Lightning-AI/pytorch-lightning · error · MisconfigurationException

With `def training_step(self, dataloader_iter)`, `self.log(.

Error message

With `def training_step(self, dataloader_iter)`, `self.log(..., batch_size=...)` should be provided.

What it means

When training_step is declared as def training_step(self, dataloader_iter), Lightning cannot infer the batch size from a batch argument. self.log uses batch size for correct metric aggregation/logging, so you must pass it explicitly whenever logging in that mode.

Source

Thrown at src/lightning/pytorch/core/module.py:518

                    raise MisconfigurationException(
                        "Could not find the `LightningModule` attribute for the `torchmetrics.Metric` logged."
                        " You can fix this by setting an attribute for the metric in your `LightningModule`."
                    )
            # try to find the passed metric in the LightningModule
            metric_attribute = self._metric_attributes.get(id(value), None)
            if metric_attribute is None:
                raise MisconfigurationException(
                    "Could not find the `LightningModule` attribute for the `torchmetrics.Metric` logged."
                    f" You can fix this by calling `self.log({name}, ..., metric_attribute=name)` where `name` is one"
                    f" of {list(self._metric_attributes.values())}"
                )

        if (
            trainer.training
            and is_param_in_hook_signature(self.training_step, "dataloader_iter", explicit=True)
            and batch_size is None
        ):
            raise MisconfigurationException(
                "With `def training_step(self, dataloader_iter)`, `self.log(..., batch_size=...)` should be provided."
            )

        if logger and trainer.logger is None:
            rank_zero_warn(
                f"You called `self.log({name!r}, ..., logger=True)` but have no logger configured. You can enable one"
                " by doing `Trainer(logger=ALogger(...))`"
            )
        if logger is None:
            # we could set false here if there's no configured logger, however, we still need to compute the "logged"
            # metrics anyway because that's what the evaluation loops use as return value
            logger = True

        results.log(
            self._current_fx_name,
            name,
            value,
            prog_bar=prog_bar,

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Pass the batch size explicitly: self.log('loss', loss, batch_size=batch_size) where batch_size comes from your extracted batch
2. Extract batch via batch, _ = next(dataloader_iter) and compute its size, then pass it to every self.log call in training_step
3. Alternatively revert to def training_step(self, batch, batch_idx) so the Trainer infers batch_size

Example fix

# before
def training_step(self, dataloader_iter):
    batch, _ = next(dataloader_iter)
    loss = self.step(batch)
    self.log('loss', loss)  # raises

# after
def training_step(self, dataloader_iter):
    batch, _ = next(dataloader_iter)
    loss = self.step(batch)
    self.log('loss', loss, batch_size=len(batch['x']))

if batch_size is None:
    batch_size = infer_batch_size(batch)  # e.g. batch['x'].shape[0] or len(next(iter))[0])
self.log('loss', loss, batch_size=batch_size)

def uses_dataloader_iter(module) -> bool:
    from lightning.pytorch.utilities.model_helpers import is_param_in_hook_signature
    return is_param_in_hook_signature(module.training_step, 'dataloader_iter', explicit=True)

When it happens

Trigger: Defining training_step(self, dataloader_iter) (iterable-style dataloader) and calling self.log('loss', loss) without batch_size=... during training.

Common situations: User switched to dataloader_iter signature to unpack data lazily (e.g. streaming, fused dataloaders) and kept old self.log calls; tutorial code migrated from batch-style signature.

Related errors

• Device should be CPU, got {device} instead.
• The `PredictionWriterCallback` does not support using `datal
• You are trying to `self.log()` but the loop's result collect
• You are trying to `self.log()` but it is not managed by the
• You called `self.log` with the key `{name}` but it should no

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/1ebe823b5389192b. Report an issue: GitHub.