v1.0 merge

[bghira]

Here's a summary of the major changes in version 1.0. These are huge changes, so it would be nice if others can retrieve the main branch and run some old configurations and ensure things still look as you'd expect.

Refactoring and Enhancements:

1. Refactor train.py into a Trainer Class:

   • The core logic of train.py has been restructured into a Trainer class, improving modularity and maintainability.
   • Exposes an SDK for reuse elsewhere.

2. Model Family Unification:

   • References to specific model types (--sd3, --flux, etc.) have been replaced with a unified --model_family argument, streamlining model specification and reducing clutter in configurations.

3. Configuration System Overhaul:

   • Switched from .env configuration files to JSON (config.json), with multiple backends supporting JSON configuration loading. This allows more flexible and readable configuration management.
   • Updated the configuration loader to auto-detect the best backend when launching.

4. Enhanced Argument Handling:

   • Deprecated old argument references and moved argument parsing to helpers/configuration/cmd_args.py for better organization.
   • Introduced support for new arguments such as --model_card_safe_for_work, --flux_schedule_shift, and --disable_bucket_pruning.

5. Improved Hugging Face Integration:

   • Modified configure.py to avoid asking for Hugging Face model name details unless required.
   • Added the ability to pass the SFW (safe-for-work) argument into the training script.

6. Optimizations and Bug Fixes:

   • Fixed several references to learning rate (lr) initialization and corrected --optimizer usage.
   • Addressed issues with attention masks swapping and fixed the persistence of text encoders in RAM after refactoring.

7. Training and Validation Enhancements:

   • Added better dataset examples with support for multiple resolutions and mixed configurations.
   • Configured training scripts to disable gradient accumulation steps by default and provided better control over training options via the updated config files.

8. Enhanced Logging and Monitoring:

   • Improved the handling of Weights & Biases (wandb) logs and updated tracker argument references.

9. Documentation Updates:

   • Revised documentation to reflect changes in model family handling, argument updates, and configuration management.
   • Added guidance on setting up the new configuration files and examples for multi-resolution datasets.

10. Miscellaneous Improvements:

    • Enabled support for NSFW tags in model cards enabled by default.
    • Updated train.sh to minimal requirements, reducing complexity and streamlining the training process.