Skip to content

fix: align channel loss with sequence parallel labels #9977

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
hjh0119 merged 1 commit into from
Sep 7, 2026
+20 −5
Merged

fix: align channel loss with sequence parallel labels #9977

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
20 changes: 16 additions & 4 deletions swift/trainers/seq2seq_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -163,7 +163,15 @@ def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=N
if (self.args.enable_dft_loss or loss_scale is not None or self.args.enable_channel_loss
or self.template.sequence_parallel_size > 1):
if self.template.sequence_parallel_size > 1:
outputs.loss = per_token_loss_func_sp(outputs, labels, enable_dft_loss=self.args.enable_dft_loss)
sp_loss = per_token_loss_func_sp(
outputs,
labels,
enable_dft_loss=self.args.enable_dft_loss,
return_labels=self.args.enable_channel_loss)
if self.args.enable_channel_loss:
outputs.loss, channel_labels = sp_loss
else:
outputs.loss = sp_loss
if loss_scale is not None:
position_ids = sequence_parallel.real_position_ids
if position_ids is not None:
Expand All @@ -183,15 +191,19 @@ def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=N

if self.args.enable_channel_loss:
metrics = self.custom_metrics[mode]
masks = torch.roll(labels, shifts=-1, dims=-1).view(-1) != -100
if self.template.sequence_parallel_size == 1:
channel_labels = torch.roll(labels, shifts=-1, dims=-1)
channel_loss = outputs.loss.view(-1)
masks = channel_labels.view(-1) != -100
if self.template.padding_free:
cu_seqlens = self.get_cu_seqlens(text_position_ids, inputs.get('logits_to_keep'))
else:
cu_seqlens = torch.arange(0, labels.shape[0] + 1) * labels.shape[1]
seq_len = channel_labels.numel() // labels.shape[0]
cu_seqlens = torch.arange(0, labels.shape[0] + 1) * seq_len
for i in range(cu_seqlens.shape[0] - 1):
channel = None if channels is None else channels[i]
slice_ = slice(cu_seqlens[i], cu_seqlens[i + 1])
metrics[f'loss_{channel}'].update(outputs.loss[slice_][masks[slice_]])
metrics[f'loss_{channel}'].update(channel_loss[slice_][masks[slice_]])

unwrapped_model = self.accelerator.unwrap_model(model)
if is_peft_available() and isinstance(unwrapped_model, PeftModel):
Expand Down
5 changes: 4 additions & 1 deletion swift/trainers/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -164,7 +164,7 @@ def is_instance_of_ms_model(model: Module) -> bool:
return False


def per_token_loss_func_sp(outputs, labels, enable_dft_loss=False, **kwargs) -> torch.Tensor:
def per_token_loss_func_sp(outputs, labels, enable_dft_loss=False, return_labels=False, **kwargs):
"""Common loss function for sequence parallel training"""
if hasattr(outputs, 'logits'):
logits = outputs.logits
Expand Down Expand Up @@ -192,7 +192,10 @@ def per_token_loss_func_sp(outputs, labels, enable_dft_loss=False, **kwargs) ->
if position_ids is not None and position_ids.min() == -1:
_pos_mask = position_ids >= 0
loss = loss[_pos_mask].contiguous()
labels = labels[_pos_mask].contiguous()

if return_labels:
return loss, labels
return loss


Expand Down
Loading