[Fix] Introduce audio channels spec (#31595)

Signed-off-by: Jeremy Teboul <jeremyte@meta.com>

vllm/model_executor/models/qwen2_5_omni_thinker.py

@@ -226,6 +226,10 @@ class Qwen2_5OmniThinkerProcessingInfo(
         assert isinstance(feature_extractor, WhisperFeatureExtractor)
         return feature_extractor
 
+    def get_target_channels(self) -> int:
+        """Return target audio channels for Qwen2.5 Omni models (mono)."""
+        return 1
+
     def get_supported_mm_limits(self) -> Mapping[str, int | None]:
         return {"audio": None, "image": None, "video": None}
 
@@ -310,6 +314,7 @@ class Qwen2_5OmniThinkerMultiModalProcessor(
         return Qwen2_5OmniThinkerMultiModalDataParser(
             spatial_merge_size=self.info.get_hf_config().vision_config.spatial_merge_size,
             target_sr=feature_extractor.sampling_rate,
+            target_channels=self.info.get_target_channels(),
         )
 
     def _call_hf_processor(

vllm/model_executor/models/qwen2_audio.py

@@ -140,6 +140,10 @@ class Qwen2AudioProcessingInfo(BaseProcessingInfo):
         assert isinstance(feature_extractor, WhisperFeatureExtractor)
         return feature_extractor
 
+    def get_target_channels(self) -> int:
+        """Return target audio channels for Qwen2 Audio models (mono)."""
+        return 1
+
     def get_supported_mm_limits(self) -> Mapping[str, int | None]:
         return {"audio": None}
 
@@ -201,7 +205,10 @@ class Qwen2AudioMultiModalDataParser(MultiModalDataParser):
 class Qwen2AudioMultiModalProcessor(BaseMultiModalProcessor[Qwen2AudioProcessingInfo]):
     def _get_data_parser(self) -> MultiModalDataParser:
         feature_extractor = self.info.get_feature_extractor()
-        return Qwen2AudioMultiModalDataParser(target_sr=feature_extractor.sampling_rate)
+        return Qwen2AudioMultiModalDataParser(
+            target_sr=feature_extractor.sampling_rate,
+            target_channels=self.info.get_target_channels(),
+        )
 
     def _call_hf_processor(
         self,

vllm/model_executor/models/ultravox.py

@@ -133,6 +133,10 @@ class UltravoxProcessingInfo(BaseProcessingInfo):
         assert isinstance(feature_extractor, WhisperFeatureExtractor)
         return feature_extractor
 
+    def get_target_channels(self) -> int:
+        """Return target audio channels for Ultravox models (mono)."""
+        return 1
+
     def get_supported_mm_limits(self) -> Mapping[str, int | None]:
         return {"audio": None}
 
@@ -169,7 +173,10 @@ class UltravoxDummyInputsBuilder(BaseDummyInputsBuilder[UltravoxProcessingInfo])
 class UltravoxMultiModalProcessor(BaseMultiModalProcessor[UltravoxProcessingInfo]):
     def _get_data_parser(self) -> MultiModalDataParser:
         feature_extractor = self.info.get_feature_extractor()
-        return MultiModalDataParser(target_sr=feature_extractor.sampling_rate)
+        return MultiModalDataParser(
+            target_sr=feature_extractor.sampling_rate,
+            target_channels=self.info.get_target_channels(),
+        )
 
     def _call_hf_processor(
         self,

vllm/model_executor/models/whisper.py

@@ -690,6 +690,10 @@ class WhisperProcessingInfo(BaseProcessingInfo):
         assert isinstance(feature_extractor, WhisperFeatureExtractor)
         return feature_extractor
 
+    def get_target_channels(self) -> int:
+        """Return target audio channels for Whisper models (mono)."""
+        return 1
+
     def get_num_audio_tokens(self) -> int:
         return self.get_hf_config().max_source_positions
 
@@ -724,7 +728,10 @@ class WhisperDummyInputsBuilder(BaseDummyInputsBuilder[WhisperProcessingInfo]):
 class WhisperMultiModalProcessor(EncDecMultiModalProcessor[WhisperProcessingInfo]):
     def _get_data_parser(self) -> MultiModalDataParser:
         feature_extractor = self.info.get_feature_extractor()
-        return MultiModalDataParser(target_sr=feature_extractor.sampling_rate)
+        return MultiModalDataParser(
+            target_sr=feature_extractor.sampling_rate,
+            target_channels=self.info.get_target_channels(),
+        )
 
     @property
     def pad_dummy_encoder_prompt(self) -> bool: