[WIP][Bugfix][MiniCPM-o] Fix Full duplex bugs for minicpm-challenge (#5496)

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

.buildkite/cuda/test-merge.yml

@@ -165,6 +165,31 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
+      - label: "Omni · MiniCPM-o 4.5 Duplex Test"
+        source_file_dependencies:
+          - tests/assets/minicpmo_4_5/response_required_16k.wav
+          - tests/assets/minicpmo_4_5/soft_interrupt_16k.wav
+          - tests/e2e/online_serving/test_minicpmo_4_5_duplex.py
+          - tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
+          - tests/e2e/online_serving/test_minicpmo_realtime_duplex_drivers.py
+          - tests/e2e/online_serving/minicpmo_realtime_duplex_scenarios.py
+          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py
+          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py
+          - tests/helpers/minicpmo_4_5_duplex.py
+          - vllm_omni/deploy/minicpmo_4_5_duplex.yaml
+          - vllm_omni/experimental/fullduplex/
+          - vllm_omni/model_executor/models/minicpmo_4_5/
+          - vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py
+          - vllm_omni/core/sched/
+          - vllm_omni/outputs/multimodal_accumulation.py
+          - vllm_omni/engine/stage_init_utils.py
+          - vllm_omni/engine/async_omni_engine.py
+          - vllm_omni/entrypoints/cli/serve.py
+        timeout_in_minutes: 50
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex.py -m 'advanced_model and cuda' --run-level 'advanced_model'
+        mirror_hardwares: h100_2
+
       - label: "TTS · VoxCPM2 Test"
         source_file_dependencies:
           - tests/e2e/offline_inference/test_voxcpm2_tts.py

.buildkite/cuda/test-nightly.yml

@@ -15,7 +15,13 @@ steps:
       - label: ":full_moon: Omni · Function Test with H100"
         timeout_in_minutes: 120
         commands:
-          - pytest -sv tests/e2e --run-level "full_model" -m "full_model and H100 and omni" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py
+          - pytest -sv tests/e2e --run-level "full_model" -m "full_model and H100 and omni" --ignore=tests/e2e/accuracy --ignore=tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py --ignore=tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
+        mirror_hardwares: h100_2
+
+      - label: ":full_moon: Omni · MiniCPM-o 4.5 Duplex Test"
+        timeout_in_minutes: 50
+        commands:
+          - pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py -m "full_model and cuda and H100 and omni" --run-level "full_model"
         mirror_hardwares: h100_2
 
       - label: ":full_moon: Omni · Doc Test with L4"

.buildkite/cuda/test-ready.yml

@@ -210,6 +210,33 @@ steps:
                       path: /mnt/hf-cache
                       type: DirectoryOrCreate
 
+      - label: "Omni · MiniCPM-o 4.5 Duplex Test"
+        source_file_dependencies:
+          - tests/assets/minicpmo_4_5/response_required_16k.wav
+          - tests/assets/minicpmo_4_5/soft_interrupt_16k.wav
+          - tests/e2e/online_serving/test_minicpmo_4_5_duplex.py
+          - tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py
+          - tests/e2e/online_serving/test_minicpmo_realtime_duplex_drivers.py
+          - tests/e2e/online_serving/minicpmo_realtime_duplex_scenarios.py
+          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py
+          - tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py
+          - tests/helpers/minicpmo_4_5_duplex.py
+          - vllm_omni/deploy/minicpmo_4_5_duplex.yaml
+          - vllm_omni/experimental/fullduplex/
+          - vllm_omni/model_executor/models/minicpmo_4_5/
+          - vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py
+          - vllm_omni/core/sched/
+          - vllm_omni/outputs/multimodal_accumulation.py
+          - vllm_omni/engine/stage_init_utils.py
+          - vllm_omni/engine/async_omni_engine.py
+          - vllm_omni/entrypoints/cli/serve.py
+        commands:
+          - |
+            timeout 20m bash -c "
+              pytest -s -v tests/e2e/online_serving/test_minicpmo_4_5_duplex.py -m 'core_model and cuda' --run-level 'core_model'
+            "
+        mirror_hardwares: h100_2
+
       - label: "TTS · VoxCPM2 Test"
         source_file_dependencies:
           - tests/e2e/online_serving/test_voxcpm2_tts.py

examples/online_serving/minicpmo/realtime_web/app/static/playback_worklet.js

@@ -6,6 +6,7 @@ class FullDuplexPcmPlayback extends AudioWorkletProcessor {
     this.playedFrames = 0;
     this.underrunFrames = 0;
     this.drain = null;
+    this.terminalFadeFrames = 0;
     this.started = false;
     this.activeResponseId = null;
     this.initialBufferFrames = Math.round(sampleRate * 0.2);
@@ -31,6 +32,7 @@ class FullDuplexPcmPlayback extends AudioWorkletProcessor {
       }
     } else if (message.type === 'drain') {
       this.drain = { responseId: message.responseId || null };
+      this.terminalFadeFrames = Math.min(this.fadeFrames, this.bufferedFrames());
       if (!this.started && this.bufferedFrames() > 0) {
         this.bufferWaitFrames = 0;
         this.startPlayback();
@@ -42,6 +44,7 @@ class FullDuplexPcmPlayback extends AudioWorkletProcessor {
       this.playedFrames = 0;
       this.underrunFrames = 0;
       this.drain = null;
+      this.terminalFadeFrames = 0;
       this.started = false;
       this.activeResponseId = null;
       this.bufferWaitFrames = this.initialBufferFrames;
@@ -76,6 +79,7 @@ class FullDuplexPcmPlayback extends AudioWorkletProcessor {
     this.playedFrames = 0;
     this.underrunFrames = 0;
     this.drain = null;
+    this.terminalFadeFrames = 0;
     this.started = false;
     this.activeResponseId = null;
     this.bufferWaitFrames = this.initialBufferFrames;
@@ -116,6 +120,7 @@ class FullDuplexPcmPlayback extends AudioWorkletProcessor {
     while (target < output.length && this.queue.length > 0) {
       const pcm = this.queue[0];
       const count = Math.min(output.length - target, pcm.length - this.offset);
+      const remainingBeforeChunk = this.bufferedFrames();
       for (let index = 0; index < count; index += 1) {
         let sample = pcm[this.offset + index] / 32768;
         if (this.fadeInFrames > 0) {
@@ -123,6 +128,16 @@ class FullDuplexPcmPlayback extends AudioWorkletProcessor {
           sample *= elapsed / this.fadeFrames;
           this.fadeInFrames -= 1;
         }
+        const remainingFrames = remainingBeforeChunk - index;
+        if (
+          this.drain
+          && this.terminalFadeFrames > 0
+          && remainingFrames <= this.terminalFadeFrames
+        ) {
+          sample *= this.terminalFadeFrames === 1
+            ? 0
+            : (remainingFrames - 1) / (this.terminalFadeFrames - 1);
+        }
         output[target + index] = sample;
       }
       target += count;

tests/core/sched/test_omni_ar_scheduler_streaming.py

@@ -242,9 +242,52 @@ def test_explicit_streaming_payload_replaces_placeholder_prompt() -> None:
         "tts_token_ids": [10, 20],
         "meta": {"replace_streaming_prompt": True},
     }
+    update.model_intermediate_buffer = {
+        "ids": {"tts": [41, 42, 99]},
+        "meta": {"turn_eos_token_id": 99},
+    }
 
     sched._update_request_as_session(session, update)
 
     assert session.prompt_token_ids == [10, 20]
     assert session.additional_information == update.additional_information
+    assert session.model_intermediate_buffer == {
+        "ids": {"tts": [41, 42, 99]},
+        "meta": {"turn_eos_token_id": 99},
+    }
+    assert session.status == RequestStatus.WAITING
+
+
+def test_model_intermediate_streaming_payload_replaces_computed_prompt() -> None:
+    sched = _make_scheduler(stage_id=1)
+    sched.chunk_transfer_adapter = SimpleNamespace(
+        receives_chunks=False,
+        segment_finished_requests=set(),
+    )
+    session = _make_request()
+    session.status = RequestStatus.WAITING_FOR_STREAMING_REQ
+    session.prompt_token_ids = [0] * 59
+    session._all_token_ids.clear()
+    session._all_token_ids.extend(session.prompt_token_ids)
+    session.num_prompt_tokens = 59
+    session.num_computed_tokens = 59
+    update = _make_update([0] * 10)
+    update.additional_information = None
+    update.model_intermediate_buffer = {
+        "ids": {"tts": list(range(8))},
+        "hidden_states": {"tts": [[0.0]] * 8},
+        "meta": {
+            "replace_streaming_prompt": True,
+            "next_stage_prompt_len": 10,
+        },
+    }
+
+    sched._update_request_as_session(session, update)
+
+    assert session.prompt_token_ids == [0] * 10
+    assert list(session._all_token_ids) == [0] * 10
+    assert session.num_prompt_tokens == 10
+    assert session.num_computed_tokens == 0
+    assert session.additional_information is None
+    assert session.model_intermediate_buffer == update.model_intermediate_buffer
     assert session.status == RequestStatus.WAITING

tests/core/sched/test_omni_generation_scheduler_update_session.py

@@ -54,6 +54,18 @@ class _SchedulerStub(OmniGenerationScheduler):
         raise AssertionError("unexpected enqueue for skipped_waiting miss")
 
 
+class _AsyncChunkStopSchedulerStub(OmniGenerationScheduler):
+    def __init__(self) -> None:
+        self.num_waiting_for_streaming_input = 0
+        self.chunk_transfer_adapter = SimpleNamespace(receives_chunks=True)
+        self.enqueued: list[Request] = []
+        self.enqueued_statuses: list[RequestStatus] = []
+
+    def _enqueue_waiting_request(self, session: Request) -> None:
+        self.enqueued.append(session)
+        self.enqueued_statuses.append(session.status)
+
+
 def _make_request(**kwargs) -> Request:
     sp = SamplingParams(max_tokens=8)
     defaults = dict(
@@ -133,6 +145,21 @@ def test_resumable_generation_stop_marks_segment_boundary() -> None:
     assert output.is_segment_finished is True
 
 
+def test_async_chunk_resumable_stop_rearms_connector_polling() -> None:
+    sched = _AsyncChunkStopSchedulerStub()
+    session = _make_request(request_id="req-async-chunk-next-segment")
+    session.status = RequestStatus.FINISHED_STOPPED
+    session.resumable = True
+
+    finished = sched._handle_stopped_request(session)
+
+    assert finished is False
+    assert sched.enqueued == [session]
+    assert sched.enqueued_statuses == [RequestStatus.WAITING]
+    assert session.status == RequestStatus.WAITING
+    assert sched.num_waiting_for_streaming_input == 0
+
+
 class TestReplaceSessionWithStreamingUpdate:
     def test_resets_tokens_and_prompt_from_update(self) -> None:
         sched = _SchedulerStub()

tests/dfx/reliability/invalid_param_test/test_invalid_minicpmo_4_5_omni.py

@@ -22,10 +22,7 @@ _MINICPMO_SERVER = [
             model="openbmb/MiniCPM-o-4_5",
             stage_config_path=_MINICPMO_DEPLOY,
             use_stage_cli=True,
-            server_args=[
-                "--trust-remote-code",
-                "--no-async-chunk",
-            ],
+            server_args=["--trust-remote-code"],
         ),
         id="minicpmo_4_5",
     )

tests/distributed/omni_connectors/test_chunk_transfer_adapter.py

@@ -71,6 +71,30 @@ def test_streaming_payload_can_replace_placeholder_prompt(mocker: MockerFixture)
     request.update_block_hashes.assert_called_once_with()
 
 
+def test_streaming_payload_can_append_exact_prompt_length(mocker: MockerFixture) -> None:
+    request = SimpleNamespace(
+        _all_token_ids=[0, 0, 7, 8],
+        _output_token_ids=[7, 8],
+        prompt_token_ids=[0, 0],
+        num_computed_tokens=4,
+        num_prompt_tokens=2,
+        update_block_hashes=mocker.Mock(),
+    )
+    payload = {
+        "ids": {"prompt": [1, 2, 3]},
+        "meta": {"next_stage_prompt_len": 3},
+    }
+
+    construct_next_stage_streaming_input_prompt(payload, request)
+
+    assert request.prompt_token_ids == [0, 0, 7, 8, 0, 0, 0]
+    assert request._all_token_ids == [0, 0, 7, 8, 0, 0, 0]
+    assert request._output_token_ids == []
+    assert request.num_computed_tokens == 4
+    assert request.num_prompt_tokens == 7
+    request.update_block_hashes.assert_called_once_with()
+
+
 @pytest.fixture
 def build_adapter(monkeypatch, mocker: MockerFixture):
     def _build(
@@ -252,6 +276,37 @@ def test_save_async_uses_confirmed_tokens_for_async_scheduler_watermark(build_ad
     assert len(adapter._pending_save_reqs) == 1
 
 
+def test_send_single_request_terminal_chunk_still_flushes_processor(build_adapter, monkeypatch):
+    """A terminal stop is not a segment boundary (#5383), but the producer-side
+    processor must still receive the flush signal on the terminal chunk.
+    Passing only ``is_segment_finished`` starved processors of their final
+    accumulated payload once terminal stops stopped setting it (#5413: the
+    downstream stage got ``meta.finished`` with the tail data missing).
+    """
+    adapter, connector = build_adapter(stage_id=0)
+    request = _req("req-terminal", RequestStatus.FINISHED_STOPPED, external_req_id="ext-terminal")
+
+    seen_flush_flags = []
+
+    def recording_processor(**kwargs):
+        seen_flush_flags.append(kwargs["is_finished"])
+        return OmniPayloadStruct(
+            codes=CodesStruct(audio=torch.tensor([1, 2, 3], dtype=torch.long)),
+        )
+
+    adapter.custom_process_next_stage_input_func = recording_processor
+    monkeypatch.setattr(adapter, "cleanup", lambda *a, **kw: None)
+
+    adapter._send_single_request(
+        {"multimodal_output": None, "request": request, "is_finished": True, "is_segment_finished": False}
+    )
+
+    assert seen_flush_flags == [True]
+    sent_payload = connector.put.call_args.kwargs["data"]
+    assert bool(sent_payload.meta.finished.item()) is True
+    assert bool(sent_payload.meta.is_segment_finished.item()) is False
+
+
 def test_send_single_request_struct_without_meta_does_not_crash(build_adapter, monkeypatch):
     """Producer may return a struct with ``meta=None`` (e.g. payload that
     carries only ``embed`` or ``codes``). The sender's ``meta is not None``
@@ -314,6 +369,23 @@ def test_send_single_request_struct_preserves_segment_finished(build_adapter, mo
     assert sent_payload.meta.is_segment_finished.item() is True
 
 
+def test_send_single_request_respects_processor_receiver_boundary(build_adapter, monkeypatch):
+    adapter, connector = build_adapter(stage_id=1)
+    request = _req("req-stream", RequestStatus.WAITING, external_req_id="ext-stream")
+
+    adapter.custom_process_next_stage_input_func = lambda **kwargs: OmniPayloadStruct(
+        meta=MetaStruct(is_segment_finished=torch.tensor(False, dtype=torch.bool))
+    )
+    monkeypatch.setattr(adapter, "cleanup", lambda *a, **kw: None)
+
+    adapter._send_single_request(
+        {"multimodal_output": None, "request": request, "is_finished": False, "is_segment_finished": True}
+    )
+
+    sent_payload = connector.put.call_args.kwargs["data"]
+    assert sent_payload.meta.is_segment_finished.item() is False
+
+
 def test_save_async_skips_stale_resumable_chunk_until_dedup_is_reset(build_adapter):
     adapter, _ = build_adapter(stage_id=1)
     request = _req("req-stream", RequestStatus.WAITING, external_req_id="ext-stream")
@@ -463,7 +535,7 @@ def test_fifo_promotion(build_adapter):
     adapter.process_pending_chunks(waiting_queue, running_queue)
 
     assert list(adapter._active_streams) == ["req-1", "req-2"]
-    assert waiting_queue == reqs[2:]
+    assert waiting_queue == []
     assert reqs[0].status == RequestStatus.WAITING_FOR_CHUNK
     assert reqs[1].status == RequestStatus.WAITING_FOR_CHUNK
     assert reqs[2].status == RequestStatus.WAITING
@@ -473,6 +545,7 @@ def test_fifo_promotion(build_adapter):
     # (commit c4d95fd9 — otherwise the terminal chunk deadlocks at c=8 K=2).
     # Simulate it here so promotion can pick up the freed slot.
     adapter._evict_finished_active_streams({"req-1"})
+    adapter.restore_queues(waiting_queue, running_queue)
     adapter.process_pending_chunks(waiting_queue, running_queue)
 
     assert list(adapter._active_streams) == ["req-2", "req-3"]
@@ -481,6 +554,22 @@ def test_fifo_promotion(build_adapter):
     assert reqs[2].status == RequestStatus.WAITING_FOR_CHUNK
 
 
+def test_non_active_waiting_request_is_held_off_scheduler(build_adapter):
+    adapter, _ = build_adapter(stage_id=2, max_num_seqs=2, active_stream_window=1)
+    active = _req("req-active", RequestStatus.WAITING)
+    non_active = _req("req-non-active", RequestStatus.WAITING)
+    waiting_queue = DummyWaitingQueue([active, non_active])
+    running_queue = []
+
+    adapter.process_pending_chunks(waiting_queue, running_queue)
+
+    assert waiting_queue == []
+    assert active.status == RequestStatus.WAITING_FOR_CHUNK
+    assert non_active.status == RequestStatus.WAITING
+    assert list(adapter._pending_load_reqs) == [active]
+    assert list(adapter.waiting_for_chunk_waiting_requests) == [active, non_active]
+
+
 def test_legacy_k0(build_adapter):
     adapter, _ = build_adapter(stage_id=1, max_num_seqs=1, active_stream_window=0)
     waiting_req = _req("waiting", RequestStatus.WAITING)
@@ -509,12 +598,13 @@ def test_finished_releases_slot(build_adapter):
 
     adapter.process_pending_chunks(waiting_queue, running_queue)
     assert list(adapter._active_streams) == ["req-1"]
-    assert waiting_queue == [req_2]
+    assert waiting_queue == []
 
     adapter.finished_requests.add("req-1")
     # Eviction is deferred to postprocess_scheduler_output in the runtime path
     # (commit c4d95fd9). Simulate it so promotion can pick up the freed slot.
     adapter._evict_finished_active_streams({"req-1"})
+    adapter.restore_queues(waiting_queue, running_queue)
     adapter.process_pending_chunks(waiting_queue, running_queue)
 
     assert list(adapter._active_streams) == ["req-2"]

tests/e2e/offline_inference/test_minicpmo_4_5.py

@@ -83,7 +83,7 @@ def test_video_to_text(omni_runner, omni_runner_handler) -> None:
 @hardware_test(res={"cuda": "H100", "npu": "A2"}, num_cards=2)
 @pytest.mark.parametrize("omni_runner", test_params, indirect=True)
 def test_text_to_audio(omni_runner, omni_runner_handler) -> None:
-    """Test processing text, generating audio output through the talker token2wav path."""
+    """Test processing text and generating audio through Talker and Code2Wav."""
     request_config = {"prompts": get_question("text"), "modalities": ["audio"]}
     omni_runner_handler.send_omni_request(request_config)
 

tests/e2e/online_serving/minicpmo_realtime_duplex_scenarios.py

@@ -1174,10 +1174,13 @@ async def _send_listen_only_overlap_pair(
     before_second_listen = state.model_listen_count
     overlap_started_while_active = not state.response_done(first_response_id)
     model_unit_ready_while_active = False
+    second_unit_event_index: int | None = None
 
     def record_model_unit_ready() -> None:
-        nonlocal model_unit_ready_while_active
+        nonlocal model_unit_ready_while_active, second_unit_event_index
         model_unit_ready_while_active = not state.response_done(first_response_id)
+        if model_unit_ready_while_active:
+            second_unit_event_index = len(state.events)
 
     await _send_pcm16(
         ws,
@@ -1239,14 +1242,13 @@ async def _send_listen_only_overlap_pair(
             label=transcripts[1],
         )
 
-    second_speech_index = _nth_event_index(state, "input_audio_buffer.speech_started", 2)
     first_done_index = _event_index_for_response(state, "response.done", first_response_id)
     overlap_ok = (
         overlap_started_while_active
         and model_unit_ready_while_active
-        and second_speech_index is not None
+        and second_unit_event_index is not None
         and first_done_index is not None
-        and second_speech_index < first_done_index
+        and second_unit_event_index < first_done_index
     )
     return [first_response_id, second_response_id], ["speak", second_outcome], overlap_ok
 
@@ -1380,6 +1382,17 @@ async def run_demo(args: argparse.Namespace) -> dict[str, object]:
             await ws.send(json.dumps({"type": "session.close"}))
             await _wait_for(state, lambda: state.count("session.closed") > 0, timeout_s=20, label="session.closed")
         finally:
+            if state.count("session.created") > 0 and state.count("session.closed") == 0:
+                try:
+                    await ws.send(json.dumps({"type": "session.close"}))
+                    await _wait_for(
+                        state,
+                        lambda: state.count("session.closed") > 0,
+                        timeout_s=min(args.timeout_s, 5),
+                        label="session.closed during cleanup",
+                    )
+                except (ConnectionClosed, TimeoutError):
+                    pass
             stop.set()
             reader.cancel()
             try:

tests/e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py

@@ -19,7 +19,11 @@ SCRIPT_DIR = Path(__file__).resolve().parent
 if str(SCRIPT_DIR) not in sys.path:
     sys.path.insert(0, str(SCRIPT_DIR))
 
-from minicpmo_realtime_duplex_scenarios import _url_with_model, run_demo  # noqa: E402
+from minicpmo_realtime_duplex_scenarios import (  # noqa: E402
+    _ref_audio_data_url,
+    _url_with_model,
+    run_demo,
+)
 
 
 def _with_resume_mode(url: str) -> str:
@@ -114,7 +118,12 @@ async def _open_admission_session(
     args: argparse.Namespace,
     session_id: str,
 ) -> tuple[object, dict[str, object]]:
-    url = _url_with_model(args.url, args.model, session_id=session_id)
+    url = _url_with_model(
+        args.url,
+        args.model,
+        autostart=False if getattr(args, "ref_audio", None) else None,
+        session_id=session_id,
+    )
     ws = await websockets.connect(url, max_size=64 * 1024 * 1024)
     await ws.send(
         json.dumps(
@@ -125,6 +134,7 @@ async def _open_admission_session(
                     "model": args.model,
                     "modalities": ["audio", "text"],
                     "extra_body": {"minicpmo45_native_duplex": True},
+                    **({"ref_audio": _ref_audio_data_url(args.ref_audio)} if getattr(args, "ref_audio", None) else {}),
                 },
             }
         )
@@ -151,7 +161,12 @@ async def _admission_probe(args: argparse.Namespace, *, limit: int) -> dict[str,
             accepted.append(await _open_admission_session(args, f"{prefix}-accepted-{index}"))
 
         overflow_id = f"{prefix}-overflow"
-        overflow_url = _url_with_model(args.url, args.model, session_id=overflow_id)
+        overflow_url = _url_with_model(
+            args.url,
+            args.model,
+            autostart=False if getattr(args, "ref_audio", None) else None,
+            session_id=overflow_id,
+        )
         async with websockets.connect(overflow_url, max_size=64 * 1024 * 1024) as overflow:
             await overflow.send(
                 json.dumps(
@@ -162,6 +177,11 @@ async def _admission_probe(args: argparse.Namespace, *, limit: int) -> dict[str,
                             "model": args.model,
                             "modalities": ["audio", "text"],
                             "extra_body": {"minicpmo45_native_duplex": True},
+                            **(
+                                {"ref_audio": _ref_audio_data_url(args.ref_audio)}
+                                if getattr(args, "ref_audio", None)
+                                else {}
+                            ),
                         },
                     }
                 )
@@ -207,16 +227,27 @@ async def _resume_probe(
     session_id: str,
     expect_expired: bool = False,
 ) -> dict[str, object]:
-    url = _url_with_model(args.url, args.model, session_id=session_id)
+    url = _url_with_model(
+        args.url,
+        args.model,
+        autostart=False if getattr(args, "ref_audio", None) else None,
+        session_id=session_id,
+    )
     async with websockets.connect(url, max_size=64 * 1024 * 1024) as first:
         await first.send(
             json.dumps(
                 {
                     "type": "session.update",
                     "session": {
+                        "session_id": session_id,
                         "model": args.model,
                         "modalities": ["audio", "text"],
                         "extra_body": {"minicpmo45_native_duplex": True},
+                        **(
+                            {"ref_audio": _ref_audio_data_url(args.ref_audio)}
+                            if getattr(args, "ref_audio", None)
+                            else {}
+                        ),
                     },
                 }
             )
@@ -306,7 +337,12 @@ async def _takeover_probe(
     *,
     session_id: str,
 ) -> dict[str, object]:
-    url = _url_with_model(args.url, args.model, session_id=session_id)
+    url = _url_with_model(
+        args.url,
+        args.model,
+        autostart=False if getattr(args, "ref_audio", None) else None,
+        session_id=session_id,
+    )
     first = await websockets.connect(url, max_size=64 * 1024 * 1024)
     second = None
     try:
@@ -315,9 +351,15 @@ async def _takeover_probe(
                 {
                     "type": "session.update",
                     "session": {
+                        "session_id": session_id,
                         "model": args.model,
                         "modalities": ["audio", "text"],
                         "extra_body": {"minicpmo45_native_duplex": True},
+                        **(
+                            {"ref_audio": _ref_audio_data_url(args.ref_audio)}
+                            if getattr(args, "ref_audio", None)
+                            else {}
+                        ),
                     },
                 }
             )
@@ -451,20 +493,7 @@ async def run_multi_session(args: argparse.Namespace) -> dict[str, object]:
             args,
             session_id=f"takeover-{args.takeover_session_index}-{uuid.uuid4().hex}",
         )
-    expiry_result = None
-    if args.expire_session_index is not None:
-        if not 0 <= args.expire_session_index < args.sessions:
-            raise ValueError("--expire-session-index is outside the session range")
-        expiry_result = await _resume_probe(
-            args,
-            session_id=f"expire-{args.expire_session_index}-{uuid.uuid4().hex}",
-            expect_expired=True,
-        )
-    admission_result = (
-        await _admission_probe(args, limit=args.verify_admission_limit)
-        if args.verify_admission_limit is not None
-        else None
-    )
+    lifecycle_result = await run_lifecycle_probes(args)
 
     session_results = await asyncio.gather(
         *(run_demo(_demo_args(args, index)) for index in range(args.sessions)),
@@ -487,16 +516,15 @@ async def run_multi_session(args: argparse.Namespace) -> dict[str, object]:
             and semantic_isolation_ok
             and (resume_result is None or resume_result.get("ok") is True)
             and (takeover_result is None or takeover_result.get("ok") is True)
-            and (expiry_result is None or expiry_result.get("ok") is True)
-            and (admission_result is None or admission_result.get("ok") is True)
+            and lifecycle_result["ok"] is True
         ),
         "session_count": args.sessions,
         "identity_isolation_ok": identity_isolation_ok,
         "semantic_isolation_ok": semantic_isolation_ok,
         "resume": resume_result,
         "takeover": takeover_result,
-        "expiry": expiry_result,
-        "admission": admission_result,
+        "expiry": lifecycle_result["expiry"],
+        "admission": lifecycle_result["admission"],
         "failures": failures,
         "sessions": completed,
     }
@@ -507,6 +535,34 @@ async def run_multi_session(args: argparse.Namespace) -> dict[str, object]:
     return result
 
 
+async def run_lifecycle_probes(args: argparse.Namespace) -> dict[str, object]:
+    """Run expiry and admission probes without requiring model output."""
+    if args.sessions < 1:
+        raise ValueError("--sessions must be positive")
+    expiry_result = None
+    if args.expire_session_index is not None:
+        if not 0 <= args.expire_session_index < args.sessions:
+            raise ValueError("--expire-session-index is outside the session range")
+        expiry_result = await _resume_probe(
+            args,
+            session_id=f"expire-{args.expire_session_index}-{uuid.uuid4().hex}",
+            expect_expired=True,
+        )
+    admission_result = (
+        await _admission_probe(args, limit=args.verify_admission_limit)
+        if args.verify_admission_limit is not None
+        else None
+    )
+    return {
+        "ok": (
+            (expiry_result is None or expiry_result.get("ok") is True)
+            and (admission_result is None or admission_result.get("ok") is True)
+        ),
+        "expiry": expiry_result,
+        "admission": admission_result,
+    }
+
+
 def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser()
     parser.add_argument("--url", default="ws://127.0.0.1:8113/v1/realtime?duplex=1")

tests/e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py

@@ -2,8 +2,8 @@
 
 This E2E driver runs the public ``realtime_duplex_demo.py`` against a live
 duplex backend. Arbitrary audio defaults to the model-policy lifecycle contract.
-The stronger response-required mode binds the two-response contract to a known
-input checksum and expected second response.
+The stronger response-required mode binds the multi-response contract to a
+known input checksum and expected follow-up response.
 """
 
 from __future__ import annotations
@@ -39,6 +39,17 @@ def _validate_input_sha256(path: Path, expected: str) -> str:
     return actual
 
 
+def _input_duration_s(path: Path) -> float:
+    with wave.open(str(path), "rb") as wav_file:
+        return wav_file.getnframes() / wav_file.getframerate()
+
+
+def _client_process_timeout_s(input_wav: Path, protocol_timeout_s: float) -> float:
+    # The child first streams the WAV in real time, then can independently
+    # exhaust its post-commit and session-close protocol waits.
+    return _input_duration_s(input_wav) + 2 * protocol_timeout_s + 30.0
+
+
 def _read_json(path: Path) -> dict[str, object]:
     if not path.is_file():
         return {}
@@ -211,7 +222,7 @@ def summarize_artifacts(
     validation_mode: str,
     min_responses: int,
     min_audio_deltas_per_response: int,
-    expect_second_response_substring: str | None,
+    expect_followup_response_substring: str | None,
 ) -> dict[str, object]:
     if validation_mode not in {"model-policy", "response-required"}:
         raise ValueError(f"unsupported validation mode: {validation_mode}")
@@ -274,12 +285,20 @@ def summarize_artifacts(
     )
     final_listen_after_commit = commit_index is not None and any(index > commit_index for index in listen_indices)
     transcript = "".join(str(summary.get("transcript") or "") for summary in response_summaries)
-    second_response_transcript = (
-        str(response_summaries[1].get("transcript") or "") if len(response_summaries) >= 2 else ""
+    followup_response_transcripts = [
+        str(summary.get("transcript") or "")
+        for summary in response_summaries[1:]
+        if commit_index is not None
+        and isinstance(summary.get("created_index"), int)
+        and summary["created_index"] < commit_index
+    ]
+    followup_response_transcript_ok = any(
+        bool(_normalize_text(transcript)) for transcript in followup_response_transcripts
+    )
+    followup_response_transcript_expectation_ok = not expect_followup_response_substring or any(
+        _normalize_text(expect_followup_response_substring) in _normalize_text(transcript)
+        for transcript in followup_response_transcripts
     )
-    second_response_transcript_expectation_ok = not expect_second_response_substring or _normalize_text(
-        expect_second_response_substring
-    ) in _normalize_text(second_response_transcript)
     error_events = [
         event
         for event in events
@@ -310,9 +329,8 @@ def summarize_artifacts(
     mode_contract_ok = validation_mode == "model-policy" or (
         second_response_before_final_commit
         and listen_before_first_response
-        and listen_between_responses
         and listen_after_last_done
-        and second_response_transcript_expectation_ok
+        and followup_response_transcript_ok
     )
     ok = common_contract_ok and mode_contract_ok
     return {
@@ -337,9 +355,10 @@ def summarize_artifacts(
         "listen_after_response_before_commit": listen_after_response_before_commit,
         "final_listen_after_commit": final_listen_after_commit,
         "transcript": transcript,
-        "second_response_transcript": second_response_transcript,
-        "expect_second_response_substring": expect_second_response_substring,
-        "second_response_transcript_expectation_ok": second_response_transcript_expectation_ok,
+        "followup_response_transcripts": followup_response_transcripts,
+        "followup_response_transcript_ok": followup_response_transcript_ok,
+        "expect_followup_response_substring": expect_followup_response_substring,
+        "followup_response_transcript_expectation_ok": followup_response_transcript_expectation_ok,
         "error_count": len(error_events),
         "cancelled_count": cancelled_count,
         "compact_sequence": _compact_sequence(events),
@@ -388,7 +407,7 @@ async def run_soft_interrupt(args: argparse.Namespace) -> dict[str, object]:
     try:
         stdout_bytes, stderr_bytes = await asyncio.wait_for(
             process.communicate(),
-            timeout=args.timeout_s + 30.0,
+            timeout=_client_process_timeout_s(input_wav, args.timeout_s),
         )
     except TimeoutError:
         process.kill()
@@ -406,7 +425,7 @@ async def run_soft_interrupt(args: argparse.Namespace) -> dict[str, object]:
         validation_mode=args.validation_mode,
         min_responses=args.min_responses,
         min_audio_deltas_per_response=args.min_audio_deltas_per_response,
-        expect_second_response_substring=args.expect_second_response_substring,
+        expect_followup_response_substring=args.expect_followup_response_substring,
     )
     summary.update(
         {
@@ -444,14 +463,12 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--min-responses", type=int, default=2)
     parser.add_argument("--min-audio-deltas-per-response", type=int, default=2)
     parser.add_argument("--input-sha256")
-    parser.add_argument("--expect-second-response-substring")
+    parser.add_argument("--expect-followup-response-substring")
     args = parser.parse_args()
     if args.validation_mode == "response-required" and args.min_responses < 2:
         parser.error("--min-responses must be at least 2")
     if args.validation_mode == "response-required" and not args.input_sha256:
         parser.error("--input-sha256 is required in response-required mode")
-    if args.validation_mode == "response-required" and not args.expect_second_response_substring:
-        parser.error("--expect-second-response-substring is required in response-required mode")
     if args.min_audio_deltas_per_response < 1:
         parser.error("--min-audio-deltas-per-response must be positive")
     return args

tests/e2e/online_serving/test_minicpmo_4_5.py

@@ -1,9 +1,7 @@
-"""
-E2E Online tests for MiniCPM-o 4.5 model with multimodal input and audio / text output.
+"""E2E online tests for MiniCPM-o 4.5 multimodal input and audio/text output.
 
-MiniCPM-o 4.5 has ``async_chunk: false``, ``max_num_seqs: 1`` on both stages,
-and the vocoder runs in-process inside the talker stage rather than as a separate
-Code2Wav stage.
+The batching deployment uses async chunk transfer across separate Thinker,
+Talker, and Code2Wav stages.
 """
 
 import os
@@ -26,10 +24,7 @@ test_params = [
             model=_MODEL,
             stage_config_path=_CI_DEPLOY,
             use_stage_cli=True,
-            server_args=[
-                "--trust-remote-code",
-                "--no-async-chunk",
-            ],
+            server_args=["--trust-remote-code"],
         ),
         id="default",
     )
@@ -99,7 +94,7 @@ def test_text_to_text_001(omni_server, openai_client) -> None:
 def test_text_to_audio_001(omni_server, openai_client) -> None:
     """
     Test text-only input generating text + audio output via OpenAI API.
-    This exercises the talker TTS region detection and in-process token2wav vocoder.
+    This exercises Talker TTS region detection and the Code2Wav stage.
     Deploy Setting: default 2GPU
     Input Modal: text
     Output Modal: text + audio
@@ -133,13 +128,14 @@ def test_audio_to_text_audio_001(omni_server, openai_client) -> None:
     messages = dummy_messages_from_mix_data(
         system_prompt=get_system_prompt(),
         audio_data_url=audio_data_url,
-        content_text=get_prompt("mix"),
+        content_text=get_prompt(),
     )
 
     request_config = {
         "model": omni_server.model,
         "messages": messages,
         "stream": True,
+        "key_words": {"text": ["Beijing"]},
     }
 
     openai_client.send_omni_request(request_config, request_num=get_max_batch_size())

tests/e2e/online_serving/test_minicpmo_4_5_duplex.py

@@ -0,0 +1,147 @@
+"""CI coverage for the MiniCPM-o 4.5 native-duplex Realtime API."""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import uuid
+from pathlib import Path
+
+import pytest
+import websockets
+
+from tests.e2e.online_serving.minicpmo_realtime_duplex_scenarios import (
+    _ref_audio_data_url,
+    run_demo,
+)
+from tests.e2e.online_serving.run_minicpmo_realtime_duplex_multi_session import (
+    run_multi_session,
+)
+from tests.helpers.mark import hardware_test
+from tests.helpers.minicpmo_4_5_duplex import (
+    CORE_SERVER_PARAMS,
+    SERVER_PARAMS,
+    demo_args,
+    multi_session_args,
+    realtime_url,
+    resolve_ref_audio,
+    validated_input_wav,
+)
+from vllm_omni.experimental.fullduplex.client import build_realtime_url
+
+pytestmark = pytest.mark.omni
+
+
+async def _receive_protocol_events(ws, required_types: set[str], *, timeout_s: float) -> list[dict[str, object]]:
+    async def receive() -> list[dict[str, object]]:
+        events: list[dict[str, object]] = []
+        seen: set[str] = set()
+        while not required_types.issubset(seen):
+            raw = await ws.recv()
+            if not isinstance(raw, str):
+                continue
+            event = json.loads(raw)
+            if not isinstance(event, dict):
+                continue
+            events.append(event)
+            event_type = event.get("type")
+            if event_type == "error":
+                raise AssertionError(f"WebSocket protocol smoke received an error: {event}")
+            if isinstance(event_type, str):
+                seen.add(event_type)
+        return events
+
+    return await asyncio.wait_for(receive(), timeout=timeout_s)
+
+
+async def _run_protocol_smoke(*, url: str, model: str, ref_audio: Path) -> list[dict[str, object]]:
+    session_id = f"duplex-ci-protocol-{uuid.uuid4().hex}"
+    websocket_url = build_realtime_url(url, model, autostart=False, session_id=session_id)
+    async with websockets.connect(websocket_url, max_size=64 * 1024 * 1024) as ws:
+        await ws.send(
+            json.dumps(
+                {
+                    "type": "session.update",
+                    "session": {
+                        "session_id": session_id,
+                        "model": model,
+                        "modalities": ["audio", "text"],
+                        "ref_audio": _ref_audio_data_url(str(ref_audio)),
+                        "extra_body": {"minicpmo45_native_duplex": True},
+                    },
+                }
+            )
+        )
+        events = await _receive_protocol_events(
+            ws,
+            {"session.created", "session.updated"},
+            timeout_s=60,
+        )
+        await ws.send(json.dumps({"type": "session.close"}))
+        events.extend(await _receive_protocol_events(ws, {"session.closed"}, timeout_s=60))
+    return events
+
+
+@pytest.mark.core_model
+@hardware_test(res={"cuda": "H100"}, num_cards=2)
+@pytest.mark.parametrize("omni_server", CORE_SERVER_PARAMS, indirect=True)
+def test_duplex_websocket_protocol_smoke(omni_server, model_prefix: str) -> None:
+    ref_audio = resolve_ref_audio(model_prefix)
+    events = asyncio.run(
+        _run_protocol_smoke(
+            url=realtime_url(omni_server),
+            model=omni_server.model,
+            ref_audio=ref_audio,
+        )
+    )
+    event_types = [event.get("type") for event in events]
+    assert "session.created" in event_types
+    assert "session.updated" in event_types
+    assert event_types[-1] == "session.closed"
+
+
+@pytest.mark.advanced_model
+@hardware_test(res={"cuda": "H100"}, num_cards=2)
+@pytest.mark.parametrize("omni_server", SERVER_PARAMS, indirect=True)
+def test_duplex_single_session_response_required(omni_server, model_prefix: str, tmp_path: Path) -> None:
+    result = asyncio.run(
+        run_demo(
+            demo_args(
+                omni_server=omni_server,
+                input_wav=validated_input_wav(),
+                ref_audio=resolve_ref_audio(model_prefix),
+                output_dir=tmp_path / "single_session",
+            )
+        )
+    )
+    assert result["ok"] is True
+    assert result["audio_delta_count"] > 0
+    assert result["done_count"] == 1
+    assert result["error_count"] == 0
+    assert result["all_audio_responses_have_transcript"] is True
+    assert result["transcript_delta_done_ok"] is True
+
+
+@pytest.mark.advanced_model
+@hardware_test(res={"cuda": "H100"}, num_cards=2)
+@pytest.mark.parametrize("omni_server", SERVER_PARAMS, indirect=True)
+def test_duplex_two_sessions_resume_and_takeover(omni_server, model_prefix: str, tmp_path: Path) -> None:
+    result = asyncio.run(
+        run_multi_session(
+            multi_session_args(
+                omni_server=omni_server,
+                input_wav=validated_input_wav(),
+                ref_audio=resolve_ref_audio(model_prefix),
+                output_dir=tmp_path / "multi_session",
+                response_required=True,
+            )
+        )
+    )
+    assert result["ok"] is True
+    assert result["session_count"] == 2
+    assert result["resume"]["ok"] is True
+    assert result["takeover"]["ok"] is True
+    assert not result["failures"]
+    assert all(session["audio_delta_count"] > 0 for session in result["sessions"])
+    assert all(session["done_count"] == 1 for session in result["sessions"])
+    assert all(session["error_count"] == 0 for session in result["sessions"])

tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py

@@ -0,0 +1,86 @@
+"""Nightly lifecycle coverage for the MiniCPM-o 4.5 native-duplex API."""
+
+from __future__ import annotations
+
+import asyncio
+import json
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+from tests.e2e.online_serving.run_minicpmo_realtime_duplex_multi_session import (
+    run_lifecycle_probes,
+)
+from tests.e2e.online_serving.run_minicpmo_realtime_duplex_soft_interrupt import (
+    run_soft_interrupt,
+)
+from tests.helpers.mark import hardware_test
+from tests.helpers.minicpmo_4_5_duplex import (
+    SERVER_PARAMS,
+    SOFT_INTERRUPT_SHA256,
+    multi_session_args,
+    realtime_url,
+    resolve_ref_audio,
+    validated_input_wav,
+    validated_soft_interrupt_wav,
+)
+
+pytestmark = [pytest.mark.full_model, pytest.mark.omni]
+
+
+@hardware_test(res={"cuda": "H100"}, num_cards=2)
+@pytest.mark.parametrize("omni_server", SERVER_PARAMS, indirect=True)
+def test_duplex_admission_and_expiry_reaper(omni_server, model_prefix: str, tmp_path: Path) -> None:
+    args = multi_session_args(
+        omni_server=omni_server,
+        input_wav=validated_input_wav(),
+        ref_audio=resolve_ref_audio(model_prefix),
+        output_dir=tmp_path / "admission_expiry",
+        response_required=False,
+    )
+    args.sessions = 1
+    args.disconnect_session_index = None
+    args.takeover_session_index = None
+    args.expire_session_index = 0
+    args.verify_admission_limit = 2
+    result = asyncio.run(run_lifecycle_probes(args))
+
+    assert result["ok"] is True, json.dumps(result, ensure_ascii=False, indent=2)
+    assert result["expiry"]["ok"] is True
+    assert result["expiry"]["error_code"] == "session_resume_expired"
+    assert result["admission"]["ok"] is True
+    assert result["admission"]["overflow_error_code"] == "resource_exhausted"
+
+
+@hardware_test(res={"cuda": "H100"}, num_cards=2)
+@pytest.mark.parametrize("omni_server", SERVER_PARAMS, indirect=True)
+def test_duplex_soft_interrupt(omni_server, model_prefix: str, tmp_path: Path) -> None:
+    input_wav = validated_soft_interrupt_wav()
+    result = asyncio.run(
+        run_soft_interrupt(
+            SimpleNamespace(
+                url=realtime_url(omni_server),
+                model=omni_server.model,
+                input_wav=str(input_wav),
+                ref_audio=str(resolve_ref_audio(model_prefix)),
+                output_dir=str(tmp_path / "soft_interrupt"),
+                summary_output=None,
+                chunk_ms=200,
+                timeout_s=180.0,
+                require_audio=True,
+                no_realtime_pacing=False,
+                validation_mode="response-required",
+                min_responses=2,
+                min_audio_deltas_per_response=2,
+                input_sha256=SOFT_INTERRUPT_SHA256,
+                expect_followup_response_substring=None,
+            )
+        )
+    )
+
+    assert result["ok"] is True, json.dumps(result, ensure_ascii=False, indent=2)
+    assert result["error_count"] == 0
+    assert result["response_lifecycle_ok"] is True
+    assert result["response_audio_contract_ok"] is True
+    assert result["followup_response_transcript_ok"] is True

tests/e2e/online_serving/test_minicpmo_4_5_expansion.py

@@ -1,7 +1,7 @@
-"""
-E2E Online expansion tests for MiniCPM-o 4.5 covering modality combinations,
-in-process token2wav vocoder behavior, max_num_seqs=1 serial execution,
-and edge-case validation.
+"""E2E online expansion tests for MiniCPM-o 4.5.
+
+These cover modality combinations, separate Code2Wav behavior, sequential
+request isolation, longer audio output, and Chinese speech.
 """
 
 import os
@@ -94,10 +94,8 @@ def test_text_video_to_text_001(omni_server, openai_client) -> None:
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_sequential_requests_independent(omni_server, openai_client) -> None:
     """
-    Verify that sequential requests produce independent results and state
-    does not leak between requests. MiniCPM-o 4.5 has max_num_seqs=1 so
-    requests are processed one-at-a-time; the second request must not
-    receive the first request's audio or token state.
+    Verify that sequential requests produce independent results and that the
+    second request does not receive the first request's audio or token state.
     Deploy Setting: default 2GPU
     Input Modal: text (two different prompts)
     Output Modal: text + audio (both)
@@ -138,7 +136,7 @@ def test_sequential_requests_independent(omni_server, openai_client) -> None:
 def test_text_to_audio_long_output_001(omni_server, openai_client) -> None:
     """
     Test text input generating a longer audio output to exercise the
-    token2wav decoder across multiple frames.
+    Code2Wav stage across multiple frames.
     Deploy Setting: default 2GPU
     Input Modal: text (longer prompt)
     Output Modal: text + audio
@@ -146,14 +144,14 @@ def test_text_to_audio_long_output_001(omni_server, openai_client) -> None:
     """
     messages = dummy_messages_from_mix_data(
         system_prompt=get_system_prompt(),
-        content_text="Tell me a short story about a cat in about 50 words.",
+        content_text="What is the capital of China? Answer in 40 words.",
     )
 
     request_config = {
         "model": omni_server.model,
         "messages": messages,
         "stream": True,
-        "key_words": {"audio": ["cat"]},
+        "key_words": {"audio": ["Beijing"]},
     }
 
     openai_client.send_omni_request(request_config, request_num=get_max_batch_size())

tests/e2e/online_serving/test_minicpmo_realtime_duplex_drivers.py

@@ -12,14 +12,11 @@ import pytest
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
-DEMO_PATH = Path(__file__).resolve().parents[1] / "e2e/online_serving/minicpmo_realtime_duplex_scenarios.py"
-MULTI_DEMO_PATH = (
-    Path(__file__).resolve().parents[1] / "e2e/online_serving/run_minicpmo_realtime_duplex_multi_session.py"
-)
-PAIR_DEMO_PATH = Path(__file__).resolve().parents[1] / "e2e/online_serving/run_minicpmo_realtime_duplex_demo_pair.py"
-SOFT_INTERRUPT_DEMO_PATH = (
-    Path(__file__).resolve().parents[1] / "e2e/online_serving/run_minicpmo_realtime_duplex_soft_interrupt.py"
-)
+DRIVER_DIR = Path(__file__).resolve().parent
+DEMO_PATH = DRIVER_DIR / "minicpmo_realtime_duplex_scenarios.py"
+MULTI_DEMO_PATH = DRIVER_DIR / "run_minicpmo_realtime_duplex_multi_session.py"
+PAIR_DEMO_PATH = DRIVER_DIR / "run_minicpmo_realtime_duplex_demo_pair.py"
+SOFT_INTERRUPT_DEMO_PATH = DRIVER_DIR / "run_minicpmo_realtime_duplex_soft_interrupt.py"
 
 
 def _load_demo_module():
@@ -338,14 +335,7 @@ def test_realtime_duplex_soft_interrupt_accepts_explicit_ref_audio(monkeypatch):
     assert args.validation_mode == "model-policy"
 
 
-@pytest.mark.parametrize(
-    "extra_args",
-    [
-        ["--expect-second-response-substring", "一加一等于二"],
-        ["--input-sha256", "a" * 64],
-    ],
-)
-def test_realtime_duplex_soft_interrupt_response_required_needs_bound_fixture(monkeypatch, extra_args):
+def test_realtime_duplex_soft_interrupt_response_required_needs_bound_fixture(monkeypatch):
     demo = _load_soft_interrupt_demo_module()
     monkeypatch.setattr(
         demo.sys,
@@ -360,7 +350,6 @@ def test_realtime_duplex_soft_interrupt_response_required_needs_bound_fixture(mo
             "ref.wav",
             "--validation-mode",
             "response-required",
-            *extra_args,
         ],
     )
 
@@ -412,7 +401,6 @@ def test_realtime_duplex_soft_interrupt_accepts_multi_delta_handoff_sequence(tmp
             "_client_received_at_s": 2.6,
         },
         {"type": "response.done", "response_id": first_response_id, "_client_received_at_s": 3.0},
-        {"type": "response.listen", "_client_received_at_s": 3.2},
         {"type": "response.created", "response": {"id": second_response_id}, "_client_received_at_s": 4.0},
         {"type": "response.speak", "response_id": second_response_id, "_client_received_at_s": 4.0},
         {
@@ -458,10 +446,11 @@ def test_realtime_duplex_soft_interrupt_accepts_multi_delta_handoff_sequence(tmp
         validation_mode="response-required",
         min_responses=2,
         min_audio_deltas_per_response=2,
-        expect_second_response_substring="一加一等于二",
+        expect_followup_response_substring="一加一等于二",
     )
 
     assert summary["ok"] is True
+    assert summary["listen_between_responses"] is False
     assert summary["second_response_before_final_commit"] is True
     assert summary["final_listen_after_commit"] is True
     assert summary["response_audio_contract_ok"] is True
@@ -498,7 +487,7 @@ def test_realtime_duplex_soft_interrupt_model_policy_accepts_single_response(tmp
         validation_mode="model-policy",
         min_responses=2,
         min_audio_deltas_per_response=2,
-        expect_second_response_substring=None,
+        expect_followup_response_substring=None,
     )
 
     assert summary["ok"] is True
@@ -537,14 +526,14 @@ def test_realtime_duplex_soft_interrupt_response_required_rejects_single_respons
         validation_mode="response-required",
         min_responses=2,
         min_audio_deltas_per_response=2,
-        expect_second_response_substring="一加一等于二",
+        expect_followup_response_substring="一加一等于二",
     )
 
     assert summary["ok"] is False
     assert summary["enough_responses"] is False
 
 
-def test_realtime_duplex_soft_interrupt_checks_expected_text_in_second_response(tmp_path):
+def test_realtime_duplex_soft_interrupt_reports_text_expectation_without_gating(tmp_path):
     demo = _load_soft_interrupt_demo_module()
     output = tmp_path / "second_response_text"
     output.mkdir()
@@ -611,11 +600,32 @@ def test_realtime_duplex_soft_interrupt_checks_expected_text_in_second_response(
         validation_mode="response-required",
         min_responses=2,
         min_audio_deltas_per_response=2,
-        expect_second_response_substring="一加一等于二",
+        expect_followup_response_substring="一加一等于二",
     )
 
-    assert summary["ok"] is False
-    assert summary["second_response_transcript_expectation_ok"] is False
+    assert summary["ok"] is True
+    assert summary["followup_response_transcript_ok"] is True
+    assert summary["followup_response_transcript_expectation_ok"] is False
+
+    events = [
+        event
+        for event in events
+        if event.get("response_id") != second_response_id or event["type"] != "response.audio_transcript.delta"
+    ]
+    (output / "events.jsonl").write_text(
+        "".join(demo.json.dumps(event) + "\n" for event in events),
+        encoding="utf-8",
+    )
+    missing_transcript_summary = demo.summarize_artifacts(
+        output_dir=output,
+        validation_mode="response-required",
+        min_responses=2,
+        min_audio_deltas_per_response=2,
+        expect_followup_response_substring="一加一等于二",
+    )
+
+    assert missing_transcript_summary["ok"] is False
+    assert missing_transcript_summary["followup_response_transcript_ok"] is False
 
 
 def test_realtime_duplex_multi_session_resume_url_disables_autostart():
@@ -1523,7 +1533,7 @@ def test_realtime_duplex_demo_waits_at_each_model_unit_and_stops_after_speak():
     assert len(ws.messages) == 10
 
 
-def test_realtime_duplex_demo_listen_only_overlap_sends_next_turn_before_first_done(monkeypatch):
+def test_realtime_duplex_demo_listen_only_overlap_accepts_silence_unit_before_first_done(monkeypatch):
     demo = _load_demo_module()
     state = demo.DemoState()
     send_calls = 0
@@ -1548,7 +1558,6 @@ def test_realtime_duplex_demo_listen_only_overlap_sends_next_turn_before_first_d
         del args
         assert not state.response_done("resp-first")
         kwargs["on_model_unit_ready"]()
-        state.add({"type": "input_audio_buffer.speech_started", "turn": 2})
         _add_response_transcript(state, "resp-first", transcript="第一轮完成", audio=False)
         state.add({"type": "response.created", "response": {"id": "resp-second"}})
         _add_response_transcript(state, "resp-second", transcript="第二轮完成", audio=True)
@@ -1568,7 +1577,7 @@ def test_realtime_duplex_demo_listen_only_overlap_sends_next_turn_before_first_d
             ws,
             state,
             b"\x00\x00",
-            b"\x01\x00",
+            b"\x00\x00",
             transcripts=("first", "second"),
             durations_ms=(None, None),
             chunk_ms=200,

tests/engine/test_multimodal_accumulation.py

@@ -18,6 +18,7 @@ def test_chunk_accumulation_policy_replaces_snapshots_and_drains_delta_state():
     accumulated = MultimodalPayload.from_dict(
         {
             "audio": torch.tensor([1.0]),
+            "meta.segment_end": torch.tensor([0]),
             "meta.tts_is_last_chunk": torch.tensor([0]),
             "meta.turn_end": torch.tensor([0]),
             "meta.stable_request_value": "keep",
@@ -26,6 +27,7 @@ def test_chunk_accumulation_policy_replaces_snapshots_and_drains_delta_state():
     incoming = MultimodalPayload.from_dict(
         {
             "audio": torch.tensor([2.0]),
+            "meta.segment_end": torch.tensor([1]),
             "meta.tts_is_last_chunk": torch.tensor([1]),
             "meta.turn_end": torch.tensor([1]),
         }
@@ -41,6 +43,7 @@ def test_chunk_accumulation_policy_replaces_snapshots_and_drains_delta_state():
     drain_delta_payload(merged)
 
     assert "audio" not in merged
+    assert "meta.segment_end" not in merged
     assert "meta.tts_is_last_chunk" not in merged
     assert "meta.turn_end" not in merged
     assert merged.metadata["meta.stable_request_value"] == "keep"

tests/engine/test_orchestrator.py

@@ -1813,6 +1813,56 @@ async def test_stage_pool_submit_update_refreshes_output_processor_state() -> No
 
 
 @pytest.mark.asyncio
+async def test_handle_streaming_update_unknown_request_is_dropped() -> None:
+    """A streaming update for an untracked request must not resurrect it.
+
+    Updates always follow the first-chunk add_request through the same
+    ordered submission queue, so an unknown id can only mean the request
+    already finished or was aborted (e.g. the client disconnected).
+    Falling back to add_request created headless sessions that kept
+    cycling through the stages (issue #4271).
+    """
+
+    class RecordingPool:
+        def __init__(self) -> None:
+            self.calls: list[tuple[str, Any]] = []
+
+        async def submit_update(self, request_id, req_state, request, *, prompt_text=None) -> int:
+            self.calls.append((request_id, prompt_text))
+            return 0
+
+    pool = RecordingPool()
+    orchestrator = object.__new__(Orchestrator)
+    orchestrator.async_chunk = False
+    orchestrator.request_states = {}
+    orchestrator.stage_pools = [pool]
+    add_request_calls: list[str] = []
+
+    async def record_add_request(msg) -> None:
+        add_request_calls.append(msg.request_id)
+
+    orchestrator._handle_add_request = record_add_request
+
+    await orchestrator._handle_streaming_update(
+        StageSubmissionMessage(
+            type="streaming_update",
+            request_id="req-unknown",
+            prompt=SimpleNamespace(request_id="req-unknown", prompt_token_ids=[1], resumable=True),
+            original_prompt={"prompt": "segment-2"},
+            output_prompt_text="segment-2",
+            sampling_params_list=[_sampling_params()],
+            final_stage_id=0,
+            preprocess_ms=0.0,
+            request_timestamp=time.time(),
+            enqueue_ts=time.perf_counter(),
+        )
+    )
+
+    assert pool.calls == []
+    assert add_request_calls == []
+    assert "req-unknown" not in orchestrator.request_states
+
+
 async def test_handle_streaming_update_passes_prompt_text_to_stage_pool() -> None:
     class RecordingPool:
         def __init__(self) -> None:

tests/engine/test_orchestrator_stage_input_bridge.py

@@ -7,15 +7,24 @@ import asyncio
 import queue
 from types import SimpleNamespace
 from typing import Any
-from unittest.mock import AsyncMock
+from unittest.mock import ANY, AsyncMock, MagicMock
 
 import janus
 import pytest
 from vllm.outputs import CompletionOutput, RequestOutput
 from vllm.sampling_params import SamplingParams
 
-from vllm_omni.engine.orchestrator import Orchestrator, OrchestratorRequestState
+from vllm_omni.engine.orchestrator import (
+    Orchestrator,
+    OrchestratorRequestState,
+    _OrchestratorDuplexStagePort,
+)
 from vllm_omni.engine.stage_pool import StagePool
+from vllm_omni.experimental.fullduplex.engine.contracts import (
+    DuplexStageRequestContext,
+    DuplexStageSubmission,
+)
+from vllm_omni.experimental.fullduplex.engine.messages import DuplexFence
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
@@ -107,11 +116,50 @@ class FakePrewarmPool:
 
     async def submit_initial(self, _request_id, _req_state, request, prompt_text=None):
         self.submitted.append(request)
+        return 0
 
     def get_bound_replica_id(self, _request_id):
         return 0
 
 
+def _duplex_stage_port_submission():
+    stage_pools = []
+    for stage_id in range(3):
+        pool = SimpleNamespace(
+            stage_client=SimpleNamespace(default_sampling_params=SamplingParams(max_tokens=1)),
+            stage_vllm_config=SimpleNamespace(model_config=SimpleNamespace(max_model_len=64)),
+            submit_initial=AsyncMock(return_value=stage_id + 10),
+            submit_update=AsyncMock(return_value=stage_id + 20),
+        )
+        stage_pools.append(pool)
+    request_states: dict[str, OrchestratorRequestState] = {}
+    prewarm = AsyncMock()
+    port = _OrchestratorDuplexStagePort(
+        stage_pools=stage_pools,
+        request_states=request_states,
+        running_counter=None,
+        cleanup_request_ids=AsyncMock(),
+        async_chunk=True,
+        prewarm_async_chunk_stages=prewarm,
+    )
+    context = DuplexStageRequestContext(
+        request_id="req-duplex",
+        session_id="session-duplex",
+        fence=DuplexFence("session-duplex"),
+        stage_id=0,
+        final_stage_id=2,
+        config_generation=0,
+        sampling_params=tuple(SamplingParams(max_tokens=1) for _ in range(3)),
+    )
+    port.ensure_request(context)
+    submission = DuplexStageSubmission(
+        context=context,
+        prompt={"prompt_token_ids": [1, 2]},
+        already_submitted=False,
+    )
+    return port, stage_pools, request_states, prewarm, submission
+
+
 def _request_output(request_id: str) -> RequestOutput:
     completion = CompletionOutput(
         index=0,
@@ -199,11 +247,13 @@ async def test_async_prewarm_skips_outgoing_only_stage() -> None:
     stage2 = FakePrewarmPool("receiver")
     orchestrator.stage_pools = [stage0, stage1, stage2]
     orchestrator._emit_tx_edge = lambda **_kwargs: None
+    orchestrator._record_duplex_stage_submission = MagicMock()
     req_state = OrchestratorRequestState(
         request_id="req-prewarm",
         prompt={"prompt_token_ids": [1, 2]},
         sampling_params_list=[SamplingParams(max_tokens=1) for _ in range(3)],
         final_stage_id=2,
+        duplex_identity=SimpleNamespace(),
     )
 
     await orchestrator._prewarm_async_chunk_stages(
@@ -216,6 +266,23 @@ async def test_async_prewarm_skips_outgoing_only_stage() -> None:
     assert len(stage2.submitted) == 1
     assert 1 not in req_state.stage_submit_ts
     assert 2 in req_state.stage_submit_ts
+    orchestrator._record_duplex_stage_submission.assert_called_once_with(
+        2,
+        "req-prewarm",
+        0,
+        req_state,
+    )
+
+
+@pytest.mark.asyncio
+async def test_duplex_prewarm_runs_after_first_stage0_submission() -> None:
+    port, stage_pools, request_states, prewarm, submission = _duplex_stage_port_submission()
+
+    result = await port.submit(submission)
+
+    assert result.stage_id == 0
+    stage_pools[0].submit_initial.assert_awaited_once()
+    prewarm.assert_awaited_once_with("req-duplex", ANY, request_states["req-duplex"])
 
 
 @pytest.mark.asyncio

tests/engine/test_output_processor.py

@@ -14,7 +14,7 @@ from vllm.v1.engine import FinishReason
 from vllm.v1.engine.output_processor import OutputProcessor as VLLMOutputProcessor
 
 from vllm_omni.outputs import output_processor
-from vllm_omni.outputs.output_modality import OutputModalityNames
+from vllm_omni.outputs.output_modality import OutputModality, OutputModalityNames
 from vllm_omni.outputs.output_processor import MultimodalOutputProcessor, OmniRequestState
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
@@ -452,6 +452,127 @@ def test_no_detokenizer_make_request_output():
     assert AUDIO in result.outputs[0].multimodal_output
 
 
+def test_no_detokenizer_process_outputs_returns_nonterminal_audio_chunk(monkeypatch):
+    """Generation-stage MM-only chunks must reach the orchestrator.
+
+    StagePool registers generation requests without a per-request collector,
+    so the processor return value is the only path from Stage2 to
+    Orchestrator._handle_processed_outputs().
+    """
+    state = _make_no_detok_state(RequestOutputKind.DELTA)
+    processor = object.__new__(MultimodalOutputProcessor)
+    processor.output_modality = OutputModality.AUDIO
+    processor.request_states = {"r": state}
+    upstream_result = SimpleNamespace(request_outputs=[], reqs_to_abort=[])
+    monkeypatch.setattr(
+        VLLMOutputProcessor,
+        "process_outputs",
+        lambda *_args, **_kwargs: upstream_result,
+    )
+    engine_output = SimpleNamespace(
+        request_id="r",
+        multimodal_output={
+            "model_outputs": torch.arange(8, dtype=torch.float32),
+            "sr": torch.tensor(24000, dtype=torch.int32),
+        },
+        output_type="audio",
+        pooling_output=None,
+        new_token_ids=[],
+        finish_reason=None,
+        stop_reason=None,
+        kv_transfer_params=None,
+        routed_experts=None,
+        num_cached_tokens=0,
+    )
+
+    result = processor.process_outputs([engine_output])
+
+    assert len(result.request_outputs) == 1
+    assert result.request_outputs[0].request_id == "r"
+    assert AUDIO in result.request_outputs[0].outputs[0].multimodal_output
+
+
+def _make_mm_only_output_processor(monkeypatch):
+    processor = object.__new__(MultimodalOutputProcessor)
+    processor.output_modality = OutputModality.AUDIO
+    processor.request_states = {"r": _make_no_detok_state(RequestOutputKind.DELTA)}
+    monkeypatch.setattr(
+        VLLMOutputProcessor,
+        "process_outputs",
+        lambda *_args, **_kwargs: SimpleNamespace(
+            request_outputs=[],
+            reqs_to_abort=[],
+        ),
+    )
+
+    def finish_request(req_state):
+        processor.request_states.pop(req_state.request_id, None)
+
+    processor._finish_request = finish_request
+    return processor
+
+
+def _audio_engine_output(*, is_segment_finished: bool, is_last_chunk: bool):
+    return SimpleNamespace(
+        request_id="r",
+        multimodal_output={
+            "model_outputs": torch.arange(8, dtype=torch.float32),
+            "meta.tts_is_last_chunk": torch.tensor([int(is_last_chunk)]),
+        },
+        output_type="audio",
+        pooling_output=None,
+        new_token_ids=[],
+        finish_reason=FinishReason.STOP,
+        stop_reason=None,
+        kv_transfer_params=None,
+        routed_experts=None,
+        num_cached_tokens=0,
+        is_segment_finished=is_segment_finished,
+    )
+
+
+def test_mm_only_segment_finish_retains_request_state_for_next_audio_chunk(monkeypatch):
+    processor = _make_mm_only_output_processor(monkeypatch)
+
+    first = processor.process_outputs([_audio_engine_output(is_segment_finished=True, is_last_chunk=False)])
+
+    assert "r" in processor.request_states
+    assert len(first.request_outputs) == 1
+    assert first.request_outputs[0].finished is False
+
+    second = processor.process_outputs([_audio_engine_output(is_segment_finished=False, is_last_chunk=True)])
+
+    assert len(second.request_outputs) == 1
+    assert second.request_outputs[0].finished is True
+    assert "r" not in processor.request_states
+
+
+def test_mm_only_nonfinal_audio_retains_state_without_segment_marker(monkeypatch):
+    processor = _make_mm_only_output_processor(monkeypatch)
+
+    first = processor.process_outputs([_audio_engine_output(is_segment_finished=False, is_last_chunk=False)])
+
+    assert "r" in processor.request_states
+    assert len(first.request_outputs) == 1
+    assert first.request_outputs[0].finished is False
+
+    second = processor.process_outputs([_audio_engine_output(is_segment_finished=False, is_last_chunk=True)])
+
+    assert len(second.request_outputs) == 1
+    assert second.request_outputs[0].finished is True
+    assert "r" not in processor.request_states
+
+
+def test_mm_only_terminal_finish_removes_request_state(monkeypatch):
+    processor = _make_mm_only_output_processor(monkeypatch)
+
+    result = processor.process_outputs([_audio_engine_output(is_segment_finished=False, is_last_chunk=True)])
+
+    assert len(result.request_outputs) == 1
+    assert result.request_outputs[0].finished is True
+    assert "r" not in processor.request_states
+
+
 def test_no_detokenizer_make_request_output_with_routed_experts():
     """make_request_output accepts the routed_experts arg that the multimodal
     output channel (_process_mm_only_outputs) passes positionally, and attaches

tests/engine/test_single_stage_mode.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import os
+import socket
 from contextlib import contextmanager
 from types import SimpleNamespace
 from typing import Any
@@ -11,6 +12,7 @@ import pytest
 from pytest_mock import MockerFixture
 from vllm.v1.engine.utils import EngineZmqAddresses
 
+from vllm_omni.config.stage_config import DuplexSessionRuntimeConfig
 from vllm_omni.engine.async_omni_engine import AsyncOmniEngine
 from vllm_omni.engine.stage_engine_core_client import StageEngineCoreClientBase
 from vllm_omni.engine.stage_engine_startup import (
@@ -189,6 +191,34 @@ class TestOmniMasterServerAllocation:
         output_port = int(alloc.output_bind_address.split(":")[-1])
         assert len({handshake_port, input_port, output_port}) == 3
 
+    def test_route_ports_remain_os_reserved_until_master_releases_them(self):
+        server = OmniMasterServer(master_address="127.0.0.1", master_port=15005, stage_ids=[0])
+        alloc = server.get_allocation(0)
+        ports = [
+            int(alloc.handshake_bind_address.rsplit(":", 1)[-1]),
+            int(alloc.input_bind_address.rsplit(":", 1)[-1]),
+            int(alloc.output_bind_address.rsplit(":", 1)[-1]),
+        ]
+
+        for port in ports:
+            with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as probe:
+                with pytest.raises(OSError):
+                    probe.bind(("127.0.0.1", port))
+
+        server.release_route_port_reservations(0, handshake=True, data=False)
+        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as probe:
+            probe.bind(("127.0.0.1", ports[0]))
+        for port in ports[1:]:
+            with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as probe:
+                with pytest.raises(OSError):
+                    probe.bind(("127.0.0.1", port))
+
+        server.stop()
+
+        for port in ports[1:]:
+            with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as probe:
+                probe.bind(("127.0.0.1", port))
+
     def test_get_zmq_addresses_returns_bind_addresses(self):
         server = OmniMasterServer(master_address="127.0.0.1", master_port=15002, stage_ids=[0])
         alloc = server.get_allocation(0)
@@ -397,6 +427,25 @@ class TestSingleStageModeDetection:
         assert engine.single_stage_mode is False
         assert engine._single_stage_id_filter is None
 
+    def test_stage_configs_path_loads_duplex_runtime_config(self, mocker: MockerFixture):
+        duplex_session = DuplexSessionRuntimeConfig(max_sessions=2)
+        mocker.patch(
+            "vllm_omni.engine.async_omni_engine.StageConfigFactory.get_pipeline_config",
+            return_value=None,
+        )
+        load_deploy_config = mocker.patch(
+            "vllm_omni.engine.async_omni_engine.load_deploy_config",
+            return_value=SimpleNamespace(duplex_session=duplex_session),
+        )
+
+        engine = self._make_engine_no_thread(
+            mocker,
+            stage_configs_path="/fake/duplex.yaml",
+        )
+
+        load_deploy_config.assert_called_once_with("/fake/duplex.yaml")
+        assert engine.duplex_session_config is duplex_session
+
     def test_single_stage_mode_without_stage_id_has_no_filter(self, mocker: MockerFixture):
         engine = self._make_engine_no_thread(
             mocker,

tests/entrypoints/openai_api/test_duplex_handler.py

@@ -2520,6 +2520,52 @@ def test_duplex_auto_response_tts_scheduler_eos_fallback():
     assert results[0]["abort_data_plane_request"] is True
 
 
+def test_duplex_auto_response_rearms_tts_boundary_for_next_segment_in_same_turn():
+    data_plane = _test_data_plane()
+    request_id = "duplex-sid-next-segment-same-turn-e0-stage0"
+    session = DuplexSession(
+        session_id="sid-next-segment-same-turn",
+        config=DuplexSessionConfig(extra_body={"auto_response": True}),
+    )
+    session.bind_response_turn(0)
+    data_plane.begin_request(request_id)
+
+    first_segment = _duplex_tts_output(
+        request_id=request_id,
+        samples=24000,
+        finished=False,
+        tts_is_last_chunk=True,
+        turn_id=0,
+    )
+    first_results = _project_data_plane(
+        data_plane,
+        {"data_plane_outputs": [first_segment]},
+        session=session,
+    )
+
+    assert len(first_results) == 1
+    assert first_results[0]["abort_data_plane_request"] is True
+
+    data_plane.begin_request(request_id)
+    second_segment = _duplex_tts_output(
+        request_id=request_id,
+        samples=48000,
+        finished=False,
+        text="hello again",
+        tts_is_last_chunk=True,
+        turn_id=0,
+    )
+    second_results = _project_data_plane(
+        data_plane,
+        {"data_plane_outputs": [second_segment]},
+        session=session,
+    )
+
+    assert len(second_results) == 1
+    assert second_results[0]["audio_data"] == "wav-24000"
+    assert second_results[0]["abort_data_plane_request"] is True
+
+
 def test_duplex_turn_end_is_not_swallowed_by_finished_segment_fallback():
     data_plane = _test_data_plane()
     request_id = "duplex-sid-turn-end-after-segment-e0-stage0"
@@ -6202,19 +6248,27 @@ async def test_minicpmo_native_auto_response_keeps_request_bound_for_segment_con
         config_timeout_s=0.1,
         idle_timeout_s=1,
     )
-    ws = TimedWebSocket()
+
+    def on_send(ws: TimedWebSocket, data: dict[str, Any]) -> None:
+        if data.get("type") != "session.created":
+            return
+        session = handler._registry.get("sid-native-auto-continuation")
+        assert session is not None
+        session.capabilities.chunk_period_ms = 50
+        ws.put(
+            {
+                "type": "input_audio_buffer.append",
+                "audio": _pcm_f32_b64(800),
+                "format": "pcm_f32le",
+                "sample_rate_hz": 16000,
+            }
+        )
+        ws.put({"type": "input_audio_buffer.commit", "final": True})
+
+    ws = TimedWebSocket(on_send=on_send)
     event = _native_session_create("sid-native-auto-continuation")
     event["session"]["extra_body"]["auto_response"] = True
     ws.put(event)
-    ws.put(
-        {
-            "type": "input_audio_buffer.append",
-            "audio": _pcm_f32_b64(16000),
-            "format": "pcm_f32le",
-            "sample_rate_hz": 16000,
-        }
-    )
-    ws.put({"type": "input_audio_buffer.commit", "final": True})
 
     await handler.handle_session(ws)
 
@@ -6225,6 +6279,112 @@ async def test_minicpmo_native_auto_response_keeps_request_bound_for_segment_con
     assert final is False
 
 
+@pytest.mark.asyncio
+async def test_minicpmo_native_auto_response_real_input_preempts_unsent_silence():
+    request_id = "duplex-sid-native-real-before-silence-e0-stage0"
+
+    class SequencedAppendEngine(FakeEngineClient):
+        def __init__(self) -> None:
+            control_result = {
+                "operation": "append",
+                "session_id": "sid-native-real-before-silence",
+                "ok": True,
+                "unsupported_count": 0,
+                "error_count": 0,
+                "stage_results": [
+                    {
+                        "result": {
+                            "data_plane_append": True,
+                            "request_id": request_id,
+                            "response_stage_id": 1,
+                        }
+                    }
+                ],
+            }
+            terminal_segment = _duplex_tts_output(
+                request_id=request_id,
+                samples=0,
+                finished=True,
+                tts_is_last_chunk=True,
+                token_ids=[151645],
+            )
+            super().__init__(
+                control_result=control_result,
+                collect_outputs=[[terminal_segment], [], []],
+                collect_delay_s=0.01,
+            )
+            self.append_sequence: list[str] = []
+            self.first_real_started = asyncio.Event()
+            self.second_real_started = asyncio.Event()
+            self.real_count = 0
+
+        async def append_duplex_input_async(self, session_id: str, **kwargs):
+            kwargs.pop("expected_epoch", None)
+            result = await super().append_duplex_input_async(session_id, **kwargs)
+            payload = kwargs.get("payload")
+            is_silence = (
+                isinstance(payload, dict)
+                and payload.get("audio") == OmniDuplexSessionHandler._NATIVE_SILENCE_UNIT_PAYLOAD_AUDIO
+            )
+            if is_silence:
+                self.append_sequence.append("silence")
+            else:
+                self.real_count += 1
+                self.append_sequence.append(f"real-{self.real_count}")
+                if self.real_count == 1:
+                    self.first_real_started.set()
+                elif self.real_count == 2:
+                    self.second_real_started.set()
+            return result
+
+    session_created = asyncio.Event()
+    engine = SequencedAppendEngine()
+    handler = OmniDuplexSessionHandler(
+        chat_service=FakeChatService(engine),
+        config_timeout_s=0.1,
+        idle_timeout_s=1,
+    )
+
+    def on_send(ws: TimedWebSocket, data: dict[str, Any]) -> None:
+        if data.get("type") != "session.created":
+            return
+        session = handler._registry.get("sid-native-real-before-silence")
+        assert session is not None
+        session.capabilities.chunk_period_ms = 50
+        ws.put(
+            {
+                "type": "input_audio_buffer.append",
+                "audio": _pcm_f32_b64(800),
+                "format": "pcm_f32le",
+                "sample_rate_hz": 16000,
+            }
+        )
+        session_created.set()
+
+    ws = TimedWebSocket(on_send=on_send, receive_timeout_s=2.0)
+    event = _native_session_create("sid-native-real-before-silence")
+    event["session"]["extra_body"]["auto_response"] = True
+    ws.put(event)
+    handler_task = asyncio.create_task(handler.handle_session(ws))
+
+    await asyncio.wait_for(session_created.wait(), timeout=1)
+    await asyncio.wait_for(engine.first_real_started.wait(), timeout=1)
+    await asyncio.sleep(0.02)
+    ws.put(
+        {
+            "type": "input_audio_buffer.append",
+            "audio": _pcm_f32_b64(800, value=0.07),
+            "format": "pcm_f32le",
+            "sample_rate_hz": 16000,
+        }
+    )
+    await asyncio.wait_for(engine.second_real_started.wait(), timeout=1)
+    ws.put({"type": "session.close"})
+    await asyncio.wait_for(handler_task, timeout=2)
+
+    assert engine.append_sequence[:2] == ["real-1", "real-2"]
+
+
 @pytest.mark.asyncio
 async def test_minicpmo_native_auto_response_real_input_waits_for_submitted_silence_tail():
     request_id = "duplex-sid-native-silence-tail-e0-stage0"
@@ -6294,18 +6454,26 @@ async def test_minicpmo_native_auto_response_real_input_waits_for_submitted_sile
         config_timeout_s=0.1,
         idle_timeout_s=1,
     )
-    ws = TimedWebSocket(receive_timeout_s=2.0)
+
+    def on_send(ws: TimedWebSocket, data: dict[str, Any]) -> None:
+        if data.get("type") != "session.created":
+            return
+        session = handler._registry.get("sid-native-silence-tail")
+        assert session is not None
+        session.capabilities.chunk_period_ms = 50
+        ws.put(
+            {
+                "type": "input_audio_buffer.append",
+                "audio": _pcm_f32_b64(800, value=0.05),
+                "format": "pcm_f32le",
+                "sample_rate_hz": 16000,
+            }
+        )
+
+    ws = TimedWebSocket(on_send=on_send, receive_timeout_s=2.0)
     event = _native_session_create("sid-native-silence-tail")
     event["session"]["extra_body"]["auto_response"] = True
     ws.put(event)
-    ws.put(
-        {
-            "type": "input_audio_buffer.append",
-            "audio": _pcm_f32_b64(16000, value=0.05),
-            "format": "pcm_f32le",
-            "sample_rate_hz": 16000,
-        }
-    )
     handler_task = asyncio.create_task(handler.handle_session(ws))
 
     await asyncio.wait_for(engine.silence_append_started.wait(), timeout=1)
@@ -6330,6 +6498,135 @@ async def test_minicpmo_native_auto_response_real_input_waits_for_submitted_sile
     assert engine.append_sequence == ["real-1", "silence-started", "silence-returned", "real-2"]
 
 
+@pytest.mark.asyncio
+async def test_minicpmo_native_auto_response_preserves_silence_continuations_across_append_tails():
+    request_id = "duplex-sid-native-continuation-tails-e0-stage0"
+
+    class SequencedAppendEngine(FakeEngineClient):
+        def __init__(self) -> None:
+            control_result = {
+                "operation": "append",
+                "session_id": "sid-native-continuation-tails",
+                "ok": True,
+                "unsupported_count": 0,
+                "error_count": 0,
+                "stage_results": [
+                    {
+                        "result": {
+                            "data_plane_append": True,
+                            "request_id": request_id,
+                            "response_stage_id": 1,
+                        }
+                    }
+                ],
+            }
+            super().__init__(control_result=control_result)
+            self.real_append_started = asyncio.Event()
+            self.release_real_append = asyncio.Event()
+            self.first_silence_started = asyncio.Event()
+            self.release_first_silence = asyncio.Event()
+            self.second_silence_started = asyncio.Event()
+            self.release_collect = asyncio.Event()
+            self.silence_count = 0
+
+        async def append_duplex_input_async(self, session_id: str, **kwargs):
+            kwargs.pop("expected_epoch", None)
+            payload = kwargs.get("payload")
+            assert isinstance(payload, dict)
+            result = await super().append_duplex_input_async(session_id, **kwargs)
+            if payload.get("audio") != OmniDuplexSessionHandler._NATIVE_SILENCE_UNIT_PAYLOAD_AUDIO:
+                self.real_append_started.set()
+                await self.release_real_append.wait()
+                return result
+
+            self.silence_count += 1
+            if self.silence_count == 1:
+                self.first_silence_started.set()
+                await self.release_first_silence.wait()
+            else:
+                self.second_silence_started.set()
+            return result
+
+        async def collect_duplex_data_plane_outputs_async(
+            self,
+            request_id: str,
+            *,
+            response_stage_id: int | None = None,
+            timeout: float | None = None,
+        ) -> list[object]:
+            del timeout
+            self.collected.append((request_id, response_stage_id))
+            await self.release_collect.wait()
+            return []
+
+    session_created = asyncio.Event()
+
+    def on_send(_ws: TimedWebSocket, data: dict[str, Any]) -> None:
+        if data.get("type") == "session.created":
+            session_created.set()
+
+    engine = SequencedAppendEngine()
+    handler = OmniDuplexSessionHandler(
+        chat_service=FakeChatService(engine),
+        config_timeout_s=0.1,
+        idle_timeout_s=1,
+    )
+    ws = TimedWebSocket(on_send=on_send, receive_timeout_s=2.0)
+    event = _native_session_create("sid-native-continuation-tails")
+    event["session"]["extra_body"]["auto_response"] = True
+    ws.put(event)
+    ws.put(
+        {
+            "type": "input_audio_buffer.append",
+            "audio": _pcm_f32_b64(16000, value=0.05),
+            "format": "pcm_f32le",
+            "sample_rate_hz": 16000,
+        }
+    )
+    handler_task = asyncio.create_task(handler.handle_session(ws))
+
+    await asyncio.wait_for(session_created.wait(), timeout=1)
+    await asyncio.wait_for(engine.real_append_started.wait(), timeout=1)
+    session = handler._registry.get("sid-native-continuation-tails")
+    assert session is not None
+    session.capabilities.chunk_period_ms = 50
+    session.bind_request(request_id)
+    response_id = session.begin_response(turn_id=session.turn_id)
+    native = handler._minicpmo_session_state(session)
+    scheduler = native.silence_continuation_scheduler
+    assert scheduler is not None
+
+    payload = handler._native_silence_unit_payload()
+    payload["duplex_turn_id"] = session.turn_id
+    kwargs = {
+        "request_id": request_id,
+        "owner_id": f"response:{response_id}",
+        "response_id": response_id,
+        "response_owned": True,
+        "expected_epoch": session.epoch,
+        "expected_incarnation": session.incarnation,
+        "expected_model_turn_id": None,
+        "send_json": ws.send_json,
+    }
+    assert await scheduler(payload, **kwargs) is True
+
+    next_schedule = asyncio.create_task(scheduler(payload, **kwargs))
+    await asyncio.sleep(0.05)
+    assert not next_schedule.done()
+
+    engine.release_real_append.set()
+    await asyncio.wait_for(engine.first_silence_started.wait(), timeout=1)
+    assert not next_schedule.done()
+
+    engine.release_first_silence.set()
+    assert await asyncio.wait_for(next_schedule, timeout=1) is True
+    await asyncio.wait_for(engine.second_silence_started.wait(), timeout=1)
+    ws.put({"type": "session.close"})
+    await asyncio.wait_for(handler_task, timeout=2)
+
+    assert engine.silence_count == 2
+
+
 @pytest.mark.asyncio
 async def test_minicpmo_native_duplex_uses_segment_text_metadata_for_transcript_cursor():
     request_id = "duplex-sid-native-repeat-text-e0-stage0"

tests/entrypoints/test_serve.py

@@ -222,6 +222,7 @@ def test_run_headless_llm_registers_with_auto_assigned_replica_id(mocker: Mocker
     from vllm_omni.engine.stage_engine_startup import StageRegistrationResponse
 
     stage_cfg = _make_stage_cfg(0, stage_type="llm")
+    stage_cfg.engine_args["async_chunk"] = True
     parallel_config = SimpleNamespace(
         data_parallel_size_local=1,
         data_parallel_rank=0,
@@ -241,7 +242,10 @@ def test_run_headless_llm_registers_with_auto_assigned_replica_id(mocker: Mocker
         "vllm_omni.distributed.omni_connectors.utils.initialization.resolve_omni_kv_config_for_stage",
         return_value=(None, None, None),
     )
-    mocker.patch("vllm_omni.engine.stage_init_utils.get_stage_connector_spec", return_value={})
+    mock_connector_spec = mocker.patch(
+        "vllm_omni.engine.stage_init_utils.get_stage_connector_spec",
+        return_value={},
+    )
     mocker.patch("vllm_omni.engine.stage_init_utils.build_engine_args_dict", return_value={})
     mocker.patch(
         "vllm_omni.engine.stage_init_utils.build_vllm_config",
@@ -277,6 +281,7 @@ def test_run_headless_llm_registers_with_auto_assigned_replica_id(mocker: Mocker
     assert kwargs["omni_stage_config"] is stage_cfg
     assert kwargs["replica_id"] is None
     assert "socket_ownership" not in kwargs
+    assert mock_connector_spec.call_args.kwargs["async_chunk"] is True
 
     assert mock_manager_cls.call_count == 1
     mgr_kwargs = mock_manager_cls.call_args.kwargs

tests/examples/test_minicpmo_realtime_web_static.py

@@ -207,3 +207,68 @@ def test_playback_worklet_waits_before_playing_and_rebuffers_after_underrun():
         capture_output=True,
         text=True,
     )
+
+
+def test_playback_worklet_fades_terminal_drain_to_zero():
+    node = shutil.which("node")
+    if node is None:
+        pytest.skip("node is required for the AudioWorklet regression test")
+
+    script = textwrap.dedent(
+        """
+        const fs = require('fs');
+        const vm = require('vm');
+
+        global.sampleRate = 1000;
+        const messages = [];
+        global.AudioWorkletProcessor = class {
+          constructor() {
+            this.port = {
+              onmessage: null,
+              postMessage: (message) => messages.push(message),
+            };
+          }
+        };
+        let Processor = null;
+        global.registerProcessor = (_name, processor) => { Processor = processor; };
+        vm.runInThisContext(fs.readFileSync(process.argv[1], 'utf8'));
+
+        const processor = new Processor();
+        const pcm = new Int16Array(100);
+        pcm.fill(16384);
+        processor.handleMessage({
+          type: 'audio',
+          pcm,
+          responseId: 'response-terminal',
+          initialBufferMs: 400,
+        });
+        processor.handleMessage({
+          type: 'drain',
+          responseId: 'response-terminal',
+        });
+
+        const output = new Float32Array(100);
+        processor.process([], [[output]]);
+        const assert = (condition, message) => {
+          if (!condition) throw new Error(message);
+        };
+
+        assert(output.some((sample) => sample !== 0), 'terminal audio must still play');
+        assert(output[output.length - 1] === 0, 'terminal drain must fade the final sample to zero');
+        const tail = Array.from(output.slice(-5), Math.abs);
+        assert(
+          tail.every((sample, index) => index === 0 || sample <= tail[index - 1]),
+          'terminal drain fade must decrease monotonically',
+        );
+        assert(
+          messages.filter((message) => message.type === 'playback-drained').length === 1,
+          'terminal drain must be reported exactly once after playback',
+        );
+        """
+    )
+    subprocess.run(
+        [node, "-e", script, str(STATIC_ROOT / "playback_worklet.js")],
+        check=True,
+        capture_output=True,
+        text=True,
+    )

tests/fullduplex/minicpmo45/test_input.py

@@ -43,6 +43,54 @@ def test_append_emits_one_model_unit_when_multiple_units_are_buffered():
     assert buffer.pending_byte_count == 16_000 * 4
 
 
+def test_speech_marker_does_not_leak_across_irregular_chunk_boundaries():
+    buffer = MiniCPMO45PcmAppendBuffer()
+
+    assert buffer.append(pcm_payload(15_000), chunk_period_ms=1_000) is None
+    speech_unit = buffer.append(pcm_payload(2_000, speech=False), chunk_period_ms=1_000)
+    silence_unit = buffer.append(pcm_payload(15_000, speech=False), chunk_period_ms=1_000)
+
+    assert speech_unit is not None
+    assert speech_unit["is_speech"] is True
+    assert silence_unit is not None
+    assert silence_unit["is_speech"] is False
+
+
+def test_force_listen_marker_does_not_leak_across_irregular_chunk_boundaries():
+    buffer = MiniCPMO45PcmAppendBuffer()
+    forced = pcm_payload(15_000, speech=False)
+    forced["force_listen"] = True
+
+    assert buffer.append(forced, chunk_period_ms=1_000) is None
+    forced_unit = buffer.append(pcm_payload(2_000, speech=False), chunk_period_ms=1_000)
+    unforced_unit = buffer.append(pcm_payload(15_000, speech=False), chunk_period_ms=1_000)
+
+    assert forced_unit is not None
+    assert forced_unit["force_listen"] is True
+    assert unforced_unit is not None
+    assert unforced_unit["force_listen"] is False
+
+
+def test_pcm_append_rollback_restores_per_span_speech_markers():
+    buffer = MiniCPMO45PcmAppendBuffer()
+    assert buffer.append(pcm_payload(15_000), chunk_period_ms=1_000) is None
+    reservation = buffer.prepare_append(
+        pcm_payload(2_000, speech=False),
+        operation_id="mixed-unit",
+        chunk_period_ms=1_000,
+    )
+
+    assert reservation is not None
+    reservation.rollback()
+    first = buffer.append(pcm_payload(15_000, speech=False), chunk_period_ms=1_000)
+    second = buffer.flush(chunk_period_ms=1_000)
+
+    assert first is not None
+    assert first["is_speech"] is True
+    assert second is not None
+    assert second["is_speech"] is False
+
+
 def test_commit_without_speech_does_not_synthesize_terminal_audio():
     buffer = MiniCPMO45PcmAppendBuffer()
     buffer.append(pcm_payload(8_000, speech=False), chunk_period_ms=1_000)

tests/helpers/minicpmo_4_5_duplex.py

@@ -0,0 +1,179 @@
+"""Shared fixtures and argument builders for MiniCPM-o 4.5 duplex E2E tests."""
+
+from __future__ import annotations
+
+import hashlib
+import uuid
+import wave
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+from huggingface_hub import snapshot_download
+
+from tests.helpers.runtime import OmniServerParams
+from tests.helpers.stage_config import get_deploy_config_path, modify_stage_config
+
+MODEL = "openbmb/MiniCPM-o-4_5"
+DEPLOY_CONFIG = modify_stage_config(
+    get_deploy_config_path("minicpmo_4_5_duplex.yaml"),
+    updates={
+        "base_config": get_deploy_config_path("minicpmo_4_5.yaml"),
+        "stages": {
+            0: {"kv_cache_memory_bytes": 6 * 1024 * 1024 * 1024},
+            1: {"kv_cache_memory_bytes": 512 * 1024 * 1024},
+            2: {"kv_cache_memory_bytes": 256 * 1024 * 1024},
+        },
+    },
+)
+CORE_DEPLOY_CONFIG = modify_stage_config(
+    DEPLOY_CONFIG,
+    updates={
+        "stages": {
+            0: {"enforce_eager": True},
+            1: {"enforce_eager": True},
+        }
+    },
+)
+ASSET_DIR = Path(__file__).resolve().parents[1] / "assets" / "minicpmo_4_5"
+RESPONSE_REQUIRED_WAV = ASSET_DIR / "response_required_16k.wav"
+RESPONSE_REQUIRED_SHA256 = "2e5fd4eb3ee434ce107ee3a0591fa624a33f7683c7462f45fe651c443c9af941"
+SOFT_INTERRUPT_WAV = ASSET_DIR / "soft_interrupt_16k.wav"
+SOFT_INTERRUPT_SHA256 = "cadae6d0ddc510310f16f8775d6379f30a5195369ccb54ff10a8e3f9a1f4a2ea"
+REF_AUDIO_RELATIVE_PATH = Path("assets") / "HT_ref_audio.wav"
+
+SERVER_PARAMS = [
+    pytest.param(
+        OmniServerParams(
+            model=MODEL,
+            stage_config_path=DEPLOY_CONFIG,
+            use_stage_cli=True,
+            server_args=["--trust-remote-code"],
+        ),
+        id="three-stage-single-gpu",
+    )
+]
+CORE_SERVER_PARAMS = [
+    pytest.param(
+        OmniServerParams(
+            model=MODEL,
+            stage_config_path=CORE_DEPLOY_CONFIG,
+            use_stage_cli=True,
+            server_args=["--trust-remote-code"],
+        ),
+        id="three-stage-single-gpu",
+    )
+]
+
+
+def validated_wav(path: Path, expected_sha256: str) -> Path:
+    actual_sha256 = hashlib.sha256(path.read_bytes()).hexdigest()
+    if actual_sha256 != expected_sha256:
+        raise AssertionError(
+            f"MiniCPM-o duplex fixture SHA256 mismatch: expected {expected_sha256}, got {actual_sha256}"
+        )
+    with wave.open(str(path), "rb") as wav_file:
+        actual_format = (
+            wav_file.getframerate(),
+            wav_file.getnchannels(),
+            wav_file.getsampwidth(),
+            wav_file.getcomptype(),
+        )
+    expected_format = (16_000, 1, 2, "NONE")
+    if actual_format != expected_format:
+        raise AssertionError(f"MiniCPM-o duplex fixture must be 16 kHz mono PCM16, got {actual_format}")
+    return path
+
+
+def validated_input_wav() -> Path:
+    return validated_wav(RESPONSE_REQUIRED_WAV, RESPONSE_REQUIRED_SHA256)
+
+
+def validated_soft_interrupt_wav() -> Path:
+    return validated_wav(SOFT_INTERRUPT_WAV, SOFT_INTERRUPT_SHA256)
+
+
+def resolve_ref_audio(model_prefix: str) -> Path:
+    if model_prefix:
+        model_root = Path(model_prefix) / MODEL
+    else:
+        model_root = Path(snapshot_download(MODEL, local_files_only=True))
+    ref_audio = model_root / REF_AUDIO_RELATIVE_PATH
+    if not ref_audio.is_file():
+        raise FileNotFoundError(f"MiniCPM-o checkpoint ref audio is missing: {ref_audio}")
+    return ref_audio
+
+
+def realtime_url(omni_server) -> str:
+    return f"ws://{omni_server.host}:{omni_server.port}/v1/realtime?duplex=1"
+
+
+def demo_args(
+    *,
+    omni_server,
+    input_wav: Path,
+    ref_audio: Path,
+    output_dir: Path,
+) -> SimpleNamespace:
+    return SimpleNamespace(
+        url=realtime_url(omni_server),
+        model=omni_server.model,
+        session_id=f"duplex-ci-single-{uuid.uuid4().hex}",
+        input_wav=str(input_wav),
+        ref_audio=str(ref_audio),
+        turn_input_wav=[],
+        output_dir=str(output_dir),
+        output_audio_format="pcm16",
+        chunk_ms=200,
+        realtime_input=True,
+        first_turn_ms=1400,
+        turn_duration_ms=[],
+        first_turn_transcript="duplex CI speech",
+        omit_transcript_hints=False,
+        validation_mode="response-required",
+        temperature=0.0,
+        scenario="sequential",
+        require_audio=True,
+        require_distinct_inputs=False,
+        expect_empty_turn=[],
+        short_ack_ms=350,
+        turns=1,
+        timeout_s=120.0,
+        model_policy_settle_ms=2000,
+    )
+
+
+def multi_session_args(
+    *,
+    omni_server,
+    input_wav: Path,
+    ref_audio: Path,
+    output_dir: Path,
+    response_required: bool,
+) -> SimpleNamespace:
+    return SimpleNamespace(
+        url=realtime_url(omni_server),
+        model=omni_server.model,
+        sessions=2,
+        input_wav=str(input_wav),
+        ref_audio=str(ref_audio),
+        session_input_wav=[],
+        session_expected_token=[],
+        turn_input_wav=[],
+        output_dir=str(output_dir),
+        realtime_input=True,
+        chunk_ms=200,
+        turns=1,
+        first_turn_ms=1400,
+        turn_duration_ms=[],
+        response_required=response_required,
+        temperature=0.0 if response_required else None,
+        disconnect_session_index=0,
+        takeover_session_index=1,
+        resume_after_ms=100,
+        expire_session_index=None,
+        expire_after_s=40.0,
+        verify_admission_limit=None,
+        model_policy_settle_ms=2000,
+        timeout_s=180.0,
+    )

tests/model_executor/models/minicpmo_4_5/test_code2wav_batching.py

@@ -19,10 +19,12 @@ class _FakeEncoder(nn.Module):
     def __init__(self):
         super().__init__()
         self.calls: list[int] = []
+        self.last_chunk_calls: list[bool] = []
 
     def forward_chunk(self, xs, last_chunk=False, cnn_cache=None, att_cache=None):
         batch, length, _ = xs.shape
         self.calls.append(batch)
+        self.last_chunk_calls.append(last_chunk)
         old_length = 0 if att_cache is None else att_cache.shape[3]
         output = xs[:, : max(1, length - 1)]
         cnn = xs[:, :1, :].transpose(1, 2).contiguous()
@@ -156,6 +158,35 @@ def _model():
     return model, token2wav
 
 
+def test_code2wav_resolves_hf_model_id_for_assets(mocker, tmp_path):
+    resolved_root = tmp_path / "snapshot"
+    resolved_root.mkdir()
+    config = _config()
+    config.model_config.model = "openbmb/MiniCPM-o-4_5"
+    config.model_config.revision = "test-revision"
+    config.model_config.stage_connector_config["extra"].pop("prompt_wav")
+    config.load_config = SimpleNamespace(download_dir="/model-cache")
+    model = MiniCPMO45Code2Wav(vllm_config=config)
+    mock_download = mocker.patch(
+        "vllm_omni.model_executor.model_loader.weight_utils.download_weights_from_hf_specific",
+        return_value=str(resolved_root),
+    )
+
+    assert model._resolve_model_root() == resolved_root
+    assert model.model_path == str(resolved_root)
+    assert model._default_prompt_wav == str(resolved_root / "assets" / "HT_ref_audio.wav")
+    mock_download.assert_called_once_with(
+        "openbmb/MiniCPM-o-4_5",
+        "/model-cache",
+        allow_patterns=[
+            "assets/HT_ref_audio.wav",
+            "assets/token2wav/*",
+        ],
+        revision="test-revision",
+        require_all=True,
+    )
+
+
 def _info(
     request_id: str,
     chunk_seq: int,
@@ -212,6 +243,18 @@ def test_adapter_runs_true_batch_cfg_and_splits_request_caches():
     assert cache1[0, 0, 0, 0, 0].item() == 20
 
 
+def test_fade_in_out_limits_overlap_to_available_previous_audio():
+    speech = torch.arange(6, dtype=torch.float32).reshape(1, -1)
+    previous = torch.full((1, 3), 2.0)
+    window = torch.hamming_window(8, periodic=False)
+
+    actual = BatchedToken2Wav._fade_in_out(speech, previous, window)
+
+    expected = speech.clone()
+    expected[..., :3] = speech[..., :3] * window[:3] + previous * window[-3:]
+    torch.testing.assert_close(actual, expected)
+
+
 def test_estimator_cache_stack_split_round_trip_preserves_cfg_rows():
     token2wav = _FakeToken2Wav()
     adapter = BatchedToken2Wav(token2wav)
@@ -259,6 +302,166 @@ def test_model_preserves_output_slots_and_prefers_runtime_codes():
     assert token2wav.flow.encoder.calls[-1] == 2
 
 
+def test_code2wav_projects_duplex_metadata_to_final_audio_output():
+    model, token2wav = _model()
+    segment = _info("duplex", 0, [10, 11])
+    segment_text_utf8 = torch.tensor(list(b"hello"), dtype=torch.uint8)
+    segment["meta"].update(
+        {
+            "duplex_epoch": 3,
+            "duplex_turn_id": 7,
+            "llm_output_text_utf8": segment_text_utf8,
+            "tts_is_last_chunk": True,
+            "turn_end": False,
+        }
+    )
+
+    segment_output = _forward(model, [segment])
+
+    assert segment_output.multimodal_outputs["meta.turn_end"][0].item() is False
+    # A Talker unit boundary only drains pending codec tokens. The official
+    # streaming path keeps Token2wav open until the assistant turn ends.
+    assert token2wav.flow.encoder.last_chunk_calls[-1] is False
+    assert "duplex" in model._states
+
+    final = _info("duplex", 1, [12, 13], last_chunk=True)
+    final["meta"].update(segment["meta"])
+    final["meta"]["chunk_seq"] = 1
+    final["meta"]["last_chunk"] = True
+    final["meta"]["turn_end"] = True
+    output = _forward(model, [final])
+
+    payload = output.multimodal_outputs
+    assert "meta" not in payload
+    assert payload["meta.duplex_epoch"][0].item() == 3
+    assert payload["meta.duplex_turn_id"][0].item() == 7
+    torch.testing.assert_close(
+        payload["meta.llm_output_text_utf8"][0],
+        segment_text_utf8,
+    )
+    assert payload["meta.tts_is_last_chunk"][0].item() is True
+    assert payload["meta.turn_end"][0].item() is True
+    assert token2wav.flow.encoder.last_chunk_calls[-1] is True
+    assert "duplex" not in model._states
+
+
+def test_initial_empty_segment_marker_initializes_stream_without_audio():
+    model, token2wav = _model()
+    boundary = _info("duplex", 0, [])
+    boundary["meta"].update(
+        {
+            "code_flat_numel": 0,
+            "tts_is_last_chunk": True,
+            "turn_end": False,
+        }
+    )
+
+    output = _forward(model, [boundary])
+
+    assert output.multimodal_outputs["model_outputs"][0].numel() == 0
+    assert "duplex" in model._states
+    assert token2wav.hift.calls == []
+
+    resumed = _info(
+        "duplex",
+        1,
+        [4218, 4218, 4218, 10, 11, 12, 13, 14],
+    )
+    output = _forward(model, [resumed])
+
+    assert output.multimodal_outputs["model_outputs"][0].numel() > 0
+    assert "duplex" in model._states
+
+
+def test_shared_runtime_prompt_recreates_missing_file_before_second_owner(tmp_path, monkeypatch):
+    monkeypatch.setattr("tempfile.gettempdir", lambda: str(tmp_path))
+    model, _ = _model()
+    reference = torch.tensor([0.0, 0.25, -0.25, 0.0])
+
+    first = _info("voice-a", 0, [10, 11])
+    first["codes"]["ref"] = reference
+    first["meta"]["ref_audio_sr"] = 16000
+    first["meta"].pop("prompt_cache_id")
+    _forward(model, [first], request_ids=["internal-a"])
+
+    prompt_key = model._request_prompt_keys["internal-a"]
+    prompt_path = Path(model._runtime_prompts[prompt_key].path)
+    prompt_path.unlink()
+
+    second = _info("voice-b", 0, [12, 13])
+    second["codes"]["ref"] = reference
+    second["meta"]["ref_audio_sr"] = 16000
+    second["meta"].pop("prompt_cache_id")
+    _forward(model, [second], request_ids=["internal-b"])
+
+    assert prompt_path.is_file()
+    assert model._runtime_prompts[prompt_key].owners == {"internal-a", "internal-b"}
+
+    model.on_requests_finished(["internal-a"])
+    assert prompt_path.is_file()
+    assert model._runtime_prompts[prompt_key].owners == {"internal-b"}
+
+    model.on_requests_finished(["internal-b"])
+    assert not prompt_path.exists()
+    assert prompt_key not in model._runtime_prompts
+
+
+def test_runtime_prompt_write_failure_does_not_publish_partial_file(tmp_path, monkeypatch):
+    monkeypatch.setattr("tempfile.gettempdir", lambda: str(tmp_path))
+    model, _ = _model()
+    reference = torch.tensor([0.0, 0.25, -0.25, 0.0])
+
+    def fail_after_partial_write(path, *_args, **_kwargs):
+        Path(path).write_bytes(b"partial")
+        raise OSError("simulated write failure")
+
+    monkeypatch.setattr(
+        "vllm_omni.model_executor.models.minicpmo_4_5.minicpmo_4_5_code2wav.sf.write",
+        fail_after_partial_write,
+    )
+
+    with pytest.raises(OSError, match="simulated write failure"):
+        model._materialize_runtime_prompt(reference, 16000)
+
+    assert len(model._runtime_prompts) == 1
+    entry = next(iter(model._runtime_prompts.values()))
+    assert not Path(entry.path).exists()
+    assert list(Path(entry.path).parent.iterdir()) == []
+
+
+def test_runtime_prompt_files_are_isolated_between_model_instances(tmp_path, monkeypatch):
+    monkeypatch.setattr("tempfile.gettempdir", lambda: str(tmp_path))
+    first_model, _ = _model()
+    second_model, _ = _model()
+    reference = torch.tensor([0.0, 0.25, -0.25, 0.0])
+
+    def runtime_ref_info(request_id: str):
+        info = _info(request_id, 0, [10, 11])
+        info["codes"]["ref"] = reference
+        info["meta"]["ref_audio_sr"] = 16000
+        info["meta"].pop("prompt_cache_id")
+        return info
+
+    _forward(first_model, [runtime_ref_info("voice-a")], request_ids=["internal-a"])
+    _forward(second_model, [runtime_ref_info("voice-b")], request_ids=["internal-b"])
+
+    first_key = first_model._request_prompt_keys["internal-a"]
+    second_key = second_model._request_prompt_keys["internal-b"]
+    first_path = Path(first_model._runtime_prompts[first_key].path)
+    second_path = Path(second_model._runtime_prompts[second_key].path)
+    assert first_key == second_key
+    assert first_path != second_path
+    assert first_path.is_file()
+    assert second_path.is_file()
+
+    first_model.on_requests_finished(["internal-a"])
+    assert not first_path.exists()
+    assert second_path.is_file()
+
+    second_model.on_requests_finished(["internal-b"])
+    assert not second_path.exists()
+
+
 def test_mixed_final_exact_buckets_keep_order_and_release_only_final_states():
     model, _ = _model()
     _forward(
@@ -485,27 +688,37 @@ def test_reference_voice_and_duplex_metadata_follow_request_lifecycle():
     model, _ = _model()
     first = _info("voice-a", 0, [1, 2])
     first["codes"]["ref"] = torch.linspace(-0.1, 0.1, 160)
+    segment_text_utf8 = torch.tensor(list(b"hello"), dtype=torch.uint8)
     first["meta"].update(
         ref_audio_sr=16000,
-        native_duplex_segment_text="hello",
+        llm_output_text_utf8=segment_text_utf8,
         duplex_turn_id=7,
         duplex_epoch=3,
     )
     first["meta"].pop("prompt_cache_id")
 
     output = _forward(model, [first])
-    prompt_cache_id, prompt_wav = model._owned_prompt_wavs["voice-a"]
-    assert prompt_cache_id.startswith("voice-a:")
+    prompt_key = model._request_prompt_keys["voice-a"]
+    prompt = model._runtime_prompts[prompt_key]
+    prompt_cache_id, prompt_wav = prompt.cache_id, prompt.path
+    assert prompt_cache_id.startswith("runtime-ref-")
     assert Path(prompt_wav).is_file()
-    assert bytes(output.multimodal_outputs["meta.llm_output_text_utf8"][0].tolist()).decode() == "hello"
+    torch.testing.assert_close(
+        output.multimodal_outputs["meta.llm_output_text_utf8"][0],
+        segment_text_utf8,
+    )
     assert output.multimodal_outputs["meta.duplex_turn_id"][0].item() == 7
     assert output.multimodal_outputs["meta.duplex_epoch"][0].item() == 3
 
     final = _info("voice-a", 1, [3, 4], last_chunk=True)
     final["meta"].pop("prompt_cache_id")
+    final["meta"]["tts_is_last_chunk"] = True
     output = _forward(model, [final])
 
     assert output.multimodal_outputs["meta.tts_is_last_chunk"][0].item() is True
-    assert "voice-a" not in model._owned_prompt_wavs
+    assert model._request_prompt_keys["voice-a"] == prompt_key
+    model.on_requests_finished(["voice-a"])
+    assert "voice-a" not in model._request_prompt_keys
+    assert prompt_key not in model._runtime_prompts
     assert not Path(prompt_wav).exists()
     assert (prompt_cache_id, prompt_wav) not in model.backend._prompt_features

tests/model_executor/models/minicpmo_4_5/test_talker_batching.py

@@ -14,6 +14,7 @@ from vllm_omni.model_executor.models.minicpmo_4_5.minicpmo_4_5_omni import (
 from vllm_omni.model_executor.models.minicpmo_4_5.minicpmo_4_5_omni_tts import (
     MiniCPMO45OmniTTSForConditionalGeneration,
     _max_audio_tokens,
+    _restore_weight_norm_weight,
 )
 from vllm_omni.utils.mm_outputs import to_payload_element
 
@@ -56,6 +57,8 @@ def _make_talker() -> MiniCPMO45OmniTTSForConditionalGeneration:
     talker._request_generators = {}
     talker._request_audio_states = {}
     talker._deferred_cleanup_ids = set()
+    talker._codec_min_tokens = 50
+    talker._codec_seed = 42
     return talker
 
 
@@ -81,6 +84,23 @@ def test_audio_token_limit_scales_with_condition_length(
     assert _max_audio_tokens(condition_tokens) == expected
 
 
+def test_weight_norm_restore_matches_checkpoint_parametrization_in_bfloat16() -> None:
+    generator = torch.Generator().manual_seed(42)
+    weight_v = torch.randn(8, 16, generator=generator, dtype=torch.bfloat16)
+    weight_g = torch.rand(8, 1, generator=generator, dtype=torch.bfloat16)
+    linear = nn.utils.parametrizations.weight_norm(
+        nn.Linear(16, 8, bias=False, dtype=torch.bfloat16),
+        dim=0,
+    )
+    with torch.no_grad():
+        linear.parametrizations.weight.original0.copy_(weight_g)
+        linear.parametrizations.weight.original1.copy_(weight_v)
+
+    restored = _restore_weight_norm_weight(weight_g, weight_v)
+
+    assert torch.equal(restored, linear.weight)
+
+
 def test_talker_emits_request_aligned_codec_deltas_after_compaction(mocker) -> None:
     talker = _make_talker()
     seen: list[tuple[str, list[float], list[int]]] = []
@@ -114,9 +134,73 @@ def test_talker_emits_request_aligned_codec_deltas_after_compaction(mocker) -> N
     assert _routed(output, 0)["codes"]["audio"].tolist() == [[2]]
     assert _routed(output, 1)["codes"]["audio"].tolist() == [[3]]
     assert _routed(output, 0)["meta"]["finished"].item() is False
+    assert set(output.multimodal_outputs["meta"]) == {"finished"}
     assert talker.compute_logits(output.text_hidden_states).argmax(dim=-1).tolist() == [0, 0]
 
 
+def test_talker_projects_request_aligned_duplex_metadata(mocker) -> None:
+    talker = _make_talker()
+    mocker.patch.object(talker, "_sample_audio_code", return_value=torch.tensor(2))
+    infos = [
+        {
+            "request_id": "req-a",
+            "native_duplex": True,
+            "duplex": {"epoch": 3, "turn_id": 7},
+            "ids": {"tts": [41]},
+            "meta": {
+                "native_duplex_segment_text": "first",
+                "turn_eos_token_id": 99,
+            },
+            "audio_codes": {"accumulated": torch.empty(0, dtype=torch.long)},
+        },
+        {
+            "request_id": "req-b",
+            "native_duplex": True,
+            "duplex": {"epoch": 4, "turn_id": 8},
+            "ids": {"tts": [42, 99]},
+            "meta": {
+                "native_duplex_segment_text": "second",
+                "turn_eos_token_id": 99,
+            },
+            "audio_codes": {"accumulated": torch.empty(0, dtype=torch.long)},
+        },
+    ]
+
+    output = talker.make_omni_output(
+        torch.ones(2, 2),
+        model_intermediate_buffer=infos,
+        request_token_spans=[(0, 1), (1, 2)],
+    )
+
+    meta = output.multimodal_outputs["meta"]
+    assert [value.item() for value in meta["native_duplex"]] == [True, True]
+    assert [value.item() for value in meta["duplex_epoch"]] == [3, 4]
+    assert [value.item() for value in meta["duplex_turn_id"]] == [7, 8]
+    assert "native_duplex_segment_text" not in meta
+    assert [bytes(value.tolist()).decode("utf-8") for value in meta["llm_output_text_utf8"]] == [
+        "first",
+        "second",
+    ]
+    assert [value.item() for value in meta["turn_end"]] == [False, True]
+
+
+def test_talker_rejects_native_duplex_without_fence_identity(mocker) -> None:
+    talker = _make_talker()
+    mocker.patch.object(talker, "_sample_audio_code", return_value=torch.tensor(2))
+    info = {
+        "request_id": "req-missing-fence",
+        "native_duplex": True,
+        "audio_codes": {"accumulated": torch.empty(0, dtype=torch.long)},
+    }
+
+    with pytest.raises(RuntimeError, match="requires non-negative integer epoch and turn_id"):
+        talker.make_omni_output(
+            torch.ones(1, 2),
+            model_intermediate_buffer=[info],
+            request_token_spans=[(0, 1)],
+        )
+
+
 def test_incomplete_prefill_emits_no_code_and_does_not_advance_state(mocker) -> None:
     talker = _make_talker()
     sample = mocker.patch.object(talker, "_sample_audio_code", return_value=torch.tensor(2))
@@ -179,7 +263,7 @@ def test_eos_is_terminal_once_and_never_enters_codec_history(mocker) -> None:
     assert talker.compute_logits(second.text_hidden_states).argmax(dim=-1).tolist() == [1]
 
 
-def test_max_token_terminal_includes_only_new_codec_delta(mocker) -> None:
+def test_max_token_terminal_drops_unconsumed_codec_delta(mocker) -> None:
     talker = _make_talker()
     mocker.patch.object(talker, "_sample_audio_code", return_value=torch.tensor(3))
     info = {
@@ -194,8 +278,10 @@ def test_max_token_terminal_includes_only_new_codec_delta(mocker) -> None:
         request_token_spans=[(0, 1)],
     )
 
-    assert info["audio_codes"]["accumulated"].tolist() == [4, 5, 3]
-    assert output.multimodal_outputs["codes"]["audio"][0].tolist() == [[3]]
+    # MiniCPMTTS.generate_chunk samples once at the max-token boundary to
+    # advance RNG state, but the sampled code is not fed into KV or returned.
+    assert info["audio_codes"]["accumulated"].tolist() == [4, 5]
+    assert output.multimodal_outputs["codes"]["audio"][0].shape == (0, 1)
     assert output.multimodal_outputs["meta"]["finished"][0].item() is True
     assert talker.compute_logits(output.text_hidden_states).argmax(dim=-1).tolist() == [1]
 
@@ -268,6 +354,94 @@ def test_empty_speech_segment_finishes_without_sampling_codes() -> None:
     assert decode_embeds.shape == (1, 4)
 
 
+def test_chunked_prefill_tail_aligns_condition_with_prompt_length(mocker) -> None:
+    talker = _make_talker()
+    talker.emb_text = nn.Embedding(1, 2)
+    condition = torch.arange(18, dtype=torch.float32).reshape(9, 2)
+    mocker.patch.object(talker, "_build_condition_embeddings", return_value=condition)
+
+    _, embeds, _ = talker.preprocess(
+        torch.zeros(9, dtype=torch.long),
+        None,
+        _omni_is_prefill=True,
+        _omni_num_computed_tokens=59,
+        _omni_prompt_len=68,
+        request_id="req-chunked-prefill",
+        tts_token_ids=torch.tensor([1]),
+        tts_hidden_states=torch.ones(1, 2),
+    )
+
+    assert torch.equal(embeds, condition)
+    state = talker._request_audio_states["req-chunked-prefill"]
+    assert state["min_tokens"] == 50
+    assert state["max_tokens"] == 64
+
+
+@pytest.mark.parametrize(
+    ("meta", "expected_min_tokens"),
+    [
+        ({"turn_start": True}, 0),
+        ({}, 26),
+        ({"turn_end": True}, 0),
+    ],
+)
+def test_native_duplex_prefill_uses_official_chunk_limits(
+    mocker,
+    meta,
+    expected_min_tokens,
+) -> None:
+    talker = _make_talker()
+    talker.emb_text = nn.Embedding(1, 2)
+    mocker.patch.object(
+        talker,
+        "_build_condition_embeddings",
+        return_value=torch.ones(3, 2),
+    )
+
+    talker.preprocess(
+        torch.zeros(3, dtype=torch.long),
+        None,
+        _omni_is_prefill=True,
+        request_id="req-duplex-chunk",
+        native_duplex=True,
+        meta=meta,
+        tts_token_ids=torch.tensor([1]),
+        tts_hidden_states=torch.ones(1, 2),
+    )
+
+    state = talker._request_audio_states["req-duplex-chunk"]
+    assert state["min_tokens"] == expected_min_tokens
+    assert state["max_tokens"] == 26
+
+
+def test_native_duplex_condition_matches_official_text_plus_audio_bos() -> None:
+    talker = _make_talker()
+    talker.emb_text = nn.Embedding(16, 2)
+    talker.projector_semantic = nn.Identity()
+    talker._normalize = False
+    talker._text_eos_id = 14
+    talker._tts_bos_id = 15
+    with torch.no_grad():
+        talker.emb_text.weight.copy_(torch.arange(32, dtype=torch.float32).reshape(16, 2))
+
+    token_ids = torch.tensor([2, 3])
+    hidden_states = torch.tensor([[0.5, 1.0], [1.5, 2.0]])
+
+    condition = talker._build_condition_embeddings(
+        token_ids,
+        hidden_states,
+        native_duplex=True,
+    )
+
+    expected_text = talker.emb_text(token_ids) + hidden_states
+    expected = torch.cat(
+        [expected_text, talker.emb_text(torch.tensor([talker._tts_bos_id]))],
+        dim=0,
+    )
+    assert torch.equal(condition, expected)
+    assert condition.shape[0] == token_ids.shape[0] + 1
+
+
 def test_request_cleanup_evicts_ar_rng_and_decode_state() -> None:
     talker = _make_talker()
     talker._request_generators["req-done"] = torch.Generator()

tests/model_executor/stage_input_processors/test_minicpmo_4_5_async_chunk.py

@@ -43,6 +43,27 @@ def _delta(*codes: int):
     }
 
 
+def _duplex_delta(
+    *codes: int,
+    epoch: int = 3,
+    turn_id: int = 7,
+    text: str = "segment",
+    turn_end: bool = False,
+):
+    text_utf8 = torch.tensor(list(text.encode("utf-8")), dtype=torch.uint8)
+    return {
+        "codes": {"audio": torch.tensor(codes, dtype=torch.long).reshape(-1, 1)},
+        "meta": {
+            "finished": torch.tensor(False),
+            "native_duplex": torch.tensor(True),
+            "duplex_epoch": torch.tensor(epoch),
+            "duplex_turn_id": torch.tensor(turn_id),
+            "llm_output_text_utf8": text_utf8,
+            "turn_end": torch.tensor(turn_end),
+        },
+    }
+
+
 def _codes(payload) -> list[int]:
     assert payload.codes is not None
     assert isinstance(payload.codes.audio, torch.Tensor)
@@ -108,29 +129,57 @@ def test_short_final_flushes_silence_prefix_and_tail() -> None:
     assert final.meta.finished.item() is True
 
 
+def test_duplex_turn_end_waits_for_terminal_codec_flush() -> None:
+    manager = _manager()
+    request = _request("req-duplex")
+
+    body = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(*range(25), turn_end=True),
+        request,
+        False,
+    )
+    final = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(turn_end=True),
+        request,
+        True,
+    )
+
+    assert body is not None
+    assert body.meta.last_chunk is False
+    assert body.meta.turn_end is False
+    assert final is not None
+    assert _codes(final) == [22, 23, 24]
+    assert final.meta.last_chunk is True
+    assert final.meta.turn_end is True
+
+
 def test_first_chunk_forwards_reference_voice_and_duplex_identity() -> None:
     manager = _manager()
     request = _request("req")
     request.additional_information = {
         "codes": {"ref": [0.1, -0.1]},
-        "meta": {
-            "ref_audio_sr": 16000,
-            "native_duplex_segment_text": "hello",
-            "segment_end": True,
-            "turn_end": True,
-        },
-        "duplex": {"epoch": 3, "model_turn_id": 7},
+        "meta": {"ref_audio_sr": 16000},
     }
 
-    payload = tts2code2wav_async_chunk(manager, _delta(*range(7)), request, True)
+    payload = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(*range(7), text="hello", turn_end=True),
+        request,
+        True,
+    )
 
     assert payload is not None
     assert payload.codes.ref.tolist() == pytest.approx([0.1, -0.1])
     assert payload.meta.ref_audio_sr == 16000
-    assert payload.meta.native_duplex_segment_text == "hello"
+    torch.testing.assert_close(
+        payload.meta.llm_output_text_utf8,
+        torch.tensor(list(b"hello"), dtype=torch.uint8),
+    )
     assert payload.meta.duplex_epoch == 3
     assert payload.meta.duplex_turn_id == 7
-    assert payload.meta.segment_end is True
+    assert payload.meta.tts_is_last_chunk is True
     assert payload.meta.turn_end is True
 
 
@@ -211,6 +260,186 @@ def test_empty_final_releases_wait_gate_once() -> None:
     assert duplicate is None
 
 
+def test_empty_duplex_boundary_uses_zero_length_transport_placeholder() -> None:
+    manager = _manager()
+
+    boundary = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(text="boundary"),
+        _request("req-duplex"),
+        True,
+    )
+
+    assert boundary is not None
+    assert _codes(boundary) == [0]
+    assert boundary.meta.code_flat_numel == 0
+    assert boundary.meta.last_chunk is False
+    assert boundary.meta.is_segment_finished.item() is False
+    torch.testing.assert_close(
+        boundary.meta.llm_output_text_utf8,
+        torch.tensor(list(b"boundary"), dtype=torch.uint8),
+    )
+
+
+def test_duplex_segments_preserve_stream_state_without_closing_turn() -> None:
+    manager = _manager()
+    request = _request("req-duplex")
+    request.additional_information = {
+        "codes": {"ref": [0.1, 0.2, 0.3]},
+        "meta": {"ref_audio_sr": 16000},
+    }
+
+    first = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(10, 11, text="first"),
+        request,
+        True,
+    )
+    second = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(12, 13, text="second"),
+        request,
+        True,
+    )
+
+    assert first is not None
+    assert second is not None
+    assert first.meta.last_chunk is False
+    assert second.meta.last_chunk is False
+    assert first.codes is not None
+    assert torch.allclose(first.codes.ref, torch.tensor([0.1, 0.2, 0.3]))
+    assert first.meta.ref_audio_sr == 16000
+    assert second.codes is not None
+    assert second.codes.ref is None
+    assert second.meta.cache_epoch == first.meta.cache_epoch
+    assert second.meta.chunk_seq == first.meta.chunk_seq + 1
+    assert second.meta.duplex_epoch == 3
+    assert second.meta.duplex_turn_id == 7
+    torch.testing.assert_close(
+        second.meta.llm_output_text_utf8,
+        torch.tensor(list(b"second"), dtype=torch.uint8),
+    )
+    assert second.meta.tts_is_last_chunk is True
+    assert second.meta.turn_end is False
+    assert first.meta.is_segment_finished.item() is False
+    assert second.meta.is_segment_finished.item() is False
+
+
+def test_duplex_short_units_wait_for_minimum_stream_body() -> None:
+    manager = _manager()
+    request = _request("req-duplex")
+
+    first = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(10, 11, 12, text="first"),
+        request,
+        True,
+    )
+    second = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(13, 14, text="first"),
+        request,
+        True,
+    )
+
+    assert first is not None
+    assert _codes(first) == [0]
+    assert first.meta.code_flat_numel == 0
+    assert first.meta.last_chunk is False
+    assert first.meta.tts_is_last_chunk is True
+    assert second is not None
+    assert _codes(second) == [4218, 4218, 4218, 10, 11, 12, 13, 14]
+    assert second.meta.code_flat_numel == 8
+    torch.testing.assert_close(
+        second.meta.llm_output_text_utf8,
+        torch.tensor(list(b"firstfirst"), dtype=torch.uint8),
+    )
+
+
+def test_duplex_empty_finish_callback_does_not_replay_previous_text() -> None:
+    manager = _manager()
+    request = _request("req-duplex")
+
+    first = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(*range(25), text="first"),
+        request,
+        False,
+    )
+    boundary = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(text="first"),
+        request,
+        True,
+    )
+    second = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(*range(25, 50), text="second"),
+        request,
+        False,
+    )
+
+    assert first is not None
+    assert boundary is not None
+    assert boundary.meta.code_flat_numel == 0
+    assert second is not None
+    torch.testing.assert_close(
+        second.meta.llm_output_text_utf8,
+        torch.tensor(list(b"second"), dtype=torch.uint8),
+    )
+
+
+def test_duplex_short_tail_does_not_replay_previous_segment_text() -> None:
+    manager = _manager()
+    request = _request("req-duplex")
+
+    def chunk(codes, text: str, finished: bool):
+        return tts2code2wav_async_chunk(
+            manager,
+            _duplex_delta(*codes, text=text),
+            request,
+            finished,
+        )
+
+    first = chunk(range(25), "和上海之间", False)
+    first_tail = chunk([25, 26], "和上海之间", True)
+    assert first is not None
+    assert first.meta.llm_output_text_utf8.tolist() == list("和上海之间".encode())
+    assert first_tail is not None
+    assert first_tail.meta.code_flat_numel == 0
+    assert chunk([27, 28, 29], "的距离大约是", False) is None
+    next_flush = chunk([], "的距离大约是", True)
+    assert next_flush is not None
+    assert next_flush.meta.llm_output_text_utf8.tolist() == list("的距离大约是".encode())
+
+
+def test_duplex_turn_end_closes_epoch_and_next_turn_restarts_sequence() -> None:
+    manager = _manager()
+    request = _request("req-duplex")
+
+    turn_end = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(14, turn_id=7, turn_end=True),
+        request,
+        True,
+    )
+    next_turn = tts2code2wav_async_chunk(
+        manager,
+        _duplex_delta(20, turn_id=8),
+        request,
+        True,
+    )
+
+    assert turn_end is not None
+    assert next_turn is not None
+    assert turn_end.meta.last_chunk is True
+    assert turn_end.meta.turn_end is True
+    assert turn_end.meta.is_segment_finished.item() is True
+    assert next_turn.meta.cache_epoch == turn_end.meta.cache_epoch + 1
+    assert next_turn.meta.chunk_seq == 0
+    assert next_turn.meta.last_chunk is False
+
+
 def test_staggered_requests_keep_accumulators_isolated() -> None:
     manager = _manager()
     req_a = _request("a")

tests/model_executor/stage_input_processors/test_minicpmo_4_5_omni.py

@@ -128,13 +128,125 @@ def test_native_duplex_speak_segment_reaches_split_talker() -> None:
     info = converted["model_intermediate_buffer"]
     assert info["native_duplex"] is True
     assert info["ids"]["tts"] == [21, 22]
-    assert converted["prompt_token_ids"] == [0, 0, 0, 0]
+    assert converted["prompt_token_ids"] == [0, 0, 0]
     assert info["meta"]["replace_streaming_prompt"] is True
+    assert info["meta"]["next_stage_prompt_len"] == 3
+    assert info["meta"]["turn_start"] is True
     assert info["meta"]["segment_end"] is True
     assert info["duplex"]["epoch"] == 3
     assert info["duplex"]["turn_id"] == 7
 
 
+def test_native_duplex_continuation_appends_only_new_talker_condition() -> None:
+    prompt_ids = [101, 102]
+    token_ids = {
+        "tts_bos_token_id": 9301,
+        "tts_eos_token_id": 9302,
+        "listen_token_id": 9303,
+        "speak_token_id": 9304,
+        "chunk_eos_token_id": 9308,
+        "chunk_tts_eos_token_id": 9309,
+        "turn_eos_token_id": 9310,
+    }
+    context = SimpleNamespace(
+        bridge_states={
+            "duplex": {
+                "epoch": 3,
+                "model_turn_id": 7,
+            }
+        }
+    )
+
+    first_ids = [9304, 21, 22, 9308]
+    first = _output(
+        prompt_ids=prompt_ids,
+        output_ids=first_ids,
+        latent=torch.arange(24, dtype=torch.float32).reshape(6, 4),
+        multimodal_output={
+            "duplex_prompt_token_ids": prompt_ids,
+            "meta": token_ids,
+        },
+    )
+    second_ids = [*first_ids, 9304, 23, 24, 9308]
+    second = _output(
+        prompt_ids=prompt_ids,
+        output_ids=second_ids,
+        latent=torch.arange(40, dtype=torch.float32).reshape(10, 4),
+        multimodal_output={
+            "duplex_prompt_token_ids": prompt_ids,
+            "meta": token_ids,
+        },
+    )
+    third_ids = [*second_ids, 9304, 25, 26, 9308]
+    third = _output(
+        prompt_ids=prompt_ids,
+        output_ids=third_ids,
+        latent=torch.arange(56, dtype=torch.float32).reshape(14, 4),
+        multimodal_output={
+            "duplex_prompt_token_ids": prompt_ids,
+            "meta": token_ids,
+        },
+    )
+
+    first_input = llm2tts([first], prompt=[{}], _streaming_context=context)[0]
+    second_input = llm2tts([second], prompt=[{}], _streaming_context=context)[0]
+    context.bridge_states["duplex"]["model_turn_id"] = 8
+    third_input = llm2tts([third], prompt=[{}], _streaming_context=context)[0]
+
+    assert first_input["model_intermediate_buffer"]["ids"]["tts"] == [21, 22]
+    assert second_input["model_intermediate_buffer"]["ids"]["tts"] == [23, 24]
+    assert third_input["model_intermediate_buffer"]["ids"]["tts"] == [25, 26]
+    assert first_input["model_intermediate_buffer"]["meta"]["turn_start"] is True
+    assert second_input["model_intermediate_buffer"]["meta"]["turn_start"] is False
+    assert third_input["model_intermediate_buffer"]["meta"]["turn_start"] is True
+    assert first_input["model_intermediate_buffer"]["meta"]["replace_streaming_prompt"] is True
+    assert "replace_streaming_prompt" not in second_input["model_intermediate_buffer"]["meta"]
+    assert third_input["model_intermediate_buffer"]["meta"]["replace_streaming_prompt"] is True
+    assert second_input["model_intermediate_buffer"]["meta"]["next_stage_prompt_len"] == 3
+    assert second_input["prompt_token_ids"] == [0, 0, 0]
+
+
+def test_native_duplex_transcript_decodes_the_talker_condition_slice() -> None:
+    prompt_ids = [101, 102]
+    metadata = {
+        "tts_bos_token_id": 9301,
+        "tts_eos_token_id": 9302,
+        "listen_token_id": 9303,
+        "speak_token_id": 9304,
+        "chunk_eos_token_id": 9308,
+        "chunk_tts_eos_token_id": 9309,
+        "turn_eos_token_id": 9310,
+    }
+    token_text = {21: "杭州", 22: "和", 23: "州和", 24: "上海之间", 25: "大"}
+
+    def output(token_ids: list[int]):
+        return _output(
+            prompt_ids=prompt_ids,
+            output_ids=token_ids,
+            latent=torch.zeros((len(prompt_ids) + len(token_ids), 1)),
+            multimodal_output={
+                "duplex_prompt_token_ids": prompt_ids,
+                "meta": metadata,
+            },
+        )
+
+    context = SimpleNamespace(
+        bridge_states={"duplex": {"epoch": 3, "model_turn_id": 7}},
+        source_token_decoder=lambda ids, **_: "".join(token_text.get(int(token_id), "") for token_id in ids),
+    )
+    first_ids = [9304, 21, 22, 9308]
+    first_info = llm2tts([output(first_ids)], prompt=[{}], _streaming_context=context)[0]["model_intermediate_buffer"]
+    second_info = llm2tts(
+        [output([*first_ids, 23, 24, 25, 9308])],
+        prompt=[{}],
+        _streaming_context=context,
+    )[0]["model_intermediate_buffer"]
+
+    assert first_info["meta"]["native_duplex_segment_text"] == "杭州和"
+    assert second_info["ids"]["tts"] == [24, 25]
+    assert second_info["meta"]["native_duplex_segment_text"] == "上海之间大"
+
+
 def test_native_duplex_requires_tokenizer_boundary_metadata() -> None:
     latent = torch.zeros((3, 4))
     source = _output(

tests/model_executor/stage_input_processors/test_qwen3_omni_streaming_helpers.py

@@ -4,6 +4,7 @@
 
 from __future__ import annotations
 
+from collections import defaultdict
 from types import SimpleNamespace
 
 import pytest
@@ -227,6 +228,47 @@ def test_talker2code2wav_full_payload_keeps_all_zero_codec_rows() -> None:
     assert "code_predictor_codes" not in payload
 
 
+def test_talker2code2wav_async_chunk_flushes_cached_tail_on_stop_token() -> None:
+    request_id = "codec_tail"
+    stop_token_id = 999
+    transfer_manager = SimpleNamespace(
+        code_prompt_token_ids=defaultdict(list),
+        put_req_chunk=defaultdict(int, {request_id: 2}),
+        connector=SimpleNamespace(
+            config={
+                "extra": {
+                    "initial_codec_chunk_frames": 4,
+                    "codec_chunk_frames": 25,
+                    "codec_left_context_frames": 25,
+                }
+            }
+        ),
+    )
+    transfer_manager.code_prompt_token_ids[request_id] = [
+        torch.tensor([[frame, frame + 100]], dtype=torch.long) for frame in range(50)
+    ]
+    request = SimpleNamespace(
+        external_req_id=request_id,
+        sampling_params=SimpleNamespace(
+            stop_token_ids=[stop_token_id],
+            stop_token_id=None,
+        ),
+    )
+
+    payload = q3.talker2code2wav_async_chunk(
+        transfer_manager,
+        {"codes": {"audio": torch.tensor([[stop_token_id, 0]], dtype=torch.long)}},
+        request,
+        is_finished=True,
+    )
+
+    assert payload is not None
+    assert payload.meta.finished.item() is True
+    # 4 initial frames and one 25-frame chunk were already emitted. The final
+    # payload contains 25 frames of left context plus the remaining 21 frames.
+    assert payload.codes.audio.numel() == (25 + 21) * 2
+
+
 def test_thinker2talker_full_payload_packs_complete_tensors() -> None:
     """Full-payload path drops the terminal thinker row before talker prefill."""
     request = SimpleNamespace(

tests/test_config_factory.py

@@ -1160,8 +1160,10 @@ class TestDeployConfigLoading:
         assert [stage.to_omegaconf().session_mode for stage in stages] == ["duplex", "duplex", "duplex"]
         assert [stage.yaml_engine_args["async_scheduling"] for stage in stages] == [False, False, False]
         assert all("Async" not in (stage.scheduler_cls or "") for stage in stages)
+        assert [stage.devices for stage in deploy.stages] == ["0", "0", "0"]
         assert deploy.stages[1].enforce_eager is False
         assert stages[1].yaml_extras["default_sampling_params"]["max_tokens"] == 2048
+        assert stages[1].yaml_extras["default_sampling_params"]["min_tokens"] == 0
         assert stages[1].yaml_extras["default_sampling_params"]["stop_token_ids"] == [1]
 
     @pytest.mark.parametrize(

tests/worker/test_native_duplex_hooks.py

@@ -276,16 +276,16 @@ def test_minicpmo_model_hook_same_speech_row_second_step_does_not_rearm_pending_
     next_logits[0, 7] = 30.0
     next_logits[0, 8] = -2.0
     next_logits[0, 10] = 20.0
+    original_next_logits = next_logits.clone()
 
     model.prepare_duplex_sampling(next_logits, SimpleNamespace(), (row,))
 
     assert state.pending_speech_context is False
     assert state.current_turn_ended is False
-    assert torch.isneginf(next_logits[0, 7])
-    assert next_logits[0, 8].item() == 30.0
+    assert torch.equal(next_logits, original_next_logits)
 
 
-def test_minicpmo_model_hook_mid_turn_speech_redirects_listen_to_tts_bos():
+def test_minicpmo_model_hook_mid_turn_speech_preserves_logits_for_post_sample_redirect():
     state = SimpleNamespace(
         current_turn_ended=False,
         last_terminator_token=None,
@@ -296,12 +296,11 @@ def test_minicpmo_model_hook_mid_turn_speech_redirects_listen_to_tts_bos():
     logits[0, 7] = 30.0
     logits[0, 8] = -2.0
     logits[0, 10] = 20.0
+    original_logits = logits.clone()
 
     model.prepare_duplex_sampling(logits, SimpleNamespace(), (row,))
 
-    assert torch.isneginf(logits[0, 7])
-    assert logits[0, 8].item() == 30.0
-    assert logits[0, 10].item() == 20.0
+    assert torch.equal(logits, original_logits)
 
 
 def test_minicpmo_model_hook_ignores_serving_new_user_turn_marker():
@@ -317,12 +316,12 @@ def test_minicpmo_model_hook_ignores_serving_new_user_turn_marker():
     logits = torch.zeros((1, 16), dtype=torch.float32)
     logits[0, 7] = 5.0
     logits[0, 10] = 20.0
+    original_logits = logits.clone()
     model.prepare_duplex_sampling(logits, SimpleNamespace(), (row,))
 
     assert state.current_turn_ended is False
     assert state.last_terminator_token is None
-    assert torch.isneginf(logits[0, 7])
-    assert logits[0, 8].item() == 5.0
+    assert torch.equal(logits, original_logits)
 
 
 def test_generic_ar_runner_builds_typed_duplex_sampling_rows():
@@ -613,6 +612,180 @@ def test_minicpmo_stage0_rejects_unknown_special_token_fallbacks():
         runtime._require_special_token_ids()
 
 
+def _minicpmo45_tokenizer_stub():
+    token_ids = {
+        token: index
+        for index, token in enumerate(
+            (
+                "<unit>",
+                "</unit>",
+                "<|listen|>",
+                "<|speak|>",
+                "<|tts_bos|>",
+                "<|tts_eos|>",
+                "<|tts_pad|>",
+                "<|chunk_eos|>",
+                "<|chunk_tts_eos|>",
+                "<|turn_eos|>",
+            ),
+            start=101,
+        )
+    }
+    return SimpleNamespace(
+        unk_token_id=0,
+        convert_tokens_to_ids=lambda token: token_ids.get(token, 0),
+        encode=lambda token, add_special_tokens=False: (
+            [token_ids[token]] if not add_special_tokens and token in token_ids else [0]
+        ),
+    )
+
+
+def _transformers_processor_stub(auto_processor):
+    class _AutoImageProcessor:
+        @staticmethod
+        def register(*_args, **_kwargs):
+            return None
+
+    return SimpleNamespace(
+        AutoImageProcessor=_AutoImageProcessor,
+        AutoProcessor=auto_processor,
+    )
+
+
+def test_minicpmo_stage0_loads_processor_from_hf_id(monkeypatch):
+    import sys
+
+    from vllm_omni.experimental.fullduplex.minicpmo45.stage0 import (
+        MiniCPMO45Stage0DuplexRuntime,
+    )
+
+    model_id = "openbmb/MiniCPM-o-4_5"
+    processor = SimpleNamespace(tokenizer=_minicpmo45_tokenizer_stub())
+    load_calls = []
+
+    class _AutoImageProcessor:
+        @staticmethod
+        def register(config_class, image_processor_class):
+            del image_processor_class
+            if isinstance(config_class, str):
+                raise AttributeError("'str' object has no attribute '__module__'")
+
+    class _AutoProcessor:
+        @classmethod
+        def from_pretrained(cls, model_path, *, trust_remote_code):
+            load_calls.append((model_path, trust_remote_code))
+            _AutoImageProcessor.register("MiniCPMVImageProcessor", object)
+            return processor
+
+    monkeypatch.setitem(
+        sys.modules,
+        "transformers",
+        SimpleNamespace(
+            AutoImageProcessor=_AutoImageProcessor,
+            AutoProcessor=_AutoProcessor,
+        ),
+    )
+
+    loaded = MiniCPMO45Stage0DuplexRuntime._load_processor_from_path(model_id)
+
+    assert loaded is processor
+    assert load_calls == [(model_id, True)]
+    with pytest.raises(AttributeError, match="__module__"):
+        _AutoImageProcessor.register("MiniCPMVImageProcessor", object)
+
+
+def test_minicpmo_stage0_processor_load_preserves_original_exception(monkeypatch):
+    import sys
+
+    from vllm_omni.experimental.fullduplex.minicpmo45.stage0 import (
+        MiniCPMO45Stage0DuplexRuntime,
+    )
+
+    original = ValueError("invalid processor config")
+
+    class _AutoProcessor:
+        @classmethod
+        def from_pretrained(cls, model_path, *, trust_remote_code):
+            del cls, model_path, trust_remote_code
+            raise original
+
+    monkeypatch.setitem(
+        sys.modules,
+        "transformers",
+        _transformers_processor_stub(_AutoProcessor),
+    )
+
+    with pytest.raises(RuntimeError, match="Failed to load MiniCPM-o duplex processor") as exc_info:
+        MiniCPMO45Stage0DuplexRuntime._load_processor_from_path("openbmb/MiniCPM-o-4_5")
+
+    assert exc_info.value.__cause__ is original
+
+
+def test_minicpmo_stage0_processor_requires_tokenizer(monkeypatch):
+    import sys
+
+    from vllm_omni.experimental.fullduplex.minicpmo45.stage0 import (
+        MiniCPMO45Stage0DuplexRuntime,
+    )
+
+    class _AutoProcessor:
+        @classmethod
+        def from_pretrained(cls, model_path, *, trust_remote_code):
+            del cls, model_path, trust_remote_code
+            return object()
+
+    monkeypatch.setitem(
+        sys.modules,
+        "transformers",
+        _transformers_processor_stub(_AutoProcessor),
+    )
+
+    with pytest.raises(RuntimeError, match="does not expose a tokenizer"):
+        MiniCPMO45Stage0DuplexRuntime._load_processor_from_path("openbmb/MiniCPM-o-4_5")
+
+
+def test_minicpmo_stage0_loaded_processor_validates_special_tokens(monkeypatch):
+    import sys
+
+    from vllm_omni.experimental.fullduplex.minicpmo45.stage0 import (
+        MiniCPMO45Stage0DuplexRuntime,
+    )
+
+    processor = SimpleNamespace(tokenizer=_minicpmo45_tokenizer_stub())
+
+    class _AutoProcessor:
+        @classmethod
+        def from_pretrained(cls, model_path, *, trust_remote_code):
+            del cls, model_path, trust_remote_code
+            return processor
+
+    monkeypatch.setitem(
+        sys.modules,
+        "transformers",
+        _transformers_processor_stub(_AutoProcessor),
+    )
+
+    runtime = MiniCPMO45Stage0DuplexRuntime(
+        SimpleNamespace(),
+        model_path="openbmb/MiniCPM-o-4_5",
+        device="cpu",
+    )
+
+    assert runtime.processor is processor
+    assert set(runtime._special_token_ids()) == {
+        "unit_token_id",
+        "unit_end_token_id",
+        "listen_token_id",
+        "speak_token_id",
+        "tts_bos_token_id",
+        "tts_eos_token_id",
+        "tts_pad_token_id",
+        "chunk_eos_token_id",
+        "chunk_tts_eos_token_id",
+        "turn_eos_token_id",
+    }
+
+
 def test_minicpmo_stage0_data_plane_prefill_matches_official_unit_format():
     import torch
 
@@ -1150,6 +1323,88 @@ def test_minicpmo_stage0_data_plane_final_first_chunk_does_not_add_silence_unit(
     )
 
 
+def test_minicpmo_stage0_final_does_not_promote_first_chunk_alignment_tail_to_unit():
+    import torch
+
+    from vllm_omni.experimental.fullduplex.minicpmo45.stage0 import (
+        MiniCPMO45Stage0DuplexRuntime,
+        _MiniCPMO45Stage0SessionState,
+    )
+
+    class _StageModel(torch.nn.Module):
+        first_chunk_ms = 10
+        chunk_ms = 8
+        sample_rate = 1000
+
+        def __init__(self):
+            super().__init__()
+            self.embed = torch.nn.Embedding(256, 2)
+
+        def get_input_embeddings(self):
+            return self.embed
+
+        def get_audio_hidden_states(self, _data):
+            return [torch.tensor([[0.5, 0.5]], dtype=torch.float32)]
+
+    class _MelProcessor:
+        sample_rate = 1000
+        is_first = True
+
+        def get_config(self):
+            return {"effective_first_chunk_ms": 9}
+
+    class _Processor:
+        def __init__(self):
+            self._streaming_mel_processor = _MelProcessor()
+
+        def get_streaming_chunk_size(self):
+            return 9 if self._streaming_mel_processor.is_first else 8
+
+        def process_audio_streaming(self, audio, **_kwargs):
+            self._streaming_mel_processor.is_first = False
+            return {"audio_features": audio, "audio_feature_lens": [[len(audio)]]}
+
+    runtime = MiniCPMO45Stage0DuplexRuntime.__new__(MiniCPMO45Stage0DuplexRuntime)
+    runtime.stage_model = _StageModel()
+    runtime.thinker = runtime.stage_model
+    runtime.tokenizer = SimpleNamespace(
+        unk_token_id=0,
+        convert_tokens_to_ids=lambda token: {
+            "<unit>": 1,
+            "</unit>": 2,
+            "<|listen|>": 3,
+            "<|speak|>": 4,
+            "<|tts_bos|>": 5,
+            "<|tts_eos|>": 6,
+            "<|tts_pad|>": 7,
+            "<|chunk_eos|>": 8,
+            "<|chunk_tts_eos|>": 9,
+            "<|turn_eos|>": 10,
+            "<|audio|>": 11,
+        }.get(token, 0),
+        encode=lambda text, add_special_tokens=False: [],
+    )
+    runtime.processor = _Processor()
+    runtime.device = "cpu"
+    runtime._init_token_ids()
+    state = _MiniCPMO45Stage0SessionState(session_id="sid-first-chunk-alignment-tail")
+
+    first = runtime._stage_prefill_embeddings_only(state, np.zeros(8, dtype=np.float32), seq=1)
+    second = runtime._stage_prefill_embeddings_only(state, np.zeros(8, dtype=np.float32), seq=2)
+    final = runtime._stage_prefill_embeddings_only(
+        state,
+        np.zeros(8, dtype=np.float32),
+        seq=3,
+        final=True,
+    )
+
+    assert first["input_token_ids"].count(1) == 1
+    assert second["input_token_ids"].count(1) == 1
+    assert final["input_token_ids"].count(1) == 1
+    assert state.audio_chunk_idx == 3
+    assert len(state.audio_buffer) == 1
+
+
 def test_minicpmo_stage0_runtime_uses_loaded_vllm_embed_tokens_when_get_input_embeddings_is_broken():
     import torch
 
@@ -1319,7 +1574,7 @@ def test_minicpmo_stage0_native_sampler_penalizes_text_from_prior_chunk():
     model.thinker = SimpleNamespace(get_tokenizer=lambda: _Tokenizer())
     session_key = ("sid-cross-chunk-repetition", 0)
     state = _MiniCPMO45Stage0SessionState(session_id=session_key[0])
-    state.generated_text_tokens = [198] * 8
+    state.generated_tokens = [198] * 8
     model._minicpmo45_duplex_data_plane_helper = SimpleNamespace(sessions={session_key: state})
     model._minicpmo45_duplex_row_sessions = {0: session_key}
 
@@ -1381,7 +1636,7 @@ def test_minicpmo_stage0_native_sampler_matches_official_negative_logit_penalty(
     state = _MiniCPMO45Stage0SessionState(session_id=session_key[0])
     repeated = 198
     alternative = 1234
-    state.generated_text_tokens = [repeated]
+    state.generated_tokens = [repeated]
     model._minicpmo45_duplex_data_plane_helper = SimpleNamespace(sessions={session_key: state})
     model._minicpmo45_duplex_row_sessions = {0: session_key}
 
@@ -1405,7 +1660,7 @@ def test_minicpmo_stage0_native_sampler_matches_official_negative_logit_penalty(
     assert sampled.sampled_token_ids.tolist() == [[repeated]]
 
 
-def test_minicpmo_stage0_records_only_bounded_text_repetition_history():
+def test_minicpmo_stage0_records_bounded_model_policy_history():
     from vllm_omni.experimental.fullduplex.minicpmo45.stage0 import (
         _MiniCPMO45Stage0SessionState,
     )
@@ -1432,15 +1687,15 @@ def test_minicpmo_stage0_records_only_bounded_text_repetition_history():
     }
 
     for sampled in range(1000, 1520):
-        model._record_minicpmo45_duplex_terminator(0, sampled, token_ids)
+        model._record_minicpmo45_duplex_generation_token(0, sampled)
 
     expected_history = list(range(1008, 1520))
-    assert state.generated_text_tokens == expected_history
+    assert state.generated_tokens == expected_history
 
     for sampled in token_ids.values():
-        model._record_minicpmo45_duplex_terminator(0, sampled, token_ids)
+        model._record_minicpmo45_duplex_generation_token(0, sampled)
 
-    assert state.generated_text_tokens == expected_history
+    assert state.generated_tokens == [*expected_history[len(token_ids) :], *token_ids.values()]
 
 
 def test_minicpmo_stage0_native_sampler_does_not_override_model_at_punctuation():

vllm_omni/core/sched/omni_ar_scheduler.py

@@ -731,9 +731,17 @@ class OmniARScheduler(OmniSchedulerMixin, VLLMScheduler):
                 if self.log_stats:
                     session.record_event(EngineCoreEventType.QUEUED)
                 return
-        update_info = getattr(update, "additional_information", None)
-        update_meta = update_info.get("meta") if isinstance(update_info, dict) else None
-        if isinstance(update_meta, dict) and update_meta.get("replace_streaming_prompt") is True:
+        update_infos = (
+            getattr(update, "model_intermediate_buffer", None),
+            getattr(update, "additional_information", None),
+        )
+        replace_streaming_prompt = any(
+            isinstance(info, dict)
+            and isinstance(info.get("meta"), dict)
+            and info["meta"].get("replace_streaming_prompt") is True
+            for info in update_infos
+        )
+        if replace_streaming_prompt:
             self._replace_streaming_session(session, update)
             return
         super()._update_request_as_session(session, update)

vllm_omni/core/sched/omni_generation_scheduler.py

@@ -54,6 +54,21 @@ class OmniGenerationScheduler(OmniSchedulerMixin, VLLMScheduler):
             )
         self._latest_omni_connector_output: OmniConnectorOutput | None = None
 
+    def _handle_stopped_request(self, request: Request) -> bool:
+        if (
+            request.resumable
+            and not request.streaming_queue
+            and self.chunk_transfer_adapter is not None
+            and self.chunk_transfer_adapter.receives_chunks
+        ):
+            # Downstream async-chunk stages receive the next segment from the
+            # connector, not from an API StreamingUpdate. Enqueue them as
+            # schedulable before the base class can park them in skipped_waiting.
+            request.status = RequestStatus.WAITING
+            self._enqueue_waiting_request(request)
+            return False
+        return super()._handle_stopped_request(request)
+
     def schedule(self, throttle_prefills: bool = False) -> SchedulerOutput:
         """One-shot generation fast path:
         - Feed all input tokens of the request at once

vllm_omni/core/sched/omni_scheduler_mixin.py

@@ -63,6 +63,11 @@ class OmniSchedulerMixin:
         session.num_computed_tokens = 0
         session.prompt_token_ids = new_prompt
         session.additional_information = update.additional_information or None
+        session.model_intermediate_buffer = getattr(
+            update,
+            "model_intermediate_buffer",
+            None,
+        )
         session.update_block_hashes()
         session.num_prompt_tokens = len(new_prompt)
         session.arrival_time = update.arrival_time

vllm_omni/deploy/minicpmo_4_5_duplex.yaml

@@ -3,7 +3,7 @@
 # The registered minicpmo_4_5 pipeline owns the fixed
 # Thinker -> Talker -> Code2Wav topology. This overlay changes session
 # lifetime and bounds all three stages to two live duplex sessions.
-base_config: minicpmo_4_5_batching.yaml
+base_config: minicpmo_4_5.yaml
 pipeline: minicpmo_4_5
 session_mode: duplex
 active_stream_window: 1
@@ -31,6 +31,7 @@ stages:
     # segment and turn boundaries travel as explicit output metadata.
     default_sampling_params:
       max_tokens: 2048
+      min_tokens: 0
       stop_token_ids: [1]
 
   - stage_id: 2

vllm_omni/distributed/omni_connectors/adapter.py

@@ -231,18 +231,18 @@ def construct_next_stage_streaming_input_prompt(payload_data: dict[str, Any], re
     ids = payload_data.get("ids", {})
     meta = payload_data.get("meta", {})
     replace_prompt = isinstance(meta, dict) and meta.get("replace_streaming_prompt") is True
-    condition_len = meta.get("next_stage_prompt_len") if replace_prompt else None
-    if isinstance(condition_len, int) and condition_len > 0:
+    next_stage_prompt_len = meta.get("next_stage_prompt_len") if isinstance(meta, dict) else None
+    if replace_prompt and isinstance(next_stage_prompt_len, int) and next_stage_prompt_len > 0:
         # Some downstream stages consume complete, independently conditioned
         # segments instead of extending an existing KV prefix. The producer
         # declares that transport behavior explicitly in payload metadata.
-        new_prompt = [0] * condition_len
+        new_prompt = [0] * next_stage_prompt_len
         request._output_token_ids.clear()
         request._all_token_ids.clear()
         request._all_token_ids.extend(new_prompt)
         request.prompt_token_ids = new_prompt
         request.num_computed_tokens = 0
-        request.num_prompt_tokens = condition_len
+        request.num_prompt_tokens = next_stage_prompt_len
         request.update_block_hashes()
         return
     prompt_token_ids = ids.get("prompt", None)
@@ -255,7 +255,10 @@ def construct_next_stage_streaming_input_prompt(payload_data: dict[str, Any], re
     assert request.prompt_token_ids is not None
     # Extend prompt with kept output tokens.
     request.prompt_token_ids.extend(kept_output_tokens)
-    next_prompt_len = max(1, compute_talker_prompt_ids_length(prompt_token_ids))
+    if isinstance(next_stage_prompt_len, int) and next_stage_prompt_len > 0:
+        next_prompt_len = next_stage_prompt_len
+    else:
+        next_prompt_len = max(1, compute_talker_prompt_ids_length(prompt_token_ids))
     new_prompt = [0] * next_prompt_len
     request._all_token_ids.extend(new_prompt or ())
     request.prompt_token_ids.extend(new_prompt or ())

vllm_omni/distributed/omni_connectors/transfer_adapter/chunk_transfer_adapter.py

@@ -283,7 +283,11 @@ class OmniChunkTransferAdapter(OmniTransferAdapterBase):
 
                 # Empty chunk with more data expected: keep polling.
                 has_new_ids = bool(new_ids.numel()) if use_tensor_codes else bool(new_ids)
-                if not has_new_ids and not payload_finished:
+                if not has_new_ids and payload_segment_finished:
+                    # Preserve an explicit scheduler boundary even when it
+                    # contains no new codec frames.
+                    request.prompt_token_ids = [0]
+                if not has_new_ids and not payload_finished and not payload_segment_finished:
                     # The base recv loop treats False as "not ready yet" and
                     # requeues the request. Do not mark an empty non-terminal
                     # chunk as ready, otherwise Stage1 can consume before the
@@ -317,7 +321,13 @@ class OmniChunkTransferAdapter(OmniTransferAdapterBase):
                     multimodal_output=multimodal_output,
                     request=request,
                     # Existing processors use is_finished as a flush signal.
-                    is_finished=is_segment_finished,
+                    # Terminal stops no longer count as segment boundaries
+                    # (is_segment_finished is False when the request finishes,
+                    # see #5383), but the processor must still flush its
+                    # accumulated tail on the terminal chunk — otherwise the
+                    # downstream stage receives the finished marker without
+                    # the final payload (#5413).
+                    is_finished=is_segment_finished or is_finished,
                 )
 
             except Exception as e:
@@ -332,7 +342,8 @@ class OmniChunkTransferAdapter(OmniTransferAdapterBase):
         if payload_data.meta is None:
             payload_data.meta = MetaStruct()
         payload_data.meta.finished = torch.tensor(is_finished, dtype=torch.bool)
-        payload_data.meta.is_segment_finished = torch.tensor(is_segment_finished, dtype=torch.bool)
+        if payload_data.meta.is_segment_finished is None:
+            payload_data.meta.is_segment_finished = torch.tensor(is_segment_finished, dtype=torch.bool)
 
         success, size, metadata = self.connector.put(
             from_stage=str(stage_id),
@@ -725,6 +736,14 @@ class OmniChunkTransferAdapter(OmniTransferAdapterBase):
         queue_snapshot = list(queue)
         for request in queue_snapshot:
             if not self._ensure_active_stream(request):
+                if target_status == RequestStatus.WAITING:
+                    # A non-active placeholder must not remain visible to the
+                    # scheduler: it has no connector payload yet, so running
+                    # it would execute the downstream model with empty
+                    # additional_information. Park it until restore_queues()
+                    # and retry admission on the next scheduler tick.
+                    queue.remove(request)
+                    waiting_for_chunk_list.append(request)
                 continue
             if request.status != RequestStatus.WAITING_FOR_CHUNK:
                 if request.request_id in self.requests_with_ready_chunks:

vllm_omni/engine/async_omni_engine.py

@@ -209,10 +209,10 @@ class AsyncOmniEngine:
                 self._omni_master_port,
             )
 
-        # Stage resolution pops deploy_config, so get the pipeline endpoint
-        # restriction beforehand. TODO (Alex) make this cleaner and refactor
-        # stage config resolution to remove kwargs hacks.
-        deploy_config_path = kwargs.get("deploy_config")
+        # Stage resolution pops deploy_config, so get pipeline-wide settings
+        # beforehand. The stage CLI exposes the same deploy YAML through
+        # stage_configs_path.
+        deploy_config_path = kwargs.get("deploy_config") or kwargs.get("stage_configs_path")
         pipeline_config = StageConfigFactory.get_pipeline_config(
             model=model,
             trust_remote_code=trust_remote_code,

vllm_omni/engine/orchestrator.py

@@ -13,7 +13,7 @@ from __future__ import annotations
 
 import asyncio
 import time as _time
-from collections.abc import Callable
+from collections.abc import Awaitable, Callable
 from dataclasses import dataclass, field
 from typing import TYPE_CHECKING, Any
 
@@ -218,11 +218,18 @@ class _OrchestratorDuplexStagePort:
         request_states: dict[str, OrchestratorRequestState],
         running_counter: OmniRequestCounter | None,
         cleanup_request_ids: Callable[..., Any],
+        async_chunk: bool,
+        prewarm_async_chunk_stages: Callable[
+            [str, Any, OrchestratorRequestState],
+            Awaitable[None],
+        ],
     ) -> None:
         self._stage_pools = stage_pools
         self._request_states = request_states
         self._running_counter = running_counter
         self._cleanup_request_ids = cleanup_request_ids
+        self._async_chunk = async_chunk
+        self._prewarm_async_chunk_stages = prewarm_async_chunk_stages
 
     @property
     def stage_count(self) -> int:
@@ -299,6 +306,12 @@ class _OrchestratorDuplexStagePort:
             replica_id = await pool.submit_update(context.request_id, request_state, request)
         else:
             replica_id = await pool.submit_initial(context.request_id, request_state, request, prompt_text=None)
+            if self._async_chunk and context.stage_id == 0:
+                await self._prewarm_async_chunk_stages(
+                    context.request_id,
+                    request,
+                    request_state,
+                )
         request_state.duplex_stage_fences[context.stage_id] = context.fence
         request_state.stage_submit_ts[context.stage_id] = _time.time()
         if not request_state.running_counter_registered and self._running_counter is not None:
@@ -388,6 +401,8 @@ class Orchestrator:
                     request_states=self.request_states,
                     running_counter=self._running_counter,
                     cleanup_request_ids=self._cleanup_request_ids,
+                    async_chunk=self.async_chunk,
+                    prewarm_async_chunk_stages=self._prewarm_async_chunk_stages,
                 ),
                 result_sink=self.rpc_async_queue,
                 lifecycle_sink=self.output_async_queue,
@@ -681,24 +696,16 @@ class Orchestrator:
         final_stage_id = msg.final_stage_id
         req_state = self.request_states.get(request_id)
         if req_state is None:
+            # Streaming updates always follow the first-chunk add_request
+            # through the same ordered submission queue, so an unknown id
+            # here can only mean the request already finished or was
+            # aborted (e.g. the client disconnected). Re-adding it would
+            # resurrect a headless session that keeps cycling through the
+            # stages with nobody consuming its outputs (issue #4271).
             logger.warning(
-                "[Orchestrator] streaming_update for unknown req=%s, falling back to add_request",
+                "[Orchestrator] streaming_update for unknown req=%s; dropping (request finished or aborted)",
                 request_id,
             )
-            fallback_msg = StageSubmissionMessage(
-                type="add_request",
-                request_id=msg.request_id,
-                prompt=msg.prompt,
-                original_prompt=msg.original_prompt,
-                output_prompt_text=msg.output_prompt_text,
-                sampling_params_list=msg.sampling_params_list,
-                final_stage_id=msg.final_stage_id,
-                final_output_stage_ids=msg.final_output_stage_ids,
-                preprocess_ms=msg.preprocess_ms,
-                request_timestamp=msg.request_timestamp,
-                enqueue_ts=msg.enqueue_ts,
-            )
-            await self._handle_add_request(fallback_msg)
             return
 
         if msg.sampling_params_list:
@@ -2025,7 +2032,7 @@ class Orchestrator:
             _t_submit_start = _time.perf_counter()
 
             if next_pool.stage_type == "diffusion":
-                await next_pool.submit_initial(
+                replica_id = await next_pool.submit_initial(
                     request_id,
                     req_state,
                     req_state.prompt,
@@ -2064,12 +2071,18 @@ class Orchestrator:
                     resumable=downstream_resumable,
                 )
                 request.external_req_id = request.request_id
-                await next_pool.submit_initial(
+                replica_id = await next_pool.submit_initial(
                     request_id,
                     req_state,
                     request,
                     prompt_text=None,
                 )
+            self._record_duplex_stage_submission(
+                next_stage_id,
+                request_id,
+                replica_id,
+                req_state,
+            )
 
             # async_chunk pre-submit fires per stage edge (N-1 -> N). Source
             # replica is stage 0's bound replica (single-replica thinker in

vllm_omni/engine/stage_engine_startup.py

@@ -183,6 +183,10 @@ class OmniMasterServer:
         # seeded with the registration port (and coordinator ROUTER port) that
         # are already bound on the same host.
         self._allocated_ports: set[int] = set()
+        # Keep route ports bound until their ZMQ owners are ready to bind them.
+        # Recording only the integer leaves a race with unrelated startup
+        # consumers (notably NCCL's own ephemeral-port selection).
+        self._port_reservations: dict[int, socket.socket] = {}
         if master_port:
             self._allocated_ports.add(int(master_port))
         # Coordinator ROUTER address echoed back in every registration reply
@@ -283,9 +287,11 @@ class OmniMasterServer:
         ``get_open_ports_list`` dedups only within its own call, so we redraw any
         port that collides with one already recorded in ``self._allocated_ports``
         (registration/coordinator ports plus every previously-allocated route
-        port) and register the winners before returning. Callers already hold the
-        relevant allocation context (``__init__`` is single-threaded;
-        registration holds ``self._alloc_lock``).
+        port). Each winner remains OS-bound until its ZMQ owner is ready, so
+        another startup subsystem cannot claim it in the allocation-to-bind
+        window. Callers already hold the relevant allocation context
+        (``__init__`` is single-threaded; registration holds
+        ``self._alloc_lock``).
         """
         picked: list[int] = []
         # Bounded retry budget so a pathological host (near port exhaustion)
@@ -294,7 +300,17 @@ class OmniMasterServer:
             for port in get_open_ports_list(count=count - len(picked)):
                 if port in self._allocated_ports:
                     continue
+                reservation = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
+                try:
+                    # Reserve on every local interface. The advertised address
+                    # can be a routable interface IP, but wildcard binding also
+                    # protects the route from local NCCL/socket allocations.
+                    reservation.bind(("", port))
+                except OSError:
+                    reservation.close()
+                    continue
                 self._allocated_ports.add(port)
+                self._port_reservations[port] = reservation
                 picked.append(port)
                 if len(picked) == count:
                     return picked
@@ -303,6 +319,33 @@ class OmniMasterServer:
             f"(host={self._address}); {len(self._allocated_ports)} already in use."
         )
 
+    def release_route_port_reservations(
+        self,
+        stage_id: int,
+        replica_id: int = 0,
+        *,
+        handshake: bool = True,
+        data: bool = True,
+    ) -> None:
+        """Release reserved route ports immediately before their ZMQ binds."""
+        alloc = self.get_allocation(stage_id, replica_id)
+        ports: list[int] = []
+        if handshake:
+            handshake_port = _port_from_zmq_address(alloc.handshake_bind_address)
+            if handshake_port is not None:
+                ports.append(handshake_port)
+        if data:
+            for address in (alloc.input_bind_address, alloc.output_bind_address):
+                port = _port_from_zmq_address(address)
+                if port is not None:
+                    ports.append(port)
+
+        with self._alloc_lock:
+            for port in ports:
+                reservation = self._port_reservations.pop(port, None)
+                if reservation is not None:
+                    reservation.close()
+
     def _next_free_replica_id(self, stage_id: int) -> int:
         """Return the next replica id to assign for an auto-assign registration.
 
@@ -437,6 +480,11 @@ class OmniMasterServer:
         self._stop_event.set()
         if self._thread is not None:
             self._thread.join(timeout=10)
+        with self._alloc_lock:
+            reservations = list(self._port_reservations.values())
+            self._port_reservations.clear()
+        for reservation in reservations:
+            reservation.close()
 
     # ------------------------------------------------------------------
     # Internal server logic
@@ -771,20 +819,34 @@ def connect_remote_engine_cores(
 
     handshake_bind_address = omni_master_server.get_allocation(stage_id, replica_id=replica_id).handshake_bind_address
 
-    with zmq_socket_ctx(handshake_bind_address, zmq.ROUTER, bind=True) as handshake_socket:
-        yield StageReplicaResources(
-            coordinator=coordinator,
-            addresses=addresses,
-        )
-        wait_for_engine_startup(
-            handshake_socket,
-            addresses,
-            engines_to_handshake,
-            vllm_config.parallel_config,
-            False,  # coordinated_dp
-            vllm_config.cache_config,
-            None,  # proc_manager (remote — no local procs)
-            None,  # coord_process
+    omni_master_server.release_route_port_reservations(
+        stage_id,
+        replica_id,
+        handshake=True,
+        data=False,
+    )
+    try:
+        with zmq_socket_ctx(handshake_bind_address, zmq.ROUTER, bind=True) as handshake_socket:
+            yield StageReplicaResources(
+                coordinator=coordinator,
+                addresses=addresses,
+            )
+            wait_for_engine_startup(
+                handshake_socket,
+                addresses,
+                engines_to_handshake,
+                vllm_config.parallel_config,
+                False,  # coordinated_dp
+                vllm_config.cache_config,
+                None,  # proc_manager (remote — no local procs)
+                None,  # coord_process
+            )
+    finally:
+        omni_master_server.release_route_port_reservations(
+            stage_id,
+            replica_id,
+            handshake=False,
+            data=True,
         )
 
 
@@ -802,17 +864,31 @@ def connect_remote_diffusion_proc(
     ).handshake_bind_address
 
     logger.info("Waiting for remote diffusion proc for stage %d replica %d", stage_id, replica_id)
-    with zmq_socket_ctx(handshake_bind_address, zmq.ROUTER, bind=True) as handshake_socket:
-        yield StageReplicaResources(addresses=addresses)
-        wait_for_engine_startup(
-            handshake_socket,
-            addresses,
-            [CoreEngine(index=0, local=False)],
-            _single_diffusion_parallel_config(local_client=False),
-            False,
-            None,
-            None,
-            None,
+    omni_master_server.release_route_port_reservations(
+        stage_id,
+        replica_id,
+        handshake=True,
+        data=False,
+    )
+    try:
+        with zmq_socket_ctx(handshake_bind_address, zmq.ROUTER, bind=True) as handshake_socket:
+            yield StageReplicaResources(addresses=addresses)
+            wait_for_engine_startup(
+                handshake_socket,
+                addresses,
+                [CoreEngine(index=0, local=False)],
+                _single_diffusion_parallel_config(local_client=False),
+                False,
+                None,
+                None,
+                None,
+            )
+    finally:
+        omni_master_server.release_route_port_reservations(
+            stage_id,
+            replica_id,
+            handshake=False,
+            data=True,
         )
 
 
@@ -918,49 +994,63 @@ def _launch_omni_core_engines(
     # Bind the pre-allocated handshake socket for this stage.
     handshake_bind_address = omni_master_server.get_allocation(stage_id, replica_id=replica_id).handshake_bind_address
 
-    with zmq_socket_ctx(handshake_bind_address, zmq.ROUTER, bind=True) as handshake_socket:
-        if omni_coordinator_address is not None:
-            # Use the omni subclass so each spawned subprocess instantiates
-            # an OmniCoordClientForStage and heartbeats to the coordinator.
-            from vllm_omni.engine.stage_engine_core_proc_manager import StageEngineCoreProcManager
+    omni_master_server.release_route_port_reservations(
+        stage_id,
+        replica_id,
+        handshake=True,
+        data=False,
+    )
+    try:
+        with zmq_socket_ctx(handshake_bind_address, zmq.ROUTER, bind=True) as handshake_socket:
+            if omni_coordinator_address is not None:
+                # Use the omni subclass so each spawned subprocess instantiates
+                # an OmniCoordClientForStage and heartbeats to the coordinator.
+                from vllm_omni.engine.stage_engine_core_proc_manager import StageEngineCoreProcManager
 
-            with scoped_spawn_device_env(stage_visible_devices, spawn_device_lock):
-                local_engine_manager: CoreEngineProcManager = StageEngineCoreProcManager(
-                    local_engine_count=local_engine_count,
-                    start_index=start_index,
-                    local_start_index=local_start_index,
-                    vllm_config=vllm_config,
-                    local_client=True,
-                    handshake_address=handshake_address,
-                    executor_class=executor_class,
-                    log_stats=log_stats,
-                    omni_stage_id=stage_id,
-                    omni_coordinator_address=omni_coordinator_address,
-                    omni_replica_base_id=replica_id,
-                )
-        else:
-            with scoped_spawn_device_env(stage_visible_devices, spawn_device_lock):
-                local_engine_manager = CoreEngineProcManager(
-                    local_engine_count=local_engine_count,
-                    start_index=start_index,
-                    local_start_index=local_start_index,
-                    vllm_config=vllm_config,
-                    local_client=True,
-                    handshake_address=handshake_address,
-                    executor_class=executor_class,
-                    log_stats=log_stats,
-                )
+                with scoped_spawn_device_env(stage_visible_devices, spawn_device_lock):
+                    local_engine_manager: CoreEngineProcManager = StageEngineCoreProcManager(
+                        local_engine_count=local_engine_count,
+                        start_index=start_index,
+                        local_start_index=local_start_index,
+                        vllm_config=vllm_config,
+                        local_client=True,
+                        handshake_address=handshake_address,
+                        executor_class=executor_class,
+                        log_stats=log_stats,
+                        omni_stage_id=stage_id,
+                        omni_coordinator_address=omni_coordinator_address,
+                        omni_replica_base_id=replica_id,
+                    )
+            else:
+                with scoped_spawn_device_env(stage_visible_devices, spawn_device_lock):
+                    local_engine_manager = CoreEngineProcManager(
+                        local_engine_count=local_engine_count,
+                        start_index=start_index,
+                        local_start_index=local_start_index,
+                        vllm_config=vllm_config,
+                        local_client=True,
+                        handshake_address=handshake_address,
+                        executor_class=executor_class,
+                        log_stats=log_stats,
+                    )
 
-        yield local_engine_manager, coordinator, addresses
-        wait_for_engine_startup(
-            handshake_socket,
-            addresses,
-            engines_to_handshake,
-            parallel_config,
-            parallel_config.data_parallel_size > 1 and vllm_config.model_config.is_moe,
-            vllm_config.cache_config,
-            local_engine_manager,
-            coordinator.proc if coordinator else None,
+            yield local_engine_manager, coordinator, addresses
+            wait_for_engine_startup(
+                handshake_socket,
+                addresses,
+                engines_to_handshake,
+                parallel_config,
+                parallel_config.data_parallel_size > 1 and vllm_config.model_config.is_moe,
+                vllm_config.cache_config,
+                local_engine_manager,
+                coordinator.proc if coordinator else None,
+            )
+    finally:
+        omni_master_server.release_route_port_reservations(
+            stage_id,
+            replica_id,
+            handshake=False,
+            data=True,
         )
 
 
@@ -1448,6 +1538,12 @@ def launch_diffusion_stage_replica(
             omni_stage_config=stage_config,
             replica_id=replica_id,
         )
+        omni_master_server.release_route_port_reservations(
+            metadata.stage_id,
+            replica_id,
+            handshake=True,
+            data=False,
+        )
         proc_manager = stage_diffusion_proc.StageDiffusionProcManager(
             model=model,
             od_config=od_config,
@@ -1461,6 +1557,12 @@ def launch_diffusion_stage_replica(
             omni_stage_id=metadata.stage_id,
             omni_replica_id=replica_id,
         )
+        omni_master_server.release_route_port_reservations(
+            metadata.stage_id,
+            replica_id,
+            handshake=False,
+            data=True,
+        )
         client = StageDiffusionClient.from_addresses(
             metadata,
             request_address=proc_manager.addresses.inputs[0],
@@ -1474,6 +1576,10 @@ def launch_diffusion_stage_replica(
             lock_fds=lock_fds,
         )
     except Exception:
+        omni_master_server.release_route_port_reservations(
+            metadata.stage_id,
+            replica_id,
+        )
         if proc_manager is not None:
             proc_manager.shutdown()
         if lock_fds:

vllm_omni/entrypoints/async_omni.py

@@ -628,6 +628,13 @@ class AsyncOmni(EngineClient, OmniBase):
 
             logger.debug(f"[AsyncOmni] Request {request_id} completed")
 
+            # The input pump can outlive a normally-completed generation
+            # (e.g. a realtime client that keeps streaming after the final
+            # output); a live pump keeps issuing streaming updates for a
+            # request the orchestrator no longer tracks.
+            if input_stream_task is not None and not input_stream_task.done():
+                input_stream_task.cancel()
+
             self._log_summary_and_cleanup(request_id)
 
         except (asyncio.CancelledError, GeneratorExit):
@@ -1059,7 +1066,10 @@ class AsyncOmni(EngineClient, OmniBase):
         """Submit request IDs to be aborted to the engine."""
         await self.engine.abort_async(request_ids)
         for rid in request_ids:
-            self.request_states.pop(rid, None)
+            state = self.request_states.pop(rid, None)
+            input_stream_task = getattr(state, "input_stream_task", None)
+            if input_stream_task is not None and not input_stream_task.done():
+                input_stream_task.cancel()
         if self.log_stats:
             logger.info("[AsyncOmni] Aborted request(s) %s", ",".join(request_ids))
 

vllm_omni/entrypoints/cli/serve.py

@@ -892,7 +892,7 @@ def run_headless(args: TrackingNamespace) -> None:
     stage_connector_spec = get_stage_connector_spec(
         omni_transfer_config=omni_transfer_config,
         stage_id=stage_id,
-        async_chunk=False,
+        async_chunk=bool(stage_cfg.engine_args.get("async_chunk", False)),
     )
 
     # ``runtime_cfg`` is mostly inherited from the parent's

vllm_omni/entrypoints/openai/realtime_connection.py

@@ -138,6 +138,7 @@ class RealtimeConnection(VllmRealtimeConnection):
             is_streaming=True,
         )
 
+        result_gen = None
         try:
             result_gen = self.engine.generate(
                 prompt=streaming_input_gen,
@@ -196,6 +197,16 @@ class RealtimeConnection(VllmRealtimeConnection):
             logger.exception("Error in generation: %s", e)
             await self.send_error(str(e), "processing_error")
         finally:
+            # Close the generator explicitly so AsyncOmni.generate's cleanup
+            # (input-pump cancellation and engine-side abort) runs now rather
+            # than whenever the event loop garbage-collects the async
+            # generator; the delay window is where a disconnected session
+            # keeps cycling through the stages (issue #4271).
+            if result_gen is not None:
+                try:
+                    await result_gen.aclose()
+                except Exception:
+                    logger.exception("Failed to close realtime result generator")
             # Always send terminal event so clients don't hang forever.
             if self._is_connected and not audio_done_sent:
                 try:
@@ -206,4 +217,10 @@ class RealtimeConnection(VllmRealtimeConnection):
                 self.audio_queue.get_nowait()
 
     async def send_json(self, payload: dict):
-        await self.websocket.send_text(json.dumps(payload))
+        try:
+            await self.websocket.send_text(json.dumps(payload))
+        except Exception:
+            # A failed send means the client is gone; flag it so the
+            # generation loop stops instead of retrying into a dead socket.
+            self._is_connected = False
+            raise

vllm_omni/experimental/fullduplex/minicpmo45/data_plane.py

@@ -65,12 +65,11 @@ class MiniCPMO45DataPlaneSession:
     def begin_request(self, request_id: str) -> None:
         state = self._requests.setdefault(request_id, _RequestState())
         state.terminal = False
-        # Preserve turn-scoped cursors for resumable requests. The legacy path
-        # only reset request-scoped terminal markers on append.
-        request_turn = state.turns.get(None)
-        if request_turn is not None:
-            request_turn.tts_eos_done = False
-            request_turn.turn_eos_done = False
+        # A resumable append starts a new Talker segment, even when it
+        # continues the same model turn. Re-arm only the segment boundary;
+        # turn EOS remains deduplicated until the model advances the turn.
+        for turn_state in state.turns.values():
+            turn_state.tts_eos_done = False
 
     def is_terminal(self, request_id: str | None) -> bool:
         if request_id is None:

vllm_omni/experimental/fullduplex/minicpmo45/input.py

@@ -2,11 +2,19 @@ from __future__ import annotations
 
 import base64
 import binascii
+from dataclasses import dataclass
 from typing import Any
 
 import numpy as np
 
 
+@dataclass(frozen=True, slots=True)
+class _PcmSpan:
+    byte_count: int
+    force_listen: bool
+    is_speech: bool
+
+
 class MiniCPMO45PcmAppendReservation:
     __slots__ = (
         "_active",
@@ -15,6 +23,7 @@ class MiniCPMO45PcmAppendReservation:
         "_owner",
         "_raw",
         "_sample_rate_hz",
+        "_spans",
         "_turn_had_speech",
         "_video_frames",
         "operation_id",
@@ -31,6 +40,7 @@ class MiniCPMO45PcmAppendReservation:
         sample_rate_hz: int,
         force_listen: bool,
         is_speech: bool,
+        spans: list[_PcmSpan] | None = None,
         turn_had_speech: bool = False,
         video_frames: list[str] | None = None,
     ) -> None:
@@ -41,6 +51,7 @@ class MiniCPMO45PcmAppendReservation:
         self._sample_rate_hz = sample_rate_hz
         self._force_listen = force_listen
         self._is_speech = is_speech
+        self._spans = list(spans or [])
         self._turn_had_speech = turn_had_speech
         self._video_frames = list(video_frames or [])
         self._active = True
@@ -97,9 +108,8 @@ class MiniCPMO45PcmAppendBuffer:
 
     def __init__(self) -> None:
         self._buffer = bytearray()
+        self._spans: list[_PcmSpan] = []
         self._sample_rate_hz: int | None = None
-        self._force_listen = False
-        self._is_speech = False
         self._turn_had_speech = False
         self._reservation_seq = 0
         self._reservations: list[MiniCPMO45PcmAppendReservation] = []
@@ -112,14 +122,13 @@ class MiniCPMO45PcmAppendBuffer:
             reservation._active = False
         self._reservations.clear()
         self._buffer.clear()
+        self._spans.clear()
         self._frame_queue.clear()
         self._sample_rate_hz = None
-        self._force_listen = False
-        self._is_speech = False
         self._turn_had_speech = False
 
     def clear_force_listen(self) -> None:
-        self._force_listen = False
+        self._spans = [_PcmSpan(span.byte_count, False, span.is_speech) for span in self._spans]
 
     def has_pending(self) -> bool:
         return bool(self._buffer)
@@ -131,6 +140,56 @@ class MiniCPMO45PcmAppendBuffer:
     def pending_byte_count(self) -> int:
         return len(self._buffer)
 
+    def _append_span(self, span: _PcmSpan) -> None:
+        if span.byte_count <= 0:
+            return
+        if (
+            self._spans
+            and self._spans[-1].force_listen == span.force_listen
+            and self._spans[-1].is_speech == span.is_speech
+        ):
+            previous = self._spans[-1]
+            self._spans[-1] = _PcmSpan(
+                previous.byte_count + span.byte_count,
+                span.force_listen,
+                span.is_speech,
+            )
+            return
+        self._spans.append(span)
+
+    def _consume_spans(self, byte_count: int) -> list[_PcmSpan]:
+        consumed: list[_PcmSpan] = []
+        remaining = byte_count
+        while remaining:
+            if not self._spans:
+                raise RuntimeError("MiniCPM-o PCM metadata is shorter than buffered audio")
+            span = self._spans.pop(0)
+            take = min(remaining, span.byte_count)
+            consumed.append(_PcmSpan(take, span.force_listen, span.is_speech))
+            if take < span.byte_count:
+                self._spans.insert(
+                    0,
+                    _PcmSpan(
+                        span.byte_count - take,
+                        span.force_listen,
+                        span.is_speech,
+                    ),
+                )
+            remaining -= take
+        return consumed
+
+    def _prepend_spans(self, spans: list[_PcmSpan]) -> None:
+        current = self._spans
+        self._spans = []
+        for span in [*spans, *current]:
+            self._append_span(span)
+
+    def _buffer_force_listen(self) -> bool:
+        return any(span.force_listen for span in self._spans)
+
+    def _buffer_is_speech(self) -> bool:
+        return any(span.is_speech for span in self._spans)
+
     def _reserve_passthrough(
         self,
         payload: dict[str, object],
@@ -180,12 +239,17 @@ class MiniCPMO45PcmAppendBuffer:
             raise ValueError("MiniCPM-o native duplex audio append sample_rate_hz changed within a session")
         self._sample_rate_hz = sample_rate_hz
         self._buffer.extend(raw)
+        self._append_span(
+            _PcmSpan(
+                len(raw),
+                bool(payload.get("force_listen", False)),
+                bool(payload.get("is_speech", False)),
+            )
+        )
         frames_in = payload.get("video_frames")
         if isinstance(frames_in, list):
             self._frame_queue.extend(frame for frame in frames_in if isinstance(frame, str) and frame)
         self._turn_had_speech = self._turn_had_speech or bool(payload.get("is_speech", False))
-        self._force_listen = self._force_listen or bool(payload.get("force_listen", False))
-        self._is_speech = self._is_speech or bool(payload.get("is_speech", False))
         if not allow_emit:
             return None
 
@@ -208,6 +272,7 @@ class MiniCPMO45PcmAppendBuffer:
             pad_samples = 0
         emit_bytes = emit_samples * 4
         reserved_raw = bytes(self._buffer[:emit_bytes])
+        reserved_spans = self._consume_spans(emit_bytes)
         emit_raw = reserved_raw + b"\x00" * (pad_samples * 4)
         del self._buffer[:emit_bytes]
 
@@ -227,11 +292,8 @@ class MiniCPMO45PcmAppendBuffer:
         if emit_samples + pad_samples >= min_samples and self._frame_queue:
             attached_frames = [self._frame_queue.pop(0)]
             out["video_frames"] = attached_frames
-        out["force_listen"] = self._force_listen
-        out["is_speech"] = self._is_speech
-        if not self._buffer:
-            self._force_listen = False
-            self._is_speech = False
+        out["force_listen"] = any(span.force_listen for span in reserved_spans)
+        out["is_speech"] = any(span.is_speech for span in reserved_spans)
         reservation = MiniCPMO45PcmAppendReservation(
             owner=self,
             operation_id=operation_id,
@@ -240,6 +302,7 @@ class MiniCPMO45PcmAppendBuffer:
             sample_rate_hz=sample_rate_hz,
             force_listen=bool(out.get("force_listen", False)),
             is_speech=bool(out.get("is_speech", False)),
+            spans=reserved_spans,
             turn_had_speech=self._turn_had_speech,
             video_frames=attached_frames,
         )
@@ -261,8 +324,8 @@ class MiniCPMO45PcmAppendBuffer:
                 "audio": "",
                 "format": "pcm_f32le",
                 "sample_rate_hz": self._sample_rate_hz or 16000,
-                "force_listen": self._force_listen,
-                "is_speech": self._is_speech,
+                "force_listen": self._buffer_force_listen(),
+                "is_speech": self._buffer_is_speech(),
             }
             reservation = self.prepare_append(
                 payload,
@@ -280,8 +343,8 @@ class MiniCPMO45PcmAppendBuffer:
                 payload=None,
                 raw=b"",
                 sample_rate_hz=self._sample_rate_hz or 0,
-                force_listen=self._force_listen,
-                is_speech=self._is_speech,
+                force_listen=self._buffer_force_listen(),
+                is_speech=self._buffer_is_speech(),
                 turn_had_speech=had_speech,
             )
             self._reservations.append(reservation)
@@ -289,8 +352,6 @@ class MiniCPMO45PcmAppendBuffer:
         # Open the next physical input generation without touching already
         # reserved appends. A rollback restores this generation's metadata.
         self._sample_rate_hz = None
-        self._force_listen = False
-        self._is_speech = False
         self._turn_had_speech = False
         return reservation
 
@@ -335,12 +396,11 @@ class MiniCPMO45PcmAppendBuffer:
         rolled_back = self._reservations[index:]
         restored = b"".join(item._raw for item in rolled_back)
         self._buffer[:0] = restored
+        self._prepend_spans([span for item in rolled_back for span in item._spans])
         restored_frames = [frame for item in rolled_back for frame in item._video_frames]
         if restored_frames:
             self._frame_queue[:0] = restored_frames
         self._sample_rate_hz = self._sample_rate_hz or reservation._sample_rate_hz
-        self._force_listen = self._force_listen or any(item._force_listen for item in rolled_back)
-        self._is_speech = self._is_speech or any(item._is_speech for item in rolled_back)
         self._turn_had_speech = self._turn_had_speech or any(item._turn_had_speech for item in rolled_back)
         for item in rolled_back:
             item._active = False
@@ -354,8 +414,8 @@ class MiniCPMO45PcmAppendBuffer:
             "audio": "",
             "format": "pcm_f32le",
             "sample_rate_hz": self._sample_rate_hz or 16000,
-            "force_listen": self._force_listen,
-            "is_speech": self._is_speech,
+            "force_listen": self._buffer_force_listen(),
+            "is_speech": self._buffer_is_speech(),
         }
         return self.append(payload, chunk_period_ms=chunk_period_ms, flush=True)
 

vllm_omni/experimental/fullduplex/minicpmo45/stage0.py

@@ -2,10 +2,10 @@ from __future__ import annotations
 
 import base64
 import copy
-import sys
 import time
 from contextlib import suppress
 from dataclasses import dataclass, field
+from threading import Lock
 from typing import Any
 
 import numpy as np
@@ -14,6 +14,7 @@ from vllm_omni.experimental.fullduplex.minicpmo45.policy import MiniCPMO45Duplex
 
 _MINICPMO45_SPECIAL_TOKEN_FIELDS = MiniCPMO45DuplexPolicy.SPECIAL_TOKEN_FIELDS
 _MINICPMO45_OPTIONAL_TOKEN_FIELDS = MiniCPMO45DuplexPolicy.OPTIONAL_TOKEN_FIELDS
+_MINICPMO45_PROCESSOR_LOAD_LOCK = Lock()
 
 
 @dataclass
@@ -35,7 +36,7 @@ class _MiniCPMO45Stage0SessionState:
     pending_speech_context: bool = False
     pending_speech_append_identity: tuple[int | None, int] | None = None
     pending_speech_response_open: bool = False
-    generated_text_tokens: list[int] = field(default_factory=list)
+    generated_tokens: list[int] = field(default_factory=list)
 
 
 class MiniCPMO45Stage0DuplexRuntime:
@@ -58,6 +59,8 @@ class MiniCPMO45Stage0DuplexRuntime:
             else getattr(stage_model, "tokenizer", None)
         )
         self._init_token_ids()
+        if self.tokenizer is not None:
+            self._require_special_token_ids()
 
     def _stage_runtime_ready(self) -> bool:
         return self.processor is not None and self.tokenizer is not None and self.thinker is not None
@@ -196,19 +199,15 @@ class MiniCPMO45Stage0DuplexRuntime:
             embed_parts.extend(state.context_embeds)
             token_ids.extend(state.context_token_ids)
 
-        # Consume every complete chunk in the buffer so the appended span and
-        # the scheduler's slot reservation for this append agree exactly. A
-        # final append may zero-pad a real residual chunk, but it must not add
-        # a whole silence unit after all input was already consumed: official
-        # duplex generation runs once per microphone unit and client commit is
-        # not an additional model decision.
+        # Consume every complete processor chunk in the buffer so the appended
+        # span and the scheduler's slot reservation agree exactly. The serving
+        # PCM buffer pads a final real residual before it reaches Stage0. Do not
+        # pad again here: the first processor chunk is hop-aligned below 1035ms
+        # and intentionally leaves a small carry that is not another model unit.
         units_built = 0
         while True:
             if len(state.audio_buffer) < chunk_size:
-                if not final or len(state.audio_buffer) == 0:
-                    break
-                pad = np.zeros(chunk_size - len(state.audio_buffer), dtype=np.float32)
-                state.audio_buffer = np.concatenate([state.audio_buffer, pad])
+                break
             audio_chunk = state.audio_buffer[:chunk_size]
             batch_feature = self._process_streaming_audio(audio_chunk, state.audio_chunk_idx, processor=processor)
             for name, value in (
@@ -224,6 +223,11 @@ class MiniCPMO45Stage0DuplexRuntime:
                 if units_built == 0:
                     return self._stage_prefill_result(False, start_time, "streaming audio embedding returned empty")
                 break
+            consumed_samples = self._consumed_audio_samples(
+                state.audio_chunk_idx,
+                chunk_size,
+                processor=processor,
+            )
             if state.audio_chunk_idx > 0:
                 # Official duplex closes every unit (finalize_unit feeds the
                 # sampled terminator + </unit>) before the next <unit> opens.
@@ -255,9 +259,7 @@ class MiniCPMO45Stage0DuplexRuntime:
             token_ids.extend(
                 [self._audio_embedding_placeholder_token_id()] * int(self._as_2d_tensor(audio_embeds).shape[0])
             )
-            state.audio_buffer = state.audio_buffer[
-                self._consumed_audio_samples(state.audio_chunk_idx, chunk_size, processor=processor) :
-            ]
+            state.audio_buffer = state.audio_buffer[consumed_samples:]
             state.audio_chunk_idx += 1
             units_built += 1
             chunk_size = self._streaming_chunk_size(processor)
@@ -610,14 +612,33 @@ class MiniCPMO45Stage0DuplexRuntime:
     def _load_processor_from_path(model_path: str | None) -> Any | None:
         if not model_path:
             return None
-        if model_path not in sys.path:
-            sys.path.insert(0, model_path)
         try:
-            from processing_minicpmo import MiniCPMOProcessor
+            from transformers import AutoImageProcessor, AutoProcessor
 
-            return MiniCPMOProcessor.from_pretrained(model_path, trust_remote_code=True)
-        except Exception:
-            return None
+            original_register = AutoImageProcessor.register
+
+            def register_image_processor(config_class, *args, **kwargs):
+                # The checkpoint's auto_map already loads this class. Its
+                # legacy string registration is incompatible with some
+                # Transformers versions and is otherwise redundant.
+                if config_class == "MiniCPMVImageProcessor":
+                    return None
+                return original_register(config_class, *args, **kwargs)
+
+            with _MINICPMO45_PROCESSOR_LOAD_LOCK:
+                AutoImageProcessor.register = staticmethod(register_image_processor)
+                try:
+                    processor = AutoProcessor.from_pretrained(
+                        model_path,
+                        trust_remote_code=True,
+                    )
+                finally:
+                    AutoImageProcessor.register = staticmethod(original_register)
+        except Exception as exc:
+            raise RuntimeError(f"Failed to load MiniCPM-o duplex processor from {model_path!r}") from exc
+        if getattr(processor, "tokenizer", None) is None:
+            raise RuntimeError(f"MiniCPM-o duplex processor loaded from {model_path!r} does not expose a tokenizer")
+        return processor
 
     def _init_token_ids(self) -> None:
         if self.tokenizer is None:

vllm_omni/experimental/fullduplex/openai/session_runner.py

@@ -503,11 +503,45 @@ class DuplexSessionRunnerMixin:
             clear_completed_pending_silence()
             pending_silence = native.pending_silence_task
             if pending_silence is not None and not pending_silence.done():
+                if pending_silence is asyncio.current_task():
+                    return False
+                try:
+                    if not await pending_silence:
+                        return False
+                except asyncio.CancelledError:
+                    current = asyncio.current_task()
+                    if current is not None and current.cancelling():
+                        raise
+                    return False
+                except Exception:
+                    return False
+                clear_completed_pending_silence()
+                pending_silence = native.pending_silence_task
+                if pending_silence is not None and not pending_silence.done():
+                    return False
+            append_tail = actor.native_append_tail
+            if (append_tail is None or append_tail.done()) and real_native_input_waiting():
                 return False
-            if actor.native_append_tail is not None and not actor.native_append_tail.done():
-                return False
-            if real_native_input_waiting():
-                return False
+            continuation_delay_s = max(
+                0.0,
+                float(session.capabilities.chunk_period_ms or 1000) / 1000.0,
+            )
+            if continuation_delay_s > 0:
+                await asyncio.sleep(continuation_delay_s)
+                if (
+                    actor.native_append_tail is not append_tail
+                    or ((append_tail is None or append_tail.done()) and real_native_input_waiting())
+                    or self._native_silence_continuation_is_stale(
+                        session,
+                        request_id=request_id,
+                        response_id=response_id,
+                        response_owned=response_owned,
+                        expected_epoch=expected_epoch,
+                        expected_incarnation=expected_incarnation,
+                        expected_model_turn_id=expected_model_turn_id,
+                    )
+                ):
+                    return False
 
             def _still_valid() -> bool:
                 return not real_native_input_waiting() and not self._native_silence_continuation_is_stale(

vllm_omni/model_executor/models/minicpmo_4_5/batched_token2wav.py

@@ -348,9 +348,16 @@ class BatchedToken2Wav(nn.Module):
         previous: torch.Tensor,
         window: torch.Tensor,
     ) -> torch.Tensor:
-        overlap = int(window.shape[0] // 2)
+        overlap = min(
+            int(window.shape[0] // 2),
+            int(speech.shape[-1]),
+            int(previous.shape[-1]),
+        )
         result = speech.clone()
-        result[..., :overlap] = result[..., :overlap] * window[:overlap] + previous[..., -overlap:] * window[overlap:]
+        if overlap > 0:
+            result[..., :overlap] = (
+                result[..., :overlap] * window[:overlap] + previous[..., -overlap:] * window[-overlap:]
+            )
         return result
 
     def decode_batch(
@@ -360,6 +367,7 @@ class BatchedToken2Wav(nn.Module):
         states: list[BatchedToken2WavState],
         *,
         last_chunk: bool,
+        flush_encoder: bool = False,
     ) -> tuple[list[torch.Tensor], list[BatchedToken2WavState]]:
         batch_size = int(tokens.shape[0])
         if batch_size != len(states):
@@ -382,7 +390,7 @@ class BatchedToken2Wav(nn.Module):
         with self._autocast(tokens.device):
             hidden, conformer_cnn, conformer_att = self._encode_chunk(
                 tokens,
-                last_chunk=last_chunk,
+                last_chunk=last_chunk or flush_encoder,
                 cnn_cache=flow_cache["conformer_cnn_cache"],
                 att_cache=flow_cache["conformer_att_cache"],
             )

vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_code2wav.py

@@ -4,12 +4,12 @@
 
 from __future__ import annotations
 
-import hashlib
 import json
 import os
 import tempfile
 from collections.abc import Iterable, Mapping, Sequence
 from dataclasses import dataclass
+from hashlib import sha256
 from pathlib import Path
 from typing import Any
 
@@ -77,6 +77,13 @@ class _RequestState:
     token2wav: BatchedToken2WavState
 
 
+@dataclass
+class _RuntimePrompt:
+    cache_id: str
+    path: str
+    owners: set[str]
+
+
 @dataclass(frozen=True)
 class _WorkItem:
     output_index: int
@@ -89,9 +96,11 @@ class _WorkItem:
     last_chunk: bool
     tokens: torch.Tensor
     previous: _RequestState | None
-    segment_text: str
-    duplex_turn_id: int | None
-    duplex_epoch: int | None
+    runtime_prompt_key: str | None
+    duplex_epoch: int
+    duplex_turn_id: int
+    segment_text_utf8: torch.Tensor
+    tts_is_last_chunk: bool
     segment_end: bool
     turn_end: bool
     has_payload: bool = True
@@ -120,12 +129,17 @@ class MiniCPMO45Code2Wav(nn.Module):
         self.model_path = str(vllm_config.model_config.model)
         self.backend: BatchedToken2Wav | None = None
         self._states: dict[str, _RequestState] = {}
-        self._owned_prompt_wavs: dict[str, tuple[str, str]] = {}
+        self._runtime_prompts: dict[str, _RuntimePrompt] = {}
+        self._request_prompt_keys: dict[str, str] = {}
+        self._runtime_prompt_dir = tempfile.TemporaryDirectory(
+            prefix="minicpmo45-runtime-prompts-",
+        )
         extra = self._extra_config()
         self._min_batch_size = int(extra.get("code2wav_min_batch_size", 1))
         if self._min_batch_size < 1:
             raise ValueError("MiniCPM-o Code2Wav code2wav_min_batch_size must be >= 1")
         self._default_prompt_id = str(extra.get("prompt_cache_id", "HT_ref_audio"))
+        self._prompt_wav_explicit = "prompt_wav" in extra
         self._default_prompt_wav = str(
             extra.get(
                 "prompt_wav",
@@ -133,6 +147,34 @@ class MiniCPMO45Code2Wav(nn.Module):
             )
         )
 
+    def _resolve_model_root(self) -> Path:
+        model_root = Path(self.model_path)
+        if model_root.is_dir():
+            return model_root
+
+        from vllm_omni.model_executor.model_loader.weight_utils import (
+            download_weights_from_hf_specific,
+        )
+
+        model_config = self.vllm_config.model_config
+        load_config = getattr(self.vllm_config, "load_config", None)
+        model_root = Path(
+            download_weights_from_hf_specific(
+                self.model_path,
+                getattr(load_config, "download_dir", None),
+                allow_patterns=[
+                    "assets/HT_ref_audio.wav",
+                    "assets/token2wav/*",
+                ],
+                revision=getattr(model_config, "revision", None),
+                require_all=True,
+            )
+        )
+        self.model_path = str(model_root)
+        if not self._prompt_wav_explicit:
+            self._default_prompt_wav = str(model_root / "assets" / "HT_ref_audio.wav")
+        return model_root
+
     def _extra_config(self) -> dict[str, Any]:
         model_config = getattr(self.vllm_config, "model_config", None)
         connector = getattr(model_config, "stage_connector_config", None)
@@ -142,63 +184,122 @@ class MiniCPMO45Code2Wav(nn.Module):
             extra = getattr(connector, "extra", None)
         return dict(extra) if isinstance(extra, Mapping) else {}
 
-    def _release_prompt(self, state_id: str) -> None:
-        owned = self._owned_prompt_wavs.pop(state_id, None)
-        if owned is None:
-            return
-        prompt_cache_id, prompt_wav = owned
-        if self.backend is not None:
-            self.backend.evict_prompt(prompt_cache_id, prompt_wav)
-        try:
-            os.unlink(prompt_wav)
-        except FileNotFoundError:
-            pass
-
-    def _write_reference_prompt(
-        self,
-        state_id: str,
-        reference: object,
-        sample_rate: object,
-    ) -> tuple[str, str]:
-        waveform = torch.as_tensor(reference, dtype=torch.float32).reshape(-1).cpu()
-        if waveform.numel() == 0:
-            return self._default_prompt_id, self._default_prompt_wav
-        sample_rate_hz = int(_scalar(sample_rate, 24000))
-        if sample_rate_hz <= 0:
-            raise _batch_error(
-                "invalid_ref_audio_sample_rate",
-                request_id=state_id,
-                sample_rate=sample_rate_hz,
-            )
-        digest = hashlib.sha256()
-        digest.update(str(sample_rate_hz).encode("ascii"))
-        digest.update(waveform.numpy().tobytes())
-        prompt_cache_id = f"{state_id}:{digest.hexdigest()}"
-        current = self._owned_prompt_wavs.get(state_id)
-        if current is not None and current[0] == prompt_cache_id and os.path.exists(current[1]):
-            return current
-
-        self._release_prompt(state_id)
-        handle = tempfile.NamedTemporaryFile(prefix="minicpmo45-ref-", suffix=".wav", delete=False)
-        prompt_wav = handle.name
-        handle.close()
-        try:
-            sf.write(prompt_wav, waveform.numpy(), sample_rate_hz, format="WAV")
-        except Exception:
-            try:
-                os.unlink(prompt_wav)
-            except FileNotFoundError:
-                pass
-            raise
-        self._owned_prompt_wavs[state_id] = (prompt_cache_id, prompt_wav)
-        return prompt_cache_id, prompt_wav
-
     def embed_input_ids(self, input_ids: torch.Tensor, **_: Any) -> torch.Tensor:
         return torch.zeros((input_ids.numel(), 1), device=input_ids.device, dtype=torch.float32)
 
     def compute_logits(self, hidden_states: Any, sampling_metadata: Any = None) -> None:
         return None
 
+    def _materialize_runtime_prompt(
+        self,
+        ref_audio: Any,
+        sample_rate: Any,
+    ) -> tuple[str, _RuntimePrompt]:
+        sample_rate_hz = int(_scalar(sample_rate, 0))
+        waveform = torch.as_tensor(ref_audio, dtype=torch.float32).reshape(-1).cpu().contiguous()
+        if sample_rate_hz <= 0:
+            raise _batch_error("invalid_ref_audio_sample_rate", sample_rate=sample_rate_hz)
+        if waveform.numel() == 0:
+            raise _batch_error("empty_ref_audio")
+        if not bool(torch.isfinite(waveform).all().item()):
+            raise _batch_error("non_finite_ref_audio")
+
+        digest = sha256()
+        digest.update(waveform.numpy().tobytes())
+        digest.update(str(sample_rate_hz).encode())
+        cache_key = digest.hexdigest()
+        cache_id = f"runtime-ref-{cache_key[:24]}-{sample_rate_hz}"
+        path = str(Path(self._runtime_prompt_dir.name) / f"minicpmo45_ref_{cache_key[:24]}_{sample_rate_hz}.wav")
+        entry = self._runtime_prompts.get(cache_key)
+        if entry is None:
+            entry = _RuntimePrompt(cache_id=cache_id, path=path, owners=set())
+            self._runtime_prompts[cache_key] = entry
+        prompt_path = Path(entry.path)
+        if not prompt_path.is_file():
+            with tempfile.NamedTemporaryFile(
+                dir=prompt_path.parent,
+                prefix=f".{prompt_path.stem}-",
+                suffix=".wav",
+                delete=False,
+            ) as handle:
+                temporary_path = Path(handle.name)
+            try:
+                sf.write(
+                    temporary_path,
+                    waveform.numpy(),
+                    sample_rate_hz,
+                    format="WAV",
+                )
+                os.replace(temporary_path, prompt_path)
+            finally:
+                temporary_path.unlink(missing_ok=True)
+        return cache_key, entry
+
+    def _resolve_prompt(
+        self,
+        state_id: str,
+        info: Mapping[str, Any],
+        meta: Mapping[str, Any],
+        previous: _RequestState | None,
+    ) -> tuple[str, str, str | None]:
+        codes = info.get("codes")
+        ref_audio = codes.get("ref") if isinstance(codes, Mapping) else None
+        if ref_audio is not None:
+            cache_key, entry = self._materialize_runtime_prompt(
+                ref_audio,
+                meta.get("ref_audio_sr"),
+            )
+            return entry.cache_id, entry.path, cache_key
+
+        if previous is not None:
+            return previous.prompt_cache_id, previous.prompt_wav, self._request_prompt_keys.get(state_id)
+
+        cache_key = self._request_prompt_keys.get(state_id)
+        entry = self._runtime_prompts.get(cache_key) if cache_key is not None else None
+        if entry is not None:
+            return entry.cache_id, entry.path, cache_key
+
+        return (
+            str(_scalar(meta.get("prompt_cache_id"), self._default_prompt_id)),
+            str(_scalar(meta.get("prompt_wav"), self._default_prompt_wav)),
+            None,
+        )
+
+    def _release_request_prompt(self, state_id: str) -> None:
+        cache_key = self._request_prompt_keys.pop(state_id, None)
+        entry = self._runtime_prompts.get(cache_key) if cache_key is not None else None
+        if entry is None:
+            return
+        entry.owners.discard(state_id)
+        if entry.owners:
+            return
+        if self.backend is not None:
+            self.backend.evict_prompt(entry.cache_id, entry.path)
+        Path(entry.path).unlink(missing_ok=True)
+        self._runtime_prompts.pop(cache_key, None)
+
+    def _commit_runtime_prompt_owners(self, items: list[_WorkItem]) -> None:
+        for item in items:
+            cache_key = item.runtime_prompt_key
+            if cache_key is None:
+                continue
+            previous_key = self._request_prompt_keys.get(item.state_id)
+            if previous_key != cache_key:
+                self._release_request_prompt(item.state_id)
+            entry = self._runtime_prompts.get(cache_key)
+            if entry is not None:
+                entry.owners.add(item.state_id)
+                self._request_prompt_keys[item.state_id] = cache_key
+
+    def _prune_unowned_runtime_prompts(self) -> None:
+        for cache_key, entry in list(self._runtime_prompts.items()):
+            if entry.owners:
+                continue
+            if self.backend is not None:
+                self.backend.evict_prompt(entry.cache_id, entry.path)
+            Path(entry.path).unlink(missing_ok=True)
+            self._runtime_prompts.pop(cache_key, None)
+
     @staticmethod
     def _split_segments(input_ids: torch.Tensor, counts: Any) -> list[torch.Tensor]:
         flat = input_ids.reshape(-1)
@@ -247,9 +348,11 @@ class MiniCPMO45Code2Wav(nn.Module):
                 last_chunk=False,
                 tokens=segment.new_empty(0, dtype=torch.long),
                 previous=None,
-                segment_text="",
-                duplex_turn_id=None,
-                duplex_epoch=None,
+                runtime_prompt_key=None,
+                duplex_epoch=-1,
+                duplex_turn_id=-1,
+                segment_text_utf8=torch.empty(0, dtype=torch.uint8),
+                tts_is_last_chunk=False,
                 segment_end=False,
                 turn_end=False,
                 has_payload=False,
@@ -266,14 +369,15 @@ class MiniCPMO45Code2Wav(nn.Module):
                 chunk_seq=chunk_seq,
             )
         last_chunk = bool(_scalar(meta.get("last_chunk"), False))
+        tts_is_last_chunk = bool(_scalar(meta.get("tts_is_last_chunk"), False))
         codes = info.get("codes")
         audio = codes.get("audio") if isinstance(codes, Mapping) else None
-        reference = codes.get("ref") if isinstance(codes, Mapping) else None
         tokens = _codec_tensor(audio, segment)
-        if last_chunk and int(_scalar(meta.get("code_flat_numel"), tokens.numel())) == 0:
+        if int(_scalar(meta.get("code_flat_numel"), tokens.numel())) == 0:
             # The generation scheduler reserves one placeholder token for an
-            # empty terminal chunk. The producer's explicit length is the
-            # authority, so do not decode that placeholder as codec data.
+            # empty terminal or segment-boundary chunk. The producer's
+            # explicit length is the authority, so do not decode that
+            # placeholder as codec data.
             tokens = segment.new_empty(0, dtype=torch.long)
         previous = self._states.get(state_id)
         if previous is None:
@@ -299,7 +403,6 @@ class MiniCPMO45Code2Wav(nn.Module):
                     cache_epoch=cache_epoch,
                     chunk_seq=chunk_seq,
                 )
-            self._release_prompt(state_id)
             previous = None
         elif chunk_seq != previous.chunk_seq + 1:
             raise _batch_error(
@@ -308,22 +411,12 @@ class MiniCPMO45Code2Wav(nn.Module):
                 expected=previous.chunk_seq + 1,
                 actual=chunk_seq,
             )
-
-        explicit_prompt_id = _scalar(meta.get("prompt_cache_id"))
-        explicit_prompt_wav = _scalar(meta.get("prompt_wav"))
-        if previous is not None and explicit_prompt_id is None and explicit_prompt_wav is None:
-            prompt_cache_id = previous.prompt_cache_id
-            prompt_wav = previous.prompt_wav
-        elif reference is not None:
-            prompt_cache_id, prompt_wav = self._write_reference_prompt(
-                state_id,
-                reference,
-                meta.get("ref_audio_sr"),
-            )
-        else:
-            prompt_cache_id = str(explicit_prompt_id or self._default_prompt_id)
-            prompt_wav = str(explicit_prompt_wav or self._default_prompt_wav)
-
+        prompt_cache_id, prompt_wav, runtime_prompt_key = self._resolve_prompt(
+            state_id,
+            info,
+            meta,
+            previous,
+        )
         if previous is not None and prompt_cache_id != previous.prompt_cache_id:
             raise _batch_error(
                 "prompt_changed_midstream",
@@ -338,8 +431,9 @@ class MiniCPMO45Code2Wav(nn.Module):
                 expected=previous.prompt_wav,
                 actual=prompt_wav,
             )
-        duplex_turn_id = _scalar(meta.get("duplex_turn_id"))
-        duplex_epoch = _scalar(meta.get("duplex_epoch"))
+        segment_text_utf8 = meta.get("llm_output_text_utf8")
+        if not isinstance(segment_text_utf8, torch.Tensor):
+            segment_text_utf8 = torch.empty(0, dtype=torch.uint8)
         return _WorkItem(
             output_index=index,
             state_id=state_id,
@@ -351,9 +445,11 @@ class MiniCPMO45Code2Wav(nn.Module):
             last_chunk=last_chunk,
             tokens=tokens,
             previous=previous,
-            segment_text=str(_scalar(meta.get("native_duplex_segment_text"), "")),
-            duplex_turn_id=int(duplex_turn_id) if duplex_turn_id is not None else None,
-            duplex_epoch=int(duplex_epoch) if duplex_epoch is not None else None,
+            runtime_prompt_key=runtime_prompt_key,
+            duplex_epoch=int(_scalar(meta.get("duplex_epoch"), -1)),
+            duplex_turn_id=int(_scalar(meta.get("duplex_turn_id"), -1)),
+            segment_text_utf8=segment_text_utf8,
+            tts_is_last_chunk=tts_is_last_chunk,
             segment_end=bool(_scalar(meta.get("segment_end"), False)),
             turn_end=bool(_scalar(meta.get("turn_end"), False)),
         )
@@ -371,6 +467,7 @@ class MiniCPMO45Code2Wav(nn.Module):
             int(item.tokens.numel()),
             cache_signature,
             item.last_chunk,
+            item.tts_is_last_chunk,
             item.cache_epoch,
         )
 
@@ -435,26 +532,37 @@ class MiniCPMO45Code2Wav(nn.Module):
                 request_ids=len(state_ids),
             )
         items: list[_WorkItem] = []
-        for index, (state_id, segment, info) in enumerate(
-            zip(state_ids, segments, runtime_additional_information, strict=True)
-        ):
-            if not isinstance(info, Mapping):
-                raise _batch_error(
-                    "invalid_runtime_info",
-                    output_index=index,
-                    value_type=type(info).__name__,
-                )
-            items.append(self._parse_item(index, str(state_id), segment, info))
+        try:
+            for index, (state_id, segment, info) in enumerate(
+                zip(state_ids, segments, runtime_additional_information, strict=True)
+            ):
+                if not isinstance(info, Mapping):
+                    raise _batch_error(
+                        "invalid_runtime_info",
+                        output_index=index,
+                        value_type=type(info).__name__,
+                    )
+                items.append(self._parse_item(index, str(state_id), segment, info))
+        except Exception:
+            self._prune_unowned_runtime_prompts()
+            raise
         state_ids = [item.state_id for item in items]
         if len(state_ids) != len(set(state_ids)):
+            self._prune_unowned_runtime_prompts()
             raise _batch_error("duplicate_request_in_forward", request_ids=state_ids)
         outputs = [empty for _ in segments]
         sentinels = [item for item in items if item.last_chunk and item.tokens.numel() == 0]
+        segment_markers = [
+            item for item in items if not item.last_chunk and item.tts_is_last_chunk and item.tokens.numel() == 0
+        ]
         compute_items = [item for item in items if item.tokens.numel() > 0]
         invalid_empty = [
-            item.request_id for item in items if item.has_payload and not item.last_chunk and item.tokens.numel() == 0
+            item.request_id
+            for item in items
+            if item.has_payload and not item.last_chunk and not item.tts_is_last_chunk and item.tokens.numel() == 0
         ]
         if invalid_empty:
+            self._prune_unowned_runtime_prompts()
             raise _batch_error("empty_nonfinal_chunk", request_ids=invalid_empty)
 
         buckets: dict[tuple[Any, ...], list[_WorkItem]] = {}
@@ -470,6 +578,7 @@ class MiniCPMO45Code2Wav(nn.Module):
             if len(bucket) < self._min_batch_size
         ]
         if undersized:
+            self._prune_unowned_runtime_prompts()
             raise _batch_error(
                 "exact_shape_bucket_below_minimum",
                 minimum=self._min_batch_size,
@@ -477,6 +586,58 @@ class MiniCPMO45Code2Wav(nn.Module):
             )
 
         pending: dict[str, _RequestState | None] = {item.state_id: None for item in sentinels}
+        pending.update(
+            {
+                item.state_id: _RequestState(
+                    cache_epoch=item.cache_epoch,
+                    chunk_seq=item.chunk_seq,
+                    prompt_cache_id=item.prompt_cache_id,
+                    prompt_wav=item.prompt_wav,
+                    token2wav=item.previous.token2wav,
+                )
+                for item in segment_markers
+                if item.previous is not None
+            }
+        )
+        initial_marker_buckets: dict[tuple[str, str], list[_WorkItem]] = {}
+        for item in segment_markers:
+            if item.previous is None:
+                initial_marker_buckets.setdefault(
+                    (item.prompt_cache_id, item.prompt_wav),
+                    [],
+                ).append(item)
+        for bucket in initial_marker_buckets.values():
+            try:
+                features = self.backend.prepare_prompt(
+                    bucket[0].prompt_cache_id,
+                    bucket[0].prompt_wav,
+                )
+                states = self.backend.setup_batch(features, len(bucket))
+            except Exception as exc:
+                self._prune_unowned_runtime_prompts()
+                if isinstance(exc, RuntimeError) and str(exc).startswith("MiniCPMO45Code2WavBatchError "):
+                    raise
+                raise _batch_error(
+                    "backend_unsupported_or_failed",
+                    request_ids=[item.request_id for item in bucket],
+                    error_type=type(exc).__name__,
+                    error=str(exc),
+                ) from exc
+            if len(states) != len(bucket):
+                self._prune_unowned_runtime_prompts()
+                raise _batch_error(
+                    "backend_result_size_mismatch",
+                    expected=len(bucket),
+                    states=len(states),
+                )
+            for item, state in zip(bucket, states, strict=True):
+                pending[item.state_id] = _RequestState(
+                    cache_epoch=item.cache_epoch,
+                    chunk_seq=item.chunk_seq,
+                    prompt_cache_id=item.prompt_cache_id,
+                    prompt_wav=item.prompt_wav,
+                    token2wav=state,
+                )
         for bucket in buckets.values():
             batch_size = len(bucket)
             try:
@@ -496,6 +657,7 @@ class MiniCPMO45Code2Wav(nn.Module):
                     last_chunk=bucket[0].last_chunk,
                 )
             except Exception as exc:
+                self._prune_unowned_runtime_prompts()
                 if isinstance(exc, RuntimeError) and str(exc).startswith("MiniCPMO45Code2WavBatchError "):
                     raise
                 raise _batch_error(
@@ -505,6 +667,7 @@ class MiniCPMO45Code2Wav(nn.Module):
                     error=str(exc),
                 ) from exc
             if len(audios) != batch_size or len(next_states) != batch_size:
+                self._prune_unowned_runtime_prompts()
                 raise _batch_error(
                     "backend_result_size_mismatch",
                     expected=batch_size,
@@ -525,10 +688,10 @@ class MiniCPMO45Code2Wav(nn.Module):
                     )
                 )
 
+        self._commit_runtime_prompt_owners(items)
         for request_id, state in pending.items():
             if state is None:
                 self._states.pop(request_id, None)
-                self._release_prompt(request_id)
             else:
                 self._states[request_id] = state
         sample_rate_tensor = torch.as_tensor(sample_rate, dtype=torch.int32)
@@ -537,24 +700,15 @@ class MiniCPMO45Code2Wav(nn.Module):
             multimodal_outputs={
                 "model_outputs": outputs,
                 "sr": [sample_rate_tensor.clone() for _ in outputs],
-                "meta.llm_output_text_utf8": [
-                    torch.tensor(list(item.segment_text.encode("utf-8")), dtype=torch.uint8) for item in items
-                ],
-                "meta.tts_is_last_chunk": [torch.tensor(item.last_chunk, dtype=torch.bool) for item in items],
+                # Generation runner wire payloads are flat and tensor-only.
+                # Dotted metadata keys are unflattened again by the output
+                # processor before the full-duplex data plane consumes them.
+                "meta.duplex_epoch": [torch.tensor(item.duplex_epoch, dtype=torch.int32) for item in items],
+                "meta.duplex_turn_id": [torch.tensor(item.duplex_turn_id, dtype=torch.int32) for item in items],
+                "meta.llm_output_text_utf8": [item.segment_text_utf8 for item in items],
+                "meta.tts_is_last_chunk": [torch.tensor(item.tts_is_last_chunk, dtype=torch.bool) for item in items],
                 "meta.segment_end": [torch.tensor(item.segment_end, dtype=torch.bool) for item in items],
                 "meta.turn_end": [torch.tensor(item.turn_end, dtype=torch.bool) for item in items],
-                "meta.duplex_turn_id": [
-                    torch.tensor(item.duplex_turn_id, dtype=torch.int64)
-                    if item.duplex_turn_id is not None
-                    else torch.empty(0, dtype=torch.int64)
-                    for item in items
-                ],
-                "meta.duplex_epoch": [
-                    torch.tensor(item.duplex_epoch, dtype=torch.int64)
-                    if item.duplex_epoch is not None
-                    else torch.empty(0, dtype=torch.int64)
-                    for item in items
-                ],
             },
         )
 
@@ -562,7 +716,7 @@ class MiniCPMO45Code2Wav(nn.Module):
         for request_id in finished_req_ids:
             state_id = str(request_id)
             self._states.pop(state_id, None)
-            self._release_prompt(state_id)
+            self._release_request_prompt(state_id)
 
     def make_omni_output(self, model_outputs: Any, **_: Any) -> OmniOutput:
         if isinstance(model_outputs, OmniOutput):
@@ -600,10 +754,11 @@ class MiniCPMO45Code2Wav(nn.Module):
             from stepaudio2.token2wav import Token2wav
 
         extra = self._extra_config()
+        model_root = self._resolve_model_root()
         prompt_path = Path(self._default_prompt_wav)
         if not prompt_path.is_file():
             raise FileNotFoundError(f"MiniCPM-o Code2Wav prompt audio not found: {prompt_path}")
-        token2wav_path = Path(self.model_path) / "assets" / "token2wav"
+        token2wav_path = model_root / "assets" / "token2wav"
         if not token2wav_path.is_dir():
             raise FileNotFoundError(f"MiniCPM-o Code2Wav assets not found: {token2wav_path}")
         use_float16 = bool(extra.get("token2wav_float16", False))

vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_omni.py

@@ -175,7 +175,6 @@ class MiniCPMO45OmniForConditionalGeneration(nn.Module, SupportsMultiModal, Supp
 
         token_ids = self._minicpmo45_native_duplex_token_ids()
         listen_id = int(token_ids.get("listen_token_id", -1))
-        tts_bos_id = int(token_ids.get("tts_bos_token_id", -1))
         turn_eos_id = int(token_ids.get("turn_eos_token_id", -1))
         if listen_id < 0 or listen_id >= logits.shape[-1]:
             return
@@ -199,7 +198,6 @@ class MiniCPMO45OmniForConditionalGeneration(nn.Module, SupportsMultiModal, Supp
                 continue
             force_listen = payload.get("force_listen") is True
             is_speech = payload.get("is_speech")
-            redirect_listen = False
             segment_key = (row.request_id, row.seq if row.seq is not None else -1)
             session_key = (row.session_id, row.incarnation) if row.session_id is not None else None
             if turn_eos_id >= 0 and session_key is not None:
@@ -209,8 +207,6 @@ class MiniCPMO45OmniForConditionalGeneration(nn.Module, SupportsMultiModal, Supp
                 )
                 if is_speech is True:
                     if state is not None:
-                        if not getattr(state, "current_turn_ended", True):
-                            redirect_listen = True
                         with suppress(Exception):
                             state.last_terminator_token = None
                 else:
@@ -218,7 +214,7 @@ class MiniCPMO45OmniForConditionalGeneration(nn.Module, SupportsMultiModal, Supp
                     if turn_ended and not pending_speech_context:
                         force_listen = True
 
-            if not force_listen and not redirect_listen:
+            if not force_listen:
                 continue
             if force_listen and segment_key in force_listen_segments:
                 continue
@@ -226,10 +222,6 @@ class MiniCPMO45OmniForConditionalGeneration(nn.Module, SupportsMultiModal, Supp
                 logits[row_idx, :] = float("-inf")
                 logits[row_idx, listen_id] = 0.0
                 force_listen_segments.add(segment_key)
-            elif 0 <= tts_bos_id < logits.shape[-1]:
-                if logits[row_idx, listen_id] > logits[row_idx, tts_bos_id]:
-                    logits[row_idx, tts_bos_id] = logits[row_idx, listen_id]
-                logits[row_idx, listen_id] = float("-inf")
 
     # -------------------- Device utilities --------------------
     @staticmethod
@@ -730,6 +722,10 @@ class MiniCPMO45OmniForConditionalGeneration(nn.Module, SupportsMultiModal, Supp
         output_token_ids = getattr(sampling_metadata, "output_token_ids", None) or []
         raw_recent_tokens = output_token_ids[row_idx] if row_idx < len(output_token_ids) else []
         recent_tokens = [int(token_id) for token_id in raw_recent_tokens if isinstance(token_id, int) and token_id >= 0]
+        temperature = float(self._sampling_metadata_value(sampling_metadata, "temperature", row_idx, 0.7))
+        top_k = int(self._sampling_metadata_value(sampling_metadata, "top_k", row_idx, 100))
+        top_p = float(self._sampling_metadata_value(sampling_metadata, "top_p", row_idx, 0.8))
+        state = self._minicpmo45_duplex_state_for_row(row_idx)
         if chunk_eos_id >= 0 and chunk_eos_id < logits.shape[-1]:
             max_speak_tokens = int(
                 getattr(
@@ -765,23 +761,19 @@ class MiniCPMO45OmniForConditionalGeneration(nn.Module, SupportsMultiModal, Supp
             if valid_forbidden:
                 logits[:, valid_forbidden] = float("-inf")
 
-        special_ids = self._minicpmo45_native_special_token_ids(token_ids)
-        state = self._minicpmo45_duplex_state_for_row(row_idx)
-        generated_text_tokens = getattr(state, "generated_text_tokens", None)
-        repetition_tokens = generated_text_tokens if generated_text_tokens else recent_tokens
+        generated_tokens = getattr(state, "generated_tokens", None)
+        repetition_tokens = generated_tokens if generated_tokens else recent_tokens
         repetition_penalty = 1.05
         if repetition_penalty != 1.0 and repetition_tokens:
             history_size = MiniCPMO45DuplexPolicy.REPETITION_HISTORY_SIZE
             for token_id in set(repetition_tokens[-history_size:]):
-                if token_id in special_ids or token_id < 0 or token_id >= logits.shape[-1]:
+                if token_id < 0 or token_id >= logits.shape[-1]:
                     continue
                 logits[0, token_id] /= repetition_penalty
 
-        temperature = float(self._sampling_metadata_value(sampling_metadata, "temperature", row_idx, 0.7))
-        top_k = int(self._sampling_metadata_value(sampling_metadata, "top_k", row_idx, 100))
-        top_p = float(self._sampling_metadata_value(sampling_metadata, "top_p", row_idx, 0.8))
         if getattr(sampling_metadata, "all_greedy", False) or temperature <= 0:
             sampled = int(torch.argmax(logits, dim=-1).item())
+            self._record_minicpmo45_duplex_generation_token(row_idx, sampled)
             sampled = self._maybe_cut_minicpmo45_native_duplex_text_chunk(
                 sampled,
                 recent_tokens,
@@ -793,6 +785,7 @@ class MiniCPMO45OmniForConditionalGeneration(nn.Module, SupportsMultiModal, Supp
         logits = self._top_k_top_p_filter(logits, top_k=top_k, top_p=top_p)
         probs = F.softmax(logits, dim=-1)
         sampled = int(torch.multinomial(probs, num_samples=1, generator=generator).item())
+        self._record_minicpmo45_duplex_generation_token(row_idx, sampled)
         sampled = self._maybe_cut_minicpmo45_native_duplex_text_chunk(
             sampled,
             recent_tokens,
@@ -901,6 +894,28 @@ class MiniCPMO45OmniForConditionalGeneration(nn.Module, SupportsMultiModal, Supp
             return int(tts_bos_id)
         return int(sampled)
 
+    def _record_minicpmo45_duplex_generation_token(self, row_idx: int, sampled: int) -> None:
+        """Track tokens returned by the model-policy decoder.
+
+        The released StreamDecoder is constructed without special_token_ids, so
+        every normally decoded token participates in repetition penalty. Forced
+        listen bypasses that decoder, while chunk_eos boundary decisions return
+        before this method is called.
+        """
+        state = self._minicpmo45_duplex_state_for_row(row_idx)
+        if state is None:
+            return
+        payload = self._minicpmo45_duplex_payload_for_row(row_idx)
+        if isinstance(payload, dict) and payload.get("force_listen") is True:
+            return
+        generated_tokens = getattr(state, "generated_tokens", None)
+        if not isinstance(generated_tokens, list):
+            generated_tokens = []
+            state.generated_tokens = generated_tokens
+        generated_tokens.append(int(sampled))
+        history_size = MiniCPMO45DuplexPolicy.REPETITION_HISTORY_SIZE
+        del generated_tokens[:-history_size]
+
     def _record_minicpmo45_duplex_terminator(self, row_idx: int, sampled: int, token_ids: dict[str, int]) -> None:
         """Remember sampled unit state for the next append.
 
@@ -928,15 +943,6 @@ class MiniCPMO45OmniForConditionalGeneration(nn.Module, SupportsMultiModal, Supp
                 with suppress(Exception):
                     state.pending_speech_response_open = False
             return
-        special_ids = self._minicpmo45_native_special_token_ids(token_ids)
-        if sampled not in special_ids:
-            generated_text_tokens = getattr(state, "generated_text_tokens", None)
-            if not isinstance(generated_text_tokens, list):
-                generated_text_tokens = []
-                state.generated_text_tokens = generated_text_tokens
-            generated_text_tokens.append(int(sampled))
-            history_size = MiniCPMO45DuplexPolicy.REPETITION_HISTORY_SIZE
-            del generated_text_tokens[:-history_size]
         if (
             sampled == tts_bos_id
             and getattr(state, "current_turn_ended", True)

vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_omni_tts.py

@@ -43,6 +43,7 @@ _CODEC_TOP_K = 25
 _CODEC_TOP_P = 0.85
 _CODEC_REPETITION_PENALTY = 1.05
 _CODEC_MIN_TOKENS = 50
+_DUPLEX_CODEC_TOKENS_PER_CHUNK = 26
 
 
 def _max_audio_tokens(condition_tokens: int) -> int:
@@ -61,12 +62,7 @@ def _max_audio_tokens(condition_tokens: int) -> int:
 
 def _restore_weight_norm_weight(weight_g: torch.Tensor, weight_v: torch.Tensor) -> torch.Tensor:
     """Materialize ``weight_norm(..., dim=0)`` checkpoint parameters."""
-    row_norm = torch.linalg.vector_norm(
-        weight_v.float(),
-        dim=1,
-        keepdim=True,
-    ).clamp_min(torch.finfo(torch.float32).tiny)
-    return weight_v * (weight_g.to(dtype=weight_v.dtype) / row_norm.to(dtype=weight_v.dtype))
+    return torch._weight_norm(weight_v, weight_g, dim=0)
 
 
 def _apply_repetition_penalty(
@@ -218,6 +214,8 @@ class MiniCPMO45OmniTTSForConditionalGeneration(nn.Module, SupportsPP):
         self,
         tts_token_ids: torch.Tensor,
         tts_hidden_states: torch.Tensor,
+        *,
+        native_duplex: bool = False,
     ) -> torch.Tensor:
         if tts_token_ids.numel() == 0 or tts_hidden_states.numel() == 0:
             # The thinker can legally emit an empty speech segment (<|tts_bos|>
@@ -239,7 +237,12 @@ class MiniCPMO45OmniTTSForConditionalGeneration(nn.Module, SupportsPP):
         hidden_embeds = self.projector_semantic(hidden)
         if self._normalize:
             hidden_embeds = F.normalize(hidden_embeds, p=2, dim=-1)
-        return torch.cat([text_embeds + hidden_embeds, self._boundary_embeddings()], dim=0)
+        audio_bos = self.emb_text(torch.tensor([self._tts_bos_id], device=device, dtype=torch.long))
+        condition = text_embeds + hidden_embeds
+        if native_duplex:
+            # Match MiniCPMTTS.generate_chunk's streaming condition.
+            return torch.cat([condition, audio_bos], dim=0)
+        return torch.cat([condition, self._boundary_embeddings()], dim=0)
 
     def preprocess(
         self,
@@ -281,18 +284,22 @@ class MiniCPMO45OmniTTSForConditionalGeneration(nn.Module, SupportsPP):
                     "MiniCPM-o Talker received an empty condition (request %s); this request produces no audio.",
                     info_dict.get("request_id"),
                 )
-            full_embeds = self._build_condition_embeddings(token_ids, hidden_states)
+            native_duplex = bool(info_dict.get("native_duplex", False))
+            full_embeds = self._build_condition_embeddings(
+                token_ids,
+                hidden_states,
+                native_duplex=native_duplex,
+            )
             offset = int(info_dict.get("_omni_num_computed_tokens", 0))
-            if offset == 0 and span_len > full_embeds.shape[0]:
-                # Async-chunk prewarms Stage 1 with placeholder token IDs. Two
-                # Thinker handoffs can arrive before the first one executes,
-                # so the scheduler prompt may already include the newer
-                # placeholder prefix while additional_information still
-                # carries the preceding condition. Materialize that prefix as
-                # the embeddings of its actual zero token IDs and keep the
-                # complete Talker condition tail-aligned (audio BOS remains
-                # the final prefill token).
-                prefix_len = span_len - full_embeds.shape[0]
+            request_id = str(info_dict.get("request_id", "0"))
+            meta = info_dict.get("meta")
+            # The handoff rebuilds only the tail-aligned Talker condition.
+            # Materialize zero-token embeddings for any scheduler prompt
+            # prefix so chunked prefill can slice from a non-zero offset.
+            prompt_len = info_dict.get("_omni_prompt_len")
+            target_len = int(prompt_len) if prompt_len is not None else offset + span_len
+            prefix_len = target_len - full_embeds.shape[0]
+            if prefix_len > 0:
                 placeholder_ids = torch.zeros(
                     prefix_len,
                     dtype=torch.long,
@@ -308,13 +315,21 @@ class MiniCPMO45OmniTTSForConditionalGeneration(nn.Module, SupportsPP):
                     f"tts_ids={token_ids.shape[0]} tts_hidden={hidden_states.shape[0]} "
                     f"prompt_len={info_dict.get('_omni_prompt_len')}"
                 )
-            max_tokens = _max_audio_tokens(int(token_ids.numel()))
+            duplex_boundary = isinstance(meta, dict) and (
+                bool(meta.get("turn_start", False)) or bool(meta.get("turn_end", False))
+            )
+            if native_duplex:
+                max_tokens = _DUPLEX_CODEC_TOKENS_PER_CHUNK
+                min_tokens = 0 if duplex_boundary else _DUPLEX_CODEC_TOKENS_PER_CHUNK
+            else:
+                max_tokens = _max_audio_tokens(int(token_ids.numel()))
+                min_tokens = self._codec_min_tokens
             state = {
                 "step": 0,
                 "max_tokens": max_tokens,
+                "min_tokens": min_tokens,
                 "finished": empty_condition,
             }
-            request_id = str(info_dict.get("request_id", "0"))
             request_states = getattr(self, "_request_audio_states", None)
             if request_states is None:
                 request_states = {}
@@ -370,7 +385,12 @@ class MiniCPMO45OmniTTSForConditionalGeneration(nn.Module, SupportsPP):
             penalty=self._codec_repetition_penalty,
             window_size=_REPETITION_WINDOW,
         )
-        if step < self._codec_min_tokens:
+        request_states = getattr(self, "_request_audio_states", {})
+        state = request_states.get(request_id)
+        min_tokens = (
+            int(state.get("min_tokens", self._codec_min_tokens)) if isinstance(state, dict) else self._codec_min_tokens
+        )
+        if step < min_tokens:
             logits[..., eos_id] = float("-inf")
         logits = _apply_top_k_top_p(
             logits,
@@ -404,12 +424,61 @@ class MiniCPMO45OmniTTSForConditionalGeneration(nn.Module, SupportsPP):
             raise RuntimeError(
                 f"MiniCPM-o continuous Talker received {len(sample_eligible)} sampling flags for {len(infos)} requests"
             )
+        emit_duplex_metadata = any(isinstance(info, dict) and info.get("native_duplex") is True for info in infos)
 
         stop_rows: list[torch.Tensor] = []
         codec_deltas: list[torch.Tensor] = []
         terminal_flags: list[torch.Tensor] = []
+        native_duplex_flags: list[torch.Tensor] = []
+        duplex_epochs: list[torch.Tensor] = []
+        duplex_turn_ids: list[torch.Tensor] = []
+        segment_texts_utf8: list[torch.Tensor] = []
+        turn_end_flags: list[torch.Tensor] = []
         empty_delta = hidden.new_empty((0, 1), dtype=torch.long)
         for index, info in enumerate(infos):
+            info_dict = info if isinstance(info, dict) else {}
+            native_duplex = info_dict.get("native_duplex") is True
+            if emit_duplex_metadata:
+                duplex_info = info_dict.get("duplex")
+                if not isinstance(duplex_info, dict):
+                    duplex_info = {}
+                epoch = duplex_info.get("epoch", -1)
+                turn_id = duplex_info.get("turn_id", -1)
+                if native_duplex and not all(
+                    isinstance(value, int) and not isinstance(value, bool) and value >= 0 for value in (epoch, turn_id)
+                ):
+                    raise RuntimeError(
+                        "MiniCPM-o native duplex Talker requires non-negative integer "
+                        f"epoch and turn_id, got epoch={epoch!r}, turn_id={turn_id!r}"
+                    )
+                meta_info = info_dict.get("meta")
+                if not isinstance(meta_info, dict):
+                    meta_info = {}
+                segment_text = meta_info.get("native_duplex_segment_text", "") if native_duplex else ""
+                if not isinstance(segment_text, str):
+                    segment_text = ""
+                turn_eos_id = meta_info.get("turn_eos_token_id")
+                ids_info = info_dict.get("ids")
+                tts_ids = ids_info.get("tts") if native_duplex and isinstance(ids_info, dict) else None
+                if isinstance(tts_ids, torch.Tensor):
+                    contains_turn_eos = isinstance(turn_eos_id, int) and bool(
+                        torch.any(tts_ids.reshape(-1) == turn_eos_id).item()
+                    )
+                elif isinstance(tts_ids, (list, tuple)):
+                    contains_turn_eos = isinstance(turn_eos_id, int) and turn_eos_id in tts_ids
+                else:
+                    contains_turn_eos = False
+                native_duplex_flags.append(torch.tensor(native_duplex, dtype=torch.bool))
+                duplex_epochs.append(torch.tensor(epoch if isinstance(epoch, int) else -1, dtype=torch.long))
+                duplex_turn_ids.append(torch.tensor(turn_id if isinstance(turn_id, int) else -1, dtype=torch.long))
+                segment_texts_utf8.append(
+                    torch.tensor(
+                        list(segment_text.encode("utf-8")),
+                        dtype=torch.uint8,
+                    )
+                )
+                turn_end_flags.append(torch.tensor(native_duplex and contains_turn_eos, dtype=torch.bool))
+
             if not isinstance(info, dict):
                 stop_rows.append(hidden.new_tensor([0.0, float("-inf")]))
                 codec_deltas.append(empty_delta)
@@ -453,12 +522,15 @@ class MiniCPMO45OmniTTSForConditionalGeneration(nn.Module, SupportsPP):
                 codes = codes.to(device=hidden.device, dtype=torch.long).reshape(-1)
             step = int(state.get("step", 0))
             sampled = self._sample_audio_code(hidden[end - 1 : end], codes, request_id, step)
-            is_eos = int(sampled.item()) == self._num_audio_tokens - 1
+            sampled_id = int(sampled.item())
+            is_eos = sampled_id == self._num_audio_tokens - 1
             state["step"] = int(state.get("step", 0)) + 1
             reached_limit = int(state["step"]) >= int(state.get("max_tokens", 2048))
             finished = is_eos or reached_limit
             state["finished"] = finished
-            if not is_eos:
+            # MiniCPMTTS.generate_chunk consumes the boundary sample but
+            # returns only codes that were fed into the retained KV state.
+            if not is_eos and not reached_limit:
                 codes = torch.cat([codes[-(_REPETITION_WINDOW - 1) :], sampled.reshape(1)])
                 delta = sampled.reshape(1, 1)
             else:
@@ -476,9 +548,20 @@ class MiniCPMO45OmniTTSForConditionalGeneration(nn.Module, SupportsPP):
         self._batch_stop_logits = torch.stack(stop_rows, dim=0) if stop_rows else hidden.new_empty((0, 2))
         # Lists are deliberate: the runner routes element i to request i,
         # preserving compaction alignment while emitting only this step's code.
+        meta_outputs = {"finished": terminal_flags}
+        if emit_duplex_metadata:
+            meta_outputs.update(
+                {
+                    "native_duplex": native_duplex_flags,
+                    "duplex_epoch": duplex_epochs,
+                    "duplex_turn_id": duplex_turn_ids,
+                    "llm_output_text_utf8": segment_texts_utf8,
+                    "turn_end": turn_end_flags,
+                }
+            )
         multimodal_outputs: dict[str, Any] = {
             "codes": {"audio": codec_deltas},
-            "meta": {"finished": terminal_flags},
+            "meta": meta_outputs,
         }
         return OmniOutput(
             text_hidden_states=hidden,

vllm_omni/model_executor/stage_input_processors/minicpmo_4_5_omni.py

@@ -21,6 +21,7 @@ logger = logging.getLogger(__name__)
 _MINICPMO45_ASYNC_STATE = "_minicpmo45_async_codec_state"
 _MINICPMO45_STREAM_RECORD = "_minicpmo45_async_stream_record"
 _MINICPMO45_SILENCE_CODE = 4218
+_MINICPMO45_MIN_STREAM_BODY_FRAMES = 5
 
 
 class _MiniCPMO45MetaStruct(MetaStruct):
@@ -28,6 +29,7 @@ class _MiniCPMO45MetaStruct(MetaStruct):
 
     ref_audio_sr: int | None = None
     native_duplex_segment_text: str | None = None
+    llm_output_text_utf8: torch.Tensor | None = None
     duplex_turn_id: int | None = None
     duplex_epoch: int | None = None
     segment_end: bool | None = None
@@ -212,6 +214,16 @@ def tts2code2wav_async_chunk(
     internal_id = getattr(request, "request_id", None)
     request_id = str(external_id if external_id is not None else internal_id)
     internal_id = str(internal_id if internal_id is not None else request_id)
+    output_meta = multimodal_output.get("meta") if isinstance(multimodal_output, Mapping) else None
+    output_meta = output_meta if isinstance(output_meta, Mapping) else {}
+    native_duplex = bool(_coerce_int(output_meta.get("native_duplex")))
+    duplex_epoch = _coerce_int(output_meta.get("duplex_epoch"))
+    duplex_turn_id = _coerce_int(output_meta.get("duplex_turn_id"))
+    segment_text_utf8 = output_meta.get("llm_output_text_utf8")
+    if not isinstance(segment_text_utf8, torch.Tensor):
+        segment_text_utf8 = None
+    turn_end = bool(_coerce_int(output_meta.get("turn_end")))
+    duplex_turn_key = (duplex_epoch, duplex_turn_id)
 
     request_payload = getattr(transfer_manager, "request_payload", None)
     if request_payload is None:
@@ -227,7 +239,9 @@ def tts2code2wav_async_chunk(
         record = {
             "internal_id": internal_id,
             "cache_epoch": 0,
+            "chunk_seq": 0,
             "retired_internal_ids": set(),
+            "last_terminal_turn": None,
         }
         container[_MINICPMO45_STREAM_RECORD] = record
     elif internal_id in record["retired_internal_ids"]:
@@ -236,6 +250,8 @@ def tts2code2wav_async_chunk(
         record["retired_internal_ids"].add(record["internal_id"])
         record["internal_id"] = internal_id
         record["cache_epoch"] = int(record["cache_epoch"]) + 1
+        record["chunk_seq"] = 0
+        record["last_terminal_turn"] = None
         _drop_codec_state(transfer_manager, request_id)
 
     if _is_aborted(request):
@@ -243,44 +259,46 @@ def tts2code2wav_async_chunk(
         _drop_codec_state(transfer_manager, request_id)
         return None
 
+    if native_duplex and turn_end and record.get("last_terminal_turn") == duplex_turn_key:
+        return None
+
     state = container.get(_MINICPMO45_ASYNC_STATE)
     if not isinstance(state, dict):
-        request_info = getattr(request, "additional_information", None)
-        if not isinstance(request_info, Mapping):
-            request_info = {}
-        codes_info = request_info.get("codes")
-        if not isinstance(codes_info, Mapping):
-            codes_info = {}
-        meta_info = request_info.get("meta")
-        if not isinstance(meta_info, Mapping):
-            meta_info = {}
-        duplex_info = request_info.get("duplex")
-        if not isinstance(duplex_info, Mapping):
-            duplex_info = {}
         state = {
             "internal_id": internal_id,
             "pending": [],
+            "pending_text_utf8": [],
+            "segment_text_recorded": False,
             "left_context": [],
             "codec_end": 0,
-            "ref_audio": codes_info.get("ref"),
-            "ref_audio_sr": meta_info.get("ref_audio_sr"),
-            "segment_text": meta_info.get("native_duplex_segment_text"),
-            "segment_end": bool(meta_info.get("segment_end", False)),
-            "turn_end": bool(meta_info.get("turn_end", False)),
-            "duplex_turn_id": duplex_info.get("model_turn_id", duplex_info.get("turn_id")),
-            "duplex_epoch": duplex_info.get("epoch"),
         }
         container[_MINICPMO45_ASYNC_STATE] = state
 
     pending = state["pending"]
     pending.extend(_extract_codec_delta(multimodal_output, request_id))
+    pending_text_utf8 = state.setdefault("pending_text_utf8", [])
+    current_text_utf8 = (
+        segment_text_utf8.detach().to(device="cpu", dtype=torch.uint8).reshape(-1).tolist()
+        if isinstance(segment_text_utf8, torch.Tensor)
+        else []
+    )
+    if native_duplex and current_text_utf8 and not state.get("segment_text_recorded", False):
+        # Talker repeats the unit text on every codec step. Queue it once for
+        # the first Code2Wav payload that can carry audio for this unit.
+        pending_text_utf8.extend(current_text_utf8)
+        state["segment_text_recorded"] = True
     request_finished = getattr(request, "is_finished", None)
     finished = bool(is_finished or (callable(request_finished) and request_finished()))
     chunk_frames, left_context_frames = _codec_config(transfer_manager)
-    if not finished and len(pending) < chunk_frames:
+    flush_pending = finished
+    last_chunk = bool(flush_pending and (not native_duplex or turn_end))
+    if not flush_pending and len(pending) < chunk_frames:
         return None
 
-    new_token_count = len(pending) if finished else chunk_frames
+    hold_short_unit = (
+        native_duplex and flush_pending and not last_chunk and 0 < len(pending) < _MINICPMO45_MIN_STREAM_BODY_FRAMES
+    )
+    new_token_count = 0 if hold_short_unit else (len(pending) if flush_pending else chunk_frames)
     new_codes = pending[:new_token_count]
     del pending[:new_token_count]
     codec_start = int(state["codec_end"])
@@ -292,53 +310,79 @@ def tts2code2wav_async_chunk(
         else:
             context = list(state["left_context"])
         output_codes = [*context, *new_codes]
-        history = [*state["left_context"], *new_codes]
+        history = output_codes
         state["left_context"] = history[-left_context_frames:] if left_context_frames else []
-    elif finished and codec_start > 0 and state["left_context"]:
+    elif last_chunk and codec_start > 0 and state["left_context"]:
         context = list(state["left_context"])
         output_codes = context
     else:
         context = []
         output_codes = []
     state["codec_end"] = codec_end
+    code_flat_numel = len(output_codes)
+    if native_duplex and code_flat_numel > 0:
+        segment_text_utf8 = torch.tensor(pending_text_utf8, dtype=torch.uint8)
+        pending_text_utf8.clear()
+    if native_duplex and finished:
+        state["segment_text_recorded"] = False
+    if flush_pending and not last_chunk and code_flat_numel == 0:
+        # Keep the generic generation connector model-agnostic: a real token
+        # makes this control-only TTS boundary schedulable, while the explicit
+        # zero length tells Code2Wav to discard the placeholder.
+        output_codes = [0]
 
-    last_chunk = bool(finished and not pending)
-    if last_chunk:
+    if last_chunk and not native_duplex:
         record["retired_internal_ids"].add(internal_id)
         _drop_codec_state(transfer_manager, request_id)
 
-    chunk_seq = int(getattr(transfer_manager, "put_req_chunk", {}).get(request_id, 0))
+    chunk_seq = int(record["chunk_seq"])
+    record["chunk_seq"] = chunk_seq + 1
+    ref_audio = None
+    ref_audio_sr = None
+    if int(record["cache_epoch"]) == 0 and chunk_seq == 0:
+        request_info = getattr(request, "additional_information", None)
+        if isinstance(request_info, Mapping):
+            codes_info = request_info.get("codes")
+            meta_info = request_info.get("meta")
+            raw_ref_audio = codes_info.get("ref") if isinstance(codes_info, Mapping) else None
+            raw_ref_audio_sr = meta_info.get("ref_audio_sr") if isinstance(meta_info, Mapping) else None
+            ref_audio_sr = _coerce_int(raw_ref_audio_sr)
+            if raw_ref_audio is not None:
+                ref_audio = torch.as_tensor(raw_ref_audio, dtype=torch.float32).reshape(-1).cpu()
     finished_tensor = torch.tensor(last_chunk, dtype=torch.bool)
-    ref_audio = state.get("ref_audio") if codec_start == 0 else None
-    return OmniPayloadStruct(
+    payload = OmniPayloadStruct(
         codes=CodesStruct(
             audio=torch.tensor(output_codes, dtype=torch.long),
-            ref=torch.as_tensor(ref_audio, dtype=torch.float32).reshape(-1) if ref_audio is not None else None,
+            ref=ref_audio,
         ),
         meta=_MiniCPMO45MetaStruct(
             request_id=request_id,
             chunk_seq=chunk_seq,
             cache_epoch=int(record["cache_epoch"]),
-            code_flat_numel=len(output_codes),
+            code_flat_numel=code_flat_numel,
             codec_chunk_frames=new_token_count,
             codec_left_context_frames=len(context),
             left_context_size=len(context),
             last_chunk=last_chunk,
             stream_finished=finished_tensor,
             finished=finished_tensor,
+            is_segment_finished=finished_tensor,
             req_id=[request_id],
-            ref_audio_sr=_coerce_int(state.get("ref_audio_sr")),
-            native_duplex_segment_text=(
-                str(state["segment_text"]) if isinstance(state.get("segment_text"), str) else None
-            ),
-            duplex_turn_id=_coerce_int(state.get("duplex_turn_id")),
-            duplex_epoch=_coerce_int(state.get("duplex_epoch")),
-            segment_end=bool(state.get("segment_end", False)),
-            turn_end=bool(state.get("turn_end", False)),
-            tts_is_last_chunk=last_chunk,
+            duplex_epoch=duplex_epoch,
+            duplex_turn_id=duplex_turn_id,
+            llm_output_text_utf8=segment_text_utf8,
+            tts_is_last_chunk=flush_pending,
+            turn_end=turn_end and last_chunk,
+            ref_audio_sr=ref_audio_sr,
         ),
         request_id=request_id,
     )
+    if last_chunk and native_duplex:
+        record["last_terminal_turn"] = duplex_turn_key
+        record["cache_epoch"] = int(record["cache_epoch"]) + 1
+        record["chunk_seq"] = 0
+        _drop_codec_state(transfer_manager, request_id)
+    return payload
 
 
 def tts2code2wav_full_payload(
@@ -496,13 +540,33 @@ def _native_tts_boundary_token_ids(special_token_ids):
     }
 
 
+def _decode_native_duplex_token_ids(
+    token_ids: list[int],
+    streaming_context,
+    *,
+    request_id: str,
+) -> str | None:
+    decode_token_ids = getattr(streaming_context, "source_token_decoder", None)
+    if not isinstance(decode_token_ids, Callable):
+        return None
+    decode_ids = [int(token_id) for token_id in token_ids]
+    try:
+        try:
+            return str(decode_token_ids(decode_ids, skip_special_tokens=True))
+        except TypeError:
+            return str(decode_token_ids(decode_ids))
+    except Exception:
+        logger.exception("Failed to decode MiniCPM-o duplex token delta for request_id=%s", request_id)
+        return ""
+
+
 def _native_duplex_segment_output_ids(
     output_ids: list[int],
     output_text: str,
     streaming_context,
     *,
     request_id: str,
-) -> tuple[list[int], str]:
+) -> tuple[list[int], str, bool]:
     """Slice the cumulative thinker output down to the current segment.
 
     Tracks how many output tokens were already handed to the talker in the
@@ -511,7 +575,7 @@ def _native_duplex_segment_output_ids(
     """
     bridge_states = getattr(streaming_context, "bridge_states", None)
     if not isinstance(bridge_states, dict):
-        return output_ids, output_text
+        return output_ids, output_text, True
     state = bridge_states.setdefault("minicpmo45_tts_handoff", {})
     duplex_state = bridge_states.get("duplex")
     turn_id = duplex_state.get("model_turn_id", duplex_state.get("turn_id")) if isinstance(duplex_state, dict) else None
@@ -521,17 +585,12 @@ def _native_duplex_segment_output_ids(
         state["request_id"] = request_id
         state["sent_output_len"] = 0
         state["sent_output_ids"] = []
-        state["acc_tts_ids"] = []
-        state["acc_tts_hidden"] = []
+    previous_turn_id = state.get("turn_id")
     sent_len = state.get("sent_output_len", 0)
     prev_output_ids = state.get("sent_output_ids", [])
-    prev_turn_id = state.get("turn_id")
     if not isinstance(sent_len, int) or sent_len < 0 or sent_len > len(output_ids):
-        # Shrunken cumulative output = epoch reset after barge-in: the talker
-        # condition history is stale too.
+        # Shrunken cumulative output = epoch reset after barge-in.
         sent_len = 0
-        state["acc_tts_ids"] = []
-        state["acc_tts_hidden"] = []
     elif sent_len and isinstance(prev_output_ids, list):
         prev_prefix = prev_output_ids[:sent_len]
         current_prefix = output_ids[:sent_len]
@@ -540,27 +599,15 @@ def _native_duplex_segment_output_ids(
             # others keep returning cumulative ids. Detect restart from the
             # token prefix instead of clearing the cursor at turn_eos.
             sent_len = 0
-            state["acc_tts_ids"] = []
-            state["acc_tts_hidden"] = []
-    if isinstance(prev_turn_id, int) and isinstance(turn_id, int) and prev_turn_id != turn_id:
-        # The thinker output list is cumulative across clean turns, so the
-        # output cursor must stay put. The talker condition is per assistant
-        # turn, though; carrying it across turn_id boundaries makes stage 1
-        # replay the previous turn after its consumed cursor has been reset.
-        state["acc_tts_ids"] = []
-        state["acc_tts_hidden"] = []
+    turn_start = sent_len == 0 or previous_turn_id != turn_id
     segment_ids = output_ids[sent_len:]
-    decode_token_ids = getattr(streaming_context, "source_token_decoder", None)
-    if isinstance(decode_token_ids, Callable):
-        decode_ids = [int(token_id) for token_id in segment_ids]
-        try:
-            try:
-                segment_text = str(decode_token_ids(decode_ids, skip_special_tokens=True))
-            except TypeError:
-                segment_text = str(decode_token_ids(decode_ids))
-        except Exception:
-            logger.exception("Failed to decode MiniCPM-o duplex token delta for request_id=%s", request_id)
-            segment_text = ""
+    decoded_segment_text = _decode_native_duplex_token_ids(
+        segment_ids,
+        streaming_context,
+        request_id=request_id,
+    )
+    if decoded_segment_text is not None:
+        segment_text = decoded_segment_text
     elif sent_len == 0:
         # Non-orchestrator unit tests and legacy callers may not install a
         # decoder. Never slice cumulative text with a stale character cursor.
@@ -575,22 +622,7 @@ def _native_duplex_segment_output_ids(
     state["sent_output_len"] = len(output_ids)
     state["sent_output_ids"] = list(output_ids)
     state["turn_id"] = turn_id
-    return segment_ids, segment_text
-
-
-def _reset_native_tts_handoff(streaming_context) -> None:
-    bridge_states = getattr(streaming_context, "bridge_states", None)
-    if not isinstance(bridge_states, dict):
-        return
-    state = bridge_states.get("minicpmo45_tts_handoff")
-    if not isinstance(state, dict):
-        return
-    # Keep the output cursor: the thinker can keep reporting cumulative
-    # output after turn_eos on the same resumable request. If a runtime really
-    # restarts token ids, _native_duplex_segment_output_ids detects the prefix
-    # mismatch and resets the slice cursor there.
-    state["acc_tts_ids"] = []
-    state["acc_tts_hidden"] = []
+    return segment_ids, segment_text, turn_start
 
 
 def _native_duplex_data_plane_metadata(streaming_context) -> dict[str, object] | None:
@@ -623,32 +655,6 @@ def _native_duplex_data_plane_metadata(streaming_context) -> dict[str, object] |
     return metadata
 
 
-def _accumulate_native_tts_handoff(streaming_context, new_ids, new_hidden):
-    """Hand the talker the FULL accumulated condition on every handoff.
-
-    The runner's streaming buffer update is not merge-safe for a resumable
-    stage-1 request: in-place updates merge sub-keys, but a resume prefill
-    REPLACES the buffer, silently dropping every earlier segment's tts
-    tokens/hiddens (observed losing alternating reply segments — the talker
-    vocalized text it never saw between islands it did). Accumulate here so
-    the latest handoff always carries the complete history and downstream
-    replace semantics are lossless; the talker consumes by cursor.
-    """
-    bridge_states = getattr(streaming_context, "bridge_states", None)
-    if not isinstance(bridge_states, dict):
-        return new_ids, new_hidden
-    state = bridge_states.setdefault("minicpmo45_tts_handoff", {})
-    acc_ids = state.setdefault("acc_tts_ids", [])
-    acc_hidden = state.setdefault("acc_tts_hidden", [])
-    if new_ids:
-        acc_ids.extend(int(t) for t in new_ids)
-        if new_hidden:
-            acc_hidden.extend(new_hidden)
-    if not acc_ids:
-        return None, None
-    return list(acc_ids), list(acc_hidden)
-
-
 def _build_tts_scheduler_prompt_token_ids(
     tts_token_ids: torch.Tensor | None,
     llm_output_ids: list[int],
@@ -723,6 +729,7 @@ def llm2tts(
         # buffer overflows.
         llm_output_ids = list(llm_output_ids)
         thinker_text = getattr(output, "text", "") or ""
+        native_turn_start = False
         if _has_native_duplex_prompt_metadata(mm_output):
             # The thinker's resumable duplex request reports cumulative
             # output ids/text, but earlier segments are already folded into
@@ -732,7 +739,7 @@ def llm2tts(
             # alignment holds, the talker prompt grows linearly with new
             # tokens instead of quadratically, and downstream transcripts
             # carry per-unit deltas instead of re-sending the whole reply.
-            llm_output_ids, thinker_text = _native_duplex_segment_output_ids(
+            llm_output_ids, thinker_text, native_turn_start = _native_duplex_segment_output_ids(
                 llm_output_ids,
                 thinker_text,
                 _streaming_context,
@@ -875,6 +882,18 @@ def llm2tts(
                         .to(torch.float32)
                         .contiguous()
                     )
+        handoff_ids = _coerce_token_id_list(tts_token_ids_slice) if tts_token_ids_slice is not None else None
+        if is_native_duplex_handoff and handoff_ids:
+            handoff_text = _decode_native_duplex_token_ids(
+                handoff_ids,
+                _streaming_context,
+                request_id=str(llm_output.request_id),
+            )
+            if handoff_text is not None:
+                # Match the released streaming_generate contract: transcript
+                # text comes from total_ids_in_unit, the same slice that
+                # conditions the Talker.
+                thinker_text = handoff_text
         model_intermediate_buffer = build_duplex_intermediate_buffer(
             request_id=str(llm_output.request_id),
             prompt_token_ids=prompt_token_ids,
@@ -900,6 +919,7 @@ def llm2tts(
                 # The talker detects turn end from <|turn_eos|> inside the
                 # handed condition (official conditions on its embedding).
                 meta["turn_eos_token_id"] = int(turn_eos_id)
+            meta["turn_start"] = native_turn_start
             if native_segment_end:
                 meta["segment_end"] = True
         req_mm_data = multi_modal_data.get(llm_output.request_id)
@@ -911,30 +931,27 @@ def llm2tts(
         if ref_audio is not None:
             ref_waveform, ref_sr = ref_audio
             set_ref_audio(model_intermediate_buffer, _to_transport_list(ref_waveform), ref_sr)
-        handoff_ids = _coerce_token_id_list(tts_token_ids_slice) if tts_token_ids_slice is not None else None
         handoff_hidden = _to_transport_list(tts_hidden_slice) if tts_hidden_slice is not None else None
         native_turn_end_handoff = False
         if is_native_duplex_handoff:
             turn_eos_id = special_token_ids.get("turn_eos_token_id")
             native_turn_end_handoff = turn_eos_id is not None and handoff_ids is not None and turn_eos_id in handoff_ids
-            handoff_ids, handoff_hidden = _accumulate_native_tts_handoff(
-                _streaming_context,
-                handoff_ids,
-                handoff_hidden,
-            )
             if not handoff_ids:
                 continue
         set_tts_handoff(model_intermediate_buffer, handoff_ids, handoff_hidden)
         if native_turn_end_handoff:
             model_intermediate_buffer.setdefault("meta", {})["turn_end"] = True
-            _reset_native_tts_handoff(_streaming_context)
 
         if handoff_ids is not None and handoff_hidden is not None:
-            condition_length = max(len(handoff_ids), len(handoff_hidden)) + 2
+            condition_suffix_length = 1 if is_native_duplex_handoff else 2
+            condition_length = max(len(handoff_ids), len(handoff_hidden)) + condition_suffix_length
             scheduler_prompt_token_ids = [0] * condition_length
             handoff_meta = model_intermediate_buffer.setdefault("meta", {})
-            handoff_meta["replace_streaming_prompt"] = True
             handoff_meta["next_stage_prompt_len"] = condition_length
+            # Native duplex resumes one Talker request within a turn, but a new
+            # assistant turn must discard the previous turn's prompt and KV.
+            if not is_native_duplex_handoff or native_turn_start:
+                handoff_meta["replace_streaming_prompt"] = True
         else:
             scheduler_prompt_token_ids = _build_tts_scheduler_prompt_token_ids(
                 tts_token_ids_slice,

vllm_omni/model_executor/stage_input_processors/qwen3_omni.py

@@ -685,19 +685,32 @@ def talker2code2wav_async_chunk(
     """
     Multimodal output version.
     """
-    if not isinstance(multimodal_output, Mapping):
-        return None
-    talker_codes = multimodal_output.get("codes", {})
-    if not isinstance(talker_codes, dict):
-        return None
-    code_predictor_codes = talker_codes.get("audio")
-    if code_predictor_codes is None:
-        return None
+    request_id = request.external_req_id
+    code_predictor_codes = None
+    if isinstance(multimodal_output, Mapping):
+        talker_codes = multimodal_output.get("codes", {})
+        if isinstance(talker_codes, dict):
+            code_predictor_codes = talker_codes.get("audio")
 
-    if code_predictor_codes.numel() == 0:
-        return None
+    sampling_params = getattr(request, "sampling_params", None)
+    stop_token_ids = set(getattr(sampling_params, "stop_token_ids", None) or [])
+    stop_token_id = getattr(sampling_params, "stop_token_id", None)
+    if stop_token_id is not None:
+        stop_token_ids.add(stop_token_id)
 
-    if not code_predictor_codes.any():
+    append_codes = (
+        isinstance(code_predictor_codes, torch.Tensor)
+        and code_predictor_codes.numel() > 0
+        and bool(code_predictor_codes.any())
+    )
+    if append_codes:
+        first_codebook = int(code_predictor_codes[0, 0].item())
+        if first_codebook in stop_token_ids:
+            logger.debug("skip stop-token codec frame: first_codebook=%s", first_codebook)
+            append_codes = False
+    if append_codes:
+        transfer_manager.code_prompt_token_ids[request_id].append(code_predictor_codes)
+    elif not is_finished:
         return None
 
     connector = getattr(transfer_manager, "connector", None)
@@ -707,20 +720,10 @@ def talker2code2wav_async_chunk(
     left_context_size_config = int(cfg.get("codec_left_context_frames", 25))
     configured_initial_chunk_size = int(cfg.get("initial_codec_chunk_frames") or 0)
 
-    sampling_params = getattr(request, "sampling_params", None)
-    stop_token_ids = set(getattr(sampling_params, "stop_token_ids", None) or [])
-    stop_token_id = getattr(sampling_params, "stop_token_id", None)
-    if stop_token_id is not None:
-        stop_token_ids.add(stop_token_id)
-    first_codebook = int(code_predictor_codes[0, 0].item())
-    if first_codebook in stop_token_ids:
-        logger.debug("skip stop-token codec frame: first_codebook=%s", first_codebook)
-        return None
-
-    request_id = request.external_req_id
     chunk_id = transfer_manager.put_req_chunk[request_id]
-    transfer_manager.code_prompt_token_ids[request_id].append(code_predictor_codes)
     length = len(transfer_manager.code_prompt_token_ids[request_id])
+    if length <= 0:
+        return None
 
     if configured_initial_chunk_size > 0:
         if chunk_id == 0:
@@ -732,6 +735,9 @@ def talker2code2wav_async_chunk(
     if chunk_length != 0 and not is_finished:
         return None
 
+    if is_finished and not append_codes and chunk_length == 0:
+        return None
+
     context_length = chunk_length if chunk_length != 0 else chunk_size_config
     # ensure left context does not exceed available length
     if configured_initial_chunk_size > 0 and chunk_id == 1:

vllm_omni/outputs/multimodal_accumulation.py

@@ -11,6 +11,7 @@ CHUNK_METADATA_KEYS: frozenset[str] = frozenset(
         "duplex_epoch",
         "duplex_turn_id",
         "llm_output_text_utf8",
+        "segment_end",
         "tts_is_last_chunk",
         "turn_end",
     }

vllm_omni/outputs/output_processor.py

@@ -523,24 +523,27 @@ class MultimodalOutputProcessor(VLLMOutputProcessor):
                 upstream_outputs.append(eco)
 
         # Handle multimodal-only outputs (generation stages) locally.
-        self._process_mm_only_outputs(mm_only_outputs)
+        mm_request_outputs = self._process_mm_only_outputs(mm_only_outputs)
 
         # Delegate text/pooling outputs to upstream.
-        return super().process_outputs(
+        processed = super().process_outputs(
             upstream_outputs,
             engine_core_timestamp=engine_core_timestamp,
             iteration_stats=iteration_stats,
         )
+        processed.request_outputs.extend(mm_request_outputs)
+        return processed
 
     def _process_mm_only_outputs(
         self,
         engine_core_outputs: list[EngineCoreOutput],
-    ) -> None:
+    ) -> list[OmniRequestOutput | PoolingRequestOutput]:
         """Handle outputs from generation stages that have no detokenizer.
 
         These cannot go through upstream process_outputs because it asserts
         detokenizer is not None when pooling_output is None.
         """
+        request_outputs: list[OmniRequestOutput | PoolingRequestOutput] = []
         for eco in engine_core_outputs:
             req_state = self.request_states.get(eco.request_id)
             if req_state is None or not isinstance(req_state, OmniRequestState):
@@ -554,6 +557,11 @@ class MultimodalOutputProcessor(VLLMOutputProcessor):
             req_state.num_cached_tokens = eco.num_cached_tokens
             req_state.is_prefilling = False
 
+            is_non_final_audio_chunk = (
+                finish_reason is not None
+                and req_state.output_kind == RequestOutputKind.DELTA
+                and is_non_final_delta_audio_chunk(req_state.mm_accumulated, req_state.mm_type)
+            )
             if request_output := req_state.make_request_output(
                 new_token_ids,
                 None,  # pooling_output
@@ -564,9 +572,13 @@ class MultimodalOutputProcessor(VLLMOutputProcessor):
             ):
                 if req_state.queue is not None:
                     req_state.queue.put(request_output)
+                else:
+                    request_outputs.append(request_output)
 
-            if finish_reason is not None:
+            is_segment_finished = bool(getattr(eco, "is_segment_finished", False))
+            if finish_reason is not None and not is_segment_finished and not is_non_final_audio_chunk:
                 self._finish_request(req_state)
+        return request_outputs
 
     def _update_stats_from_output(
         self,