ArcForge: Fine-Tuning on Intel Battlemage GPUs

ArcForge is an open-source toolkit for fine-tuning large language models (LLMs) using Intel Arc™ Battlemage and other next-generation Intel GPUs. It integrates cutting-edge tools like HuggingFace Transformers, Intel IPEX-LLM, and QLoRA/LoRA methods to deliver scalable, efficient training pipelines optimized for the Intel XPU architecture.

🚀 Project Goals

• Enable high-performance fine-tuning of open-weight LLMs on Intel GPUs
• Provide end-to-end workflows for QLoRA and LoRA with checkpointing, evaluation, and inference
• Support for modern instruction-tuning datasets (Alpaca, OpenOrca, etc.)
• Serve as a foundational framework for building ML infra on Intel's Battlemage GPUs

🧩 Key Features

• ✅ LoRA and QLoRA fine-tuning on LLaMA 2/3 models
• ✅ HuggingFace Transformers integration
• ✅ Intel IPEX-LLM acceleration for Arc GPUs
• ✅ Alpaca-compatible training datasets
• ✅ Efficient checkpointing, evaluation, and inference

📦 Installation

1. Clone the repository:

git clone https://github.com/arcforge-tune/bmg-lora.git

2. Install dependencies using conda and pip:

conda create -n fine-tune python=3.11.13
conda activate fine-tune
cd bmg-lora
pip install -r requirements.txt

🛠️ Usage

To run training with LLaMA-2 on the Alpaca dataset examples:

1. Clone the repository and install dependencies.

2. Configure your training parameters in src/config/

File Name	Model ID	Model Name
gpt2_lora_finetune_config.yaml	gpt2	GPT-2
llama2_hf_qlora_xpu_config.yaml	meta-llama/Llama-2-7b-hf	Llama 2 (7B)
llamma2_chat_hf_qlora_xpu_config.yaml	meta-llama/Llama-2-7b-chat-hf	Llama 2 (7B Chat)
mistral-7B-v0.1_xpu_config.yaml	mistralai/Mistral-7B-v0.1	Mistral (7B)
llama3.18B_qlora_config.yaml	meta-llama/Llama-3.1-8B-Instruct	llama 3.1
deepseek_r1_distill_llama_8b_config.yaml	deepseek-ai/DeepSeek-R1-Distill-Llama-8B	DeepSeek R1 Distill Llama 8B

3. Run the training for your model using the provided configuration:

Start training:

python src/main.py --config .\src\config\gpt2_lora_finetune_config.yaml    

Resume training from a checkpoint:

python src/main.py --config .\src\config\gpt2_lora_finetune_config.yaml --resume .\outputs\lora_llama3_1_8b_instruct_xpu\checkpoint-epoch1-step11\

For a clean run or resume output use the PS script:

.\Run‑FineTune.ps1 -config .\src\config\gpt2_lora_finetune_config.yaml  

Requirements

• Python 3.6+
• PyTorch
• HuggingFace Transformers
• Intel IPEX
• ONEDNN library

For GPU support, ensure you have the correct Intel GPU drivers installed