Lightning-AI/pytorch-lightning · error · NotImplementedError

The `{type(self).__name__}` does not support setting a `Chec

Error message

The `{type(self).__name__}` does not support setting a `CheckpointIO` plugin.

What it means

FSDPStrategy does not accept a custom CheckpointIO plugin; assigning to the checkpoint_io setter raises NotImplementedError because FSDP checkpointing is handled natively and cannot be swapped out.

Source

Thrown at src/lightning/fabric/strategies/fsdp.py:196

            self._fsdp_kwargs["device_mesh"] = device_mesh

        self._activation_checkpointing_kwargs = _activation_checkpointing_kwargs(
            activation_checkpointing, activation_checkpointing_policy
        )
        self._state_dict_type = state_dict_type
        self.sharding_strategy = _init_sharding_strategy(sharding_strategy, self._fsdp_kwargs)
        self.cpu_offload = _init_cpu_offload(cpu_offload)
        self.mixed_precision = mixed_precision

    @property
    @override
    def checkpoint_io(self) -> CheckpointIO:
        raise NotImplementedError(f"The `{type(self).__name__}` does not use the `CheckpointIO` plugin interface.")

    @checkpoint_io.setter
    @override
    def checkpoint_io(self, io: CheckpointIO) -> None:
        raise NotImplementedError(f"The `{type(self).__name__}` does not support setting a `CheckpointIO` plugin.")

    @property
    @override
    def root_device(self) -> torch.device:
        assert self.parallel_devices is not None
        return self.parallel_devices[self.local_rank]

    @property
    def num_nodes(self) -> int:
        return self._num_nodes

    @num_nodes.setter
    def num_nodes(self, num_nodes: int) -> None:
        self._num_nodes = num_nodes

    @property
    def num_processes(self) -> int:
        return len(self.parallel_devices) if self.parallel_devices is not None else 0

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Remove the checkpoint_io assignment for FSDP; use strategy.save_checkpoint/load_checkpoint
2. Implement remote/async saving outside the strategy (e.g. after save_checkpoint, copy files to storage)
3. Branch configuration: only set checkpoint_io for strategies that support it (not FSDP)

Example fix

# before
strategy.checkpoint_io = AsyncCheckpointIO()

# after
# FSDPStrategy handles checkpointing itself:
strategy.save_checkpoint(path, state)

from lightning.fabric.strategies import FSDPStrategy
if not isinstance(strategy, FSDPStrategy) and custom_io is not None:
    strategy.checkpoint_io = custom_io

from lightning.fabric.strategies import FSDPStrategy
def accepts_checkpoint_io(strategy) -> bool:
    return not isinstance(strategy, FSDPStrategy)

When it happens

Trigger: strategy.checkpoint_io = MyCheckpointIO() or passing checkpoint_io=... to a constructor/utility that assigns it, when strategy is FSDPStrategy.

Common situations: Reusing configuration meant for the parallel/ddp strategy (e.g. AsyncCheckpointIO, S3CheckpointIO) on FSDP; shared training harness that injects checkpoint IO for all strategies.

Related errors

• The `{type(self).__name__}` does not use the `CheckpointIO`
• Loading a single optimizer object from a checkpoint is not s
• Loading a single module or optimizer object from a checkpoin
• XLAFSDP only supports a single model instance with 'model' a
• `FSDPStrategy.save_checkpoint(..., storage_options=...)` is

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/5c80f4b8ce160ff7. Report an issue: GitHub.