sgl-project/sglang · error · ValueError

SANA-WM forward_long requires timestep.

Error message

SANA-WM forward_long requires timestep.

What it means

forward_long requires the diffusion timestep for each chunk; None raises immediately. Chunked streaming still denoises, so a timestep per chunk is mandatory.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/dits/sana_wm.py:758

        chunk_plucker: Optional[torch.Tensor] = None,
        *,
        kv_cache: Optional[list] = None,
        save_kv_cache: bool = True,
        start_f: Optional[int] = None,
        end_f: Optional[int] = None,
        frame_index: Optional[torch.Tensor] = None,
        **kwargs,
    ) -> Tuple[torch.Tensor, list]:
        """Streaming autoregressive forward over a chunk of latent frames.

        RoPE / camera / plücker are windowed to the chunk's GLOBAL frame range
        ``[start_f, end_f)``; a per-block 10-slot ``kv_cache`` carries recurrent
        state / concat-windows across chunks. Returns ``(out, new_cache)``.
        """
        if encoder_hidden_states is None:
            raise ValueError("SANA-WM forward_long requires encoder_hidden_states.")
        if timestep is None:
            raise ValueError("SANA-WM forward_long requires timestep.")

        if kv_cache is None:
            kv_cache = [[None] * _NUM_STREAM_CACHE_SLOTS for _ in self.blocks]

        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
        start = 0 if start_f is None else int(start_f)
        end = start + T if end_f is None else int(end_f)

        x = self.x_embedder(hidden_states.to(dtype=self.x_embedder.proj.weight.dtype))

        # Timestep AdaLN-single: force the framewise (B, 1, T) path so blocks
        # always apply per-frame modulation.
        if timestep.dim() == 1:
            timestep = timestep[:, None, None].expand(-1, 1, T)

View on GitHub (pinned to 0132848349)

Solutions

1. Pass the scheduler timestep tensor for the current denoising step to every forward_long chunk call
2. Keep the same t across chunks within one denoising step

Example fix

# before
out, cache = model.forward_long(h, ehs, start_f=sf, end_f=ef)
# after
out, cache = model.forward_long(h, t, encoder_hidden_states=ehs, start_f=sf, end_f=ef)

assert timestep is not None at the top of your chunk loop

When it happens

Trigger: Calling forward_long without timestep or with timestep=None while iterating chunks.

Common situations: Streaming wrapper that supplies timestep only once, or per-pixel-frame instead of per-chunk; porting from a wrapper where t defaulted.

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

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/b0a28892e8eda375. Report an issue: GitHub.