GGUF usage with llama.cpp

Llama.cpp allows you to download and run inference on a GGUF simply by providing a path to the Hugging Face repo path and the file name. llama.cpp downloads the model checkpoint and automatically caches it. The location of the cache is defined by LLAMA_CACHE environment variable; read more about it here.

You can install llama.cpp through brew (works on Mac and Linux), or you can build it from source. There are also pre-built binaries and Docker images that you can check in the official documentation.

Option 1: Install with brew

brew install llama.cpp

Option 2: build from source

Step 1: Clone llama.cpp from GitHub.

git clone https://github.com/ggerganov/llama.cpp

Step 2: Move into the llama.cpp folder and build it. You can also add hardware-specific flags (for ex: -DGGML_CUDA=1 for Nvidia GPUs).

cd llama.cpp
cmake -B build   # optionally, add -DGGML_CUDA=ON to activate CUDA
cmake --build build --config Release

Note: for other hardware support (for ex: AMD ROCm, Intel SYCL), please refer to llama.cpp’s build guide

Once installed, you can use the llama-cli or llama-server as follows:

llama-cli -hf bartowski/Llama-3.2-3B-Instruct-GGUF:Q8_0

Note: You can explicitly add -no-cnv to run the CLI in raw completion mode (non-chat mode).

Additionally, you can invoke an OpenAI spec chat completions endpoint directly using the llama.cpp server:

llama-server -hf bartowski/Llama-3.2-3B-Instruct-GGUF:Q8_0

After running the server you can simply utilise the endpoint as below:

curl http://localhost:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer no-key" \
-d '{
"messages": [
    {
        "role": "system",
        "content": "You are an AI assistant. Your top priority is achieving user fulfilment via helping them with their requests."
    },
    {
        "role": "user",
        "content": "Write a limerick about Python exceptions"
    }
  ]
}'

Replace -hf with any valid Hugging Face hub repo name - off you go! 🦙