[Core][Frontend] Unified Full-duplex Framework (#7413)

Signed-off-by: chickeyton <ngton2014@gmail.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

tests/model_executor/models/minicpmo_4_5/duplex/test_commit_policy.py

@@ -5,7 +5,7 @@ from __future__ import annotations
 
 import pytest
 
-from vllm_omni.entrypoints.duplex.commit_policy import (
+from vllm_omni.engine.duplex.session.commit_policy import (
     CommitAction,
     CommitSnapshot,
     decide_commit_action,
@@ -24,7 +24,7 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
                 speech_since_commit=True,
                 active_response_id=None,
                 overlap_speech_ms=0,
-                native_response_in_progress=False,
+                response_in_progress=False,
                 playback_active=False,
             ),
             CommitAction.START_AUTO_RESPONSE,
@@ -36,7 +36,7 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
                 speech_since_commit=True,
                 active_response_id=None,
                 overlap_speech_ms=800,
-                native_response_in_progress=True,
+                response_in_progress=True,
                 playback_active=True,
             ),
             CommitAction.START_AUTO_RESPONSE,
@@ -48,7 +48,7 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
                 speech_since_commit=False,
                 active_response_id=None,
                 overlap_speech_ms=0,
-                native_response_in_progress=True,
+                response_in_progress=True,
                 playback_active=False,
             ),
             CommitAction.COMMIT_ONLY,
@@ -60,7 +60,7 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
                 speech_since_commit=True,
                 active_response_id=None,
                 overlap_speech_ms=800,
-                native_response_in_progress=True,
+                response_in_progress=True,
                 playback_active=False,
             ),
             CommitAction.DEFER_ACTIVE_RESPONSE,
@@ -72,7 +72,7 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
                 speech_since_commit=True,
                 active_response_id=None,
                 overlap_speech_ms=800,
-                native_response_in_progress=False,
+                response_in_progress=False,
                 playback_active=True,
             ),
             CommitAction.DEFER_ACTIVE_RESPONSE,
@@ -84,7 +84,7 @@ pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
                 speech_since_commit=True,
                 active_response_id=None,
                 overlap_speech_ms=0,
-                native_response_in_progress=False,
+                response_in_progress=False,
                 playback_active=False,
             ),
             CommitAction.COMMIT_ONLY,

tests/model_executor/models/minicpmo_4_5/duplex/test_data_plane_audio_cursor.py

@@ -0,0 +1,78 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+"""The cumulative Stage1 audio cursor.
+
+Stage1 hands the data plane a snapshot of every sample it has generated for the
+request so far, and only the new tail may reach the client. The cursor that
+makes that possible is the one piece of per-request state whose loss is silent
+and expensive: rewinding it re-sends the whole session as a single delta.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+import pytest
+
+from vllm_omni.model_executor.models.minicpmo_4_5.duplex.data_plane import (
+    MiniCPMO45DataPlaneSession,
+)
+
+pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
+
+REQUEST_ID = "duplex-s.c2Vzc2lvbg==.e.0.r.stage0"
+
+
+def _session() -> MiniCPMO45DataPlaneSession:
+    return MiniCPMO45DataPlaneSession(lambda *args, **kwargs: "")
+
+
+def _snapshot(num_samples: int) -> np.ndarray:
+    return np.arange(num_samples, dtype=np.float32)
+
+
+def _sliced_len(session: MiniCPMO45DataPlaneSession, num_samples: int) -> int:
+    sliced = session.slice_cumulative_audio(REQUEST_ID, _snapshot(num_samples))
+    if sliced is None:
+        return 0
+    return int(np.asarray(sliced).reshape(-1).size)
+
+
+def test_a_growing_snapshot_yields_only_the_new_tail() -> None:
+    session = _session()
+    assert _sliced_len(session, 24000) == 24000
+    assert _sliced_len(session, 48000) == 24000
+    assert _sliced_len(session, 48000) == 0  # unchanged snapshot emits nothing
+
+
+def test_a_placeholder_unit_does_not_rewind_the_cursor() -> None:
+    """The failure behind a 300s benchmark case producing 1373s of audio.
+
+    A continuation unit arrives as a one-sample snapshot. Taking that as "the
+    producer restarted" drops the cursor to 1, so the next real snapshot is
+    sliced from the start and the client is handed the entire session again.
+    """
+    session = _session()
+    assert _sliced_len(session, 637440) == 637440
+    assert _sliced_len(session, 1) == 1
+    assert _sliced_len(session, 1) == 1
+    assert _sliced_len(session, 1) == 1
+
+    # 649920 - 637440, not 649920 - 1.
+    assert _sliced_len(session, 649920) == 12480
+
+
+def test_a_genuinely_restarted_producer_is_still_picked_up() -> None:
+    """Two consecutive below-cursor snapshots that grow are a new buffer."""
+    session = _session()
+    assert _sliced_len(session, 48000) == 48000
+    assert _sliced_len(session, 1000) == 1000  # head of the new buffer
+    assert _sliced_len(session, 3000) == 2000  # its tail, not 3000 - 48000
+    assert _sliced_len(session, 5000) == 2000
+
+
+def test_close_stream_releases_the_cursor_for_the_next_stream() -> None:
+    session = _session()
+    assert _sliced_len(session, 48000) == 48000
+    session.close_stream(REQUEST_ID)
+    assert _sliced_len(session, 24000) == 24000

tests/model_executor/models/minicpmo_4_5/test_pipeline.py

@@ -55,20 +55,18 @@ class TestRegistryDeclaration:
         assert pipeline.model_type == _PIPELINE_KEY
         assert pipeline.model_arch == "MiniCPMO45OmniForConditionalGeneration"
 
-    def test_native_duplex_control_is_explicit_without_a_fixed_session_cap(self) -> None:
+    def test_duplex_plugin_is_declared_without_a_fixed_session_cap(self) -> None:
         pipeline = OMNI_PIPELINES[_PIPELINE_KEY]
-        assert pipeline.duplex_control_enabled is True
-        assert pipeline.duplex_serving_adapter == (
-            "vllm_omni.model_executor.models.minicpmo_4_5.duplex.serving_adapter.MiniCPMO45ServingRuntimeAdapter"
+        assert pipeline.duplex_plugin == (
+            "vllm_omni.model_executor.models.minicpmo_4_5.duplex.plugin.MiniCPMO45DuplexPlugin"
         )
         assert not hasattr(pipeline, "max_native_duplex_sessions")
 
-    def test_ordinary_pipeline_defaults_to_no_duplex_control(self) -> None:
+    def test_ordinary_pipeline_declares_no_duplex_plugin(self) -> None:
         pipeline = PipelineConfig(
             model_type="ordinary", stages=(StagePipelineConfig(stage_id=0, model_stage="a", final_output=True),)
         )
-        assert pipeline.duplex_control_enabled is False
-        assert pipeline.duplex_serving_adapter is None
+        assert pipeline.duplex_plugin is None
         assert not hasattr(pipeline, "max_native_duplex_sessions")
 
 

tests/model_executor/models/nemotron_voicechat/duplex/conftest.py

@@ -0,0 +1,20 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+"""Nemotron VoiceChat duplex is not on the duplex plugin framework yet.
+
+RFC vllm-omni#7181 splits the port: this PR lands the unified framework with
+MiniCPM-o 4.5, and the follow-up PR moves Nemotron VoiceChat onto
+``DuplexModelPlugin``. Until then the modules under test still import the
+pre-framework duplex runtime / serving adapter that this PR removed, so
+collecting them would fail at import time and abort the whole pytest session.
+The tests themselves are unchanged and come back with the port.
+
+Listed per file rather than globbed, so the coverage loss is visible and shrinks
+file by file as the port lands. ``test_input.py`` imports nothing that was
+removed and keeps running.
+"""
+
+#: Both import ``DuplexFence`` from ``engine.duplex.messages``, which no longer
+#: exports it; ``test_data_plane`` also imports ``engine.duplex.runtime``.
+collect_ignore = ["test_data_plane.py", "test_runtime_contract.py"]

tests/model_executor/models/nemotron_voicechat/duplex/test_data_plane.py

@@ -7,9 +7,9 @@ from types import SimpleNamespace
 
 import numpy as np
 import pytest
+from vllm_omni.engine.duplex.runtime import duplex_resource_request_id
 
 from vllm_omni.engine.duplex.messages import DuplexFence
-from vllm_omni.engine.duplex.runtime import duplex_resource_request_id
 from vllm_omni.model_executor.models.nemotron_voicechat.duplex.data_plane import (
     NemotronVoiceChatDataPlaneContext,
     NemotronVoiceChatDataPlaneSession,

tests/model_executor/models/personaplex/duplex/conftest.py

@@ -0,0 +1,21 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
+"""PersonaPlex duplex is not on the duplex plugin framework yet.
+
+RFC vllm-omni#7181 splits the port: this PR lands the unified framework with
+MiniCPM-o 4.5, and the follow-up PR moves PersonaPlex onto ``DuplexModelPlugin``.
+Until then the modules under test still import the pre-framework duplex
+runtime / serving adapter that this PR removed, so collecting them would fail
+at import time and abort the whole pytest session. The tests themselves are
+unchanged and come back with the port.
+
+Listed per file rather than globbed, so the coverage loss is visible and shrinks
+file by file as the port lands.
+"""
+
+#: ``test_unified_runtime`` imports ``engine.duplex.runtime`` and
+#: ``entrypoints.duplex.runtime_adapter`` directly. ``test_stage0_runtime`` fails
+#: transitively: ``personaplex.duplex.runtime_extension`` imports ``DuplexFence``
+#: from ``engine.duplex.messages``, which no longer exports it.
+collect_ignore = ["test_stage0_runtime.py", "test_unified_runtime.py"]

tests/model_executor/models/personaplex/duplex/test_unified_runtime.py

@@ -8,6 +8,12 @@ from types import SimpleNamespace
 import numpy as np
 import pytest
 from vllm.sampling_params import SamplingParams
+from vllm_omni.engine.duplex.runtime import (
+    DuplexInputMode,
+)
+from vllm_omni.entrypoints.duplex.runtime_adapter import (
+    ServingRuntimeConfigError,
+)
 
 from tests.e2e.online_serving import personaplex_realtime_duplex as e2e_driver
 from vllm_omni.config.stage_config import (
@@ -15,12 +21,6 @@ from vllm_omni.config.stage_config import (
     merge_pipeline_deploy,
 )
 from vllm_omni.engine.duplex.messages import DuplexFence
-from vllm_omni.engine.duplex.runtime import (
-    DuplexInputMode,
-)
-from vllm_omni.entrypoints.duplex.runtime_adapter import (
-    ServingRuntimeConfigError,
-)
 from vllm_omni.model_executor.models.personaplex.duplex.config import DEFAULT_PERSONA
 from vllm_omni.model_executor.models.personaplex.duplex.data_plane import (
     PersonaPlexDataPlaneContext,