Lightning-AI/pytorch-lightning · error · TypeError

You need to set up the model first before you can call `fabr

Error message

You need to set up the model first before you can call `fabric.no_backward_sync()`: `model = fabric.setup(model, ...)`

What it means

fabric.no_backward_sync(model) requires the model to be a wrapped _FabricModule, because the context manager needs access to the strategy's backward-sync control on the wrapped module. Passing a raw nn.Module (never set up) fails the isinstance check and raises TypeError with remediation instructions.

Source

Thrown at src/lightning/fabric/fabric.py:771

            becomes a no-op. For single-device strategies, it is always a no-op.

        Example::

            # Accumulate gradients over 8 batches
            for batch_idx, batch in enumerate(dataloader):
                with fabric.no_backward_sync(model, enabled=(batch_idx % 8 != 0)):
                    output = model(batch)
                    loss = criterion(output, target)
                    fabric.backward(loss)

                if batch_idx % 8 == 0:
                    optimizer.step()
                    optimizer.zero_grad()

        """
        module, _ = _unwrap_compiled(module)
        if not isinstance(module, _FabricModule):
            raise TypeError(
                "You need to set up the model first before you can call `fabric.no_backward_sync()`:"
                " `model = fabric.setup(model, ...)`"
            )
        if isinstance(self._strategy, (SingleDeviceStrategy, XLAStrategy)):
            return nullcontext()
        if self._strategy._backward_sync_control is None:
            rank_zero_warn(
                f"The `{self._strategy.__class__.__name__}` does not support skipping the gradient synchronization."
                f" Remove `.no_backward_sync()` from your code or choose a different strategy.",
                category=PossibleUserWarning,
            )
            return nullcontext()

        forward_module, _ = _unwrap_compiled(module._forward_module)
        return self._strategy._backward_sync_control.no_backward_sync(forward_module, enabled)

    def sharded_model(self) -> AbstractContextManager:
        r"""Instantiate a model under this context manager to prepare it for model-parallel sharding.

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Use the return value of setup: model = fabric.setup(model), then fabric.no_backward_sync(model)
2. Note: on single-device or XLA strategies this context is a nullcontext anyway — the call is only meaningful for DDP/FSDP

Example fix

# before
model = MyModel()
fabric.setup(model)
with fabric.no_backward_sync(model):  # raw module
    ...
# after
model = MyModel()
model = fabric.setup(model)
with fabric.no_backward_sync(model):  # wrapped _FabricModule
    ...

from lightning.fabric.wrappers import _FabricModule
assert isinstance(model, _FabricModule), 'setup the model first'

from lightning.fabric.wrappers import _FabricModule
def is_setup(model):
    return isinstance(model, _FabricModule)

When it happens

Trigger: Calling fabric.no_backward_sync(model) with the original unwrapped module instead of the object returned by fabric.setup(model)/setup_module(model). Also happens after _FabricModule.unwrap() when re-wrapping was forgotten.

Common situations: Using gradient accumulation with no_sync optimization; keeping references to the pre-setup model around and passing the stale reference; unwrapping for checkpointing then continuing training.

Related errors

• No models were set up for backward. Did you forget to call `
• To use Fabric with more than one device, you must call `.lau
• Received multiple values for {', '.join(duplicated_plugin_ke
• Received both `precision={precision_input}` and `plugins={se
• accelerator set through both strategy class and accelerator

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/d5cd731cc278f8ed. Report an issue: GitHub.