llama.cpp

llama.cpp is the best backend in two important scenarios:

You don't have a GPU.
You want to run a model that doesn't fit into your GPU.

Setting up the models

Pre-converted

Download the ggml model directly into your text-generation-webui/models folder, making sure that its name contains ggml somewhere and ends in .bin. It's a single file.

q4_K_M quantization is recommended.

Convert Llama yourself

Follow the instructions in the llama.cpp README to generate a ggml: https://github.com/ggerganov/llama.cpp#prepare-data--run

GPU acceleration

Enabled with the --n-gpu-layers parameter.

If you have enough VRAM, use a high number like --n-gpu-layers 1000 to offload all layers to the GPU.
Otherwise, start with a low number like --n-gpu-layers 10 and then gradually increase it until you run out of memory.

This feature works out of the box for NVIDIA GPUs. For other GPUs, you need to uninstall llama-cpp-python with

pip uninstall -y llama-cpp-python

and then recompile it using the commands here: https://pypi.org/project/llama-cpp-python/

macOS

For macOS, these are the commands:

pip uninstall -y llama-cpp-python
CMAKE_ARGS="-DLLAMA_METAL=on" FORCE_CMAKE=1 pip install llama-cpp-python --no-cache-dir

1.3 KiB Raw Blame History