sgl-project/sglang · error · ValueError

SANA-WM refiner decoding expected a sink frame plus refined

Error message

SANA-WM refiner decoding expected a sink frame plus refined frames, got temporal length {frames.shape[2]}.

What it means

After the 5D check, decode() requires frames.shape[2] > 1: the decoded clip must contain the sink frame plus at least one refined frame. A temporal length of exactly 1 means only the sink survived (refiner produced nothing) and dropping it would yield an empty video.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/sana_wm/refiner.py:788

            self._drop_refiner_sink = True

    @torch.no_grad()
    def decode(
        self,
        latents: torch.Tensor,
        server_args: ServerArgs,
        *,
        vae_dtype: torch.dtype,
    ) -> torch.Tensor:
        frames = super().decode(latents, server_args, vae_dtype=vae_dtype)
        log_sana_wm_tensor_stats("refiner.decode.frames_with_sink", frames)
        if frames.ndim != 5:
            raise ValueError(
                "SANA-WM refiner decoding expects decoded video shaped "
                f"(B, C, T, H, W), got {tuple(frames.shape)}."
            )
        if frames.shape[2] <= 1:
            raise ValueError(
                "SANA-WM refiner decoding expected a sink frame plus refined "
                f"frames, got temporal length {frames.shape[2]}."
            )
        if not getattr(self, "_drop_refiner_sink", True):
            log_sana_wm_tensor_stats("refiner.decode.frames_output", frames)
            return frames
        # Match NVlabs `inference_sana_wm.py`: decode with the clean sink anchor,
        # then drop the first frame from the returned video.
        frames = frames[:, :, 1:].contiguous()
        log_sana_wm_tensor_stats("refiner.decode.frames_output", frames)
        return frames

View on GitHub (pinned to 0132848349)

Solutions

1. Ensure the refined latent contains sink + >=1 frames before decode (see error 2582's constraint)
2. Audit any slicing between refine and decode that could drop the refined segment
3. Verify the VAE decodes the full temporal extent

Example fix

# before
latents = z[:, :, :1]  # accidentally kept only the sink frame
# after
latents = z  # sink + refined frames retained

assert frames.ndim == 5 and frames.shape[2] > 1, f"decoded clip lacks refined frames: {tuple(frames.shape)}"

def has_sink_plus_frames(t: torch.Tensor) -> bool:
    return t.ndim == 5 and t.shape[2] > 1

null

When it happens

Trigger: Refining a latent where all non-sink frames were lost/empty; sink_size >= frame count so the refiner emitted only the sink; a VAE that decodes only the first frame.

Common situations: Misconfigured sink_size; an upstream slicing bug trimming refined frames to zero; testing with a minimal 1-frame latent end-to-end.

Related errors

• SANA-WM refiner requires a string prompt or one prompt per b
• Stage-1 latent has {z.shape[2]} frames but sink_size={sink_s
• SANA-WM refiner requires batch.latents from stage 1.
• SANA-WM refiner expects 5D latents shaped (B, C, T, H, W), g
• SANA-WM refiner decoding expects decoded video shaped (B, C,

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/89df9173cc80bfbb. Report an issue: GitHub.