Lightning-AI/pytorch-lightning · error · AttributeError

'{type(self).__name__}' object has no attribute '{name}'

Error message

'{type(self).__name__}' object has no attribute '{name}'

What it means

This AttributeError comes from `_UnloadedTensor` (Lightning's lazy checkpoint loading via `_lazy_load`), which wraps tensors that live on disk in a checkpoint instead of memory. It forwards most attribute access to the underlying meta tensor or materializes the tensor for a few special names (`contiguous`, `cuda`, `half`, `data`, `to`), but any other tensor attribute/method on an unloaded tensor cannot be served and raises.

Source

Thrown at src/lightning/fabric/utilities/load.py:176

            "grad_fn",
            "is_meta",
            "layout",
            "names",
            "ndim",
            "output_nr",
            "requires_grad",
            "retains_grad",
            "size",
            "shape",
            "volatile",
        }:
            return getattr(self.metatensor, name)

        # materializing these is needed for quantization (see lit-gpt)
        if name in {"contiguous", "cuda", "half", "data", "to"}:
            return getattr(self._load_tensor(), name)

        raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")

    def __repr__(self) -> str:
        return f"{self.__class__.__name__}({repr(self.metatensor)})"


# Modified from https://github.com/lernapparat/torchhacks by Thomas Viehmann
class _LazyLoadingUnpickler(pickle.Unpickler):
    def __init__(self, file: IO, file_reader: torch.PyTorchFileReader) -> None:
        super().__init__(file)
        self.file_reader = file_reader

    @override
    def find_class(self, module: str, name: str) -> Any:
        if module == "torch._utils" and name == "_rebuild_tensor_v2":
            return partial(_NotYetLoadedTensor.rebuild_tensor_v2, archiveinfo=self)
        if module == "torch._tensor" and name == "_rebuild_from_type_v2":
            return partial(_NotYetLoadedTensor.rebuild_from_type_v2, archiveinfo=self)
        if module == "torch._utils" and name == "_rebuild_parameter":

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Materialize the tensor first via `.to(torch.float32)`, `.cuda`, `.contiguous()`, or `.data`, which are handled by the wrapper
2. Load the checkpoint fully (non-lazy `torch.load`) if you need arbitrary tensor operations
3. Restructure code to call `.to(...)`/`.data` before any other tensor method

Example fix

// before
w = state["model"]["proj.weight"]
x = w.float()  # AttributeError on _UnloadedTensor

// after
w = state["model"]["proj.weight"]
w = w.to(torch.float32)  # materializes
x = w.float()

def materialize(t):
    return t.to(t.metatensor.dtype) if hasattr(t, "metatensor") else t

def is_unloaded_tensor(t) -> bool:
    return hasattr(t, "metatensor") and hasattr(t, "_load_tensor")

When it happens

Trigger: Calling arbitrary tensor methods/attributes (e.g. `.shape` is fine via metatensor, but e.g. `.float()`, `.view(...)`, `.t()`, `.numpy()`) directly on tensors obtained from a lazily-loaded checkpoint (`_lazy_load`, quantized/lazy checkpoint workflows like lit-gpt).

Common situations: Inspecting or manipulating weights right after lazy-loading a large/quantized checkpoint without materializing it; code written for regular tensors reused against lazily loaded state dicts.

Related errors

• Path {str(filename)!r} does not exist or is not a file.
• Received multiple values for {', '.join(duplicated_plugin_ke
• Received both `precision={precision_input}` and `plugins={se
• accelerator set through both strategy class and accelerator
• precision set through both strategy class and plugins, choos

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/d6be95307d5d134d. Report an issue: GitHub.