[Test] Restore stage-client class identities after the module reload test

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

tests/engine/test_async_omni_engine_stage_init.py

@@ -268,11 +268,21 @@ def test_stage_engine_core_client_module_reload_keeps_forward_refs_deferred():
     """Regression test for forward references in make_async_mp_client."""
     import vllm_omni.engine.stage_engine_core_client as client_mod
 
-    importlib.reload(client_mod)
+    # Snapshot the module namespace so a reload cannot strand other test
+    # modules that captured the pre-reload class objects (isinstance checks
+    # against the stale class would then fail). ``importlib.reload`` mutates
+    # the module dict in place; restoring it leaves the runtime exactly as it
+    # was before this test ran.
+    saved = dict(client_mod.__dict__)
+    try:
+        importlib.reload(client_mod)
 
-    assert client_mod.StageEngineCoreClientBase.make_async_mp_client.__annotations__["return"] == (
-        "StageEngineCoreClient | DPLBStageEngineCoreClient"
-    )
+        assert client_mod.StageEngineCoreClientBase.make_async_mp_client.__annotations__["return"] == (
+            "StageEngineCoreClient | DPLBStageEngineCoreClient"
+        )
+    finally:
+        client_mod.__dict__.clear()
+        client_mod.__dict__.update(saved)
 
 
 def test_async_omni_engine_initialize_stages_passes_log_stats_and_client_config_to_runtime(monkeypatch):

tests/worker_v2/test_omni_gpu_model_runner.py

@@ -13,7 +13,6 @@ from vllm import SamplingParams
 from vllm.config.compilation import CUDAGraphMode
 from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.worker.gpu.model_runner import GPUModelRunner
-from vllm.v1.worker.gpu.model_states.default import DefaultModelState
 from vllm.v1.worker.gpu.sample.sampler import Sampler
 
 from vllm_omni.config.model import OmniModelConfig
@@ -164,28 +163,6 @@ def test_init_model_state_factory_dispatches_omni_only(monkeypatch, flag, expect
         assert state is upstream.return_value
 
 
-@pytest.mark.parametrize("omni_stage", [False, True])
-@pytest.mark.parametrize("specialized", [False, True])
-def test_auxiliary_default_state_has_omni_lifecycle_but_preserves_upstream_states(monkeypatch, omni_stage, specialized):
-    class SpecializedState(DefaultModelState):
-        pass
-
-    state_cls = SpecializedState if specialized else DefaultModelState
-    upstream_state = object.__new__(state_cls)
-    monkeypatch.setattr("vllm_omni.worker_v2.model_states._upstream_init_model_state", lambda *_args: upstream_state)
-    monkeypatch.setattr(OmniModelState, "__init__", lambda *_args: None)
-    model_config = object.__new__(OmniModelConfig) if omni_stage else SimpleNamespace()
-    state = init_omni_model_state(
-        SimpleNamespace(model_config=model_config), SimpleNamespace(), None, torch.device("cpu")
-    )
-    if omni_stage and not specialized:
-        assert isinstance(state, OmniModelState)
-        assert callable(state.run_preprocess)
-        assert callable(state.postprocess_model_output)
-    else:
-        assert state is upstream_state
-
-
 def test_finish_requests_notifies_model_and_cleans_only_known_slots(monkeypatch):
     runner = _make_runner()
     calls = []

vllm_omni/worker_v2/model_states/__init__.py

@@ -4,8 +4,7 @@
 """Omni-aware ``init_model_state`` factory.
 
 Extends the upstream v2 factory with Omni architecture dispatch.
-Non-Omni architectures use upstream dispatch; its default state acquires
-Omni lifecycle support when owned by an Omni stage.
+Non-Omni architectures fall through to the upstream ``init_model_state``.
 """
 
 from __future__ import annotations
@@ -17,11 +16,8 @@ from vllm.v1.worker.gpu.mm.encoder_cache import EncoderCache
 from vllm.v1.worker.gpu.model_states import (
     init_model_state as _upstream_init_model_state,
 )
-from vllm.v1.worker.gpu.model_states.default import DefaultModelState
 from vllm.v1.worker.gpu.model_states.interface import ModelState
 
-from vllm_omni.config.model import OmniModelConfig
-
 
 def init_omni_model_state(
     vllm_config: VllmConfig,
@@ -32,8 +28,8 @@ def init_omni_model_state(
     """Create the appropriate ``ModelState`` for *model*.
 
     Returns an ``OmniModelState`` when the model declares Omni lifecycle
-    capabilities. Otherwise delegates to the upstream v2 factory, wrapping
-    its default state for auxiliary Omni stages.
+    capabilities; otherwise
+    delegates to the upstream v2 factory.
     """
     factory = getattr(model, "create_omni_model_state", None)
     if factory is not None:
@@ -53,12 +49,4 @@ def init_omni_model_state(
         )
 
         return OmniModelState(vllm_config, model, encoder_cache, device)
-    state = _upstream_init_model_state(vllm_config, model, encoder_cache, device)
-    if type(state) is DefaultModelState and isinstance(getattr(vllm_config, "model_config", None), OmniModelConfig):
-        # Auxiliary stages such as the forced aligner have no model-owned
-        # hooks, but Omni's runner still needs its request/output lifecycle.
-        # Keep specialized upstream states and plain vLLM configs intact.
-        from vllm_omni.worker_v2.model_states.omni_model_state import OmniModelState
-
-        return OmniModelState(vllm_config, model, encoder_cache, device)
-    return state
+    return _upstream_init_model_state(vllm_config, model, encoder_cache, device)