Skip to content

[fix]: correct condition for restoring layer embeddings in TransformerDecorator/AdaptiveLayerLoss #3560

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

Open
wants to merge 2 commits into
base: master
Choose a base branch
from
Open

[fix]: correct condition for restoring layer embeddings in TransformerDecorator/AdaptiveLayerLoss #3560

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e01b31a
fix: correct condition for restoring layer embeddings in TransformerD…
emapco Nov 4, 2025
7bb9dc9
fix: include all layer embeddings in call_use_cache if requested
emapco Nov 4, 2025
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
7 changes: 5 additions & 2 deletions sentence_transformers/losses/AdaptiveLayerLoss.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,13 +71,16 @@ def call_grow_cache(self, features: dict[str, Tensor]) -> dict[str, Tensor]:
# Restore original setting
self.transformer.auto_model.config.output_hidden_states = original_output_hidden_states

if original_output_hidden_states:
if not original_output_hidden_states:
del output["all_layer_embeddings"]

return output

def call_use_cache(self, features: dict[str, Tensor]) -> dict[str, Tensor]:
return {**self.features[self.call_idx], "token_embeddings": self.embeddings[self.call_idx][self.layer_idx]}
output = {**self.features[self.call_idx], "token_embeddings": self.embeddings[self.call_idx][self.layer_idx]}
if self.transformer.auto_model.config.output_hidden_states and "all_layer_embeddings" in features:
output["all_layer_embeddings"] = features["all_layer_embeddings"]
Comment on lines +81 to +82
Copy link

Copilot AI Nov 4, 2025

Choose a reason for hiding this comment

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

The call_use_cache method returns cached data without calling original_forward, so the input features parameter will not contain all_layer_embeddings. This check will always fail because features is the raw input to the forward method, not the output. If all_layer_embeddings needs to be included when output_hidden_states is enabled, it should be reconstructed from the cached self.embeddings data instead of reading from features.

Suggested change
if self.transformer.auto_model.config.output_hidden_states and "all_layer_embeddings" in features:
output["all_layer_embeddings"] = features["all_layer_embeddings"]
if self.transformer.auto_model.config.output_hidden_states:
# Reconstruct all_layer_embeddings from cached self.embeddings
output["all_layer_embeddings"] = self.embeddings[self.call_idx]

Copilot uses AI. Check for mistakes.
Copy link
Contributor Author

@emapco emapco Nov 4, 2025

Choose a reason for hiding this comment

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

self.embeddings doesn't contain all of the hidden states/layer embeddings. Perhaps call_grow_cache() should also storeself.all_layer_embeddings

return output


class ForwardDecorator:
Expand Down