Lightning-AI/pytorch-lightning · critical · ValueError

The instantiator {instantiator_path!r} from the checkpoint i

Error message

The instantiator {instantiator_path!r} from the checkpoint is not in the allowlist of trusted instantiators and was blocked to prevent arbitrary code execution. If you trust this checkpoint, add the path to `lightning.pytorch.core.saving._ALLOWED_INSTANTIATORS` before loading.

What it means

Security guard in _load_state: checkpoints may embed an `_instantiator` entry naming a callable (module.path.to.fn) that Lightning will import and call to rebuild the object. Because that is arbitrary code execution, the path must appear in the explicit allowlist `lightning.pytorch.core.saving._ALLOWED_INSTANTIATORS` or loading aborts.

Source

Thrown at src/lightning/pytorch/core/saving.py:166

        cls_kwargs_loaded.update(checkpoint.get(cls.CHECKPOINT_HYPER_PARAMS_KEY, {}))

        # 3. Ensure that `cls_kwargs_old` has the right type, back compatibility between dict and Namespace
        cls_kwargs_loaded = _convert_loaded_hparams(cls_kwargs_loaded, checkpoint.get(cls.CHECKPOINT_HYPER_PARAMS_TYPE))

        # 4. Update cls_kwargs_new with cls_kwargs_old, such that new has higher priority
        args_name = checkpoint.get(cls.CHECKPOINT_HYPER_PARAMS_NAME)
        if args_name and args_name in cls_init_args_name:
            cls_kwargs_loaded = {args_name: cls_kwargs_loaded}

    _cls_kwargs = {}
    _cls_kwargs.update(cls_kwargs_loaded)
    _cls_kwargs.update(cls_kwargs_new)

    instantiator = None
    instantiator_path = _cls_kwargs.pop("_instantiator", None)
    if instantiator_path is not None:
        if instantiator_path not in _ALLOWED_INSTANTIATORS:
            raise ValueError(
                f"The instantiator {instantiator_path!r} from the checkpoint is not in the allowlist of trusted"
                " instantiators and was blocked to prevent arbitrary code execution. If you trust this checkpoint,"
                " add the path to `lightning.pytorch.core.saving._ALLOWED_INSTANTIATORS` before loading."
            )
        module_path, name = instantiator_path.rsplit(".", 1)
        instantiator = getattr(__import__(module_path, fromlist=[name]), name)

    if not cls_spec.varkw:
        # filter kwargs according to class init unless it allows any argument via kwargs
        _cls_kwargs = {k: v for k, v in _cls_kwargs.items() if k in cls_init_args_name}

    obj = instantiator(cls, _cls_kwargs) if instantiator else cls(**_cls_kwargs)

    if isinstance(obj, pl.LightningDataModule):
        if obj.__class__.__qualname__ in checkpoint:
            obj.load_state_dict(checkpoint[obj.__class__.__qualname__])
        return obj

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Inspect checkpoint['hyper_parameters']['_instantiator'] with torch.load to see the path
2. If you trust the source, append the path: from lightning.pytorch.core.saving import _ALLOWED_INSTANTIATORS; _ALLOWED_INSTANTIATORS.add(path) before loading
3. Otherwise strip the key: pop '_instantiator' from the checkpoint's hyper_parameters and load without it

Example fix

// before
model = MyModel.load_from_checkpoint(ckpt)  # ValueError: instantiator blocked
// after
from lightning.pytorch.core.saving import _ALLOWED_INSTANTIATORS
_ALLOWED_INSTANTIATORS.add("litdata.streaming.ini")
model = MyModel.load_from_checkpoint(ckpt)

ckpt = torch.load(path, map_location="cpu", weights_only=False)
inst = ckpt.get("hyper_parameters", {}).get("_instantiator")
if inst is not None and inst not in _ALLOWED_INSTANTIATORS:
    # decide: trust it or strip it
    ckpt["hyper_parameters"].pop("_instantiator")

from lightning.pytorch.core.saving import _ALLOWED_INSTANTIATORS
try:
    model = cls.load_from_checkpoint(p)
except ValueError as e:
    if "_ALLOWED_INSTANTIATORS" in str(e) and trust_source:
        _ALLOWED_INSTANTIATORS.add(extract_path_from(e))
    else:
        raise

When it happens

Trigger: Loading a checkpoint saved by Lightning AI's studio/app flows (or a crafted one) whose hyperparameters contain `_instantiator='some.module.fn'` where 'some.module.fn' is not in the allowlist.

Common situations: Downloading or receiving a checkpoint created elsewhere (studio export, another org, tutorial artifact) and calling load_from_checkpoint on it; the embedded instantiator path is not one Lightning trusts by default.

Related errors

• Could not find a distributed model in the provided checkpoin
• Found multiple distributed models in the given state. Loadin
• The path {str(path)!r} does not point to a valid checkpoint.
• Failed to load checkpoint directly into the model. The given
• The model contains a key '{full_param_name}' that does not e

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/de68aa0033e5e93b. Report an issue: GitHub.