Skip to content

Fail early for unsupported PEFT + Liger Kernel in DPO #5709

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
albertvillanova merged 4 commits into from
May 6, 2026
Merged

Fail early for unsupported PEFT + Liger Kernel in DPO #5709

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
2706112
Make DPO fail fast for peft + liger-kernel
albertvillanova May 6, 2026
62a66af
Test DPO fail fast for peft + liger-kernel
albertvillanova May 6, 2026
70540c1
Remove error raising from _compute_loss_liger
albertvillanova May 6, 2026
623f168
Merge remote-tracking branch 'upstream/main' into dpo-fail-fast-peft-…
albertvillanova May 6, 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
19 changes: 19 additions & 0 deletions tests/test_dpo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -662,6 +662,25 @@ def test_train_with_liger(self):
new_param = trainer.model.get_parameter(n)
assert not torch.equal(param, new_param), f"Parameter {n} has not changed."

@require_liger_kernel
@require_peft
def test_init_fails_with_peft_and_liger(self):
dataset = load_dataset("trl-internal-testing/zen", "standard_preference", split="train")

training_args = DPOConfig(
output_dir=self.tmp_dir,
use_liger_kernel=True,
report_to="none",
)

with pytest.raises(NotImplementedError, match="Liger DPO loss is not implemented for PEFT models."):
DPOTrainer(
model="trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
args=training_args,
train_dataset=dataset,
peft_config=LoraConfig(),
)

def test_train_with_iterable_dataset(self):
dataset = load_dataset("trl-internal-testing/zen", "standard_preference", split="train", streaming=True)

Expand Down
19 changes: 9 additions & 10 deletions trl/trainer/dpo_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -709,6 +709,8 @@ def __init__(
"Liger DPO loss does not support precomputing reference log probabilities. Either disable "
"`precompute_ref_log_probs` or set `use_liger_kernel` to False."
)
if is_peft_model(model):
raise NotImplementedError("Liger DPO loss is not implemented for PEFT models.")
Comment on lines +712 to +713

@qgallouedec qgallouedec May 6, 2026 •
edited
Loading

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

maybe we could remove this then?

trl/trl/trainer/dpo_trainer.py

Lines 1101 to 1102 in 62a66af

if is_peft_model(model):
raise NotImplementedError("Liger DPO loss is not implemented for PEFT models.")

@albertvillanova albertvillanova May 6, 2026

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I thought the same! But codex told me to keep it... I should have followed my intuition.


# Dataset
# Skip dataset preparation if it's a VLM, where preprocessing (e.g., image-to-pixel conversion) is too costly
Expand Down Expand Up @@ -1096,16 +1098,13 @@ def _compute_loss_liger(self, model, inputs, return_outputs):
weight = lm_head.weight
bias = lm_head.bias

if is_peft_model(model):
raise NotImplementedError("Liger DPO loss is not implemented for PEFT models.")
else:
with torch.no_grad(), disable_gradient_checkpointing(self.model, self.args.gradient_checkpointing_kwargs):
ref_decoder = self.ref_model.get_decoder()
ref_outputs = ref_decoder(input_ids, attention_mask=attention_mask, use_cache=False)
ref_lm_head = self.ref_model.get_output_embeddings()
ref_hidden_states = ref_outputs.last_hidden_state[:, :-1].contiguous()
ref_weight = ref_lm_head.weight
ref_bias = ref_lm_head.bias
with torch.no_grad(), disable_gradient_checkpointing(self.model, self.args.gradient_checkpointing_kwargs):
ref_decoder = self.ref_model.get_decoder()
ref_outputs = ref_decoder(input_ids, attention_mask=attention_mask, use_cache=False)
ref_lm_head = self.ref_model.get_output_embeddings()
ref_hidden_states = ref_outputs.last_hidden_state[:, :-1].contiguous()
ref_weight = ref_lm_head.weight
ref_bias = ref_lm_head.bias

shift_completion_mask = completion_mask[:, 1:].contiguous()
labels = input_ids[:, 1:].clone()
Expand Down
Loading