Skip to content

fix LigerORPOTrainer bug in none fsdp setup #1230

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

Draft
Ola-Vish wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

fix LigerORPOTrainer bug in none fsdp setup #1230

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
26 changes: 19 additions & 7 deletions src/liger_kernel/transformers/trainer/orpo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -74,16 +74,28 @@ def orpo_partial(lm_head, last_hidden_state, concatenated_labels, nll_target):
lm_head.weight, last_hidden_state, concatenated_labels, lm_head.bias, nll_target=nll_target
)

orpo_loss, aux_outputs = _FSDPForwardRedirection()(
model,
orpo_partial,
model.lm_head,
outputs.last_hidden_state[:, :-1] if not self.is_encoder_decoder else outputs.last_hidden_state,
last_hidden_state = (
outputs.last_hidden_state[:, :-1] if not self.is_encoder_decoder else outputs.last_hidden_state
)
concatenated_labels = (
concatenated_batch["concatenated_labels"][:, 1:]
if not self.is_encoder_decoder
else concatenated_batch["concatenated_labels"],
labels[:, 1:] if not self.is_encoder_decoder else labels,
else concatenated_batch["concatenated_labels"]
)
nll_target = labels[:, 1:] if not self.is_encoder_decoder else labels

if isinstance(model, FullyShardedDataParallel):
orpo_loss, aux_outputs = _FSDPForwardRedirection()(
model,
orpo_partial,
model.lm_head,
last_hidden_state,
concatenated_labels,
nll_target,
)
else:
orpo_loss, aux_outputs = orpo_partial(model.lm_head, last_hidden_state, concatenated_labels, nll_target)

# if aux_loss_enabled, add the aux_loss to the orpo_loss
if self.aux_loss_enabled:
orpo_loss += self.aux_loss_coef * outputs.aux_loss
Expand Down