Skip to content

handle updated gradient accumulation fixes from transformers #322

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

Closed
winglian wants to merge 1 commit into from
Closed

handle updated gradient accumulation fixes from transformers #322

Show file tree
Hide file tree
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
4 changes: 2 additions & 2 deletions src/liger_kernel/transformers/fused_linear_cross_entropy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,13 +9,13 @@ class LigerFusedLinearCrossEntropyLoss(CrossEntropyLoss):
def __init__(self, *args, **kwargs):
super(LigerFusedLinearCrossEntropyLoss, self).__init__(*args, **kwargs)

def forward(self, lin_weight, _input, target, bias=None):
def forward(self, lin_weight, _input, target, bias=None, reduction=None):
return LigerFusedLinearCrossEntropyFunction.apply(
_input,
lin_weight,
target,
bias,
self.ignore_index,
self.label_smoothing,
self.reduction,
reduction or self.reduction,
)
9 changes: 8 additions & 1 deletion src/liger_kernel/transformers/model/llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,8 @@ def lce_forward(
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
cache_position: Optional[torch.LongTensor] = None,
num_logits_to_keep: int = 0,
**loss_kwargs,
) -> Union[Tuple, CausalLMOutputWithPast]:
r"""
Copy paste llama forward but replace torch cross entropy with liger fused linear cross entropy
Expand Down Expand Up @@ -106,7 +108,12 @@ def lce_forward(
shift_labels = shift_labels.view(-1)

lce = LigerFusedLinearCrossEntropyLoss()
loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels)
lce_kwargs = {}
if "num_items_in_batch" in loss_kwargs:
lce_kwargs["reduction"] = "sum"
loss = lce(self.lm_head.weight, shift_hidden_states, shift_labels, **lce_kwargs)
if "num_items_in_batch" in loss_kwargs:
loss = loss / loss_kwargs["num_items_in_batch"]

else:
if self.config.pretraining_tp > 1:
Expand Down