[Bugfix] Fix Qwen2.5-Omni and Qwen3-Omni mixed-modality embed regression (#35368)

Signed-off-by: linyueqian <linyueqian@outlook.com>

vllm/model_executor/models/qwen3_omni_moe_thinker.py

@@ -1904,15 +1904,17 @@ class Qwen3OmniMoeThinkerForConditionalGeneration(
                 num_audio,
             )
 
-        # Default: standard merge (no interleaving)
-        inputs_embeds = _merge_multimodal_embeddings(
-            inputs_embeds=inputs_embeds,
+        # Default: standard merge (no interleaving), same as parent class.
+        # multimodal_embeddings may have been updated above (deepstack
+        # main-scale). Use super() to stay consistent with the parent
+        # implementation and avoid issues seen in Qwen2.5-Omni (#34506).
+        return super().embed_input_ids(
+            input_ids,
             multimodal_embeddings=multimodal_embeddings,
             is_multimodal=is_multimodal,
+            handle_oov_mm_token=handle_oov_mm_token,
         )
 
-        return inputs_embeds
-
     def forward(
         self,
         input_ids: torch.Tensor | None,