[V1][Molmo] Fix get_multimodal_embeddings() in molmo.py (#14161)

vllm/model_executor/models/ultravox.py

@@ -476,7 +476,9 @@ class UltravoxModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA):
 
         return result
 
-    def get_multimodal_embeddings(self, **kwargs) -> Optional[NestedTensors]:
+    def get_multimodal_embeddings(
+        self, **kwargs
+    ) -> Union[list[torch.Tensor], torch.Tensor, tuple[torch.Tensor, ...]]:
         audio_input = self._parse_and_validate_audio_input(**kwargs)
         if audio_input is None:
             return None

vllm/model_executor/models/whisper.py

@@ -692,7 +692,9 @@ class WhisperForConditionalGeneration(nn.Module, SupportsTranscription,
         )
         return decoder_outputs
 
-    def get_multimodal_embeddings(self, **kwargs) -> Optional[NestedTensors]:
+    def get_multimodal_embeddings(
+        self, **kwargs
+    ) -> Union[list[torch.Tensor], torch.Tensor, tuple[torch.Tensor, ...]]:
         # TODO: This method does not obey the interface for SupportsMultiModal.
         # Refactor this once encoder/decoder support is implemented in V1.
         audio_input = self._parse_and_validate_audio_input(**kwargs)