sgl-project/sglang · error · ValueError

{context}video block token count must be positive

Error message

{context}video block token count must be positive

What it means

Raised when a video temporal block's token count is <= 0 while emitting the timestamped VIDEO_PAD vision blocks. Each block must contain at least one VIDEO pad token for the vision encoder to have content.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/presentation.py:101

def _timestamped_video_blocks(
    presentation: _Presentation,
    tokenizer: Any,
    *,
    counts: Sequence[int],
    timestamps: Sequence[float],
    context: str,
    video_token_id: int | None,
) -> None:
    """Emit per-temporal-block ``<{t:.1f} seconds>`` text + VIDEO vision."""

    counts = [int(value) for value in counts]
    timestamps = [float(value) for value in timestamps]
    if not counts or len(counts) != len(timestamps):
        raise ValueError(f"{context}video block token counts and timestamps must align")
    for count, timestamp in zip(counts, timestamps):
        if count <= 0:
            raise ValueError(f"{context}video block token count must be positive")
        presentation.text(_text_ids(tokenizer, f"<{timestamp:.1f} seconds>"))
        presentation.vision(
            _vision_block_ids(tokenizer, VIDEO_PAD, count),
            video_token_id=video_token_id,
        )


def minimax_h3_text_only_ids(tokenizer: Any, prompt: str) -> torch.Tensor:
    """t2va presentation: verbatim prompt, no special tokens."""
    if not prompt:
        raise ValueError("prompt must be non-empty")
    return torch.tensor(_text_ids(tokenizer, prompt), dtype=torch.long)


def minimax_h3_multi_image_presentation(
    tokenizer: Any,
    *,
    prompt: str,

View on GitHub (pinned to 0132848349)

Solutions

1. Filter out or clamp zero-count blocks before calling the API
2. Fix the token-count computation so each retained segment yields >= 1 vision token
3. Drop the corresponding timestamp as well so lengths stay aligned

Example fix

// before
counts = [0, 196]
// after
counts = [max(1, c) for c in counts]

counts = [c for c in counts if c > 0]
# then re-pair timestamps accordingly

def positive_blocks(counts: list[int], ts: list[float]) -> bool:
    return all(c >= 1 for c in counts)

When it happens

Trigger: Passing a video_block_token_counts entry of 0 or a negative number (e.g. [0, 196]) to minimax_h3_ref2va_video_presentation for a video reference.

Common situations: Patch-count arithmetic that floors to zero for very short/thin video frames; default-initialized arrays containing 0; unit conversions producing 0 tokens for tiny segments.

Related errors

• {context}video block token counts and timestamps must align
• MiniMax-H3 adaln_t_table must have shape [N, D] with N >= 2,
• MiniMax H3 AdaLN cache has invalid timestep plans
• TP size must be positive.
• num_attention_heads must be positive.

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/d12682ecc6a355db. Report an issue: GitHub.