sgl-project/sglang · error · TypeError

Expected Hunyuan3D2PipelineConfig, got {type(config)}

Error message

Expected Hunyuan3D2PipelineConfig, got {type(config)}

What it means

Raised by Hunyuan3D2Pipeline.load_modules when server_args.pipeline_config is not a Hunyuan3D2PipelineConfig instance. The pipeline reads component paths (shape model, texture model) exclusively from this typed config, so any other config type (or None) is rejected early with a TypeError rather than failing later with attribute errors.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines/hunyuan3d_pipeline.py:505

            )
            components["delight_scheduler"] = (
                EulerAncestralDiscreteScheduler.from_config(delight_scheduler_config)
            )

        return components

    # Module loading override
    def load_modules(
        self,
        server_args: ServerArgs,
        loaded_modules: dict[str, torch.nn.Module] | None = None,
    ) -> dict[str, Any]:
        """Load Hunyuan3D shape and optional texture components."""
        del loaded_modules

        config = server_args.pipeline_config
        if not isinstance(config, Hunyuan3D2PipelineConfig):
            raise TypeError(f"Expected Hunyuan3D2PipelineConfig, got {type(config)}")

        model_path = config.shape_model_path or server_args.model_path

        logger.info("Loading Hunyuan3D shape models from %s", model_path)

        config_path, ckpt_path = self._resolve_shape_dir(
            model_path,
            config.shape_subfolder,
            config.shape_use_safetensors,
            config.shape_variant,
        )

        with open(config_path, "r") as f:
            model_config = yaml.safe_load(f)

        ckpt = self._load_and_split_checkpoint(ckpt_path, config.shape_use_safetensors)

        dtype = torch.float16

View on GitHub (pinned to 0132848349)

Solutions

1. Set server_args.pipeline_config to a Hunyuan3D2PipelineConfig instance before launching the Hunyuan3D pipeline (configure shape_model_path/texture settings on it)
2. If constructing ServerArgs manually, import the config from the same module/version of sglang as the pipeline to avoid class-identity mismatches
3. Add an isinstance assertion in your launcher right after assigning pipeline_config to fail fast with your own message

Example fix

// before
server_args.pipeline_config = None  # or generic config
pipeline.load_modules(server_args)

// after
from sglang.multimodal_gen.runtime.pipelines.hunyuan3d_pipeline import Hunyuan3D2PipelineConfig
server_args.pipeline_config = Hunyuan3D2PipelineConfig(shape_model_path=...)
assert isinstance(server_args.pipeline_config, Hunyuan3D2PipelineConfig)
pipeline.load_modules(server_args)

from sglang.multimodal_gen.runtime.pipelines.hunyuan3d_pipeline import Hunyuan3D2PipelineConfig
if not isinstance(server_args.pipeline_config, Hunyuan3D2PipelineConfig):
    server_args.pipeline_config = Hunyuan3D2PipelineConfig(shape_model_path=server_args.model_path)

def is_hunyuan3d_config(cfg: object) -> TypeGuard[Hunyuan3D2PipelineConfig]:
    return isinstance(cfg, Hunyuan3D2PipelineConfig)

When it happens

Trigger: Calling load_modules(server_args) on the Hunyuan3D pipeline when server_args.pipeline_config is None, a generic PipelineConfig, or a config belonging to a different model family; e.g. launching the server with --pipeline hunyuan3d but without constructing Hunyuan3D2PipelineConfig in server args.

Common situations: Server args built programmatically or reused from another model's launch path; a default pipeline_config left as None; tests instantiating the pipeline with bare ServerArgs; version changes that renamed or moved the config class so an old import produces a lookalike object.

Related errors

• JoyEchoPipeline requires JoyEchoPipelineConfig, got {type(co
• Unknown serve backend {name!r}. Available values: {available
• Multiple distributions register serve backend {name!r}: {pro
• Failed to load serve backend {name!r} from {self._entry_poin
• Serve backend {name!r} factory returned {type(backend).__nam

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/95c0a05bc9da96c9. Report an issue: GitHub.