sgl-project/sglang · error · ValueError

Unsupported dims: {self.dims}

Error message

Unsupported dims: {self.dims}

What it means

LatentUpsampler.forward only supports 3D (video: b c f h w) and modified-4D/5D rearrange paths keyed by self.dims and upscale_factors; any other dims value falls through to this ValueError. The module is built for a fixed spatial/temporal upsampling geometry.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/upsampler/latent_upsampler.py:92

                p1=self.upscale_factors[0],
                p2=self.upscale_factors[1],
                p3=self.upscale_factors[2],
            )
        elif self.dims == 2:
            return rearrange(
                x,
                "b (c p1 p2) h w -> b c (h p1) (w p2)",
                p1=self.upscale_factors[0],
                p2=self.upscale_factors[1],
            )
        elif self.dims == 1:
            return rearrange(
                x,
                "b (c p1) f h w -> b c (f p1) h w",
                p1=self.upscale_factors[0],
            )
        else:
            raise ValueError(f"Unsupported dims: {self.dims}")


class ResBlock(torch.nn.Module):
    """Residual block with two conv layers, group norm, and SiLU activation."""

    def __init__(
        self, channels: int, mid_channels: Optional[int] = None, dims: int = 3
    ):
        super().__init__()
        if mid_channels is None:
            mid_channels = channels
        conv = torch.nn.Conv2d if dims == 2 else torch.nn.Conv3d
        self.conv1 = conv(channels, mid_channels, kernel_size=3, padding=1)
        self.norm1 = torch.nn.GroupNorm(32, mid_channels)
        self.conv2 = conv(mid_channels, channels, kernel_size=3, padding=1)
        self.norm2 = torch.nn.GroupNorm(32, channels)
        self.activation = torch.nn.SiLU()

View on GitHub (pinned to 0132848349)

Solutions

1. Set dims to the supported value matching your data (5 for video latents b c f h w)
2. For plain 2D image latent upsampling use a different module (e.g. a VAE decoder-side upsampler or interpolate)
3. Verify tensor shape before forward: x.ndim should equal self.dims + 2

Example fix

// before
ups = LatentUpsampler(..., dims=2)
y = ups(x)  # x: b c h w
// after
ups = LatentUpsampler(..., dims=5)
y = ups(x)  # x: b c f h w video latents

assert x.ndim == ups.dims + 2 or (x.ndim == ups.dims and ups.dims in (3, 4)), f"unexpected rank {x.ndim} for dims={ups.dims}"

def is_supported_upsampler_input(x: torch.Tensor, ups) -> bool:
    return x.ndim in (ups.dims, ups.dims + 1, ups.dims + 2) and ups.dims in (3, 5)

try:
    y = ups(x)
except ValueError as e:
    if "Unsupported dims" in str(e):
        raise ValueError(f"feed {ups.dims + 2}-D latents to LatentUpsampler") from e
    raise

When it happens

Trigger: Instantiating LatentUpsampler with dims not equal to the supported values (e.g. dims=2 for plain 2D image latents) and calling forward(), or feeding tensors whose rank does not match the configured path.

Common situations: Reusing the video latent upsampler for 2D image latents, changing dims in config, or passing incorrectly rearranged tensors from a different pipeline stage.

Related errors

• Z-Image text embeddings must have shape [seq, dim] or [batch
• Z-Image caption tensor must have rank 2 or 3
• f"Expected camera embedding shape [B, C, F, H, W], got {tupl
• f"Input shape {tuple(x.shape)} must be divisible by patch_si
• Expected [F, H, W, C] uint8 video, got shape={tuple(video_ui

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/df3689e1eb484daa. Report an issue: GitHub.