deepset-ai/haystack · error · ValueError

Only single devices can be converted to PyTorch format

Error message

Only single devices can be converted to PyTorch format

What it means

ComponentDevice.to_torch raises ValueError when the component holds a multi-device map rather than a single device, since a multi-device setup cannot be expressed as one torch.device.

Source

Thrown at haystack/utils/device.py:323

        """
        if not (self._single_device is not None) ^ (self._multiple_devices is not None):
            raise ValueError(
                "The component device can neither be empty nor contain both a single device and a device map"
            )

    def to_torch(self) -> "torch.device":
        """
        Convert the component device representation to PyTorch format.

        Device maps are not supported.

        :returns:
            The PyTorch device representation.
        """
        self._validate()

        if self._single_device is None:
            raise ValueError("Only single devices can be converted to PyTorch format")

        torch_import.check()
        assert self._single_device is not None
        return torch.device(str(self._single_device))

    def to_torch_str(self) -> str:
        """
        Convert the component device representation to PyTorch string format.

        Device maps are not supported.

        :returns:
            The PyTorch device string representation.
        """
        self._validate()

        if self._single_device is None:
            raise ValueError("Only single devices can be converted to PyTorch format")

View on GitHub (pinned to e318778c9b)

Solutions

1. Use to_torch only on single-device components; inspect has_multiple_devices() first
2. For multi-device setups, access ComponentDevice._multiple_devices/DeviceMap and handle per-device entries or rely on the HF/accelerate integration instead of to_torch
3. Reconfigure the component to a single device if a single torch.device is required

Example fix

// before
comp = ComponentDevice.from_multiple(DeviceMap({"0": dev0, "1": dev1}))
torch_dev = comp.to_torch()  # raises
// after
if comp.has_multiple_devices():
    # handle the map explicitly
    devices = comp._multiple_devices.devices
else:
    torch_dev = comp.to_torch()

def torch_device_or_none(comp: ComponentDevice):
    return comp.to_torch() if not comp.has_multiple_devices() else None

def has_single_device(comp: ComponentDevice) -> bool:
    return comp._single_device is not None

try:
    dev = comp.to_torch()
except ValueError:
    # multi-device: handle per-device or delegate to HF/accelerate
    dev = None

When it happens

Trigger: Calling to_torch() on a ComponentDevice built from a DeviceMap (from_multiple/from_hf with multiple devices).

Common situations: Pipeline configured with multi-GPU device maps; component-level utility assumes single device but the model was sharded across devices (e.g. via accelerate).

Related errors

• Unknown device type string '{string}'
• Device id must be >= 0, got {id}
• Only single devices can be converted to spaCy format
• The disk device can only be used as a part of device maps
• The component device can neither be empty nor contain both a

AI-assisted analysis of deepset-ai/haystack@e318778c9b (2026-08-30). Data as JSON: /api/errors/4f2ab2f26fa96e51. Report an issue: GitHub.