sgl-project/sglang · error · ValueError

SANA-WM denoising expects 5D latents shaped (B, C, T, H, W),

Error message

SANA-WM denoising expects 5D latents shaped (B, C, T, H, W), got {tuple(batch.latents.shape)}.

What it means

Raised by the SANA-WM denoising forward when batch.latents.ndim != 5. The video denoiser expects latents shaped (B, C, T, H, W).

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/sana_wm/base.py:828

    def _combine_cfg_parallel_noise(
        noise_pred: torch.Tensor,
        guidance_scale: float,
        cfg_rank: int,
    ) -> torch.Tensor:
        if cfg_rank == 0:
            partial = guidance_scale * noise_pred
        elif cfg_rank == 1:
            partial = (1.0 - guidance_scale) * noise_pred
        else:
            partial = torch.zeros_like(noise_pred)
        return cfg_model_parallel_all_reduce(partial)

    @torch.no_grad()
    def forward(self, batch: Req, server_args: ServerArgs) -> Req:
        if batch.latents is None:
            raise ValueError("SANA-WM denoising requires initialized latents.")
        if batch.latents.ndim != 5:
            raise ValueError(
                "SANA-WM denoising expects 5D latents shaped (B, C, T, H, W), "
                f"got {tuple(batch.latents.shape)}."
            )

        device = get_local_torch_device()
        target_dtype = PRECISION_TO_TYPE.get(
            getattr(server_args.pipeline_config, "dit_precision", "bf16"),
            torch.bfloat16,
        )
        scheduler = getattr(
            batch, "scheduler", None
        ) or get_or_create_request_scheduler(batch, self.scheduler)
        self._move_scheduler_tensors_to_device(scheduler, device)
        timesteps = batch.timesteps
        if timesteps is None:
            raise ValueError("SANA-WM denoising requires prepared timesteps.")
        timesteps = timesteps.to(device=device)

View on GitHub (pinned to 0132848349)

Solutions

1. Add the batch dim: latents = latents.unsqueeze(0) if 4D
2. Verify the latent-init stage produces 5D output (B,C,T,H,W)
3. For image-only conditioning, expand T=1 rather than dropping the axis

Example fix

# before
latents = latents  # (C,T,H,W)
# after
latents = latents.unsqueeze(0)  # (1,C,T,H,W)

assert batch.latents is not None and batch.latents.ndim == 5

def latents_5d(batch) -> bool:
    l = getattr(batch, 'latents', None)
    return l is not None and l.ndim == 5

When it happens

Trigger: Latents of shape (C,T,H,W) (missing batch dim), (B,C,H,W) (image latents, missing time), or (B,T,H,W) passed into the denoising stage.

Common situations: Image-pipeline latents reused for the video pipeline; a stage dropped or never added the batch dimension; conditioning latents prepared as 4D.

Related errors

• camera trajectory must have shape (F, 4, 4); got {c2w.shape}
• unsupported intrinsics shape {arr.shape}; expected (4,), (3,
• SANA-WM denoising requires initialized latents.
• condition_image tensor must be CHW or HWC with 1, 3, or 4 ch
• camera_conditions must have shape (T,20) or (B,T,20), got {t

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/328055f1941961fd. Report an issue: GitHub.