[Rebase] Align vLLM-Omni with vLLM v0.31.0 (#8459)

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>

tests/engine/test_arg_utils_shared_fields.py

@@ -8,6 +8,7 @@ from __future__ import annotations
 from dataclasses import fields
 
 import pytest
+from vllm.config import AuxOutputConfig
 
 from vllm_omni.engine.arg_utils import (
     SHARED_FIELDS,
@@ -49,3 +50,26 @@ def test_internal_blacklist_keys_derived_from_orchestrator():
     """
     blacklist = internal_blacklist_keys()
     assert blacklist == orchestrator_field_names() - SHARED_FIELDS
+
+
+@pytest.mark.parametrize(
+    "aux_args",
+    [
+        {"enable_return_routed_experts": True},
+        {"aux_output_config": {"enable_return_routed_experts": True}},
+        {"aux_output_config": AuxOutputConfig(enable_return_routed_experts=True)},
+    ],
+)
+@pytest.mark.parametrize("use_v2_model_runner", [False, True])
+@pytest.mark.parametrize("is_npu", [False, True])
+def test_aux_outputs_require_a_supported_omni_stage_runner(aux_args, use_v2_model_runner, is_npu, monkeypatch):
+    # Upstream runner selection reads this environment variable; Omni workers
+    # select their runner from the separate per-stage flag instead.
+    monkeypatch.setenv("VLLM_USE_V2_MODEL_RUNNER", "1")
+    monkeypatch.setattr("vllm_omni.engine.arg_utils.current_omni_platform.is_npu", lambda: is_npu)
+    if use_v2_model_runner or is_npu:
+        args = OmniEngineArgs(model="unused", use_v2_model_runner=use_v2_model_runner, **aux_args)
+        assert args.aux_output_config.enabled
+    else:
+        with pytest.raises(ValueError, match="use_v2_model_runner=True for this Omni stage"):
+            OmniEngineArgs(model="unused", **aux_args)