Skip to content
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

[DPO] add reference log-prob outputs in DPO #521

Merged
merged 6 commits into from
Jan 30, 2025
Merged

[DPO] add reference log-prob outputs in DPO #521

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
5a26e54
add reference log-prob outputs in DPO
kashif Jan 14, 2025
142fa87
return chosen_rewards, rejected_rewards
kashif Jan 15, 2025
bc3d4f6
Merge branch 'main' into dpo-fix
kashif Jan 15, 2025
5fce8e4
Merge branch 'main' into dpo-fix
lancerts Jan 21, 2025
1582cda
Merge branch 'main' into dpo-fix
kashif Jan 22, 2025
a18766c
remove unused outputs back to default
kashif Jan 22, 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
7 changes: 5 additions & 2 deletions src/liger_kernel/chunked_loss/dpo_loss.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,9 +45,12 @@ def preference_loss_fn(
chosen_logratios = chosen_logps - ref_chosen_logps
rejected_logratios = rejected_logps - ref_rejected_logps

chosen_rewards = beta * (chosen_logps - ref_chosen_logps)
rejected_rewards = beta * (rejected_logps - ref_rejected_logps)

logits_diff = beta * (chosen_logratios - rejected_logratios)
loss = -F.logsigmoid(logits_diff).sum() / (full_target.shape[0] // 2)
return loss
return loss, chosen_rewards, rejected_rewards

@staticmethod
def forward(
Expand Down Expand Up @@ -99,7 +102,7 @@ def __init__(
beta: float = 0.1,
compute_nll_loss: bool = False,
compiled: bool = True,
use_ref_model: bool = False,
use_ref_model: bool = True,
):
"""
Args:
Expand Down
5 changes: 4 additions & 1 deletion test/chunked_loss/test_dpo_loss.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,9 +56,12 @@ def alignment_loss(
chosen_logratios = policy_chosen_logps - ref_chosen_logps
rejected_logratios = policy_rejected_logps - ref_rejected_logps

chosen_rewards = self.beta * (policy_chosen_logps - ref_chosen_logps)
rejected_rewards = self.beta * (policy_rejected_logps - ref_rejected_logps)

logits_diff = self.beta * (chosen_logratios - rejected_logratios)
losses = -F.logsigmoid(logits_diff)
return losses
return losses, chosen_rewards, rejected_rewards


class TorchLMHeadDPO(torch.nn.Module):
Expand Down
Loading