sgl-project/sglang · error · ValueError

SANA-WM forward requires timestep.

Error message

SANA-WM forward requires timestep.

What it means

SANA-WM forward requires a timestep tensor. Pass `timestep` when calling the world-model forward.

Solutions

1. Pass the scheduler's current timestep tensor (shape (B,) or broadcastable) to forward
2. Verify your sampler loop forwards t each step

Example fix

# before
out = model(h, encoder_hidden_states=ehs)
# after
out = model(h, timestep=t, encoder_hidden_states=ehs)

assert timestep is not None and timestep.numel() == B

def valid_timestep(t) -> bool: return isinstance(t, torch.Tensor) and t.numel() >= 1

When it happens

Trigger: Calling forward without timestep, or with timestep=None; e.g. running a single clean pass or a wrapper that only supplies hidden_states and text embeddings.

Common situations: Adapters from other DiT APIs where timestep is optional (e.g. some inference wrappers default t=0); forgetting to pass scheduler timesteps in a sampling loop.

Understand the failure class

Background: "missing required argument" and "the following required arguments were not provided": what required-argument errors mean and how to fix them — this error's family across 20 libraries.

Related errors

• SANA-WM forward_long requires timestep.
• Either chunk_index or chunk_size must be provided.
• num_frames/height/width are required when hidden_states is…
• SANA-WM forward_long requires encoder_hidden_states.
• SANA-WM forward requires encoder_hidden_states.

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/6d6923fdd706d065. Report an issue: GitHub.

Appendix: source

Thrown at python/sglang/multimodal_gen/runtime/models/dits/sana_wm.py:622

        if not torch.is_grad_enabled():
            self._plucker_emb_cache = (key, chunk_plucker, plucker_emb)
        return plucker_emb

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        timestep: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
        camera_conditions: Optional[torch.Tensor] = None,
        chunk_plucker: Optional[torch.Tensor] = None,
        guidance: Optional[torch.Tensor] = None,  # kept for compat
        **kwargs,
    ) -> torch.Tensor:
        if encoder_hidden_states is None:
            raise ValueError("SANA-WM forward requires encoder_hidden_states.")
        if timestep is None:
            raise ValueError("SANA-WM forward requires timestep.")

        B, C, T_raw, H_raw, W_raw = hidden_states.shape
        p_t, p_h, p_w = self.patch_size
        T = T_raw // p_t
        H = H_raw // p_h
        W = W_raw // p_w
        chunk_size = kwargs.get("chunk_size", self.chunk_size)
        chunk_split_strategy = kwargs.get(
            "chunk_split_strategy", self.chunk_split_strategy
        )
        chunk_index = kwargs.get("chunk_index", None)

        # Patch embed: (B, C, T, H, W) -> (B, T*H*W, D)
        x = self.x_embedder(hidden_states.to(dtype=self.x_embedder.proj.weight.dtype))

        # Timestep AdaLN-single. SANA-WM's LTX sampler passes per-frame
        # timesteps shaped (B, 1, T) so the clean first-frame condition can stay
        # at timestep 0 while remaining latent frames denoise. Keep the scalar

View on GitHub (pinned to 0132848349)