fix(qwen4-exp): keep text inference on mlx-vlm path

tests/test_text_model_dispatch.py

@@ -109,3 +109,20 @@ class _Vlm:
     with caplog.at_level(logging.ERROR, logger="vllm_mlx.text_model_from_vlm"):
         assert build_text_model(_Vlm(), tmp_path) is None
     assert not caplog.records
+
+
+@pytest.mark.parametrize("model_type", ["qwen4_exp", "qwen4_exp_text"])
+def test_qwen4_exp_stays_on_mlx_vlm_text_path(tmp_path, caplog, model_type):
+    """Qwen4-Exp is not compatible with the generic Qwen3.5 TextModel."""
+    (tmp_path / "config.json").write_text(
+        '{"text_config": {"model_type": "' + model_type + '"}}'
+    )
+
+    class _Vlm:
+        language_model = object()
+
+    with caplog.at_level(logging.INFO, logger="vllm_mlx.text_model_from_vlm"):
+        assert build_text_model(_Vlm(), tmp_path) is None
+
+    assert "mlx-vlm text path" in caplog.records[-1].getMessage()
+    assert not [record for record in caplog.records if record.levelno >= logging.ERROR]

vllm_mlx/text_model_from_vlm.py

@@ -37,6 +37,12 @@
 # that names neither the model nor the class. Hence the logging either side.
 _DEFAULT_TEXT_MODEL = ("mlx_lm.models.qwen3_5", ("TextModel", "TextModelArgs"))
 
+# These architectures are not compatible with the generic Qwen3.5 text
+# skeleton. Returning no extracted TextModel keeps SimpleEngine on the loaded
+# mlx-vlm path for text as well as media instead of silently serving incorrect
+# logits from a mechanically compatible but architecturally different model.
+_VLM_ONLY_TEXT_MODEL_PREFIXES = ("qwen4_exp",)
+
 
 def _import_text_model_classes(model_type: str):
     """Return ``(Model, ModelArgs)`` for a text config's ``model_type``."""
@@ -92,6 +98,13 @@ def build_text_model(
         config = json.loads((model_path / "config.json").read_text())
         text_config = config.get("text_config", config)
         model_type = text_config.get("model_type") or config.get("model_type", "")
+        if model_type.startswith(_VLM_ONLY_TEXT_MODEL_PREFIXES):
+            logger.info(
+                "Keeping model_type=%r on the mlx-vlm text path; no compatible "
+                "mlx-lm TextModel is registered",
+                model_type,
+            )
+            return None
         TextModel, TextModelArgs = _import_text_model_classes(model_type)
         text_model_cls = f"{TextModel.__module__}.{TextModel.__qualname__}"
         logger.debug(