Skip to content

Add average_log_prob as an init param for LigerFusedLinearDPOLoss #676

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
shivam15s merged 2 commits into from
Apr 22, 2025
Merged

Add average_log_prob as an init param for LigerFusedLinearDPOLoss #676

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8ae0ba9
Add average_log_prob as an init param for LigerFusedLinearDPOLoss
vaibhavjindal Apr 22, 2025
03aea56
Merge branch 'main' into dpo-api-fix
shivam15s Apr 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
9 changes: 8 additions & 1 deletion 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 @@ -68,6 +68,7 @@ def forward(
compute_nll_loss=False,
compiled=True,
use_ref_model=True,
average_log_prob=False,
chunk_size=1,
):
"""
Expand All @@ -85,6 +86,7 @@ def forward(
compute_nll_loss (bool): Whether to compute the NLL loss
compiled (bool): Whether to use torch compile
use_ref_model (bool): Whether to use a reference model
average_log_prob (bool): Whether to average the log probability per non-masked token
chunk_size (int): Size of chunks for processing.
Returns:
torch.Tensor: Computed loss
Expand All @@ -104,13 +106,14 @@ def forward(
ref_input=ref_input,
ref_weight=ref_weight,
ref_bias=ref_bias,
average_log_prob=average_log_prob,
chunk_size=chunk_size,
)

@staticmethod
def backward(ctx, *grad_output):
grads = LigerFusedLinearPreferenceBase.backward(ctx, grad_output)[:4]
return *grads, None, None, None, None, None, None, None, None, None
return *grads, None, None, None, None, None, None, None, None, None, None


class LigerFusedLinearDPOLoss(torch.nn.Module):
Expand All @@ -125,6 +128,7 @@ def __init__(
compute_nll_loss: bool = False,
compiled: bool = True,
use_ref_model: bool = True,
average_log_prob: bool = True,
chunk_size: int = 1,
):
"""
Expand All @@ -134,6 +138,7 @@ def __init__(
compute_nll_loss (bool): Whether to compute the NLL loss.
compiled (bool): Whether to use the torch compiled kernel.
use_ref_model (bool): Whether to use a reference model for the DPO loss.
average_log_prob (bool): Whether to average the log probability per non-masked token.
chunk_size (int): Size of chunks for processing.
"""
super().__init__()
Expand All @@ -142,6 +147,7 @@ def __init__(
self.compute_nll_loss = compute_nll_loss
self.compiled = compiled
self.use_ref_model = use_ref_model
self.average_log_prob = average_log_prob
self.chunk_size = chunk_size

def forward(
Expand All @@ -167,5 +173,6 @@ def forward(
self.compute_nll_loss,
self.compiled,
self.use_ref_model,
self.average_log_prob,
self.chunk_size,
)