How can I shut down automatically distributing data when using StreamingDataset?

[ygtxr1997]

🚀 Feature

Giving the option to shut down the automatic data distributed sampler when using StreamingDataset.

Motivation

When I use StreamingDataset in 'DDP' environment, the dataset_len of StreamingDataset seems always be original_len/world_size.
But I want the different processes (with different local_ranks) to share the totally same StreamingDataset, without any data splitting.

Pitch

How to stop the automatic data distribution when using StreamingDataset in DDP? Or could you provide a setting to this? Or could you explain why we can't shut down the distribution?

Alternatives

Additional context

[github-actions]

Hi! thanks for your contribution!, great first issue!

[tchaton]

Hey @ygtxr1997. You can override the distributed env on the dataset. It is inferred automatically from torch.

What is your use case ?

[ygtxr1997]

Hey @ygtxr1997. You can override the distributed env on the dataset. It is inferred automatically from torch.

What is your use case ?

I think overriding the litdata class should be a method to solve my above issue.

Initially, I want to use litdata to optimize my dataset which consists of ~500k small files (each is about 200KB file size). All these files are stored in a remote storage server. However, which is a bit different from other image datasets, my dataloader needs to read from two distinct files, and the indices between the two file vary in [20,50], like a sliding window. For instance, the file indices in a data batch (batch_size=4) could be like this:

batch: [100,120], [1000,1030], [1020,1070], [500,550]

According to your example of usage and distributed data loading illustration GIF, litdata seems not good at dealing with such random-read-like case, am I right? Maybe the performance depends on how the original files are merged into a litdata chunk. Maybe keeping the order of original files (from small index to large index) could result in a faster loading speed, but this potentially impact the learning of deep models?

Therefore, I don't know if litdata could help me and boost the data loading speed in my case.