sgl-project/sglang · error · ValueError

Dots note omni video preprocessing requires one request's sa

Error message

Dots note omni video preprocessing requires one request's sampling_params as a dictionary.

What it means

Raised by process_mm_data_async when request_obj.sampling_params is neither None nor a dict (e.g. a SamplingParams object). The video preprocessing path needs to read max_new_tokens from a plain dictionary.

Source

Thrown at python/sglang/srt/multimodal/processors/dots_note_omni.py:417

            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)
            audio_cap = video_config.pop("audio_cap", 1.0)
            audio_sr = video_config.pop("audio_sr", 16000)
            k_mode = video_config.pop("k_mode", "eval_ek")
            if video_config:
                raise ValueError(
                    "Unsupported dots note omni video_config fields: "
                    + ", ".join(sorted(video_config))
                )
            sampling_params = request_obj.sampling_params or {}
            if not isinstance(sampling_params, dict):
                raise ValueError(
                    "Dots note omni video preprocessing requires one request's "
                    "sampling_params as a dictionary."
                )
            max_new_tokens = sampling_params.get("max_new_tokens") or 0
            loop = asyncio.get_running_loop()
            preprocess_started = time.perf_counter()
            video_media = {}
            total_content_items = 0
            total_frames = 0
            total_audio_segments = 0
            for video_index, video in enumerate(video_data):
                content = await loop.run_in_executor(
                    self.io_executor,
                    lambda video=video: preprocess_dots_video(
                        video,
                        question,
                        tokenizer=self._tokenizer,
                        seq=seq,

View on GitHub (pinned to 0132848349)

Solutions

1. Pass sampling_params as a plain dict, e.g. {"max_new_tokens": 1024}
2. Or omit it (None) so the default `{}` / max_new_tokens=0 path is used
3. Update sglang — newer versions may serialize params to dict before the processor

Example fix

// before
request.sampling_params = SamplingParams(max_new_tokens=1024)
// after
request.sampling_params = {"max_new_tokens": 1024}

sp = getattr(request_obj, 'sampling_params', None)
if sp is not None and not isinstance(sp, dict):
    request_obj.sampling_params = {'max_new_tokens': getattr(sp, 'max_new_tokens', 0)}

def sampling_params_is_dict(req) -> bool:
    sp = getattr(req, 'sampling_params', None)
    return sp is None or isinstance(sp, dict)

When it happens

Trigger: A request object whose sampling_params is a non-dict object (string, dataclass, already-instantiated SamplingParams) while video_data is present; `sampling_params or {}` returns the truthy non-dict and the isinstance check fails.

Common situations: Engine-layer callers that pass typed sampling params objects instead of the raw per-request dict the multimodal processor expects; version drift between sglang layers after SamplingParams was introduced as a class.

Related errors

• max_new_tokens must be non-negative, got {max_new_tokens}
• seq must be positive, got {seq}
• 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/e9d677b791ff50a8. Report an issue: GitHub.