Skip to content

Fix Gemma3n audio encoder for Transformers v5 #33673

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
vllm-bot merged 2 commits into from
Feb 3, 2026
+9 −4
Merged

Fix Gemma3n audio encoder for Transformers v5 #33673

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e368fc3
Fix Gemma3n for Transformers v5
hmellor Feb 3, 2026
1d4800f
Turns out kwargs was not a no-op
hmellor Feb 3, 2026
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
13 changes: 9 additions & 4 deletions vllm/model_executor/models/gemma3n_mm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -621,10 +621,15 @@ def _process_audio_input(
# Run on padded features to enable batching
input_features = audio_input["input_features_padded"].squeeze(1)
input_features_mask = audio_input["input_features_mask"].squeeze(1)
audio_outputs, audio_mask = self.audio_tower(
input_features, ~input_features_mask
)
audio_features = self.embed_audio(inputs_embeds=audio_outputs)
audio_outputs = self.audio_tower(input_features, ~input_features_mask)
if isinstance(audio_outputs, tuple):
# Transformers v4
audio_encodings, audio_mask = audio_outputs
else:
# Transformers v5
audio_encodings = audio_outputs.last_hidden_state
audio_mask = audio_outputs.audio_mel_mask
audio_features = self.embed_audio(inputs_embeds=audio_encodings)

# The Gemma3nProcessor expects all audio will be 30s in length and
# inserts 188 audio soft tokens into the text to account for this.
Expand Down