[Bugfix] Fix Qwen3-Omni realtime playback interruption and conversation history (#7791)
Signed-off-by: psv666 <2693925048@qq.com>
This commit is contained in:
1 parent
d3396fc5fb
commit
2d0b03f717
13 files changed
+1292
-68
No files matched your search
@@ -946,6 +946,16 @@ def _project(state: RealtimeProjectionState, event: dict[str, object]) -> list[D
|
||||
if not isinstance(response_id, str) or not response_id:
|
||||
return events
|
||||
if response_is_done(state, response_id):
|
||||
# Generation may finish long before the client drains its audio.
|
||||
# A later barge-in still needs to stop that queued playback, but
|
||||
# must not produce a second response.done for the completed turn.
|
||||
playback = event.get("playback")
|
||||
if (
|
||||
not events
|
||||
and isinstance(playback, Mapping)
|
||||
and _int_or(playback.get("sent_ms")) > _int_or(playback.get("committed_ms"))
|
||||
):
|
||||
events.append(OutputAudioCleared(response_id=response_id))
|
||||
return events
|
||||
committed_ms = event.get("committed_ms")
|
||||
if isinstance(committed_ms, int | float):
|
||||
@@ -1320,7 +1330,11 @@ def resolve_cancel_response(state: RealtimeProjectionState, command: CancelRespo
|
||||
def resolve_clear_output_audio(state: RealtimeProjectionState, command: ClearOutputAudio) -> ResolvedControl:
|
||||
payload: dict[str, object] = {"type": "output_audio_buffer.clear", "reason": "output_audio_buffer.clear"}
|
||||
response_id = command.response_id or state.active_response_id or state.last_response_id
|
||||
if response_is_done(state, response_id):
|
||||
# A completed response can still own queued playback. Clear the latest
|
||||
# response in the engine, while keeping late clears away from newer turns.
|
||||
if response_is_done(state, response_id) and (
|
||||
state.active_response_id is not None or response_id != state.last_response_id
|
||||
):
|
||||
return ResolvedControl(payloads=[], events=[OutputAudioCleared(response_id=_str_or_none(response_id))])
|
||||
if isinstance(response_id, str) and response_id:
|
||||
payload["response_id"] = response_id
|
||||
|
||||
@@ -1029,6 +1029,16 @@ class DuplexEngineSession:
|
||||
last_message, audio_text_marks, _ = response_snapshot
|
||||
self._conversation.pending_item_ids[item_id] = copy.deepcopy(last_message)
|
||||
self._conversation.pending_item_input_commit_seqs[item_id] = response_snapshot[2]
|
||||
playback = self._playback_cursor_for_item_id(item_id)
|
||||
if (
|
||||
item_id in self._conversation.history_item_placeholders
|
||||
and playback is not None
|
||||
and playback.text_requires_complete_audio
|
||||
):
|
||||
# Keep the assistant turn between its user inputs even when
|
||||
# unaligned audio cannot establish any heard text. A later
|
||||
# full playback ACK fills this same slot from the snapshot.
|
||||
self._store_history_item_message(item_id, _object_dict(role="assistant", content=""))
|
||||
if audio_text_marks:
|
||||
self._conversation.pending_item_audio_text_marks[item_id] = list(copy.deepcopy(audio_text_marks))
|
||||
pending_audio_ms = self._conversation.pending_truncations_ms.get(item_id)
|
||||
@@ -1073,10 +1083,10 @@ class DuplexEngineSession:
|
||||
self._conversation.assistant_response_snapshots.pop(response_id, None)
|
||||
if message is None:
|
||||
return pending is not None or removed_placeholder
|
||||
try:
|
||||
self._conversation.messages.remove(message)
|
||||
except ValueError:
|
||||
pass
|
||||
# Distinct response items can have identical (including empty) content.
|
||||
self._conversation.messages = [
|
||||
candidate for candidate in self._conversation.messages if candidate is not message
|
||||
]
|
||||
return True
|
||||
|
||||
def truncate_history_item(
|
||||
@@ -1167,10 +1177,9 @@ class DuplexEngineSession:
|
||||
if changed and self._message_text_len(message) <= 0:
|
||||
self._conversation.item_ids.pop(item_id, None)
|
||||
self._conversation.item_audio_text_marks.pop(item_id, None)
|
||||
try:
|
||||
self._conversation.messages.remove(message)
|
||||
except ValueError:
|
||||
pass
|
||||
self._conversation.messages = [
|
||||
candidate for candidate in self._conversation.messages if candidate is not message
|
||||
]
|
||||
if item_id.startswith("item_"):
|
||||
self._conversation.assistant_response_snapshots.pop(item_id.removeprefix("item_"), None)
|
||||
elif changed and item_id.startswith("item_"):
|
||||
|
||||
@@ -1313,7 +1313,7 @@ class DuplexSessionRunner:
|
||||
{
|
||||
"type": "audio.cancelled",
|
||||
"session_id": session.session_id,
|
||||
"response_id": session.active_response_id,
|
||||
"response_id": event.get("response_id") or session.active_response_id,
|
||||
"reason": cancel_reason,
|
||||
"cancelled_epoch": session.epoch,
|
||||
"epoch": session.epoch,
|
||||
|
||||
Reference in new issue
Block a user