LFM2-8B-A1B (UV Edition)

Liquid AI's 8B model - Zero setup, just works! 🚀

⚡ Setup (10 seconds)

# Install UV (one time only)
bash setup_uv.sh

🎯 Usage (NO VENV NEEDED!)

# Run inference
uv run python run.py "What is machine learning?"

# Train LoRA
uv run python train_simple.py

# Debug model
uv run python debug/debug_all.py

That's it! No venv, no activation, no pip install. UV handles everything automatically.

🔥 Why UV?

• No venv needed - UV manages everything
• 100x faster than pip
• Zero conflicts - automatic dependency resolution
• Just works - run any Python file with uv run

📁 Clean Project Structure

├── run.py                 # Inference (1 file, 30 lines)
├── train_simple.py        # Training (1 file, 40 lines)
├── inspect_safetensors.py # Weight inspector/peek tool
├── setup_uv.sh            # UV installer
├── pyproject.toml         # Dependencies
├── models/                # 8B model (4.7GB)
├── train/                 # Data scripts
│   ├── download_data.py
│   └── prepare_data.py
└── debug/                 # Model debugging tools

🎯 Training Examples

# Download training data
uv run python train/download_data.py

# Train LoRA adapter
uv run python train_simple.py

# Debug model internals
uv run python debug/debug_all.py

🔍 Safetensors Inspector (Peek Tool)

Deep inspection tool for .safetensors files - like dotPeek but for ML models!

# Quick inspection
uv run python inspect_safetensors.py your_model.safetensors

# Interactive mode
uv run python inspect_safetensors.py

📖 Full documentation: README_INSPECTOR.md

Features: tensor analysis, weight visualization, checkpoint comparison, NumPy export, bfloat16 support, and more!

Model Info

• Size: 4.7GB (quantized)
• Memory: ~5-6GB RAM
• Speed: 70-80 tokens/sec on M1/M2/M3
• Architecture: MoE with 4-bit/8-bit quantization

Files

├── run.py                    # Main inference script
├── models/
│   └── LFM2-8B-A1B-mlx/     # Quantized model (4.7GB)
└── train/                    # Training scripts
    ├── train_dense_lora.py   # Dense LoRA training
    └── README.md             # Training guide

Requirements

• Apple Silicon Mac (M1/M2/M3) recommended
• 16GB+ RAM
• Python 3.9+
• MLX framework

---

Model: LiquidAI/LFM2-8B-A1B