add cpu support

[swap357]

This pull request introduces the --use_cpu option to the main script, allowing users to run training and inference on the CPU. The --use_cpu option can be specified when running the script to enable CPU usage. By default, the script will use the GPU (if available).

Changes:

• Added an argument parser for the --use_cpu option to specify CPU usage for training and inference.
• Updated the load_base_model function to adjust load_in_8bit, torch_dtype, and device_map based on the use_cpu argument.
• Modified the TrainingArguments to set fp16 based on the use_cpu argument.
• Removed the .cuda() call from the generate_text function when using the CPU.

This update provides flexibility for users who want to run the script on systems without a GPU. The option to use the CPU can be helpful for testing, debugging, and using the script on machines with limited GPU resources.

[lxe]

Since I refactored the whole thing, you might need to adjust it a bit. Sorry

[Alex-Pavlyuk]

The most expected pull request for now) to be able to work with llm on laptops

[swap357]

I went through the latest code and seems like the updated code already takes care of running on cpu if cuda is unavailable. I believe the trainer should work out of the box on cpu now.

here: config.py -
HAS_CUDA = torch.cuda.is_available()
DEVICE = torch.device('cuda' if HAS_CUDA else 'cpu')

Let me know if otherwise.