sgl-project/sglang · error · ImportError

ModelOpt export functionality is not available. Please ensur

Error message

ModelOpt export functionality is not available. Please ensure you have the latest version of modelopt installed.

What it means

The ModelOpt export path (_maybe_export_modelopt) imports `export_hf_checkpoint` from modelopt.torch.export; older modelopt releases do not have that module, so the import fails and is re-raised as this ImportError. It means your installed modelopt version is too old (or too minimal) to export a HuggingFace checkpoint.

Source

Thrown at python/sglang/srt/model_loader/loader.py:3870

    ) -> None:
        """
        Export the quantized model to HuggingFace format using ModelOpt export API.

        Args:
            model: The quantized model to export
            export_path: Directory path to export the model to
            model_path: Path to the original model (for tokenizer export)
            trust_remote_code: Whether to trust remote code for tokenizer loading

        Raises:
            ImportError: If ModelOpt export functionality is not available
            Exception: If export fails
        """
        try:
            from modelopt.torch.export import export_hf_checkpoint
            from transformers import AutoTokenizer
        except ImportError as e:
            raise ImportError(
                "ModelOpt export functionality is not available. "
                "Please ensure you have the latest version of modelopt installed."
            ) from e

        # Create export directory if it doesn't exist
        os.makedirs(export_path, exist_ok=True)

        # Export the quantized model
        export_hf_checkpoint(model, export_dir=export_path)

        # Export the tokenizer if model_path is provided
        if model_path:
            try:
                tokenizer = AutoTokenizer.from_pretrained(
                    model_path, trust_remote_code=trust_remote_code
                )
                tokenizer.save_pretrained(export_path)
                rank0_log(f"Tokenizer exported to {export_path}")

View on GitHub (pinned to 0132848349)

Solutions

Upgrade modelopt: pip install -U nvidia-modelopt (export API exists in recent releases)
Verify: python -c "from modelopt.torch.export import export_hf_checkpoint"
Ensure transformers is installed/importable since it's imported in the same try block
If you don't need export, remove the export path argument

Example fix

# before
pip install nvidia-modelopt==0.09  # lacks export API
python -m sglang.launch_server --load-format modelopt --modelopt-export-path ./out

# after
pip install -U nvidia-modelopt
python -m sglang.launch_server --load-format modelopt --modelopt-export-path ./out

Defensive patterns

Strategy: validation

Validate before calling

def modelopt_export_available() -> bool:
    try:
        from modelopt.torch.export import export_hf_checkpoint  # noqa
        import transformers  # noqa
        return True
    except ImportError:
        return False

if export_path and not modelopt_export_available():
    raise SystemExit("Upgrade nvidia-modelopt to use --modelopt-export-path")

Try / catch

try:
    server.launch(args_with_export)
except ImportError as e:
    if "ModelOpt export" in str(e): print("pip install -U nvidia-modelopt"); raise

Prevention

Check for the export API before enabling export_path
Keep modelopt upgraded in lockstep with sglang
Don't request export on clusters with old pinned modelopt wheels

When it happens

Trigger: Running sglang with a ModelOpt load format plus an export path (e.g. --modelopt-export-path / --json-model-override-args with export_path) while the installed nvidia-modelopt lacks modelopt.torch.export.export_hf_checkpoint.

Common situations: Pinning an old modelopt version; using a modelopt wheel built without export extras; transformers missing is also caught by the same try block.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/22016c9b5c493289. Report an issue: GitHub.