sgl-project/sglang · error · ValueError

image_token_count must be positive

Error message

image_token_count must be positive

What it means

Each image in the fl2va multi-image presentation needs a positive IMAGE_PAD token count; a zero or negative count cannot form a valid vision block.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/presentation.py:127

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,
    image_token_counts: list[int],
) -> tuple[torch.Tensor, torch.Tensor]:
    if not image_token_counts:
        raise ValueError("image_token_counts must be non-empty")
    presentation = _Presentation()
    for index, count in enumerate(image_token_counts, start=1):
        if int(count) <= 0:
            raise ValueError("image_token_count must be positive")
        presentation.text(_text_ids(tokenizer, f"<Picture {index}>: "))
        presentation.vision(_vision_block_ids(tokenizer, IMAGE_PAD, count))
    presentation.text(_text_ids(tokenizer, prompt))
    return presentation.build()


def minimax_h3_ref2va_presentation(
    tokenizer: Any,
    *,
    prompt: str,
    condition_labels: list[tuple[str, int]],
    image_token_count: int | list[int] | None,
) -> tuple[torch.Tensor, torch.Tensor]:
    """ref2va positive presentation:

    per condition in request order — image i: ``<Picture i>: `` label followed
    by the vision block; audio j: ``<Audio j>: `` label only (audio content
    never enters Qwen) — then the verbatim prompt. Returns ``(ids, token_tags)``

View on GitHub (pinned to 0132848349)

Solutions

1. Recompute the offending entry's token budget from image dimensions
2. Clamp to at least 1 or drop that image and its `<Picture N>` reference
3. Add pre-call validation of all counts

Example fix

// before
image_token_counts=[196, 0]
// after
image_token_counts=[196, 196]

assert all(int(c) > 0 for c in image_token_counts), "image counts must be positive"

def valid_image_counts(cs: list[int]) -> bool:
    return all(isinstance(c, int) and c > 0 for c in cs)

When it happens

Trigger: Passing image_token_counts containing a 0 or negative entry, e.g. [196, 0].

Common situations: Resolution-derived token math flooring to 0 for tiny images; default zero-filled arrays; index bugs mapping image sizes to counts.

Related errors

• image_token_counts must be non-empty
• 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/f339c22f9bb819af. Report an issue: GitHub.