Docker Build with llama.cpp

Dockerfile

RUN apt-get update && apt-get install -y \
    bash curl htop stress-ng \
    build-essential cmake git \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /experiment

CMD ["/bin/bash"]

docker-compose.yml

  pixel-pro-sim:
    build: .
    image: pixel-pro-sim
    container_name: pixel-pro-sim
    mem_limit: 16g # change this
    memswap_limit: 16g # change this
    cpus: 6
    stdin_open: true
    tty: true
    volumes:
      - ../llama.cpp:/experiment/llama.cpp. # change this
      - ~/.cache/huggingface/hub:/models # change this

Build & Run

cd pixel-pro-sim
docker compose build
docker compose run pixel-pro-sim bash
cd /experiment/llama.cpp

# compile optimized (-O2) but keep debug symbols and keep #ifndef NDEBUG blocks active
cmake -B build -DCMAKE_BUILD_TYPE=RelWithDebInfo -DGGML_CUDA=OFF

# skip CUDA backend — no GPU in this container
cmake --build build -j6

# no mmap
./build/bin/llama-cli   --model /models/models--unsloth--Llama-3.3-70B-Instruct-GGUF/snapshots/8f14c5c5d06fca109ff16f94496147a8479711de/Llama-3.3-70B-Instruct-UD-IQ1_S.gguf   --prompt "Hello"   --n-predict 1   --no-mmap   -ngl 0  -v   2>&1 | tee /experiment/llama.cpp/run.log

# mmap
./build/bin/llama-cli   --model /models/models--unsloth--Llama-3.3-70B-Instruct-GGUF/snapshots/8f14c5c5d06fca109ff16f94496147a8479711de/Llama-3.3-70B-Instruct-UD-IQ1_S.gguf   --prompt "Hello"   --n-predict 1   --mmap   -ngl 0   -v   2>&1 | tee /experiment/llama.cpp/run_mmap.log