Potential Issue on data loader in distributed setting.

[dingyuan-shi]

Hello,
It seems that the dataloader is not adapted to distributed setting (Line 881 at train.py).
The data entries will be repeatedly loaded and trained by different processes.
Maybe a sampler should be added, code as below:

train_dataloader = torch.utils.data.DataLoader(
        train_dataset,
        shuffle=(args.split=='train'),
        collate_fn=collate_fn,
        batch_size=args.train_batch_size,
        num_workers=args.dataloader_num_workers,
        drop_last=True, 
        sampler=torch.utils.data.distributed.DistributedSampler(train_dataset), 
    )

[visionxyz]

I tried running it, but it seems that I need to remove shuffle=(args.split=='train') because it conflicts with sampler.

[avinabsaha]

@dingyuan-shi @visionxyz , could you share a sample configuration file to launch training across multiple nodes using "Accelerate Launch"?

My training is stuck after data loads into 1 GPU per node.

[visionxyz]

@dingyuan-shi @visionxyz , could you share a sample configuration file to launch training across multiple nodes using "Accelerate Launch"?

My training is stuck after data loads into 1 GPU per node.

Hello, I just use 'accelerate launch --multi_gpu train.py', and make a config with 'accelerate config' first.

[avinabsaha]

Hi, @visionxyz @dingyuan-shi have you folks fine-tuned the SDXL model? I completed one iteration of SDXL fine-tuning and all the images coming out the model are very over-saturated To verify, I also ran the models provided by the authors, and it is working correctly with no oversaturation problems