[Core] Graduate MiniCPM-o 4.5 and PersonaPlex full-duplex serving out of experimental (#6196)

Signed-off-by: chickeyton <ngton2014@gmail.com>
Signed-off-by: Gao Han <hgaoaf@connect.ust.hk>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Gao Han <hgaoaf@connect.ust.hk>

tests/engine/test_duplex_import_boundary.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from __future__ import annotations
 
@@ -29,7 +29,9 @@ def _assert_isolated_import_succeeds(script: str) -> None:
         env=env,
         capture_output=True,
         text=True,
-        timeout=60,
+        # The CLI probe imports the whole vllm + benchmark graph; on a loaded
+        # shared host that alone can take over a minute.
+        timeout=180,
         check=False,
     )
     assert result.returncode == 0, result.stderr
@@ -43,7 +45,11 @@ def test_isolated_import_ignores_conflicting_gpu_visibility(monkeypatch: pytest.
     _assert_isolated_import_succeeds("import vllm_omni.outputs")
 
 
-def test_stable_engine_imports_do_not_load_experimental_duplex() -> None:
+def test_stable_engine_imports_load_duplex_kernel_eagerly() -> None:
+    # The duplex kernel is imported eagerly by the stable engine modules.
+    # Model-specific duplex adapters must still load only via the
+    # dotted-string plugin paths, and the client-side vllm_omni.clients
+    # package must never enter the stable runtime import graph.
     _assert_isolated_import_succeeds("""
 import sys
 
@@ -51,14 +57,84 @@ import vllm_omni.engine.async_omni_engine
 import vllm_omni.engine.orchestrator
 import vllm_omni.entrypoints.async_omni
 
+expected_eager = (
+    "vllm_omni.engine.duplex.contracts",
+    "vllm_omni.engine.duplex.control_plane",
+    "vllm_omni.entrypoints.duplex_request_client",
+    "vllm_omni.outputs.duplex",
+)
+missing = sorted(name for name in expected_eager if name not in sys.modules)
+if missing:
+    raise SystemExit("duplex kernel modules not imported eagerly: " + ", ".join(missing))
+
+forbidden_prefixes = (
+    "vllm_omni.clients",
+    "vllm_omni.model_executor.models.minicpmo_4_5.duplex",
+    "vllm_omni.model_executor.models.nemotron_voicechat.duplex",
+    "vllm_omni.model_executor.models.personaplex.duplex",
+)
 loaded = sorted(
     name
     for name in sys.modules
-    if name == "vllm_omni.experimental.fullduplex"
-    or name.startswith("vllm_omni.experimental.fullduplex.")
+    if any(name == prefix or name.startswith(prefix + ".") for prefix in forbidden_prefixes)
 )
 if loaded:
-    raise SystemExit("stable imports loaded experimental duplex: " + ", ".join(loaded))
+    raise SystemExit("stable imports loaded plugin-only duplex modules: " + ", ".join(loaded))
+""")
+
+
+def test_cli_import_graph_stays_clear_of_the_client_package() -> None:
+    # `vllm-omni serve` imports the CLI package, whose __init__ eagerly pulls
+    # in the benchmark patch module and every benchmark subcommand. None of
+    # that may drag the client-side vllm_omni.clients package into the server
+    # process; the benchmark runtime imports it lazily at execution time.
+    _assert_isolated_import_succeeds("""
+import sys
+
+import vllm_omni.entrypoints.cli.main
+
+loaded = sorted(
+    name
+    for name in sys.modules
+    if name == "vllm_omni.clients" or name.startswith("vllm_omni.clients.")
+)
+if loaded:
+    raise SystemExit("CLI import graph loaded the client package: " + ", ".join(loaded))
+""")
+
+
+def test_client_package_stays_clear_of_the_server_import_graph() -> None:
+    # vllm_omni.clients.duplex advertises a lightweight dependency set
+    # (pybase64 + websockets). Importing it and exercising the code paths
+    # that previously reached into the server tree (timing_summary once
+    # imported vllm_omni.metrics, dragging in prometheus_client) must not
+    # load any vllm_omni module outside the client package.
+    _assert_isolated_import_succeeds("""
+import base64
+import sys
+
+import vllm_omni  # the package __init__ loads its own baseline graph
+
+baseline = set(sys.modules)
+
+from vllm_omni.clients.duplex import EventCollector
+
+collector = EventCollector()
+chunk = base64.b64encode(b"\\x00\\x00" * 2400).decode("ascii")
+collector.add({"type": "response.created", "response_id": "r1"}, received_at_s=1.0)
+collector.add({"type": "response.audio.delta", "response_id": "r1", "delta": chunk}, received_at_s=1.1)
+collector.add({"type": "response.audio.delta", "response_id": "r1", "delta": chunk}, received_at_s=1.2)
+summary = collector.timing_summary(after_s=0.0, response_id="r1")
+if "audio_output" not in summary or "request_metrics" not in summary:
+    raise SystemExit("timing_summary produced no audio_output/request_metrics sections")
+
+added = sorted(
+    name
+    for name in set(sys.modules) - baseline
+    if name.startswith("vllm_omni.") and not name.startswith("vllm_omni.clients")
+)
+if added:
+    raise SystemExit("the client package pulled in server modules: " + ", ".join(added))
 """)
 
 
@@ -95,26 +171,43 @@ if "duplex_output_decision" in fields:
 """)
 
 
-def test_stable_model_executor_does_not_expose_duplex_helper_module() -> None:
+def test_duplex_sampling_helper_lives_in_model_executor() -> None:
     _assert_isolated_import_succeeds("""
 import importlib.util
 
-if importlib.util.find_spec("vllm_omni.model_executor.duplex") is not None:
-    raise SystemExit("stable model_executor still exposes duplex helper module")
+if importlib.util.find_spec("vllm_omni.model_executor.duplex_sampling") is None:
+    raise SystemExit("model_executor duplex_sampling helper module is missing")
 """)
+    assert not (REPO_ROOT / "vllm_omni" / "experimental" / "fullduplex" / "model_executor.py").exists()
 
 
 def test_runtime_package_does_not_bundle_the_browser_demo() -> None:
     assert not (REPO_ROOT / "vllm_omni" / "experimental" / "fullduplex" / "web").exists()
 
 
-def test_experimental_engine_uses_canonical_contract_module_names() -> None:
-    engine_dir = REPO_ROOT / "vllm_omni" / "experimental" / "fullduplex" / "engine"
+def test_engine_duplex_uses_canonical_contract_module_names() -> None:
+    engine_dir = REPO_ROOT / "vllm_omni" / "engine" / "duplex"
     core_dir = REPO_ROOT / "vllm_omni" / "experimental" / "fullduplex" / "core"
 
-    assert (engine_dir / "contracts.py").is_file()
-    assert (engine_dir / "lease.py").is_file()
-    assert (engine_dir / "messages.py").is_file()
+    for name in (
+        "contracts.py",
+        "lease.py",
+        "messages.py",
+        "session.py",
+        "control_plane.py",
+        "control_client.py",
+        "runtime.py",
+        "intermediate.py",
+    ):
+        assert (engine_dir / name).is_file()
+    # the duplex_ prefix is dropped inside the duplex package
+    assert not (engine_dir / "duplex_session.py").exists()
     assert not (engine_dir / "duplex_lease.py").exists()
     assert not (engine_dir / "duplex_types.py").exists()
+    # the experimental package retains only the joyvl framework; stale
+    # __pycache__ leftovers may exist, so assert on source files, not the dir
+    old_engine_dir = REPO_ROOT / "vllm_omni" / "experimental" / "fullduplex" / "engine"
+    assert not list(old_engine_dir.glob("*.py"))
+    old_personaplex_dir = REPO_ROOT / "vllm_omni" / "experimental" / "fullduplex" / "personaplex"
+    assert not list(old_personaplex_dir.glob("**/*.py"))
     assert not (core_dir / "identity.py").exists()