Lightning-AI/pytorch-lightning · error · TypeError

The XLA strategy can only work with the `XLAPrecision` plugi

Error message

The XLA strategy can only work with the `XLAPrecision` plugin, found {precision_plugin}

What it means

The XLA strategies require the XLAPrecision precision plugin (it handles TPU-specific mixed precision via torch_xla). The precision_plugin setter rejects any other Precision implementation with TypeError.

Source

Thrown at src/lightning/pytorch/strategies/single_xla.py:88

    def checkpoint_io(self, io: Optional[CheckpointIO]) -> None:
        if io is not None and not isinstance(io, (XLACheckpointIO, _WrappingCheckpointIO)):
            raise TypeError(f"The XLA strategy can only work with the `XLACheckpointIO` plugin, found {io}")
        self._checkpoint_io = io

    @property
    @override
    def precision_plugin(self) -> XLAPrecision:
        plugin = self._precision_plugin
        if plugin is not None:
            assert isinstance(plugin, XLAPrecision)
            return plugin
        return XLAPrecision()

    @precision_plugin.setter
    @override
    def precision_plugin(self, precision_plugin: Optional[Precision]) -> None:
        if precision_plugin is not None and not isinstance(precision_plugin, XLAPrecision):
            raise TypeError(f"The XLA strategy can only work with the `XLAPrecision` plugin, found {precision_plugin}")
        self._precision_plugin = precision_plugin

    @override
    def setup(self, trainer: "pl.Trainer") -> None:
        if self.debug:
            os.environ["PT_XLA_DEBUG"] = str(1)

        assert self.accelerator is not None
        self.accelerator.setup(trainer)

        assert self.model is not None
        self.precision_plugin.convert_module(self.model)

        shared_params = find_shared_parameters(self.model)
        self.model_to_device()
        set_shared_parameters(self.model, shared_params)

        self.model = self._setup_model(self.model)

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Use XLAPrecision (default) — configure precision through Trainer(precision=...) instead of a custom plugin
2. If customizing, subclass XLAPrecision
3. Remove precision_plugin from the strategy kwargs

Example fix

# before
from lightning.pytorch.plugins.precision import MixedPrecision
strategy = SingleDeviceXLAStrategy(precision_plugin=MixedPrecision())

# after
strategy = SingleDeviceXLAStrategy()
trainer = L.Trainer(strategy=strategy, precision="16-mixed")

assert precision_plugin is None or isinstance(precision_plugin, XLAPrecision)

def is_valid_xla_precision(p) -> bool:
    from lightning.pytorch.plugins.precision.xla import XLAPrecision
    return p is None or isinstance(p, XLAPrecision)

When it happens

Trigger: Assigning strategy.precision_plugin = MixedPrecision() or Precision(), or constructing SingleDeviceXLAStrategy/XLAStrategy with precision_plugin=<non-XLAPrecision plugin>; often from reusing a config built for CUDA strategies.

Common situations: Setting 16-bit precision via a generic plugin object; sharing strategy kwargs across GPU and TPU runs.

Related errors

• The XLA strategy can only work with the `XLAPrecision` plugi
• The XLA strategy can only work with the `XLACheckpointIO` pl
• The XLA strategy can only work with the `XLACheckpointIO` pl
• To spawn processes with the `{type(self.strategy).__name__}`
• The `{type(self._strategy).__name__}` requires the model and

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/a121ccd6a74c6cb2. Report an issue: GitHub.