Manual `Dataloader` reloading during training

[FilippoOlivo]

Description & Motivation

Currently, PyTorch Lightning assumes that Dataset and DataLoader parameters (e.g., sequence length, unrolling steps $N→N+m$, or sampling ratios) remain static throughout the lifecycle of a fit() call (unless you set a static policy using reload_dataloader_every_n_epochs in Trainer). When a user needs to adjust these parameters dynamically based on real-time metrics (such as curriculum learning or adaptive unrolling), they are forced into two suboptimal patterns:

• Manual Script Restarts: Stopping execution, saving a checkpoint, updating the configuration, and resuming. This introduces significant overhead and breaks the continuity of the training loop.
• Dataset Over-engineering: Building complex, stateful Dataset objects that internally track metrics and handle logic.

Pitch

I propose introducing a formal mechanism to trigger in-place DataLoader reconfiguration without exiting the fit() loop. This would likely manifest as a new hook or an enhancement to the existing reload_dataloaders_every_n_epochs functionality, allowing for:

• Metric-driven triggers: Re-initializing the DataLoader when a specific monitor value hits a threshold.
• Parameter Injection: A way to pass updated hyper-parameters from the LightningModule or a Callback directly into the datamodule during the reload process.

Alternatives

No response

Additional context

Here a snapshot of how a callback could be implemented if the requested feature will be accepted

def on_train_batch_end(self, outputs, batch, batch_idx):
    if self.trainer.callback_metrics['train_loss'] < 0.5:  # this is just an example (more complex logic will be fine)
        # Trigger a reload with updated sequence length
        self.trainer.datamodule.unroll_steps += 1
        self.trainer.manual_dataloader_reload()

I already have a draft code for this improvement and it will be a pleasure for me to open a PR is the feature will be of your interest

cc @lantiga