sgl-project/sglang · error · ValueError

Dots note omni requires a text prompt for multimodal request

Error message

Dots note omni requires a text prompt for multimodal requests

What it means

Raised by process_mm_data_async when a multimodal request (image/audio/video data present) has a non-string input_text (e.g. a chat-messages list). The Dots Note Omni processor requires the already-rendered text prompt.

Source

Thrown at python/sglang/srt/multimodal/processors/dots_note_omni.py:388

            pass
        return "".join(rendered), ordered_images, ordered_audios

    async def process_mm_data_async(
        self,
        input_text: list[int] | str,
        request_obj: GenerateReqInput,
        max_req_input_len: int,
        *args,
        image_data: list | None = None,
        audio_data: list | None = None,
        video_data=None,
        **kwargs,
    ):
        video_data = request_obj.video_data or video_data
        if not image_data and not audio_data and not video_data:
            return None
        if not isinstance(input_text, str):
            raise ValueError(  # noqa: TRY004 - preserve the processor API contract
                "Dots note omni requires a text prompt for multimodal requests"
            )

        request_videos = len(video_data) if video_data else 0
        request_images = len(image_data) if image_data else 0
        request_audios = len(audio_data) if audio_data else 0
        logger.info(
            "[dots_mm] rid=%s request videos=%d images=%d audios=%d",
            request_obj.rid,
            request_videos,
            request_images,
            request_audios,
        )

        if video_data:
            video_config = dict(request_obj.video_config or {})
            question = video_config.pop("_question", "") or ""
            seq = video_config.pop("seq", 131072)

View on GitHub (pinned to 0132848349)

Solutions

1. Pre-render your chat template client-side and pass the resulting string as input_text
2. Or use the endpoint/flow intended for this model so text reaches the processor as a str
3. If building requests programmatically, assert isinstance(text, str) before attaching media

Example fix

// before
input_text = [{"role": "user", "content": "describe this"}]
// after
input_text = "<|user|>\ndescribe this<|end|>"  # rendered string

assert isinstance(input_text, str) and input_text, 'render chat template to a string before attaching media'

def is_rendered_prompt(text) -> bool:
    return isinstance(text, str) and len(text) > 0

When it happens

Trigger: Calling the /generate-style API with image_data/video_data set while passing input_text as a list of chat messages instead of a string; the processor checks isinstance(input_text, str) right after detecting mm data.

Common situations: Using the OpenAI-compatible chat endpoint path where text arrives unrendered, or passing {'text': [...messages...]} in a raw request with media attached.

Related errors

• seq must be positive, got {seq}
• max_new_tokens must be non-negative, got {max_new_tokens}
• max_new_tokens must leave room for input: max_new_tokens={ma
• audio_cap must be non-negative, got {audio_cap}
• audio_sr must be positive, got {audio_sr}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/b5bbf365fff6200f. Report an issue: GitHub.