[Bugfix] Reject /v1/audio/speech for Qwen omni models (#2763)

Signed-off-by: Bvicii <yizhanhuang2002@gmail.com>

tests/entrypoints/openai_api/test_serving_speech.py

@@ -684,6 +684,32 @@ class TestTTSMethods:
         assert server._is_tts is True
         assert server._tts_stage is mock_stage
 
+    def test_prepare_speech_rejects_non_tts_omni_model(self, mocker: MockerFixture):
+        """Multi-stage omni models (e.g. Qwen3-Omni) must not use /v1/audio/speech."""
+        mock_engine_client = mocker.MagicMock()
+        mock_engine_client.errored = False
+        mock_engine_client.tts_max_instructions_length = None
+
+        # Simulate Qwen3-Omni: multiple stages, none in _TTS_MODEL_STAGES
+        thinker = SimpleNamespace(engine_args=SimpleNamespace(model_stage="thinker"), tts_args={})
+        talker = SimpleNamespace(engine_args=SimpleNamespace(model_stage="talker"), tts_args={})
+        code2wav = SimpleNamespace(engine_args=SimpleNamespace(model_stage="code2wav"), tts_args={})
+        mock_engine_client.stage_configs = [thinker, talker, code2wav]
+
+        mock_models = mocker.MagicMock()
+        mock_models.is_base_model.return_value = True
+        server = OmniOpenAIServingSpeech(
+            engine_client=mock_engine_client,
+            models=mock_models,
+            request_logger=mocker.MagicMock(),
+        )
+        assert server._is_tts is False
+
+        request = OpenAICreateSpeechRequest(input="Hello world")
+        with pytest.raises(ValueError, match="only supported for dedicated TTS models"):
+            asyncio.run(server._prepare_speech_generation(request))
+        server.shutdown()
+
     def test_estimate_prompt_len_fallback(self, speech_server):
         """Test prompt length estimation falls back to 2048 when model is unavailable."""
         tts_params = {"text": ["Hello"], "task_type": ["CustomVoice"]}

vllm_omni/entrypoints/openai/serving_speech.py

@@ -1471,6 +1471,24 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
                 ph_len = await self._estimate_prompt_len_async(tts_params)
                 prompt = {"prompt_token_ids": [1] * ph_len, "additional_information": tts_params}
         else:
+            # Qwen omni models (Qwen3-Omni, Qwen2.5-Omni) use a "talker"
+            # stage whose preprocess requires chat-templated tokens.  The
+            # async-chunk orchestrator prewarms the talker via
+            # compute_talker_prompt_ids_length(), which scans for Qwen
+            # chat-template markers (im_start_token_id 151644).  A raw-text
+            # prompt produces a 1-token placeholder that crashes the talker's
+            # prefill/decode handoff.  Reject early with an actionable message.
+            stage_names = {
+                getattr(getattr(s, "engine_args", None), "model_stage", None) for s in self.engine_client.stage_configs
+            }
+            if "talker" in stage_names:
+                raise ValueError(
+                    "The /v1/audio/speech endpoint is only supported for "
+                    "dedicated TTS models (e.g., Qwen3-TTS, Voxtral, Fish "
+                    "Speech, CosyVoice3, OmniVoice, VoxCPM2). For omni "
+                    "models like Qwen3-Omni, use /v1/chat/completions with "
+                    '\'"modalities": ["audio"]\' instead.'
+                )
             tts_params = {}
             prompt = {"prompt": request.input}