sgl-project/sglang · error · ValueError

Either spatial_upsample or temporal_upsample must be True

Error message

Either spatial_upsample or temporal_upsample must be True

What it means

The upsampling stage of the latent upsampler must upscale in at least one dimension: spatial (H×W) or temporal (T). Passing spatial_upsample=False together with temporal_upsample=False leaves the 'else' branch with nothing to build, so __init__ raises.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/upsampler/latent_upsampler.py:233

                self.upsampler = SpatialRationalResampler(
                    mid_channels=mid_channels, scale=self.spatial_scale
                )
            else:
                self.upsampler = torch.nn.Sequential(
                    torch.nn.Conv2d(
                        mid_channels, 4 * mid_channels, kernel_size=3, padding=1
                    ),
                    PixelShuffleND(2),
                )
        elif temporal_upsample:
            self.upsampler = torch.nn.Sequential(
                torch.nn.Conv3d(
                    mid_channels, 2 * mid_channels, kernel_size=3, padding=1
                ),
                PixelShuffleND(1),
            )
        else:
            raise ValueError(
                "Either spatial_upsample or temporal_upsample must be True"
            )

        self.post_upsample_res_blocks = torch.nn.ModuleList(
            [ResBlock(mid_channels, dims=dims) for _ in range(num_blocks_per_stage)]
        )

        self.final_conv = conv(mid_channels, in_channels, kernel_size=3, padding=1)

    def forward(self, latent: torch.Tensor) -> torch.Tensor:
        b, _, f, _, _ = latent.shape

        if self.dims == 2:
            x = rearrange(latent, "b c f h w -> (b f) c h w")
            x = self.initial_conv(x)
            x = apply_group_norm_silu(x, self.initial_norm, self.initial_activation)
            for block in self.res_blocks:
                x = block(x)

View on GitHub (pinned to 0132848349)

Solutions

1. Enable at least one of spatial_upsample or temporal_upsample for every stage that instantiates this block
2. If the stage truly should not upsample, skip constructing the block entirely instead of passing both False
3. Check the stage's target resolution vs input resolution to decide which flag to set

Example fix

# before
block = UpsampleBlock(c, spatial_upsample=False, temporal_upsample=False)
# after
block = UpsampleBlock(c, spatial_upsample=True, temporal_upsample=False)

if not (spatial_upsample or temporal_upsample):
    raise ValueError("stage config must enable spatial or temporal upsampling")
# before constructing the block

def is_valid_upsample_cfg(cfg) -> bool:
    return bool(cfg.get("spatial_upsample")) or bool(cfg.get("temporal_upsample"))

When it happens

Trigger: Constructing the upsampler block with both flags False, e.g. UpsampleBlock(mid_channels, spatial_upsample=False, temporal_upsample=False) or a config where both are disabled for the final stage.

Common situations: Config generated programmatically that disables all upsampling for a 'no-op' stage; YAML/JSON config where the last stage defaults to False for both keys; editing a config and accidentally turning off the wrong flag.

Related errors

• Unsupported scale {scale}. Choose from {list(mapping.keys())
• Cosmos3 AVAE dec_strides product must equal hop_size: produc
• Invalid threshold_type for topk: {threshold_type}. Choose 'q
• Invalid threshold_type: {threshold_type}. Choose 'query_head
• SGLANG_DIFFUSION_ATTENTION_CONFIG is not set

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/3709a2bf67271c67. Report an issue: GitHub.