Add optimizations to the training loop.

To enable efficient training on GPUs and scale our repository for models with millions to billions of parameters—essential for working with large visual language models—we must implement optimization techniques. These include qLora, precision changes, and Flash Attention for Transformers.

To facilitate these enhancements, we can leverage a suite of Hugging Face libraries designed to streamline the implementation of these techniques. The recommended libraries are:

Accelerate: Enables more efficient parallel execution and training.
PEFT: Enables QLora and Adapters techniques.
bitsandbytes: Offers optimized implementations for managing model parameters and memory.

These optimizations are not currently integrated into our repository. However, incorporating them would be highly beneficial, especially for fine-tuning models on large datasets.

For a practical example of how these features could be implemented, consider the following resource: Dreambooth Fine-tuning Colab. This example provides a solid foundation for developing similar capabilities within our projects.

Cruiz102 / Advesarial_Attacks_Tests

Add optimizations to the training loop. #1