Skip to content

[fix] Use last_hidden_state key from get_image_features for llama4 #43882

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
vasqu merged 1 commit into from
Feb 10, 2026
Merged

[fix] Use last_hidden_state key from get_image_features for llama4 #43882

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
6 changes: 3 additions & 3 deletions src/transformers/models/llama4/modeling_llama4.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1080,7 +1080,7 @@ def forward(
output_hidden_states: bool | None = None,
return_dict: bool | None = None,
**kwargs,
) -> BaseModelOutput | tuple[torch.Tensor, ...]:
) -> BaseModelOutputWithPooling | tuple[torch.Tensor, ...]:
r"""

Example:
Expand Down Expand Up @@ -1166,7 +1166,7 @@ def forward(
if not return_dict:
return tuple(v for v in [hidden_state, hidden_states, attentions] if v is not None)

return BaseModelOutput(
return BaseModelOutputWithPooling(
last_hidden_state=hidden_state,
hidden_states=hidden_states,
Comment on lines -1169 to 1171
Copy link
Member

@zucchini-nlp zucchini-nlp Feb 10, 2026

Choose a reason for hiding this comment

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

i think the last hidden state is the one after layernorm_post, and pooler state is after adapter. Though it'll be a breaking change...
Fine with leaving it as is, thanks

Copy link
Member Author

@tomaarsen tomaarsen Feb 10, 2026

Choose a reason for hiding this comment

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

You're right :/ Damn, I wish I spotted that before v5, as it's indeed breaking if we improve it. I did make roughly this change for a few other architectures pre-v5.

attentions=attentions,
Expand Down Expand Up @@ -1321,7 +1321,7 @@ def forward(
pixel_values=pixel_values,
vision_feature_select_strategy=vision_feature_select_strategy,
return_dict=True,
).pooler_output
).last_hidden_state

vision_flat = image_features.view(-1, image_features.size(-1))
projected_vision_flat = self.multi_modal_projector(vision_flat).to(
Expand Down