[BUG FIX]: prevent EngineCore crash when Qwen TTS Base task is missing ref_text (#2203)

Signed-off-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com>
Co-authored-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com>

tests/entrypoints/openai_api/test_serving_speech.py

@@ -752,6 +752,26 @@ class TestTTSMethods:
         )
         assert speech_server._validate_tts_request(req) is None
 
+    @pytest.mark.parametrize(
+        "ref_text",
+        [None, "", "   "],
+        ids=["none", "empty", "whitespace"],
+    )
+    def test_validate_base_task_missing_ref_text_returns_400(self, speech_server, ref_text):
+        """Regression: Base task without ref_text must return 400, not crash EngineCore.
+
+        See https://github.com/vllm-project/vllm-omni/pull/2203
+        """
+        req = OpenAICreateSpeechRequest(
+            input="Hello",
+            task_type="Base",
+            ref_audio="data:audio/wav;base64,abc",
+            ref_text=ref_text,
+        )
+        result = speech_server._validate_tts_request(req)
+        assert result is not None, f"ref_text={ref_text!r} should be rejected"
+        assert "ref_text" in result
+
     def test_validate_tts_request_customvoice_no_speakers(self, speech_server):
         """CustomVoice on a model with no speakers returns 400 instead of crashing engine."""
         req = OpenAICreateSpeechRequest(input="Hello", task_type="CustomVoice")

vllm_omni/entrypoints/openai/serving_speech.py

@@ -919,6 +919,13 @@ class OmniOpenAIServingSpeech(OpenAIServing, AudioMixin):
                     fmt_err = self._validate_ref_audio_format(request.ref_audio)
                     if fmt_err:
                         return fmt_err
+                    if not getattr(request, "x_vector_only_mode", False) and (
+                        not request.ref_text or not request.ref_text.strip()
+                    ):
+                        return (
+                            "Base task requires non-empty 'ref_text' (transcript of "
+                            "the reference audio) unless 'x_vector_only_mode' is enabled"
+                        )
 
         # Validate cross-parameter dependencies
         if task_type != "Base":

vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_talker.py

@@ -1439,11 +1439,16 @@ class Qwen3TTSTalkerForConditionalGeneration(nn.Module):
                     )
                 if ref_ids is None:
                     ref_text = _as_singleton(info_dict.get("ref_text"))
-                    if not isinstance(ref_text, str) or not ref_text.strip():
-                        raise ValueError("Base in-context voice cloning requires `ref_text` or tokenized `ref_ids`.")
-                    ref_ids = tok(self._build_ref_text(ref_text), return_tensors="pt", padding=False)["input_ids"].to(
-                        device=input_ids.device
-                    )
+                    if isinstance(ref_text, str) and ref_text.strip():
+                        ref_ids = tok(
+                            self._build_ref_text(ref_text),
+                            return_tensors="pt",
+                            padding=False,
+                        )["input_ids"].to(device=input_ids.device)
+                    else:
+                        logger.warning("Base ICL: ref_text/ref_ids missing, falling back to x-vector-only mode.")
+                        in_context_mode = False
+            if in_context_mode:
                 icl_input_embed, trailing_text_hidden = self._generate_icl_prompt(
                     text_id=input_ids[:, 3:-5],
                     ref_id=ref_ids[:, 3:-2],