LLaMA.cpp Enhanced Docker Image for Modern GPUs

🎯 What is this?
This is an enhanced version of the official llama.cpp Docker image, specifically optimized for modern NVIDIA GPUs (RTX 30/40/50 series). It upgrades CUDA from 12.4.0 to 13.0.1 and adds RPC backend support for distributed processing.

🚀 Why use this instead of the official image?

• Better RTX 40/50 series support with CUDA 13.0.1
• RPC backend for distributed inference across multiple machines
• Smaller, faster - only targets modern GPU architectures (no legacy bloat)
• Same functionality as official ghcr.io/ggml-org/llama.cpp:full-cuda but enhanced

📦 Ready to use - No building required!
Available on Docker Hub: philglod/llamacpp-cuda13-modern-full:latest

🚀 Quick Start (Most Users Start Here!)

What You Need

• NVIDIA RTX 30/40/50 series GPU (older GPUs won't work with this optimized build)
• Docker Desktop with GPU support
• NVIDIA Container Toolkit

Get Started in 2 Minutes

1. Pull the image:

docker pull philglod/llamacpp-cuda13-modern-full:latest

2. Test it works:

docker run --rm --gpus all philglod/llamacpp-cuda13-modern-full:latest --server --help | grep -i cuda

You should see your GPU detected like: Device 0: NVIDIA GeForce RTX 4090, compute capability 8.9

3. Start using it! Choose what you want to do:

🌐 Run a Web Server

docker run --rm --gpus all -p 8080:8080 \
  philglod/llamacpp-cuda13-modern-full:latest \
  --server --host 0.0.0.0 --port 8080

Then visit http://localhost:8080 for the web interface!

📥 Download & Convert a Model from HuggingFace

mkdir ./models
docker run --rm --gpus all -v $(pwd)/models:/models \
  philglod/llamacpp-cuda13-modern-full:latest \
  --convert --hf-repo microsoft/Phi-3-mini-4k-instruct --outtype f16

🚀 Run a Complete AI Server

# After converting a model (like above), run a full server:
docker run -d --name my-ai-server --gpus all -p 8080:8080 -v $(pwd)/models:/models \
  philglod/llamacpp-cuda13-modern-full:latest \
  --server --host 0.0.0.0 --port 8080 \
  --model /models/Phi-3-mini-4k-instruct-f16.gguf \
  --ctx-size 4096 --n-gpu-layers 999

Access web UI at http://localhost:8080 or API at http://localhost:8080/v1/chat/completions

💡 What Can This Do?

This image includes everything you need for AI model work:

• 🌐 Web Server - Run models with a web interface
• 🔄 Model Conversion - Convert HuggingFace models to llama.cpp format
• 📊 Benchmarking - Test your GPU performance
• 💬 Interactive Chat - Talk to models directly
• 🔧 All Tools - Complete llama.cpp toolkit included

🎯 Who Should Use This?

✅ Perfect for you if:

• You have RTX 30/40/50 series GPU
• You want the latest CUDA performance improvements
• You need RPC support for distributed setups
• You want a ready-to-use solution (no building required)

❌ Not for you if:

• You have older GPUs (GTX 10/20 series, Tesla K80, etc.)
• You need to customize the build extensively
• You're fine with the official CUDA 12.4.0 images

🔄 Alternative: Use Official Images

For older GPUs or standard setups: ghcr.io/ggml-org/llama.cpp:full-cuda

📋 More Usage Examples

Interactive Chat with a Model

docker run --rm -it --gpus all -v $(pwd)/models:/models \
  philglod/llamacpp-cuda13-modern-full:latest \
  --run -m /models/your-model.gguf -p "Hello, how are you?"

Benchmark Your GPU

docker run --rm --gpus all -v $(pwd)/models:/models \
  philglod/llamacpp-cuda13-modern-full:latest \
  --bench -m /models/your-model.gguf

Convert Your Own Model

docker run --rm --gpus all -v $(pwd)/my-model:/input -v $(pwd)/converted:/output \
  philglod/llamacpp-cuda13-modern-full:latest \
  --convert --outtype f16 /input/ --output-dir /output/

🔧 GPU Compatibility

✅ Supported (Modern GPUs Only)

Series	Examples	CUDA Compute
RTX 30	3060, 3070, 3080, 3090	8.6
RTX 40	4060, 4070, 4080, 4090	8.9
RTX 50	5090, etc.	9.0

❌ Not Supported (Use Official Images Instead)

• GTX 10/20 series (Pascal, Turing)
• Tesla K80, P100, V100 (older data center GPUs)
• Any GPU with compute capability below 8.6

🏗️ For Developers: Building from Source

Most users don't need this section! Only if you want to customize the build.

Prerequisites

• Docker with GPU support
• Git
• This repository cloned locally

Build Process

# Clone llama.cpp source
git submodule update --init --recursive

# Build the image
docker build -t my-custom-llamacpp:latest --target full -f docker/cuda-13.0.1-custom.Dockerfile .

# Test it
docker run --rm --gpus all my-custom-llamacpp:latest --help

Publishing Your Own Version

# Tag for Docker Hub
docker tag my-custom-llamacpp:latest YOUR_USERNAME/llamacpp-custom:latest

# Push to Docker Hub
docker login
docker push YOUR_USERNAME/llamacpp-custom:latest

🔍 Technical Details

Custom CMake Configuration

Built with optimized flags for modern GPUs:

-DGGML_CUDA=ON                    # CUDA support
-DGGML_FORCE_CUBLAS=ON            # Force cuBLAS usage
-DGGML_RPC=ON                     # RPC backend support
-DCMAKE_CUDA_ARCHITECTURES="86;89;90"  # Modern GPUs only

Docker Hub Information

• Repository: philglod/llamacpp-cuda13-modern-full
• Tags: latest, 4067f07 (specific commit)
• Base Images: nvidia/cuda:13.0.1-devel-ubuntu24.04

System Requirements

• NVIDIA GPU with compute capability 8.6+
• NVIDIA Container Toolkit installed
• Docker with GPU support enabled
• Sufficient VRAM for your target models

🆘 Troubleshooting

GPU Not Detected?

# Check if NVIDIA Container Toolkit is working:
docker run --rm --gpus all nvidia/cuda:12.0-base-ubuntu20.04 nvidia-smi

Image Won't Start?

Make sure you're using --gpus all flag and have a compatible GPU (RTX 30/40/50 series)

Performance Issues?

This image is optimized for modern GPUs. For older GPUs, use the official images instead.

📜 License & Credits

Based on the official llama.cpp project. See the llama.cpp repository for licensing terms.

Special thanks to the llama.cpp team for the excellent foundation this build enhances.