[Bugfix] Fix Qwen3-Omni realtime playback interruption and conversation history (#7791)

Signed-off-by: psv666 <2693925048@qq.com>

tests/engine/duplex/test_duplex_events.py

@@ -11,7 +11,7 @@ import inspect
 import pytest
 
 from vllm_omni.engine.duplex import events as events_module
-from vllm_omni.engine.duplex.commands import TruncateItem
+from vllm_omni.engine.duplex.commands import ClearOutputAudio, TruncateItem
 from vllm_omni.engine.duplex.events import (
     REALTIME_ERROR_TYPES_BY_CODE,
     AudioDelta,
@@ -41,6 +41,7 @@ from vllm_omni.engine.duplex.events import (
 from vllm_omni.engine.duplex.realtime_events import (
     RealtimeProjectionState,
     project_internal_event,
+    resolve_clear_output_audio,
     resolve_truncate_item,
     retrieve_item_events,
 )
@@ -376,6 +377,72 @@ def test_projection_of_output_audio_buffer_clear_emits_cleared_before_terminals(
     assert state.item_truncation_cursors["item_resp_1"] == (0, 250)
 
 
+@pytest.mark.parametrize("reason", ["turn_detected", "barge_in", "output_audio_buffer_clear"])
+@pytest.mark.parametrize("committed_ms", [1000, 10000])
+def test_completed_response_cancellation_only_clears_pending_playback(reason, committed_ms):
+    state = RealtimeProjectionState(session_id="duplex-playback")
+    project_internal_event(state, {"type": "response.created", "response_id": "resp_1"})
+    project_internal_event(state, {"type": "response.done", "response_id": "resp_1"})
+
+    cancelled = project_internal_event(
+        state,
+        {
+            "type": "audio.cancelled",
+            "response_id": "resp_1",
+            "reason": reason,
+            "committed_ms": committed_ms,
+            "playback": {"sent_ms": 10000, "committed_ms": committed_ms},
+        },
+    )
+
+    if reason == "output_audio_buffer_clear" or committed_ms < 10000:
+        assert _types(cancelled) == ["output_audio_buffer.cleared"]
+        assert cancelled[0].response_id == "resp_1"
+    else:
+        assert cancelled == []
+
+
+@pytest.mark.parametrize("explicit_response_id", [False, True])
+def test_clear_latest_completed_response_reaches_playback_owner(explicit_response_id):
+    state = RealtimeProjectionState(session_id="duplex-clear")
+    project_internal_event(state, {"type": "response.created", "response_id": "resp_1"})
+    project_internal_event(state, {"type": "response.done", "response_id": "resp_1"})
+
+    control = resolve_clear_output_audio(
+        state, ClearOutputAudio(response_id="resp_1" if explicit_response_id else None, event_id="clear_1")
+    )
+
+    # Generation completion does not prove the client has drained its audio.
+    # A cleared event must follow the engine clearing its playback cursor.
+    assert control.events == []
+    assert control.payloads == [
+        {
+            "type": "output_audio_buffer.clear",
+            "reason": "output_audio_buffer.clear",
+            "response_id": "resp_1",
+            "realtime_event_id": "clear_1",
+        }
+    ]
+
+
+@pytest.mark.parametrize("new_response_finished", [False, True])
+def test_late_clear_of_completed_response_does_not_touch_new_response(new_response_finished):
+    state = RealtimeProjectionState(session_id="duplex-clear")
+    project_internal_event(state, {"type": "response.created", "response_id": "resp_1"})
+    project_internal_event(state, {"type": "response.done", "response_id": "resp_1"})
+    project_internal_event(state, {"type": "response.created", "response_id": "resp_2"})
+    if new_response_finished:
+        project_internal_event(state, {"type": "response.done", "response_id": "resp_2"})
+
+    control = resolve_clear_output_audio(state, ClearOutputAudio(response_id="resp_1"))
+
+    assert control.payloads == []
+    assert _types(control.events) == ["output_audio_buffer.cleared"]
+    assert control.events[0].response_id == "resp_1"
+    assert state.active_response_id == (None if new_response_finished else "resp_2")
+    assert state.last_response_id == "resp_2"
+
+
 def test_one_truncate_command_truncates_retrieved_transcript_once():
     state = RealtimeProjectionState(session_id="duplex-truncate")
     state.conversation_items["item_resp_1"] = {

tests/engine/duplex/test_engine_session.py

@@ -21,6 +21,7 @@ from vllm_omni.engine.duplex.config import (
 from vllm_omni.engine.duplex.contracts import DuplexFence
 from vllm_omni.engine.duplex.events import TurnEvent
 from vllm_omni.engine.duplex.session.engine_session import DuplexEngineSession, DuplexFenceMismatchError
+from vllm_omni.engine.duplex.session.playback_ledger import apply_playback_ack
 from vllm_omni.model_executor.models.minicpmo_4_5.duplex.capabilities import (
     minicpmo45_native_capabilities,
 )
@@ -241,6 +242,83 @@ def test_history_commit_with_ack_only_policy_defers_unacknowledged_text():
     assert f"item_{response_id}" in session.pending_history_item_ids
 
 
+@pytest.mark.parametrize("audio_complete", [False, True])
+def test_unaligned_response_keeps_empty_turn_before_later_user_input(audio_complete):
+    session = _session(config=DuplexSessionConfig(playback_commit_policy="ack_only"))
+    first = session.commit_audio_input(transcript="recite a poem")
+    response_id = session.begin_response()
+    session.reserve_history_item(f"item_{response_id}")
+    session.append_assistant_text("An answer the user has only partly heard")
+    session.mark_audio_sent(
+        duration_ms=10_000,
+        text_requires_complete_audio=True,
+        audio_complete=audio_complete,
+    )
+
+    assert session.end_response(commit_text=True) is None
+    session.barge_in()
+    session.clear_playback_cursor()
+    second = session.commit_audio_input(transcript="stop")
+    empty_answer = {"role": "assistant", "content": ""}
+    assert session.history == (first.message, empty_answer, second.message)
+
+    ack = apply_playback_ack(session, {"response_id": response_id, "played_ms": 6_981})
+    assert ack[0].to_realtime()["event"]["history_committed"] is False
+    assert session.history == (first.message, empty_answer, second.message)
+
+    if audio_complete:
+        ack = apply_playback_ack(session, {"response_id": response_id, "played_ms": 10_000})
+        assert ack[0].to_realtime()["event"]["history_committed"] is True
+        assert session.history == (
+            first.message,
+            {"role": "assistant", "content": "An answer the user has only partly heard"},
+            second.message,
+        )
+
+
+@pytest.mark.parametrize(("reserve_slot", "assistant_text"), [(True, ""), (False, "unplayed answer")])
+def test_unaligned_response_only_materializes_an_existing_answer_slot(reserve_slot, assistant_text):
+    session = _session(config=DuplexSessionConfig(playback_commit_policy="ack_only"))
+    first = session.commit_audio_input(transcript="first input")
+    response_id = session.begin_response()
+    if reserve_slot:
+        session.reserve_history_item(f"item_{response_id}")
+    session.append_assistant_text(assistant_text)
+    session.mark_audio_sent(duration_ms=10_000, text_requires_complete_audio=True, audio_complete=True)
+
+    assert session.end_response(commit_text=reserve_slot) is None
+    assert session.history == (first.message,)
+
+
+@pytest.mark.parametrize("operation", ["delete", "truncate"])
+def test_removing_one_empty_assistant_preserves_the_other_turn_boundary(operation):
+    session = _session(config=DuplexSessionConfig(playback_commit_policy="ack_only"))
+    user_inputs = []
+    response_ids = []
+    for transcript in ("first input", "second input"):
+        user_inputs.append(session.commit_audio_input(transcript=transcript).message)
+        response_id = session.begin_response()
+        response_ids.append(response_id)
+        session.reserve_history_item(f"item_{response_id}")
+        session.append_assistant_text("unplayed answer")
+        session.mark_audio_sent(duration_ms=10_000, text_requires_complete_audio=True, audio_complete=True)
+        session.end_response(commit_text=True)
+
+    item_id = f"item_{response_ids[1]}"
+    if operation == "delete":
+        assert session.delete_history_item(item_id) is True
+    else:
+        # A full ACK releases the pending snapshot. Truncation then edits the
+        # stored message into another empty dict before removing that item.
+        ack = apply_playback_ack(session, {"response_id": response_ids[1], "played_ms": 10_000})
+        assert ack[0].to_realtime()["event"]["history_committed"] is True
+        assert session.truncate_history_item(item_id, audio_end_ms=0, hard=True) is True
+    assert session.history == (user_inputs[0], {"role": "assistant", "content": ""}, user_inputs[1])
+    ack = apply_playback_ack(session, {"response_id": response_ids[1], "played_ms": 10_000})
+    assert ack[0].to_realtime()["error"]["code"] == "playback_item_not_found"
+    assert session.history == (user_inputs[0], {"role": "assistant", "content": ""}, user_inputs[1])
+
+
 def test_truncate_history_item_uses_response_alignment_marks():
     session = _session()
     response_id = session.begin_response()