sgl-project/sglang · error · ValueError

Dots note omni expanded prompt is too long: {len(input_ids)}

Error message

Dots note omni expanded prompt is too long: {len(input_ids)} > {max_req_input_len}

What it means

Raised by process_mm_data_async after expanding multimodal tokens into input_ids: the fully expanded prompt exceeds the request's maximum input length (max_req_input_len), so the request would be unschedulable.

Source

Thrown at python/sglang/srt/multimodal/processors/dots_note_omni.py:553

            item_start = len(input_ids)
            input_ids.extend(item_token_ids)
            local_offsets = self.get_mm_items_offset(
                torch.tensor(item_token_ids), pad_token_id
            )
            offsets = [
                (item_start + start, item_start + end) for start, end in local_offsets
            ]
            item = MultimodalDataItem(
                modality=modality,
                feature=feature,
                offsets=offsets,
                model_specific_data=model_specific_data,
            )
            item.set_pad_value()
            mm_items.append(item)

        if len(input_ids) > max_req_input_len:
            raise ValueError(
                "Dots note omni expanded prompt is too long: "
                f"{len(input_ids)} > {max_req_input_len}"
            )
        padded_input_ids = MultimodalProcessorOutput.build_padded_input_ids(
            input_ids, mm_items
        )
        return MultimodalProcessorOutput(
            mm_items=mm_items,
            input_ids=input_ids,
            padded_input_ids=padded_input_ids,
            **self.mm_token_ids,
        )

View on GitHub (pinned to 0132848349)

Solutions

1. Shorten/trim the video (fewer frames, lower resolution) or reduce the number of media items
2. Raise the server context length (--context-length) if the model supports it
3. Reduce sampling_params.max_new_tokens so seq_length = seq - max_new_tokens leaves more room for input

Example fix

# before
video_config = {"seq": 131072}
sampling_params = {"max_new_tokens": 120000}  # leaves 11108 for video tokens
# after
sampling_params = {"max_new_tokens": 1024}  # leaves ~130k for input

est_tokens = estimate_expanded_tokens(video, images, prompt)  # frames * per-frame tokens + text
assert est_tokens <= max_req_input_len, f'estimated {est_tokens} > {max_req_input_len}'

try:
    await processor.process_mm_data_async(...)
except ValueError as e:
    if 'expanded prompt is too long' in str(e):
        # trim video frames/resolution or lower max_new_tokens, then resubmit

When it happens

Trigger: A video (plus images/audio/text) expands to more token ids than max_req_input_len — long videos are the usual cause since keyframes consume thousands of tokens each; e.g. seq budget misconfigured so seq_length = seq - max_new_tokens is too small for the media.

Common situations: Sending long/high-fps videos, many videos in one prompt, or setting a small context length (--context-length) while seq defaults assume 131072. The check runs after all padding-relevant items are built.

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/d35d6faf6832b667. Report an issue: GitHub.