Add `StatefulDataLoader` support

[yzhangcs]

Add `StatefulDataLoader` support

[amyeroberts]

cc @muellerzr @lhoestq

[byi8220]

Hm, maybe I misunderstand the problem. My understanding is that what we are focused on is that when the Trainer is loading from a checkpoint, it calls skip_first_batches to skip past the beginning of the dataset until the DataLoader iterator is pointing to where it was at that checkpoint.

And for an IterableDataset, the way this is done under the hood is that it has to manually loop over the items to iterate. And StatefulDataLoader may solve this problem by allowing one to call load_state_dict somewhere in the Trainer while loading the checkpoint, and writing the StatefulDataLoader's state dict to the checkpoint.

This process involves returning self-defined classes like DataLoaderShard to handle cases involving distributed data dispatch.

Yes, it seems like DataLoaderShard and DataLoaderDispatcher are created in the prepare_data_loader function and skip_first_batches function in the accelerate library. These classes are both subclasses of DataLoader, so likely need to be modified or copied to extend from StatefulDataLoader

So IIUC, it seems maybe the implementation of this feature would involve the following steps?

1. In the accelerate library, add either refactor DataLoaderShard and DataLoaderDispatcher to compose or add new variants that inherit from a StatefulDataLoader.
2. In the Trainer class, allow dropping in StatefulDataLoader instead of a regular DataLoader
3. Also in the Trainer class, support loading and saving the state_dict to and from the checkpoint

Thanks for point this out. I still might not be understanding correctly. Maybe it's a lot more complicated than this.

[byi8220]

Hey, just giving my 2 cents since unless I'm missing something, this seems extremely simple to implement since StatefulDataLoader is a drop-in replacement for DataLoader. (I.e. literally just replace DataLoader construction w/ StatefulDataLoader construction in trainer.py)

If it's simple enough I could probably take a shot at implementing it?

The only caveat is it seems torchdata.stateful_dataloader is a beta feature only available in the nightly build. Does it make sense to officially support unreleased features?

[yzhangcs]

@byi8220 Hi, as I can see, the hf Trainer uses the accelerate library internally to prepare the dataloader. This process involves returning self-defined classes like DataLoaderShard to handle cases involving distributed data dispatch. I think it might be challenging to directly combine the Trainer with StatefulDataLoader without delving into the intricate details of the Trainer implementation.

[muellerzr]

Correct, we need to:

1. Support the StatefulDataLoader in accelerate and use it as an optional alternative in the DataLoaderConfiguration
2. Then we can move it to the Trainer!

[byi8220]

Makes sense. It also seems like there's a related issue raised in accelerate: huggingface/accelerate#2859

Regarding using it in the trainer, it feels a bit awkward. IIUC, the desired behavior is that if a StatefulDataLoader is being used, and loading from a checkpoint, then it should not call skip_first_batches at all, unless you are passing in the state dict and checkpoints to that function as well. But imo it feels like skip_first_batches and "restore from checkpoint" are two separate concepts.

[yzhangcs]

Thank you for your responses @byi8220 @muellerzr.

Yes, I agree with you that if we properly manage the states of dataloaders in the Trainer, we no longer need to use the accelerate skip_first_batches option for recovery.

As a workaround, I bypass accelerate to prepare my dataloaders by hacking the Trainer class to support stateful ones:

class Trainer(transformers.Trainer):

    def get_train_dataloader(self) -> DPAwareDataLoader:
        if self.train_dataset is None:
            raise ValueError("Trainer: training requires a train_dataset.")

        logger.info(f"Split the dataset for the node at rank {self.args.process_index} / {self.args.world_size}.")
        train_dataset = HuggingFaceDataset(self.train_dataset,
                                           self.tokenizer,
                                           self.args.context_length,
                                           self.args.process_index,
                                           self.args.world_size)
        loader = DPAwareDataLoader(rank=self.args.process_index,
                                   dataset=train_dataset,
                                   batch_size=self.args.train_batch_size,
                                   collate_fn=self.data_collator,
                                   num_workers=self.args.dataloader_num_workers,
                                   pin_memory=self.args.dataloader_pin_memory,
                                   persistent_workers=self.args.dataloader_persistent_workers)
        data_callback = DataCallback(loader)
        self.add_callback(data_callback)

        return loader

The DPAwareDataLoader is borrowed from torchtitan's impls. This pkg is also developing similar ideas. Then making use of self-defined callbacks to save/load states

class DataCallback(TrainerCallback, ExportableState):

    def __init__(self, loader):

        self.loader = loader

    def on_train_begin(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs):
        output_dir = None
        if isinstance(args.resume_from_checkpoint, bool):
            if args.resume_from_checkpoint:
                output_dir = get_last_checkpoint(args.output_dir)
        elif args.resume_from_checkpoint is not None:
            output_dir = args.resume_from_checkpoint
        if output_dir is not None:
            if args.world_size <= 1:
                data_state_pth = os.path.join(output_dir, "data_state.json")
            else:
                data_state_pth = os.path.join(output_dir, f"data_state_{args.process_index}.json")
            with open(data_state_pth, "r") as f:
                self.loader.load_state_dict(json.loads(f.read()))

    def on_save(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs):
        output_dir = os.path.join(args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}")
        if args.world_size <= 1:
            data_state_pth = os.path.join(output_dir, "data_state.json")
        else:
            data_state_pth = os.path.join(output_dir, f"data_state_{args.process_index}.json")
        with open(data_state_pth, "w") as f:
            f.write(json.dumps(self.state(), indent=2, sort_keys=True) + "\n")

    def state(self) -> dict:
        return self.loader.state_dict()

skip_first_batches is ignored by --ignore_data_skip.
I performed some minimal unit tests, and the states were successfully recovered without perceiving any delays.

This approach can be extremely useful when performing online tokenization with IterableDataset.
Some people have conducted benchmarks and observed even faster speeds than pre-tokenization in https://github.com/XinDongol/on-the-fly-tokenization-profiling.
I've tried using stateful loaders with the above ugly hacking code to train the mamba model on the subsets of 627B Slimpajama data, reducing the total training time from ~173h to ~170h.
This could also save ~3TB of disk space compared to pre-tokenized map-style data.

So I'm really looking forward to your official impls, very happy to hear about any progress :D