sgl-project/sglang · error · ValueError

SANA-WM seed list length must be 1 or match latent batch siz

Error message

SANA-WM seed list length must be 1 or match latent batch size; got {len(seed)} seeds for batch {batch_size}.

What it means

Raised by _generator_from_seed when a seed list has length other than 1 or batch_size — either one shared seed or exactly one seed per latent sample is allowed.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/sana_wm/base.py:1210

        seed: int | list[int] | tuple[int, ...] | None,
        *,
        batch_size: int,
        device: torch.device,
    ) -> torch.Generator | list[torch.Generator]:
        if seed is None:
            seed = 0
        if isinstance(seed, (list, tuple)):
            if not seed:
                raise ValueError("SANA-WM seed list must not be empty.")
            if len(seed) == 1:
                seed = seed[0]
            elif len(seed) == batch_size:
                return [
                    torch.Generator(device=device).manual_seed(int(sample_seed))
                    for sample_seed in seed
                ]
            else:
                raise ValueError(
                    "SANA-WM seed list length must be 1 or match latent batch "
                    f"size; got {len(seed)} seeds for batch {batch_size}."
                )
        return torch.Generator(device=device).manual_seed(int(seed))

    @staticmethod
    def _canonical_condition_image_tensor(image: torch.Tensor) -> torch.Tensor:
        """Return image as NCHW RGB float tensor without changing its value range."""
        image = image.float()
        if image.dim() == 5 and image.shape[2] == 1:
            image = image.squeeze(2)
        if image.dim() == 3:
            if image.shape[0] in (1, 3, 4):
                image = image.unsqueeze(0)
            elif image.shape[-1] in (1, 3, 4):
                image = image.permute(2, 0, 1).unsqueeze(0)
            else:
                raise ValueError(

View on GitHub (pinned to 0132848349)

Solutions

Use a single int (or length-1 list) for a shared seed across the batch
Or provide exactly batch_size seeds, filling unseeded samples with 0
Compute seeds from the same request list that determines batch_size

Example fix

# before
seeds = [r.seed for r in requests if r.seed is not None]  # partial
# after
seeds = [r.seed if r.seed is not None else 0 for r in requests]  # len == batch

Defensive patterns

Strategy: validation

Validate before calling

assert not isinstance(seed, (list, tuple)) or len(seed) in (1, batch_size)

Type guard

def seeds_match(s, batch: int) -> bool:
    return not isinstance(s, (list, tuple)) or len(s) in (1, batch)

Prevention

Fill unseeded samples with 0 instead of dropping them
Compute seed lists from the final request batch

When it happens

Trigger: Passing seed=[1,2,3] for a batch of 2 (or 4); seed lists whose length diverges from the latent batch size.

Common situations: Batch size recomputed after seeds set; some requests in a batch carry seeds and others don't, producing a partial list; off-by-one when chunking requests.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/8b1fa8fe42fb7b23. Report an issue: GitHub.