sgl-project/sglang · error · ValueError

SANA-WM height/width must be divisible by the LTX-2 spatial

Error message

SANA-WM height/width must be divisible by the LTX-2 spatial stride ({h_stride}, {w_stride}); got height={batch.height}, width={batch.width}.

What it means

SANA-WM's latent shape preparation requires spatial dimensions divisible by the LTX-2 VAE strides. If batch.height or batch.width isn't a multiple of the respective stride, latent dims wouldn't be integers, so prepare_latent_shape rejects the request with the exact strides and offending dimensions.

Source

Thrown at python/sglang/multimodal_gen/configs/pipeline_configs/sana_wm.py:181

        return ModelDeploymentConfig(
            dit_layerwise_offload_modes=("memory",),
            # Conservative auto-FSDP gate for the 720p world-model path. Users
            # can still force FSDP explicitly on smaller cards.
            fsdp_auto_min_available_memory_gb=60,
        )

    # --- Latent shape ---
    def prepare_latent_shape(self, batch, batch_size: int, num_frames: int):
        """
        Returns 5D latent shape: (B, 128, T_latent, H_sp, W_sp).
        T_latent = ceil((num_frames - 1) / temporal_stride) + 1
        """
        t_stride = self.vae_stride[0]
        h_stride = self.vae_stride[1]
        w_stride = self.vae_stride[2] if len(self.vae_stride) > 2 else h_stride

        if batch.height % h_stride != 0 or batch.width % w_stride != 0:
            raise ValueError(
                "SANA-WM height/width must be divisible by the LTX-2 spatial "
                f"stride ({h_stride}, {w_stride}); got "
                f"height={batch.height}, width={batch.width}."
            )

        T_latent = (num_frames - 1) // t_stride + 1
        H_sp = batch.height // h_stride
        W_sp = batch.width // w_stride
        z_dim = self.vae_config.arch_config.latent_channels  # 128

        return (batch_size, z_dim, T_latent, H_sp, W_sp)

    def adjust_num_frames(self, num_frames: int) -> int:
        """Ensure (num_frames - 1) is divisible by VAE temporal stride."""
        t_stride = self.vae_stride[0]
        if (num_frames - 1) % t_stride != 0:
            adjusted = ((num_frames - 1) // t_stride) * t_stride + 1
            logger.warning(

View on GitHub (pinned to 0132848349)

Solutions

1. Round height/width to the nearest multiple of the stride: h = round(h / h_stride) * h_stride
2. Print/inspect self.vae_stride and ensure both dimensions satisfy h % h_stride == 0 and w % w_stride == 0
3. Snap user input at the API boundary rather than letting raw values through

Example fix

# before
pipe(height=1000, width=700)  # 1000 % 32 == 16 -> error

# after
h_stride, w_stride = 32, 32
pipe(height=round(1000/h_stride)*h_stride, width=round(700/w_stride)*w_stride)  # 992x704

h_stride, w_stride = pipe.vae_stride[1], pipe.vae_stride[2] if len(pipe.vae_stride) > 2 else pipe.vae_stride[1]
if batch.height % h_stride or batch.width % w_stride:
    batch.height = round(batch.height / h_stride) * h_stride
    batch.width = round(batch.width / w_stride) * w_stride

try:
    latent_shape = pipe.prepare_latent_shape(batch, batch_size, num_frames)
except ValueError as e:
    if "divisible by" in str(e):
        batch.height = round(batch.height / h_stride) * h_stride
        batch.width = round(batch.width / w_stride) * w_stride
        latent_shape = pipe.prepare_latent_shape(batch, batch_size, num_frames)
    else:
        raise

When it happens

Trigger: Requesting video/image generation with height/width like 1000 (not divisible by a stride of e.g. 32), or odd sizes from arbitrary user input; mixing up the stride order when setting vae_stride.

Common situations: Free-form resolution fields in a UI; upscaling/scaling math producing non-multiple sizes; assuming any even number works when the LTX-2 VAE needs multiples of 32/64.

Related errors

• Invalid latent H/W computed from batch.height/width: {batch.
• Invalid spatial patching for packed token latents. Expected
• camera_conditions must have shape (T,20) or (B,T,20), got {t
• Video generation failed: {error_msg}
• Lost connection to server after {consecutive_errors} consecu

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/a25e9aff96c49248. Report an issue: GitHub.