sgl-project/sglang · error · ValueError

video reference requires block token counts and timestamps

Error message

video reference requires block token counts and timestamps

What it means

A video reference appears in condition_labels but there are fewer normalized video_block_token_counts groups than video references — no per-block data exists for this video.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/presentation.py:277

        raise ValueError("video block token counts and timestamps must align")
    image_seen = 0
    video_seen = 0
    for cond_type, ordinal in condition_labels:
        if cond_type == "image":
            image_seen += 1
            if image_seen > len(image_token_counts):
                raise ValueError("image_token_count required for an image reference")
            count = int(image_token_counts[image_seen - 1])
            if count <= 0:
                raise ValueError("image_token_count required for an image reference")
            presentation.text(_text_ids(tokenizer, f"<Picture {ordinal}>: "))
            presentation.vision(_vision_block_ids(tokenizer, IMAGE_PAD, count))
        elif cond_type == "audio":
            presentation.text(_text_ids(tokenizer, f"<Audio {ordinal}>: "))
        elif cond_type == "video":
            video_seen += 1
            if video_seen > len(video_counts_by_ref):
                raise ValueError(
                    "video reference requires block token counts and timestamps"
                )
            counts = video_counts_by_ref[video_seen - 1]
            timestamps = video_timestamps_by_ref[video_seen - 1]
            if not counts or not timestamps:
                raise ValueError(
                    "video reference requires block token counts and timestamps"
                )
            presentation.text(_text_ids(tokenizer, f"<Video {ordinal}>: "))
            _timestamped_video_blocks(
                presentation,
                tokenizer,
                counts=counts,
                timestamps=timestamps,
                context="",
                video_token_id=video_token_id,
            )
        else:

View on GitHub (pinned to 0132848349)

Solutions

Provide one counts group per video reference: [[...],[...]]
Trim extra video references from condition_labels
Regenerate conditioning data from the current plan

Example fix

// before
condition_labels=[("video",1),("video",2)], video_block_token_counts=[196]
// after
condition_labels=[("video",1),("video",2)], video_block_token_counts=[[196],[196]]

Defensive patterns

Strategy: validation

Validate before calling

n_videos = sum(1 for t, _ in condition_labels if t == "video")
assert len(video_counts_by_ref) >= n_videos

Prevention

Provide one counts group per video reference
Use nested form for multi-video plans

When it happens

Trigger: condition_labels=[("video",1),("video",2)] but video_block_token_counts=[[196]] (only one group).

Common situations: Single-video convenience form (flat list) used with a multi-video plan; forgetting per-video groups when the plan has several videos.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/5f440ea5e271ab4c. Report an issue: GitHub.