Qwen-Omni Realtime API - Invalid session.finish event, session failure on ws.close(), and VAD token consumption inquiry

[ptaejoon]

Context / Background

I am using the Qwen3.5-Omni-Flash-Realtime model over a WebSocket connection. I reviewed the sample implementation provided in your repository at:
https://github.com/aliyun/alibabacloud-bailian-speech-demo/blob/master/samples/conversation/omni/python/run_server_vad.py

Problem 1: Invalid session.finish event & Missing Graceful Shutdown

When attempting to gracefully shut down the WebSocket connection, I tried sending a session.finish event type, same as the sdk of Realtime API specifications:

{
    "event_id": "...",
    "type": "session.finish"
}

However, the server returns an error stating that this is an invalid event type.

Looking closely at your sample code (run_server_vad.py), it only calls conversation.close(), which internally just executes ws.close(). This abruptly terminates the WebSocket connection without a proper higher-level elegant handshake or session termination event.

Problem 2: Dashboard showing "Failed" sessions & Token usage tracking

Because the connection is abruptly killed via ws.close(), every single one of my Realtime API sessions is marked as "Failed" in my Qwen Cloud Console / DashScope Dashboard. This makes it extremely difficult to track my exact token usage, session histories, and accurate analytics.

• Is the rejection of session.finish expected behavior, or is there another protocol-compliant event to gracefully close a session?
• Is there a bug on the server side that records abrupt ws.close() actions as session failures?

---

Question 3: VAD Token Consumption

Additionally, I have a question regarding the Voice Activity Detection (VAD) mode.
When VAD is enabled, does the server consume any tokens during the idle state (i.e., when the client is streaming background silence/noise before the actual speech or user message is triggered/recognized)?

Environment

• Model: Qwen3.5-Omni-Flash-Realtime

Any clarification or guidance on how to properly perform a graceful shutdown and understand VAD token mechanics would be greatly appreciated. Thank you!

[lzsweb]

Hi @ptaejoon,

Problem 1 & 2:

The SDK's OmniRealtimeConversation class already supports graceful shutdown via end_session(), which sends a session.finish event and waits for the server's session.finished confirmation. The sample code you referenced (run_server_vad.py) is not part of this SDK — it operates WebSocket directly, hence the ws.close() only.

We recommend using the SDK method:

conversation.end_session(timeout=20)

Regarding your session.finish being rejected and sessions showing as "Failed" in the dashboard: we checked the server-side logs for your session and found a 500 InternalError (source: model, error_message: "Internal service error: null"). This is a server-side crash (Java NPE) that occurred ~8.5 seconds into the session, before any close attempt. The session was already in a broken state when session.finish was sent, which is why it was rejected.

We've escalated request_id: 27c9df83-0d58-9eba-89c7-31c56fd67a70 to the model service team for investigation.

Question 3 (VAD token consumption):

This is a model-level question that the model service team is better positioned to answer. We'll include it in the escalation.

[lzsweb]

Problem 1 & 2: Your session hit a server-side 500 InternalError after ~8.5s, which caused session.finish rejection and "Failed" dashboard status. We've escalated to the model team (request_id: 27c9df83-0d58-9eba-89c7-31c56fd67a70). For graceful shutdown, use conversation.end_session().

Problem 3: No token consumption during VAD idle state (silence/noise before speech is detected).

[ptaejoon]

@lzsweb

Thanks for the quick response. Could you also take a look at this connection issue?

The error message in the screenshot I uploaded is different from the ones I've usually seen. Most of the connection-related errors I've encountered were along the lines of:

"Client disconnected before task finished."

Would using session.finish() before disconnecting help prevent this issue, or is there something else I should be doing?

[songguocola]

Using conversation.end_session() would solve this issue. This demo "https://github.com/aliyun/alibabacloud-bailian-speech-demo/blob/master/samples/conversation/omni/python/run_server_vad.py" had been updated.