Skip to content

Cosmos3-edge PR 5: end a failed chat stream with an error event instead of stop #296

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
merceod wants to merge 2 commits into
base: main
Choose a base branch
from
+35 −0
Open

Cosmos3-edge PR 5: end a failed chat stream with an error event instead of stop #296

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9fb858e
chat: end a failed stream with an in-band error event
merceod Sep 18, 2026
105bc2a
Test a failed chat stream reports the error in-band
merceod Sep 18, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 13 additions & 0 deletions mstar/api_server/openai/serving_chat.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -98,5 +98,18 @@ def chunk(delta, finish=None) -> str:
yield chunk({"audio": {"id": rid("audio"), "data": base64.b64encode(c.data).decode("ascii")}})
elif c.modality == "image":
yield chunk({"content": media_io.png_to_data_url(c.data)})
elif c.modality == "error":

@NSagan271 NSagan271 Sep 26, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Note: this doesn't cover the timeout error, which directly raises an HTTPException in iter_result_chunks

# The request failed after the stream opened (an engine error mid
# generation, a delivery timeout); the HTTP status is committed, so
# the failure travels in-band the way the non-streaming path's
# error body does — not as a normal ``stop``, which a client would
# take for a complete answer.
yield sse({"error": {
"message": c.data.decode("utf-8", "replace"),
"type": "server_error",

@NSagan271 NSagan271 Sep 26, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The type is hardcoded to "server_error", but the data worker's _fail_request will return a 400 if it sees a ValueError or TypeError

"code": c.metadata.get("status", 500),
}})
yield SSE_DONE
return

@NSagan271 NSagan271 Sep 26, 2026

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This early return trips the finally in iter_result_chunks, which calls self.abort_request(request_id) and triggers an abort, even though the drain/removal of the request is already being processed.

(reproduced the behavior with a dummy example:

import asyncio

async def x():
    finished = False
    try:
        for i in range(5):
            yield i
        finished = True
    finally:
        if not finished:
            print("finally reached: not finished")


async def y():
    async for c in x():
        print(c)
        if c == 3:
            return

asyncio.run(y())

prints the "finally reached: not finished")

yield chunk({}, finish="stop")
yield SSE_DONE
22 changes: 22 additions & 0 deletions test/modular/test_openai_router.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -227,6 +227,28 @@ def test_chat_stream(client_and_stub):
assert lines[-1]["choices"][0]["finish_reason"] == "stop"


def test_chat_stream_reports_a_failed_request_in_band(client_and_stub):
"""A request that fails after the stream opened ends with an error event,
not a ``finish_reason: stop`` that reads as a complete answer."""
client, stub = client_and_stub
stub.model_name = "bagel"
stub.next_chunks = [
_Chunk("text", b"Paris"),
_Chunk("error", b"Error in worker: ValueError: q implies q_len_per_req=5", {"status": 500}),
]
text = client.post(
"/v1/chat/completions",
json={"model": "bagel", "messages": [{"role": "user", "content": "go"}], "stream": True},
).text
events = [json.loads(l[6:]) for l in text.splitlines() if l.startswith("data: ") and "[DONE]" not in l]
assert events[1]["choices"][0]["delta"]["content"] == "Paris"
assert events[-1]["error"] == {
"message": "Error in worker: ValueError: q implies q_len_per_req=5", "type": "server_error", "code": 500,
}
assert not any(e.get("choices", [{}])[0].get("finish_reason") for e in events)
assert text.rstrip().endswith("data: [DONE]")


def test_unsupported_model_404(client_and_stub):
client, stub = client_and_stub
stub.model_name = "pi05"
Expand Down
Loading