Llame Worker Example

Shows how to consume llame-worker (the llameworker library) from a plain CMake project. One-off multimodal prompts against a local GGUF model: text, image, and video.

The entire integration is two CMake lines:

add_subdirectory(externals/llame-worker)
target_link_libraries(llameworker_example PRIVATE llameworker)

Requirements

A vision-capable GGUF model and its matching multimodal projector (mmproj-*.gguf) in models/, plus input.jpg and input.mp4 in images/ (both paths are hard-coded in main.cpp). For video, ffmpeg (with ffprobe) must be on PATH.

Ready-to-use vision GGUFs (model + mmproj) live in the ggml-org multimodal collection on Hugging Face. Gemma 3 is the safe default; for stronger text/UI reading try Qwen2.5-VL, InternVL3, or Pixtral. Download with the hf CLI, then point the constants at the top of main.cpp at the files:

hf download ggml-org/Qwen2.5-VL-7B-Instruct-GGUF \
  Qwen2.5-VL-7B-Instruct-Q8_0.gguf mmproj-Qwen2.5-VL-7B-Instruct-f16.gguf \
  --local-dir models

Setup

git submodule update --init --recursive

Build & run

cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j
cd build
./llameworker_example

Build Release: a Debug build of llama.cpp is an order of magnitude slower at inference. On macOS, Metal is enabled by default; for NVIDIA GPUs, add -DGGML_CUDA=ON to the configure step. The binary takes no arguments and runs three sections: a text-only prompt, image description, and a frame-sampled video summary.