Thank you for this PR. This seems to be very simple.

However, I wonder how this works if I don't use wandb? Also some arguments (wandb and hagging face id, absolute path, etc.) would be undesirable to expose for security reasons. A filter for them would be necessary.

I also think we need an option to enable this feature as it would be a breaking change.

This wouldn't affect normal workflow if not using wandb, that is because the extra configs are passed within the ['wandb'] key of the configs. This also is not a breaking change for wandb users since the extra arguments can just be ignored on the wandb side.
However i agree that some filtering is necessary, can i revise this PR to be in line with #1123 ?

This change sends all the training parameters to wandb.ai when that is used instead of TensorBoard (by default ie. if log_tracker_config is not specified).

Now they show up in the Runs table, so it helps to get a better understanding of what the difference is between runs. It also uses the model's name as run name to give more context.

This is a re-implementation of #792.

HI
Seems like SDXL_Train.py has changed. Could you please advise how to change this code so it works again?

This is what I am seeing in Kohya_ss >> sdxl_train.py

if accelerator.is_main_process:
        init_kwargs = {}
        if args.wandb_run_name:
            init_kwargs["wandb"] = {"name": args.wandb_run_name, "config": args}
        if args.log_tracker_config is not None:
            init_kwargs = toml.load(args.log_tracker_config)
        accelerator.init_trackers("finetuning" if args.log_tracker_name is None else args.log_tracker_name, init_kwargs=init_kwargs)

Unfortunately, that code "config": args isn't sending args to wandb

thank you

@rafstahelin check out #1285