sgl-project/sglang · error · ValueError

Stage-1 latent has {z.shape[2]} frames but sink_size={sink_s

Error message

Stage-1 latent has {z.shape[2]} frames but sink_size={sink_size}.

What it means

_refine_one moves the stage-1 latent to device and validates that its temporal dimension (z.shape[2]) is strictly greater than sink_size (default 1). The refiner keeps the first sink_size frame(s) fixed and refines the remainder, so a latent with no refinable frames (e.g. exactly 1 frame with sink_size=1) is invalid.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/sana_wm/refiner.py:628

                    )

        denoised = latent_tokens.float() - velocity_tokens.float() * raw_timestep
        return denoised[:, n_context_tokens:, :].to(self.dtype)

    @torch.inference_mode()
    def _refine_one(
        self,
        latent: torch.Tensor,
        prompt: str,
        *,
        fps: float,
        seed: int,
        sink_size: int = 1,
    ) -> torch.Tensor:
        device = get_local_torch_device()
        z = latent.to(device=device, dtype=self.dtype)
        if z.shape[2] <= sink_size:
            raise ValueError(
                f"Stage-1 latent has {z.shape[2]} frames but sink_size={sink_size}."
            )
        self.log_info(
            "SANA-WM refiner start: latent=%s, fps=%.3f, seed=%d, "
            "sink_size=%d, sigmas=%s, diagnostics=%s",
            tuple(z.shape),
            fps,
            seed,
            sink_size,
            STAGE_2_DISTILLED_SIGMA_VALUES,
            "on" if sana_wm_diagnostics_enabled() else "off",
        )
        log_sana_wm_tensor_stats("refiner.input_latent", z)

        prompt_embeds, prompt_attention_mask = self._encode_prompt(prompt, device)

        sigmas = torch.tensor(
            STAGE_2_DISTILLED_SIGMA_VALUES, dtype=torch.float32, device=device

View on GitHub (pinned to 0132848349)

Solutions

1. Feed a stage-1 latent with at least sink_size+1 frames (typically a full video latent)
2. If intentionally refining single frames, lower sink_size below the frame count
3. Check the stage-1 generation config produced the expected temporal length

Example fix

# before
z = torch.randn(1, 16, 1, 32, 32)  # 1 frame, sink_size=1
refined = stage._refine_one(z, fps=25.0, seed=0)
# after
z = torch.randn(1, 16, 33, 32, 32)  # sink + 32 refinable frames
refined = stage._refine_one(z, fps=25.0, seed=0)

sink = stage.sink_size if hasattr(stage, "sink_size") else 1
assert latent.shape[2] > sink, f"need >{sink} frames, got {latent.shape[2]}"

def refinable(latent: torch.Tensor, sink_size: int = 1) -> bool:
    return latent.ndim == 5 and latent.shape[2] > sink_size

try:
    refined = stage._refine_one(latent, fps=fps, seed=seed)
except ValueError:
    refined = latent  # skip refinement for degenerate clips
    log.warning("skipping SANA-WM refine: too few frames")

When it happens

Trigger: Feeding a stage-1 latent whose frame count equals or is below sink_size (a single-frame latent with the default sink_size=1), or configuring a larger sink_size than frames-1.

Common situations: Testing the refiner with a tiny 1-frame dummy latent; still/image mode where stage 1 emitted a sink-only latent; misconfigured sink_size in the refiner options.

Related errors

• SANA-WM refiner expects 5D latents shaped (B, C, T, H, W), g
• SANA-WM refiner requires a string prompt or one prompt per b
• SANA-WM refiner requires batch.latents from stage 1.
• SANA-WM refiner decoding expects decoded video shaped (B, C,
• SANA-WM refiner decoding expected a sink frame plus refined

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/5037de58dea92e97. Report an issue: GitHub.