sgl-project/sglang · error · ValueError

chunk_plucker batch dimension must be 1 or match request bat

Error message

chunk_plucker batch dimension must be 1 or match request batch size {batch_size}, got {chunk_plucker.shape[0]}.

What it means

After normalizing chunk_plucker to 5-D, the batch dim must be 1 (broadcast to the request batch) or exactly the request batch size.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/sana_wm/base.py:2134

                )

        if chunk_plucker is not None:
            chunk_plucker = (
                chunk_plucker
                if isinstance(chunk_plucker, torch.Tensor)
                else torch.as_tensor(chunk_plucker)
            ).to(device=device, dtype=dtype)
            if chunk_plucker.dim() == 4:
                chunk_plucker = chunk_plucker.unsqueeze(0)
            if chunk_plucker.shape[0] == 1 and batch_size > 1:
                chunk_plucker = chunk_plucker.expand(batch_size, -1, -1, -1, -1)
            if chunk_plucker.dim() != 5:
                raise ValueError(
                    "chunk_plucker must have shape (48,T,H,W) or "
                    f"(B,48,T,H,W), got {tuple(chunk_plucker.shape)}"
                )
            if chunk_plucker.shape[0] != batch_size:
                raise ValueError(
                    "chunk_plucker batch dimension must be 1 or match "
                    f"request batch size {batch_size}, got "
                    f"{chunk_plucker.shape[0]}."
                )
            expected_chunk_shape = (batch_size, 48, T_lat, sp_h, sp_w)
            if tuple(chunk_plucker.shape) != expected_chunk_shape:
                raise ValueError(
                    "chunk_plucker shape mismatch for SANA-WM: expected "
                    f"{expected_chunk_shape}, got {tuple(chunk_plucker.shape)}."
                )

        if camera_conditions is not None:
            camera_conditions = camera_conditions.to(device=device, dtype=dtype)

        return camera_conditions, chunk_plucker, source

    def _prepare_timesteps(
        self,

View on GitHub (pinned to 0132848349)

Solutions

1. Slice chunk_plucker to the request batch size.
2. Or pass a single (1,48,T,H,W) volume to broadcast across requests.
3. Recompute plücker whenever the batch composition changes.

Example fix

# before
plucker = export_all(4)
# after
plucker = export_all(4)[:batch_size]

if chunk_plucker.shape[0] not in (1, batch_size):
    chunk_plucker = chunk_plucker[:batch_size]

When it happens

Trigger: Providing B plücker volumes with B not in {1, batch_size}, e.g. 4 exported ray maps for a 2-request batch.

Common situations: Caching per-video plücker tensors and replaying them against a differently-sized request batch.

Related errors

• camera_conditions batch dimension must be 1 or match request
• Prepacked latent-frame camera_conditions require chunk_pluck
• chunk_plucker must have shape (48,T,H,W) or (B,48,T,H,W), go
• chunk_plucker shape mismatch for SANA-WM: expected {expected
• SANA-WM does not support tensor parallelism yet. Use --num-g

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/83e652f8ad88ac7e. Report an issue: GitHub.