sgl-project/sglang · error · ValueError

chunk_index must be strictly increasing, got {normalized}.

Error message

chunk_index must be strictly increasing, got {normalized}.

What it means

After normalization, chunk boundaries must be strictly increasing (and end at T). Any non-increasing adjacent pair — duplicate indices or a decrease — fails this invariant before slicing.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/dits/sana_wm_components.py:417

        normalized = [int(idx) for idx in chunk_index]
        if not normalized or normalized[0] != 0:
            normalized = [0] + [idx for idx in normalized if idx > 0]
        normalized = [idx for idx in normalized if idx < T]
        if not normalized:
            normalized = [0]
    else:
        if chunk_size is None:
            raise ValueError("Either chunk_index or chunk_size must be provided.")
        normalized = _sana_wm_chunk_index_from_chunk_size(
            T,
            int(chunk_size),
            strategy=chunk_split_strategy,
        )

    if normalized[-1] != T:
        normalized.append(T)
    if any(end <= start for start, end in zip(normalized[:-1], normalized[1:])):
        raise ValueError(f"chunk_index must be strictly increasing, got {normalized}.")
    return normalized


def _sana_wm_chunk_boundaries_for_attention(
    HW: Tuple[int, int, int],
    chunk_size: Optional[int],
    chunk_split_strategy: str,
    chunk_index: Optional[List[int]],
) -> Optional[list[int]]:
    T, _, _ = HW
    if chunk_index is None and (chunk_size is None or int(chunk_size) >= T):
        return None

    boundaries = _sana_wm_normalize_chunk_index(
        chunk_index,
        T,
        chunk_size=chunk_size,
        chunk_split_strategy=chunk_split_strategy,

View on GitHub (pinned to 0132848349)

Solutions

1. Sort and deduplicate chunk_index before passing: sorted(set(chunk_index))
2. Ensure all indices are in (0, T) and the list strictly increases
3. Prefer passing chunk_size and letting the helper compute indices

Example fix

# before
idx = _sana_wm_normalize_chunk_index(T, chunk_index=[0, 8, 8, 16])
# after
idx = _sana_wm_normalize_chunk_index(T, chunk_index=sorted({0, 8, 8, 16}))

idx = sorted({int(i) for i in chunk_index if 0 < i < T})
assert all(b > a for a, b in zip(idx, idx[1:]))

When it happens

Trigger: Passing chunk_index with duplicates like [0, 8, 8, 16], an unsorted list, or an index >= T (post-filtering can leave e.g. [0,0]); also indices not starting at 0 can interact badly with normalization.

Common situations: User-supplied chunk_index from config computed with float arithmetic then int()-cast (duplicates); indices generated for a larger T reused on a smaller clip (filtered to [0]); off-by-one in boundary generation.

Understand the failure class

Background: "Must be a positive integer", "Invalid value", "Unsupported": the invalid-argument-value error family, when a library rejects the value you pass — this error's family across 35 libraries.

Related errors

• chunk_size must be > 0, got {chunk_size}.
• T must be > 0, got {T}.
• Unknown chunk_split_strategy '{strategy}'. Supported: unifor
• Either chunk_index or chunk_size must be provided.
• Unsupported content type ${header.content_type}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/dbb197b229b50d2a. Report an issue: GitHub.