[Bugfix] Fix broken internvl2 inference with v1 (vllm-project#12360)

Signed-off-by: Isotr0py <2037008807@qq.com>

Author: Isotr0py

vllm/multimodal/utils.py

@@ -503,8 +503,13 @@ def modality_group_func(mm_input: "MultiModalKwargs") -> Union[str, int]:
         if len(mm_input.modalities) > 1:
             return id(mm_input)
 
-        # Otherwise return the modality string
-        return list(mm_input.modalities)[0]
+        elif len(mm_input.modalities) == 1:
+            return list(mm_input.modalities)[0]
+
+        # FIXME(Isotr0py): Modality of mm_input from legacy pipeline is empty,
+        # this is used to make InternVL with legacy pipeline still work with v1.
+        else:
+            return ""
 
     return [
         list(group) for _, group in groupby(mm_inputs, key=modality_group_func)