sgl-project/sglang · error · ValueError

SANA-WM denoising requires prepared timesteps.

Error message

SANA-WM denoising requires prepared timesteps.

What it means

Raised by the SANA-WM denoising forward when batch.timesteps is None — the diffusion scheduler's timestep tensor was never prepared on the request.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/sana_wm/base.py:844

            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)

        latents = batch.latents.to(device=device, dtype=target_dtype)
        init_latents = latents.clone()
        condition_mask = torch.zeros_like(latents)
        condition_mask[:, :, :1] = 1

        pos_embeds = _to_device_dtype(
            _first_tensor(server_args.pipeline_config.get_pos_prompt_embeds(batch)),
            device=device,
            dtype=target_dtype,
        )
        pos_mask = _to_device_dtype(
            _first_tensor(batch.prompt_attention_mask), device=device
        )
        if pos_embeds is None:
            raise ValueError("SANA-WM denoising requires positive prompt embeds.")

View on GitHub (pinned to 0132848349)

Solutions

1. Ensure the init/before-denoising stage sets batch.timesteps from the scheduler
2. Call scheduler.set_timesteps(...) and attach the tensor to the batch before denoising
3. Check pipeline ordering so timesteps preparation precedes denoising

assert batch.timesteps is not None

def has_timesteps(batch) -> bool:
    return getattr(batch, 'timesteps', None) is not None

When it happens

Trigger: Invoking the denoising stage without a preceding stage (typically latent init) having set batch.timesteps via the scheduler; scheduler present but timesteps not copied onto the batch.

Common situations: Skipped or reordered init stage; scheduler created but set_timesteps never ran; a request path that bypasses the standard preparation stage.

Related errors

• {self.config.timestep_spacing} is not supported. Please make
• Timesteps must be provided
• Only one of `timesteps` or `sigmas` can be passed. Please ch
• The current scheduler class {scheduler.__class__}'s `set_tim
• SANA-WM denoising requires initialized latents.

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/64dc8914ac14dd29. Report an issue: GitHub.