OmniTune: Unified Efficient Fine-Tuning of 100+ LLMs & VLMs 🚀✨

---

✨ Key Features

• 🤖 Multi-Model Support: Fine-tune 100+ architectures including LLaMA 3/3.1/3.2, DeepSeek-V3/R1, Qwen 2.5/Qwen2-VL, Mistral/Mixtral, Gemma 2, Phi-3/4, InternLM, and multimodal variants.
• 🎯 Efficient Fine-Tuning Algorithms: Full parameter tuning, Freeze tuning, LoRA, QLoRA, GaLore, BAdam, DoRA, LongLoRA, and PiSSA.
• 🔄 Diverse Training Paradigms:
  • Supervised Fine-Tuning (SFT) 📝
  • Direct Preference Optimization (DPO) ⚖️
  • Reward Modeling (RM) 🏆
  • Proximal Policy Optimization (PPO) 📈
  • Odds Ratio Preference Optimization (ORPO) 🔀
  • Kahneman-Tversky Optimization (KTO) 🧠
• 🚀 High-Performance Acceleration: DeepSpeed (ZeRO-2, ZeRO-3, Offload), PyTorch FSDP2, FSDPTurbo, FlashAttention-2, Liger Kernel, and Ulysses Sequence Parallelism.
• 🎨 Zero-Code Web Interface: Launch OmniTune-Board powered by Gradio to configure hyperparameters, monitor loss curves, and evaluate checkpoints interactively.
• ⚡ Production Inference & Serving: Export merged HuggingFace weights, quantize models (AWQ, GPTQ), or serve high-throughput APIs powered by vLLM.

---

📦 Installation

📋 Prerequisites

• Python >= 3.11 🐍
• PyTorch >= 2.4.0 🔥
• CUDA >= 12.0 (for NVIDIA GPUs) or ROCm / NPU CANN 💻

🛠️ Install from Source

git clone https://github.com/ishandutta2007/OmniTune.git
cd OmniTune
pip install -e ".[torch,metrics]"

⚡ Optional Optimizers & Kernels

# FlashAttention-2
pip install flash-attn --no-build-isolation

# DeepSpeed acceleration
pip install deepspeed

# High-throughput vLLM engine
pip install vllm

---

🖥️ OmniTune-Board Web UI

Launch the zero-code browser dashboard: 🚀

omnitune webui

Or use the short command:

ot webui

Access the interface in your browser at http://localhost:7860. From the web dashboard, you can:

• 📑 Select models, templates, and datasets.
• ⚙️ Configure LoRA ranks, learning rate schedules, and gradient accumulation.
• 📊 Visualize real-time training loss and evaluation metrics.
• 💬 Chat with fine-tuned checkpoints directly inside the playground.
• 💾 Merge adapter weights and export standalone checkpoints.

---

⚡ Command-Line Interface (CLI)

OmniTune provides both single-line execution commands and clean YAML config-based execution. 💻

1. 📝 Supervised Fine-Tuning (SFT) with LoRA

Run from a configuration file:

omnitune train examples/train_lora/llama3_lora_sft.yaml

Or via direct command-line arguments:

omnitune train \
    --stage sft \
    --do_train \
    --model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct \
    --dataset alpaca_en_demo \
    --dataset_dir data \
    --template llama3 \
    --finetuning_type lora \
    --lora_target all \
    --output_dir saves/llama3-8b/lora/sft \
    --overwrite_output_dir \
    --cutoff_len 2048 \
    --learning_rate 1e-4 \
    --num_train_epochs 3.0 \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 8 \
    --lr_scheduler_type cosine \
    --logging_steps 10 \
    --save_steps 100 \
    --warmup_ratio 0.1 \
    --fp16

2. ⚖️ Direct Preference Optimization (DPO)

omnitune train \
    --stage dpo \
    --do_train \
    --model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct \
    --adapter_name_or_path saves/llama3-8b/lora/sft \
    --dataset dpo_en_demo \
    --dataset_dir data \
    --template llama3 \
    --finetuning_type lora \
    --output_dir saves/llama3-8b/lora/dpo \
    --cutoff_len 2048 \
    --learning_rate 5e-6 \
    --num_train_epochs 2.0 \
    --dpo_beta 0.1 \
    --fp16

3. 💬 Interactive Chat / Testing

omnitune chat \
    --model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct \
    --adapter_name_or_path saves/llama3-8b/lora/sft \
    --template llama3

4. 🌐 OpenAI-Compatible API Server

Deploy a production REST API compatible with the OpenAI specification:

omnitune api \
    --model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct \
    --adapter_name_or_path saves/llama3-8b/lora/sft \
    --template llama3 \
    --api_port 8000

5. 💾 Export & Merge LoRA Weights

Export standalone model weights for fast inference engines (vLLM, Ollama, TensorRT-LLM):

omnitune export \
    --model_name_or_path meta-llama/Meta-Llama-3-8B-Instruct \
    --adapter_name_or_path saves/llama3-8b/lora/sft \
    --template llama3 \
    --export_dir saves/llama3-8b-merged \
    --export_size 4 \
    --export_device cpu

---

🧠 Supported Model Families

---

📚 Documentation

• 📖 Getting Started Guide
• 🖥️ OmniTune-Board Web UI Guide
• ⚙️ Distributed Training (FSDP, DeepSpeed)
• 📋 Hyperparameters Reference

---

⭐ Star History

---

📄 License

This repository is licensed under the Apache 2.0 License. 📜