Update gemma4_audio_encoder.py

gemma4_audio_encoder.py

@@ -16,7 +16,7 @@ class Gemma4AudioEncoder(PreTrainedModel):
         """
         Args:
             input_features: Audio mel-spectrogram features.
-            input_features_mask: Attention mask for audio features.
+            input_features_mask: Attention mask for audio features (True = valid, False = padding).
             project: If True, project to LLM embedding space (2560-dim).
                      If False, return audio tower output (1536-dim).
 
@@ -24,7 +24,7 @@ class Gemma4AudioEncoder(PreTrainedModel):
             If project=True: (projected_features, attention_mask)
             If project=False: (encoder_features, attention_mask)
         """
-        output = self.audio_tower(input_features, ~input_features_mask)
+        output = self.audio_tower(input_features, input_features_mask)
         if project:
             projected = self.embed_audio(inputs_embeds=output.last_hidden_state)
             return projected, output.attention_mask