sgl-project/sglang · error · ValueError

SANA-WM refiner requires batch.latents from stage 1.

Error message

SANA-WM refiner requires batch.latents from stage 1.

What it means

The refiner's forward() requires batch.latents to be set by the preceding stage-1 generation stage. If batch.latents is None the refiner has nothing to refine, so it fails fast with this ValueError rather than producing garbage.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/sana_wm/refiner.py:711

                height=noisy.shape[3],
                width=noisy.shape[4],
                patch_size=patch_size,
                patch_size_t=patch_size_t,
            )
            log_sana_wm_tensor_stats(
                f"refiner.step_{step_idx}.velocity_current",
                velocity_5d.to(self.dtype),
            )
            log_sana_wm_tensor_stats(f"refiner.step_{step_idx}.current_latent", noisy)

        refined = torch.cat([sink, noisy], dim=2)
        log_sana_wm_tensor_stats("refiner.output_latent", refined)
        return refined

    @torch.inference_mode()
    def forward(self, batch: Req, server_args: ServerArgs) -> Req:
        if batch.latents is None:
            raise ValueError("SANA-WM refiner requires batch.latents from stage 1.")
        if batch.latents.ndim != 5:
            raise ValueError(
                "SANA-WM refiner expects 5D latents shaped (B, C, T, H, W), "
                f"got {tuple(batch.latents.shape)}."
            )

        if sana_wm_skip_refiner_enabled(batch):
            if batch.extra is None:
                batch.extra = {}
            batch.extra["sana_wm_refiner_applied"] = False
            self.log_info(
                "SANA-WM LTX-2 refiner skipped by SGLANG_SANA_WM_SKIP_REFINER."
            )
            return batch

        batch_size = int(batch.latents.shape[0])
        prompts = self._prompts_for_batch(batch, batch_size)
        fps = float(getattr(batch, "fps", 16) or 16)

View on GitHub (pinned to 0132848349)

Solutions

1. Ensure the stage-1 SANA-WM denoising stage runs before the refiner and sets batch.latents
2. Verify pipeline stage ordering in pipeline_config
3. When unit-testing the refiner, construct batch.latents explicitly (5D tensor)

Example fix

# before
batch.latents = None
refiner.forward(batch, server_args)
# after
batch.latents = stage1_denoiser.forward(batch, server_args).latents
refiner.forward(batch, server_args)

if batch.latents is None:
    raise RuntimeError("stage-1 did not produce latents; check pipeline order")

null

try:
    out = refiner.forward(batch, server_args)
except ValueError as e:
    if "batch.latents" in str(e):
        batch = stage1.forward(batch, server_args)
        out = refiner.forward(batch, server_args)
    else:
        raise

When it happens

Trigger: Running the SANA-WM refiner stage without a prior denoising stage in the pipeline, or the stage-1 output failing to write latents onto the Req (skip flag, serialization drop, or wrong pipeline wiring).

Common situations: Building a custom pipeline that omits the stage-1 denoiser; a conditioning/skip path that returns the batch before latents are assigned; debugging the refiner in isolation with a hand-built Req.

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 expects 5D latents shaped (B, C, T, H, W), g
• 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/962b288bb2d98c17. Report an issue: GitHub.