Lightning-AI/pytorch-lightning · error · RuntimeError

`{type(self).__name__}` already owns a group.

Error message

`{type(self).__name__}` already owns a group.

What it means

Raised by `Collective.create_group()` when the collective already owns a process group (`self._group is not None`). Creating a second group over the same collective would leak/duplicate groups, so it is rejected.

Source

Thrown at src/lightning/fabric/plugins/collectives/collective.py:110

    @classmethod
    @abstractmethod
    def _convert_to_native_op(cls, op: str) -> Any: ...

    def setup(self, **kwargs: Any) -> Self:
        if not self.is_initialized():
            self.init_group(**kwargs)
        return self

    def create_group(self, **kwargs: Any) -> Self:
        """Create a group.

        This assumes that :meth:`~lightning.fabric.plugins.collectives.Collective.init_group` has been
        called already by the user.

        """
        if self._group is not None:
            raise RuntimeError(f"`{type(self).__name__}` already owns a group.")
        self._group = self.new_group(**kwargs)
        return self

    def teardown(self) -> Self:
        if self._group is None:
            raise RuntimeError(f"`{type(self).__name__}` does not own a group to destroy.")
        self.destroy_group(self._group)
        self._group = None
        return self

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Call `collective.teardown()` before creating a new group
2. Use a fresh `TorchCollective()` instance for each group you need
3. Check `collective._group is None` (or track creation state) before calling create_group

Example fix

# before
collective.create_group(ranks=ranks)
collective.create_group(ranks=new_ranks)  # RuntimeError

# after
collective.create_group(ranks=ranks)
collective.teardown()
collective.create_group(ranks=new_ranks)

if collective._group is not None:
    collective.teardown()
collective.create_group(ranks=ranks)

When it happens

Trigger: Calling `collective.create_group()` twice without an intervening `teardown()`; loops over strategies that reuse one collective instance; tests that recreate groups (as seen in the many test callers).

Common situations: Retrying initialization logic on failure that already created a group; running multiple sequential distributed phases with the same Collective object; repeated test setups sharing a module-level collective.

Related errors

• `{type(self).__name__}` does not own a group. HINT: try `col
• `{type(self).__name__}` does not own a group to destroy.
• An optimizer should be passed only once to the `setup_optimi
• A dataloader should be passed only once to the `setup_datalo
• Torch distributed is not available.

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/ad4d399ac4eda1d5. Report an issue: GitHub.