sgl-project/sglang · error · ValueError

Unsupported library: {transformers_or_diffusers}

Error message

Unsupported library: {transformers_or_diffusers}

What it means

load_native dispatches on a library name ('transformers' or 'diffusers'); any other value falls through to this ValueError. The library string comes from the component/loader configuration describing which framework loads the checkpoint natively.

Source

Thrown at python/sglang/multimodal_gen/runtime/loader/component_loaders/component_loader.py:390

                config=config,
                trust_remote_code=server_args.trust_remote_code,
                revision=server_args.revision,
                **load_kwargs,
            )
        elif transformers_or_diffusers == "diffusers":
            from diffusers import AutoModel

            component_model_path = prepare_diffusers_component_path_for_loading(
                component_model_path
            )
            return AutoModel.from_pretrained(
                component_model_path,
                revision=server_args.revision,
                trust_remote_code=server_args.trust_remote_code,
                **load_kwargs,
            )
        else:
            raise ValueError(f"Unsupported library: {transformers_or_diffusers}")

    def resolve_native_transformers_model_class(self, config: PretrainedConfig) -> type:
        return transformers.AutoModel

    def load_customized(
        self, component_model_path: str, server_args: ServerArgs, component_name: str
    ):
        """
        Load the customized version component, implemented and optimized in SGL-diffusion
        """
        raise NotImplementedError(
            f"load_customized not implemented for {self.__class__.__name__}"
        )

    @classmethod
    def _ensure_loaders_registered(cls):
        """
        avoid multiple registration

View on GitHub (pinned to 0132848349)

Solutions

1. Set the library attribute to exactly 'transformers' or 'diffusers' (lowercase)
2. If a third-party loader needs another framework, implement load_customized instead of relying on load_native
3. Strip/normalize the string when reading it from config: value.strip().lower()

Example fix

# before
self.library = 'Transformers'
# after
self.library = 'transformers'

library = (getattr(loader, 'library', '') or '').strip().lower()
assert library in ('transformers', 'diffusers'), f"Unsupported library: {library}"

def is_supported_library(v) -> bool:
    return isinstance(v, str) and v.strip().lower() in ('transformers', 'diffusers')

When it happens

Trigger: A loader or component config setting library to something like 'diffuser', 'Transformers' (wrong case), 'timm', or None reaching load_native; called via _load_native_with_context during component loading.

Common situations: Typos in component library configuration; new loaders copied from an example forgetting to set the library; config serialization changing case or adding whitespace.

Related errors

• f"flash attention version {fa_ver} is not supported."
• Model config does not contain a _class_name attribute. Only
• Model config does not contain a _class_name attribute. Only
• f"Cannot parse checkpoint quantization for {component_name!r
• f"Transformers-managed {component_name!r} quantization requi

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/16afaa578de74f91. Report an issue: GitHub.