Lightning-AI/pytorch-lightning · error · RuntimeError

Cannot set the dtype explicitly. Please use module.to(new_dt

Error message

Cannot set the dtype explicitly. Please use module.to(new_dtype).

What it means

The `_DeviceDtypeModuleMixin` (used by Lightning modules to track device/dtype) defines a `dtype` property with a setter that deliberately raises, because dtype must be changed through `module.to(new_dtype)` so the change propagates to parameters and buffers. Assigning `obj.dtype = torch.float16` triggers this RuntimeError (the setter exists mainly to block the infinite-recursion path that a bare property would cause).

Source

Thrown at src/lightning/fabric/utilities/device_dtype_mixin.py:42

class _DeviceDtypeModuleMixin(Module):
    __jit_unused_properties__: list[str] = ["device", "dtype"]

    def __init__(self) -> None:
        super().__init__()
        self._dtype: Union[str, torch.dtype] = torch.get_default_dtype()
        # Workarounds from the original pytorch issue:
        # https://github.com/pytorch/pytorch/issues/115333#issuecomment-1848449687
        # `get_default_device` only honors the `torch.device` context manager from 2.8 onwards
        self._device = torch.get_default_device() if _TORCH_GREATER_EQUAL_2_8 else torch.empty(0).device

    @property
    def dtype(self) -> Union[str, torch.dtype]:
        return self._dtype

    @dtype.setter
    def dtype(self, new_dtype: Union[str, torch.dtype]) -> None:
        # necessary to avoid infinite recursion
        raise RuntimeError("Cannot set the dtype explicitly. Please use module.to(new_dtype).")

    @property
    def device(self) -> torch.device:
        device = self._device

        # make this more explicit to always include the index
        if device.type == "cuda" and device.index is None:
            return torch.device(f"cuda:{torch.cuda.current_device()}")

        return device

    @override
    def to(self, *args: Any, **kwargs: Any) -> Self:
        """See :meth:`torch.nn.Module.to`."""
        # this converts `str` device to `torch.device`
        device, dtype = torch._C._nn._parse_to(*args, **kwargs)[:2]
        _update_properties(self, device=device, dtype=dtype)
        return super().to(*args, **kwargs)

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Replace assignment with a `.to()` call: `module.to(torch.float16)`.
2. If storing a target dtype for later, keep it in a plain attribute with a different name (e.g. `self._target_dtype`).
3. Audit code for any `x.dtype = ...` on Lightning modules.

Example fix

# before
model.dtype = torch.float16

# after
model.to(torch.float16)

# there is nothing to pre-validate; simply never assign .dtype
model.to(torch.float16)  # correct way

try:
    model.dtype = torch.float16
except RuntimeError:
    model.to(torch.float16)

When it happens

Trigger: Executing `model.dtype = torch.float16` (or any dtype assignment) on a LightningModule or Fabric-managed module that mixes in `_DeviceDtypeModuleMixin`, instead of calling `.to()`.

Common situations: Porting plain PyTorch code that assigned `.dtype` on wrapper/inner modules; helper functions trying to record intended dtype on a module; stale tutorials using attribute assignment.

Related errors

• {mode!r} only works with `dtype=torch.float16`, but you chos
• Device should be CPU, got {device} instead.
• `devices` selected with `CPUAccelerator` should be an int >
• Device should be CUDA, got {device} instead.
• You requested to find {num_devices} devices but there are no

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/6baccea3f1186d26. Report an issue: GitHub.