From 2a8d84e66d19014c44155ca1ee79b4aa0227734d Mon Sep 17 00:00:00 2001
From: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Date: Tue, 3 Feb 2026 13:49:49 +0000
Subject: [PATCH] Fix Gemma3n audio encoder for Transformers v5 (#33673)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
---
 vllm/model_executor/models/gemma3n_mm.py | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/vllm/model_executor/models/gemma3n_mm.py b/vllm/model_executor/models/gemma3n_mm.py
index 1460a4586..8b5e7b8bb 100644
--- a/vllm/model_executor/models/gemma3n_mm.py
+++ b/vllm/model_executor/models/gemma3n_mm.py
@@ -621,10 +621,15 @@ class Gemma3nForConditionalGeneration(
         # 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.