multi nodes multi gpu is support ？

[YLongJin]

multi nodes multi gpu is support ？

[bghira]

never tried.

[YLongJin]

thanks your reply

[YLongJin]

@bghira can I train use the pixart dataset module？ the json file with image path and caption and prompt

[bghira]

possibly by just configuring the parquet backend to use it

[YLongJin]

the data is local, is this compatible？

[sayakpaul]

Local parquet loading shouldn't be an issue. If you face any, do report it.

[YLongJin]

ValueError: Your dataloader config must contain at least one image dataset AND at least one text_embed dataset. See this link for more information about dataset_type: https://github.com/bghira/SimpleTuner/blob/main/documentation/DATALOADER.md#configuration-options

[YLongJin]

I do not have text_embed dataset， i just have json file with image path and caption and prompt

[sayakpaul]

Oh then you should probably use the datasets library for that:

https://huggingface.co/docs/datasets/en/image_load

[YLongJin]

how can I make the embed dataset

[sayakpaul]

Cc: @bghira

[bghira]

please follow the quickstart and use the demo dataset first just so you know you have it working

[bghira]

the link to the quickstart is here for stable diffusion 3. I do not have a specific one for SDXL, but the basic concepts apply other than the difference for MODEL_NAME and STABLE_DIFFUSION_3 variable properties.

here is what the config looks like:

[
  {
    "id": "pseudo-camera-10k-sd3",
    "type": "local",
    "crop": true,
    "crop_aspect": "square",
    "crop_style": "center",
    "resolution": 0.5,
    "minimum_image_size": 0.25,
    "maximum_image_size": 1.0,
    "target_downsample_size": 1.0,
    "resolution_type": "area",
    "cache_dir_vae": "cache/vae/sd3/pseudo-camera-10k",
    "instance_data_dir": "datasets/pseudo-camera-10k",
    "disabled": false,
    "skip_file_discovery": "",
    "caption_strategy": "filename",
    "metadata_backend": "json"
  },
  {
    "id": "text-embeds",
    "type": "local",
    "dataset_type": "text_embeds",
    "default": true,
    "cache_dir": "cache/text/sd3/pseudo-camera-10k",
    "disabled": false,
    "write_batch_size": 128
  }
]

it's not anything you manually have to create - the trainer will do that for you.

this merely points to the storage location where these embeds can be stored.

it might seem needlessly complicated, and it is - because the trainer can split the storage locations of everything but the VAE cache objects, for efficiency purposes. you can store the image data locally via NVME and text embeds on a S3 storage bucket, for example.

hopefully after following the quickstart you'll have something working for your model, and then you can expand the configuration of accelerate to include multi-node.

because i don't have access to multiple nodes to train on, i have no ability to test or verify anything about configuration or runtime problems. please report any issues you do have, and we can work together on solving them.