[pytorch-finetuning] Introduction of AI Halo Reference Platform

[ldokovic-personal]

Improvements:

• Added AI Halo as a device option, supporting both Windows and Linux operating systems.
• Enhanced the user interface on the website to provide a clear overview when users select categories.
• Fixed typo on Model's directory output to match expected title for the 'saving' step.
• Updated the README to include guidelines for formatting datasets, ensuring compatibility with JSON, CSV, and Hugging Face formats.
• Emphasized the importance of maintaining an instruction-response pair in datasets and provided instructions for modifying the format_instruction() function to accommodate custom keys.

**What has not been fixed by 221bc72 commit and was addressed in Finetune LLMs with Pytorch playbook Report

• The check_os_kernel() method in ~path/to/$venv/lib/python3.13/site-packages/accelerate/utils/other.py is currently under dispute.

Issue: When setting the SFTTrainer step, the check_os_kernel() method checks the kernel version, which causes an error on the HaloBox system with kernel version

6.18+unreleased-amd64

The function expects different formats for kernels on BARE Metal Native Linux distributions.

Modification Needed

Since check_os_kernel() doesn't return anything, it should be adjusted to correctly parse the kernel version and avoid unnecessary exceptions.
Proposed change: Modify the regex to match the format 'X.Y' and ignore the rest, ensuring compatibility with kernel versions greater than 5.5.0.

     # Adjust the regex to match the format 'X.Y' and ignore the rest    
     # this will return 6.18 which is grater than 5 kernel version so, this won't throw exception as it will not return anything 
     match = re.match(r"(\d+\.\d+)", info.release) 
  
     if not match: 
         logger.error(f"Unexpected format for kernel version: {info.release}") 
         return