GLM-4.7 - Installation and Configuration for Low-End PCs

This repository contains automated scripts to install and configure the GLM-4.7 model on machines with limited resources.

📋 Minimum Requirements

Minimum Hardware (CPU Only)

• RAM: 32GB+ (64GB+ recommended)
• Disk: 200GB+ free space (SSD recommended)
• CPU: Modern multi-core processor

Recommended Hardware (with GPU)

• GPU: 8GB+ VRAM (16GB+ recommended)
• RAM: 64GB+ (128GB+ recommended)
• Disk: 300GB+ free space (NVMe SSD recommended)
• CUDA: Compatible with CUDA 11.8+ or 12.1+

🎯 Available Model Versions

GLM-4.7 is available in various quantizations for different hardware capabilities:

🚀 Quick Start

Windows (PowerShell)

# 1. Install dependencies
.\scripts\install.ps1

# 2. Download model (choose appropriate version)
.\scripts\download-model.ps1 -Version "Q4_K_S"

# 3. Run the model
.\scripts\run-llamacpp.ps1

Linux/Mac (Bash)

# 1. Install dependencies
chmod +x scripts/install.sh
./scripts/install.sh

# 2. Download model (choose appropriate version)
chmod +x scripts/download-model.sh
./scripts/download-model.sh Q4_K_S

# 3. Run the model
chmod +x scripts/run-llamacpp.sh
./scripts/run-llamacpp.sh

📁 Repository Structure

.
├── README.md                 # This file
├── scripts/
│   ├── install.sh            # Linux/Mac installation
│   ├── install.ps1           # Windows installation
│   ├── download-model.sh     # Model download (Linux/Mac)
│   ├── download-model.ps1    # Model download (Windows)
│   ├── run-llamacpp.sh       # Run with llama.cpp (Linux/Mac)
│   ├── run-llamacpp.ps1      # Run with llama.cpp (Windows)
│   ├── run-ollama.sh         # Run with Ollama (Linux/Mac)
│   └── run-ollama.ps1        # Run with Ollama (Windows)
├── config/
│   ├── hardware-config.yaml  # Hardware configuration
│   └── model-config.json     # Model settings
└── models/                   # Directory for downloaded models

⚙️ Configuration

1. Configure Hardware

Edit config/hardware-config.yaml with your machine specifications:

hardware:
  gpu:
    available: true
    vram_gb: 8
    cuda_arch: "75"  # For RTX 2060, 2070, 2080
  ram_gb: 32
  cpu_cores: 8
  disk_space_gb: 500

2. Choose Model Version

Based on your hardware, choose the appropriate version:

• Very low-end PC (32GB RAM, no GPU): Use Q4_K_S or consider smaller models
• Moderate PC (64GB RAM, 8-16GB GPU): Use Q4_K_M
• Powerful PC (128GB+ RAM, 24GB+ GPU): Use UD-Q2_K_XL or Q5_K_M

🔧 Execution Methods

Option 1: llama.cpp (Recommended for limited hardware)

llama.cpp offers better control over CPU/GPU offloading and quantization.

Advantages:

• Intelligent offloading support
• Lower memory usage
• Better for limited hardware

Option 2: Ollama (Simpler)

Ollama is easier to use, but may be less efficient on limited hardware.

Advantages:

• Simpler installation
• More user-friendly interface
• Automatic model management

📝 Usage Examples

Run with small context (saves memory)

./scripts/run-llamacpp.sh --ctx-size 4096 --threads 4

Run CPU-only

./scripts/run-llamacpp.sh --cpu-only

Run with partial CPU offloading

./scripts/run-llamacpp.sh --gpu-layers 10

🐛 Troubleshooting

Error: "Out of memory"

• Reduce --ctx-size (context size)
• Use a more quantized version of the model
• Reduce --gpu-layers to offload more to CPU

Error: "CUDA not found"

• Check if CUDA is installed: nvidia-smi
• Recompile llama.cpp with CUDA support

Model too slow

• Increase --threads (number of CPU threads)
• Use more GPU layers if you have available VRAM
• Consider using a lighter version of the model

📚 Additional Resources

• Official GLM-4.7 Documentation
• llama.cpp GitHub
• Ollama Documentation
• Quantized models on Hugging Face

📄 License

This repository contains installation and configuration scripts. The GLM-4.7 model has its own license - consult the official repository.

🤝 Contributions

Contributions are welcome! Feel free to open issues or pull requests.

⚠️ Warnings

• Large models may take a long time to download (100GB+)
• First run may be slow while the model loads
• Make sure you have enough disk space before downloading
• On very limited hardware, consider using smaller models or cloud services