[Bugfix][Qwen3-TTS] Ignore voice labels for inline Base cloning (#6974)

Signed-off-by: sc-hua <sc_hua@qq.com>

tests/e2e/online_serving/test_qwen3_tts_base_expansion.py

@@ -152,3 +152,23 @@ def test_xvector_then_icl_same_ref_audio_keeps_engine_alive(omni_server, online_
     online_client.send_audio_speech_request({**base_request, "x_vector_only_mode": True})
     online_client.send_audio_speech_request({**base_request, "x_vector_only_mode": False, "ref_text": REF_TEXT})
     online_client.send_audio_speech_request({**base_request, "x_vector_only_mode": True})
+
+
+@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@pytest.mark.parametrize("omni_server", tts_async_chunk_server_params, indirect=True)
+def test_inline_ref_audio_cache_ignores_openai_voice_label(omni_server, online_client) -> None:
+    """An OpenAI voice label must not identify an inline Base voice clone."""
+    base_request = {
+        "model": omni_server.model,
+        "input": get_prompt(),
+        "stream": False,
+        "timeout": DEFAULT_AUDIO_SPEECH_TIMEOUT_S,
+        "response_format": "wav",
+        "task_type": "Base",
+        "ref_audio": REF_AUDIO_URL,
+        "ref_text": REF_TEXT,
+        "min_audio_bytes": 1,
+    }
+
+    online_client.send_audio_speech_request({**base_request, "voice": "voice-a"})
+    online_client.send_audio_speech_request({**base_request, "voice": "voice-b"})

tests/entrypoints/openai_api/test_serving_speech.py

@@ -5507,8 +5507,9 @@ class TestTTSAsyncOffloading:
         assert prompt["additional_information"]["non_streaming_mode"] == [True]
         assert "full_utterance_decode" not in prompt["additional_information"]
 
-    def test_qwen3_repeated_ref_audio_hot_path_sends_cache_key_without_waveform(self, qwen3_tts_server):
-        """After a ref artifact is marked ready, repeated requests avoid ref_audio payload IPC."""
+    def test_qwen3_inline_ref_audio_hot_path_does_not_use_named_speaker_cache(self, qwen3_tts_server, mocker):
+        """An OpenAI-compatible voice does not identify an inline voice clone."""
+        ignored_voice_log = mocker.patch("vllm_omni.entrypoints.openai.tts_adapters.qwen3_tts.logger.info")
         wav_list = [0.0] * 48000
         artifact_key = "a" * 40
         ref_audio = "data:audio/wav;base64,same"
@@ -5531,6 +5532,7 @@ class TestTTSAsyncOffloading:
 
         request = OpenAICreateSpeechRequest(
             input="hello",
+            voice="voice-a",
             task_type="Base",
             ref_audio=ref_audio,
             ref_text="reference",
@@ -5540,7 +5542,14 @@ class TestTTSAsyncOffloading:
         )
 
         assert request_id == "req-hot"
+        assert request.voice == "voice-a"
+        ignored_voice_log.assert_called_once_with(
+            "Ignoring voice=%r for Qwen3-TTS Base request because inline ref_audio takes precedence",
+            "voice-a",
+        )
         assert "ref_audio" not in tts_params
+        assert "speaker" not in tts_params
+        assert "voice_created_at" not in tts_params
         assert tts_params["_qwen3_tts_ref_audio_cache_key"] == [artifact_key]
         assert tts_params["ref_code_length"] == [50]
         prompt = qwen3_tts_server.engine_client.generate.call_args.kwargs["prompt"]

vllm_omni/entrypoints/openai/tts_adapters/qwen3_tts.py

@@ -431,6 +431,14 @@ class Qwen3TTSAdapter(ARTTSAdapter):
         Qwen3-TTS ref-audio artifact tracked after ``generate()``.
         """
         server = self.ctx.server
+        # Inline Base cloning derives its voice from ref_audio, not the
+        # OpenAI-compatible voice label.
+        if has_inline_ref_audio and request.task_type == "Base" and request.voice is not None:
+            logger.info(
+                "Ignoring voice=%r for Qwen3-TTS Base request because inline ref_audio takes precedence",
+                request.voice,
+            )
+            request = request.model_copy(update={"voice": None})
         qwen3_ref_audio_warmup_artifact_key: str | None = None
         tts_params = self._build_tts_params(request)
         # Resolve ref_audio (explicit or auto-set for uploaded voices)