sgl-project/sglang · error · ValueError

Unsupported mc_algo {mc_algo}, available: {list(SurfaceExtra

Error message

Unsupported mc_algo {mc_algo}, available: {list(SurfaceExtractors.keys())}

What it means

enable_flashvdm_decoder validates the mc_algo argument against the SurfaceExtractors registry dict before constructing the surface extractor. Any mc_algo string not present as a key (typically 'mc' and 'dmc') raises this ValueError listing the valid options.

Source

Thrown at python/sglang/multimodal_gen/runtime/models/vaes/hunyuan3d_vae.py:1137

        grid_logits = self.volume_decoder(latents, self.geo_decoder, **kwargs)
        outputs = self.surface_extractor(grid_logits, **kwargs)
        return outputs

    def enable_flashvdm_decoder(
        self,
        enabled: bool = True,
        adaptive_kv_selection=True,
        topk_mode="mean",
        mc_algo="dmc",
    ):
        """Enable or disable FlashVDM decoder for faster inference."""
        if enabled:
            if adaptive_kv_selection:
                self.volume_decoder = FlashVDMVolumeDecoding(topk_mode)
            else:
                self.volume_decoder = HierarchicalVolumeDecoding()
            if mc_algo not in SurfaceExtractors:
                raise ValueError(
                    f"Unsupported mc_algo {mc_algo}, available: {list(SurfaceExtractors.keys())}"
                )
            self.surface_extractor = SurfaceExtractors[mc_algo]()
        else:
            self.volume_decoder = VanillaVolumeDecoder()
            self.surface_extractor = MCSurfaceExtractor()


class ShapeVAE(VectsetVAE):
    """Shape VAE for 3D mesh generation from latent codes."""

    _aliases = ["hy3dgen.shapegen.models.ShapeVAE"]

    def __init__(
        self,
        *,
        num_latents: int,
        embed_dim: int,

View on GitHub (pinned to 0132848349)

Solutions

Use one of the keys printed in the error message — normally 'mc' (scikit-image marching cubes) or 'dmc' (diso DiffDMC)
Check the SurfaceExtractors dict in hunyuan3d_vae.py for the exact registered keys in your version
If adding a custom extractor, register it: SurfaceExtractors['my_algo'] = MyExtractor before calling enable_flashvdm_decoder

Example fix

// before
model.enable_flashvdm_decoder(enabled=True, mc_algo='marching_cubes')

// after
model.enable_flashvdm_decoder(enabled=True, mc_algo='mc')

Defensive patterns

Strategy: validation

Validate before calling

from sglang.multimodal_gen.runtime.models.vaes.hunyuan3d_vae import SurfaceExtractors
mc_algo = 'mc'  # or 'dmc'
assert mc_algo in SurfaceExtractors, f'mc_algo must be one of {list(SurfaceExtractors)}'
model.enable_flashvdm_decoder(enabled=True, mc_algo=mc_algo)

Type guard

def is_valid_mc_algo(name: str) -> bool:
    from sglang.multimodal_gen.runtime.models.vaes.hunyuan3d_vae import SurfaceExtractors
    return isinstance(name, str) and name in SurfaceExtractors

Try / catch

try:
    model.enable_flashvdm_decoder(enabled=True, mc_algo=mc_algo)
except ValueError as e:
    raise ValueError(f'bad mc_algo {mc_algo!r}; choose from mc/dmc') from e

Prevention

Derive mc_algo from the SurfaceExtractors registry keys, not from docs memory
Watch for case: keys are lowercase ('dmc', not 'DMC')
Register custom extractors in SurfaceExtractors before use

When it happens

Trigger: Calling enable_flashvdm_decoder(enabled=True, mc_algo=...) with a typo'd or unsupported algorithm name, e.g. 'marching_cubes', 'DMC', or 'cuda_dmc'. The check only runs when enabled=True and adaptive_kv_selection path is configured.

Common situations: Copying an mc_algo value from a different codebase or model card; case mismatch ('DMC' vs 'dmc'); custom forks that register an extractor under a new name but forget to add it to SurfaceExtractors.

Understand the failure class

Background: "Invalid value" and "allowed values are" config errors: what your library rejected and how to fix it — this error's family across 41 libraries.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/9cecab9c2bd064b5. Report an issue: GitHub.