Lightning-AI/pytorch-lightning · error · ValueError

`{type(self).__name__}` does not support the `CombinedLoader

Error message

`{type(self).__name__}` does not support the `CombinedLoader(mode="sequential")` mode. The available modes are: {[m for m in _SUPPORTED_MODES if m != 'sequential']}

What it means

Raised in FitLoop.advance when the training CombinedLoader was created with mode='sequential'. Sequential mode concatenates multiple dataloaders one after another, which does not produce aligned batches across the epoch and is therefore unsupported for training; only the min_size/max_size/.. cycle-like modes are allowed.

Source

Thrown at src/lightning/pytorch/loops/fit_loop.py:472

            _set_sampler_epoch(dl, self.epoch_progress.current.processed)

        if not self.restarted_mid_epoch and not self.restarted_on_epoch_end:
            if not self.restarted_on_epoch_start:
                self.epoch_progress.increment_ready()

            call._call_callback_hooks(trainer, "on_train_epoch_start")
            call._call_lightning_module_hook(trainer, "on_train_epoch_start")

            self.epoch_progress.increment_started()

    def advance(self) -> None:
        """Runs one whole epoch."""
        log.debug(f"{type(self).__name__}: advancing loop")

        combined_loader = self._combined_loader
        assert combined_loader is not None
        if combined_loader._mode == "sequential":
            raise ValueError(
                f'`{type(self).__name__}` does not support the `CombinedLoader(mode="sequential")` mode.'
                f" The available modes are: {[m for m in _SUPPORTED_MODES if m != 'sequential']}"
            )
        with self.trainer.profiler.profile("run_training_epoch"):
            assert self._data_fetcher is not None
            self.epoch_loop.run(self._data_fetcher)

    def on_advance_end(self) -> None:
        trainer = self.trainer
        # inform logger the batch loop has finished
        trainer._logger_connector.epoch_end_reached()

        self.epoch_progress.increment_processed()

        # call train epoch end hooks
        # we always call callback hooks first, but here we need to make an exception for the callbacks that
        # monitor a metric, otherwise they wouldn't be able to monitor a key logged in
        # `LightningModule.on_train_epoch_end`

View on GitHub (pinned to 9fed5c27d2)

Solutions

1. Use one of the supported cycling modes: 'min_size', 'max_size', 'max_size_cycle'
2. If sequential consumption is truly needed, manually chain the datasets into a single ConcatDataset/ConcatLoader and pass one dataloader
3. If you intended evaluation over multiple loaders, use trainer.validate/test instead, where sequential is allowed

Example fix

# before
train_loader = CombinedLoader([loader_a, loader_b], mode='sequential')
trainer.fit(model, train_loader)

# after
train_loader = CombinedLoader([loader_a, loader_b], mode='max_size_cycle')
trainer.fit(model, train_loader)

from lightning.pytorch.utilities import CombinedLoader
from lightning.fabric.utilities import _SUPPORTED_MODES

assert combined.mode != 'sequential', 'training does not support sequential CombinedLoader mode'

def mode_ok_for_training(mode: str) -> bool:
    return mode != 'sequential'

When it happens

Trigger: Passing multiple train dataloaders to Trainer.fit and constructing the CombinedLoader (e.g. via LightningDataModule or CombinedLoader directly) with `CombinedLoader(mode='sequential')`, then calling trainer.fit.

Common situations: Multi-dataset training setups (e.g. domain adaptation with source/target loaders) where the developer chose sequential mode thinking it cycles; copying a prediction-loop pattern (which does support sequential) into training.

Related errors

• `trainer.predict()` only supports the `CombinedLoader(mode="
• You called `self.log` with the key `{name}` but it should no
• The loss returned in `training_step` is {loss}.
• You provided multiple `{stage.dataloader_prefix}_dataloader`
• Device should be CPU, got {device} instead.

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/f6b7c575559e6dd5. Report an issue: GitHub.