Skip to content

Refactor loss handling in gemma.py #156

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
wants to merge 1 commit into from
Closed

Refactor loss handling in gemma.py #156

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
17 changes: 3 additions & 14 deletions unsloth_zoo/temporary_patches/gemma.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -262,10 +262,6 @@ def forward(
logits_to_keep=logits_to_keep,
**lm_kwargs,
)
labels = None
# We NEVER ENTER if labels is not None: since we already accounted for it


logits = outputs.logits
loss = None
if labels is not None:
Expand All @@ -288,7 +284,7 @@ def forward(
flat_logits = shift_logits.view(-1, self.config.text_config.vocab_size)
flat_labels = shift_labels.view(-1).to(shift_logits.device)
loss = loss_fct(flat_logits, flat_labels)
loss = outputs.loss
loss = getattr(outputs, "loss", loss)

return Gemma3CausalLMOutputWithPast(
loss=loss,
Expand Down Expand Up @@ -352,14 +348,7 @@ def forward(
cache_position=cache_position,
**lm_kwargs,
)
labels = None
# We NEVER ENTER if labels is not None: since we already accounted for it

hidden_states = outputs[0]
# Only compute necessary logits, and do not upcast them to float if we are not computing the loss
slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep
logits = self.lm_head(hidden_states[:, slice_indices, :])

logits = outputs.logits
loss = None
if labels is not None:
# Upcast to float if we need to compute the loss to avoid potential precision issues
Expand All @@ -381,7 +370,7 @@ def forward(
flat_logits = shift_logits.view(-1, self.config.text_config.vocab_size)
flat_labels = shift_labels.view(-1).to(shift_logits.device)
loss = loss_fct(flat_logits, flat_labels)
loss = outputs.loss
loss = getattr(outputs, "loss", loss)

if not return_dict:
output = (logits,) + outputs[1:]
Expand Down