Using LLaVA
LLaVA / BakLLaVA can be used with LLaMA.cpp.
You can find the full llama.cpp documentation here.
Step 1 - Clone the repo
git clone https://github.com/ggerganov/llama.cpp
cd llama.cppStep 2 - Download the model
For example, we will use BakLLaVA-1 model, which is what is used on the demo instance.
Navigate to mys/ggml_bakllava-1 and download either q4 or q5 quant, as well as the mmproj-model-f16.gguf file.
The mmproj-model-f16.gguf file is necessary for the vision model.
Step 3 - Build the server
make serverStep 4 - Run the server
Read the llama.cpp documentation for more information on the server options. Or run ./server --help.
./server -t 4 -c 4096 -ngl 35 -b 512 --mlock -m models/openchat_3.5.Q5_K_M.gguf --mmproj models/mmproj-model-f16.ggufStep 5 - Enable the server in the client
settings -> Vision -> Vision Backend -> LLaMA.cppLast updated