[Bugfix][Full Duplex] Preserve shared runtime compatibility

Signed-off-by: Sy03 <1370724210@qq.com>

tests/distributed/omni_connectors/test_chunk_transfer_adapter.py

@@ -591,7 +591,7 @@ def test_save_async_drops_late_previous_segment_after_boundary_reset(build_adapt
     assert adapter._pending_save_reqs[-1]["request"] is next_segment
 
 
-def test_send_single_request_drops_queued_old_segment_after_boundary(build_adapter):
+def test_send_single_request_preserves_queued_preboundary_chunk(build_adapter):
     adapter, connector = build_adapter(stage_id=1)
     adapter.custom_process_next_stage_input_func = lambda **kwargs: OmniPayloadStruct(meta=MetaStruct())
     request = _req("req-queued-old", RequestStatus.WAITING, external_req_id="ext-queued-old")
@@ -603,11 +603,11 @@ def test_send_single_request_drops_queued_old_segment_after_boundary(build_adapt
 
     old_task = adapter._pending_save_reqs.popleft()
     adapter._send_single_request(old_task)
-    connector.put.assert_not_called()
+    connector.put.assert_called_once()
 
     boundary_task = adapter._pending_save_reqs.popleft()
     adapter._send_single_request(boundary_task)
-    connector.put.assert_called_once()
+    assert connector.put.call_count == 2
 
 
 def test_send_single_request_cleans_up_after_finished_payload(build_adapter, monkeypatch):

tests/entrypoints/openai/test_duplex_protocol.py

@@ -462,6 +462,16 @@ def test_minicpmo_native_capabilities_separate_model_state_from_core_kv_lease():
     assert caps["stage_handoff_transport"] == "scheduler_data_plane"
 
 
+def test_duplex_overlap_policy_auto_remains_a_compatibility_value():
+    from vllm_omni.experimental.fullduplex.openai.protocol import (
+        DuplexOverlapPolicy,
+        DuplexSessionConfig,
+    )
+
+    assert DuplexOverlapPolicy.AUTO.value == "auto"
+    assert DuplexSessionConfig._normalize_overlap_policy("auto") == "auto"
+
+
 def test_minicpmo_native_capabilities_do_not_overclaim_single_session_deployment():
     caps = DuplexCapabilities.minicpmo45_native(max_sessions=1).as_dict()
 

tests/entrypoints/openai_api/test_duplex_handler.py

@@ -5439,6 +5439,33 @@ async def test_minicpmo_native_duplex_open_unsupported_fails_session_create():
     assert ws.sent[0]["runtime_control"]["unsupported_count"] == 1
 
 
+@pytest.mark.asyncio
+async def test_native_runtime_rejects_server_vad_without_barge_in(monkeypatch: pytest.MonkeyPatch):
+    engine = FakeEngineClient()
+    handler = OmniDuplexSessionHandler(
+        chat_service=FakeChatService(engine),
+        config_timeout_s=0.1,
+        idle_timeout_s=1,
+    )
+    capabilities = DuplexCapabilities.minicpmo45_native()
+    capabilities.supports_barge_in = False
+    monkeypatch.setattr(
+        handler._serving_runtime_adapter,
+        "capabilities",
+        lambda *, max_sessions: capabilities,
+    )
+    event = _native_session_create("sid-native-no-barge-in")
+    event["session"]["overlap_policy"] = "barge_in_on_speech"
+    ws = TimedWebSocket()
+    ws.put(event)
+
+    await handler.handle_session(ws)
+
+    assert ws.sent_types() == ["error"]
+    assert ws.sent[0]["code"] == "unsupported_turn_detection"
+    assert engine.opened == []
+
+
 @pytest.mark.asyncio
 async def test_minicpmo_native_duplex_rejects_engine_without_fence_contract():
     class LegacyEngineClient(FakeEngineClient):

vllm_omni/distributed/omni_connectors/transfer_adapter/chunk_transfer_adapter.py

@@ -370,21 +370,6 @@ class OmniChunkTransferAdapter(OmniTransferAdapterBase):
         stage_id = self.connector.stage_id
         next_stage_id = stage_id + 1
         external_req_id = request.external_req_id
-        task_generation = task.get("segment_generation")
-        expected_generation = self._segment_generation.get(external_req_id, task_generation)
-        if (
-            isinstance(task_generation, int)
-            and isinstance(expected_generation, int)
-            and task_generation < expected_generation
-            and not is_segment_finished
-        ):
-            logger.warning(
-                "Skip queued late chunk for request %s, segment_generation=%s, expected=%s",
-                external_req_id,
-                task_generation,
-                expected_generation,
-            )
-            return
         chunk_id = self.put_req_chunk[external_req_id]
         connector_put_key = f"{external_req_id}_{stage_id}_{chunk_id}"
         # Process payload in save_loop thread

vllm_omni/experimental/fullduplex/openai/protocol.py

@@ -9,6 +9,7 @@ from uuid import uuid4
 
 
 class DuplexOverlapPolicy(str, Enum):
+    AUTO = "auto"
     LISTEN_ONLY = "listen_only"
     BARGE_IN_ON_SPEECH = "barge_in_on_speech"
 

vllm_omni/experimental/fullduplex/openai/serving.py

@@ -961,6 +961,23 @@ class OmniDuplexSessionHandler(
                     max_sessions=self._duplex_session_config.max_sessions,
                 )
             )
+            if (
+                config.overlap_policy == DuplexOverlapPolicy.BARGE_IN_ON_SPEECH.value
+                and not session.capabilities.supports_barge_in
+            ):
+                if realtime_protocol is not None:
+                    realtime_protocol.reject_realtime_turn_detection_update()
+                await send_json(
+                    {
+                        "type": "error",
+                        "error": (
+                            "server_vad with interrupt_response=true requires a runtime "
+                            "that supports barge-in"
+                        ),
+                        "code": "unsupported_turn_detection",
+                    }
+                )
+                return None
             session.replace_runtime_config(runtime_config)
         return _DuplexSessionHandshake(session=session)