Skip to content

Fix DPO IPO log-prob normalization #9925

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
taking-lying-flat wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix DPO IPO log-prob normalization #9925

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fb6e518
Fix DPO IPO log-prob normalization
taking-lying-flat Aug 16, 2026
490e336
Remove DPO loss test file
taking-lying-flat Aug 23, 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
45 changes: 36 additions & 9 deletions swift/rlhf_trainers/dpo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -130,23 +130,20 @@ def concatenated_forward(
all_logits, labels, label_pad_token_id=self.label_pad_token_id)
origin_per_token_logps = per_token_logps

loss_types = self.loss_type if isinstance(self.loss_type, list) else [self.loss_type]
if 'ipo' in loss_types:
size_completion = loss_mask.sum(dim=-1)
per_token_logps = per_token_logps / size_completion

output = {}
if self.template.padding_free:
cu_seqlens = self.get_cu_seqlens(text_position_ids, batch.get('logits_to_keep'))
num_examples = (cu_seqlens.shape[0] - 1) // 2
all_logps = per_token_logps.new_zeros((num_examples * 2, ))
completion_token_counts = loss_mask.new_zeros((num_examples * 2, ), dtype=torch.long)
completion_lengths = (cu_seqlens[1:] - cu_seqlens[:-1])
chosen_lengths = completion_lengths[:num_examples]
rejected_lengths = completion_lengths[num_examples:]
public_lengths = torch.min(chosen_lengths, rejected_lengths) # l_p in the paper

for i in range(cu_seqlens.shape[0] - 1):
start, end = cu_seqlens[i], cu_seqlens[i + 1]
completion_token_counts[i] = loss_mask[:, start:end].sum()
length = end - start
public_length = public_lengths[i % num_examples]
if self.args.ld_alpha is not None and not is_ref_model and length > public_length:
Expand All @@ -160,10 +157,13 @@ def concatenated_forward(
output['nll_loss'] = -origin_per_token_logps[:, :num_tokens][loss_mask[:, :num_tokens]].mean()
output['chosen_logps'] = all_logps[:num_examples]
output['rejected_logps'] = all_logps[num_examples:]
output['chosen_completion_token_counts'] = completion_token_counts[:num_examples]
output['rejected_completion_token_counts'] = completion_token_counts[num_examples:]
output['mean_chosen_logits'] = mean_all_logits[:, :num_tokens][loss_mask[:, :num_tokens]].mean()
output['mean_rejected_logits'] = mean_all_logits[:, num_tokens:][loss_mask[:, num_tokens:]].mean()
else:
num_examples = labels.shape[0] // 2
completion_token_counts = loss_mask.sum(dim=1)
if not is_ref_model:
output['nll_loss'] = -origin_per_token_logps[:num_examples][loss_mask[:num_examples]].mean()
if self.args.ld_alpha is not None and not is_ref_model:
Expand Down Expand Up @@ -191,6 +191,8 @@ def concatenated_forward(
all_logps = per_token_logps.sum(-1)
output['chosen_logps'] = all_logps[:num_examples]
output['rejected_logps'] = all_logps[num_examples:]
output['chosen_completion_token_counts'] = completion_token_counts[:num_examples]
output['rejected_completion_token_counts'] = completion_token_counts[num_examples:]
output['mean_chosen_logits'] = mean_all_logits[:num_examples][loss_mask[:num_examples]].mean()
output['mean_rejected_logits'] = mean_all_logits[num_examples:][loss_mask[num_examples:]].mean()
if self.aux_loss_enabled:
Expand All @@ -208,6 +210,10 @@ def compute_ref_log_probs(self, batch):
ref_model_output = self.concatenated_forward(self.ref_model, batch, is_ref_model=True)
return ref_model_output['chosen_logps'], ref_model_output['rejected_logps']

@staticmethod
def _get_ipo_sequence_logps(sequence_logps, completion_token_counts):
return sequence_logps / completion_token_counts.to(sequence_logps).clamp_min(1)

def dpo_loss(
self,
chosen_logps: torch.FloatTensor,
Expand Down Expand Up @@ -365,11 +371,23 @@ def get_batch_loss_metrics(
loss_types = self.loss_type if isinstance(self.loss_type, list) else [self.loss_type]
loss_weights = self.loss_weights if hasattr(self, 'loss_weights') and self.loss_weights else None
for idx, loss_type in enumerate(loss_types):
chosen_logps = model_output['chosen_logps']
rejected_logps = model_output['rejected_logps']
current_ref_chosen_logps = ref_chosen_logps
current_ref_rejected_logps = ref_rejected_logps
if loss_type == 'ipo':
chosen_token_counts = model_output['chosen_completion_token_counts']
rejected_token_counts = model_output['rejected_completion_token_counts']
chosen_logps = self._get_ipo_sequence_logps(chosen_logps, chosen_token_counts)
rejected_logps = self._get_ipo_sequence_logps(rejected_logps, rejected_token_counts)
current_ref_chosen_logps = self._get_ipo_sequence_logps(current_ref_chosen_logps, chosen_token_counts)
current_ref_rejected_logps = self._get_ipo_sequence_logps(current_ref_rejected_logps,
rejected_token_counts)
_losses, _chosen_rewards, _rejected_rewards = self.dpo_loss(
model_output['chosen_logps'],
model_output['rejected_logps'],
ref_chosen_logps,
ref_rejected_logps,
chosen_logps,
rejected_logps,
current_ref_chosen_logps,
current_ref_rejected_logps,
loss_type,
model_output,
)
Expand Down Expand Up @@ -399,6 +417,15 @@ def get_batch_loss_metrics(
self.accelerator.gather_for_metrics(model_output['chosen_logps']).detach().mean().item())
metrics[f'{prefix}logps/rejected'] = (
self.accelerator.gather_for_metrics(model_output['rejected_logps']).detach().mean().item())
if 'ipo' in loss_types:
chosen_mean_logps = self._get_ipo_sequence_logps(model_output['chosen_logps'],
model_output['chosen_completion_token_counts'])
rejected_mean_logps = self._get_ipo_sequence_logps(model_output['rejected_logps'],
model_output['rejected_completion_token_counts'])
metrics[f'{prefix}logps_mean/chosen'] = (
self.accelerator.gather_for_metrics(chosen_mean_logps).detach().mean().item())
metrics[f'{prefix}logps_mean/rejected'] = (
self.accelerator.gather_for_metrics(rejected_mean_logps).detach().mean().item())
metrics[f'{prefix}logits/chosen'] = (
self.accelerator.gather_for_metrics(model_output['mean_chosen_logits']).detach().mean().item())
metrics[f'{prefix}logits/rejected'] = (
Expand Down
Loading