gather_object(completions))

for i, name in enumerate(reward_func_names):
self._textual_logs['rewards'][name].extend(rewards_per_func[:, i][metrics_mask].tolist())
self._textual_logs['rewards'][name].extend(rewards_per_func[:, i].tolist())

@profiling_decorator
def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
Expand Down Expand Up @@ -1077,7 +1076,10 @@ def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=N
mean_kl = (per_token_kl * completion_mask).sum() / completions_length
metrics['kl'] = mean_kl

is_clipped = (coef_1 < (1 - self.epsilon_low)) | (coef_1 > (1 + self.epsilon_high))
is_clipped = ((coef_1 < 1 - self.epsilon_low) &
(advantages.unsqueeze(1) < 0)) | ((coef_1 > 1 + self.epsilon_high) &
(advantages.unsqueeze(1) > 0))

clip_ratio = (is_clipped * completion_mask).sum() / completions_length
metrics['clip_ratio'] = clip_ratio

Expand Down