Dockerfile

# Dockerfile
FROM ubuntu:22.04

# Faster noninteractive apt
ENV DEBIAN_FRONTEND=noninteractive

# --- build deps
RUN apt-get update && apt-get install -y --no-install-recommends \
    build-essential cmake git curl ca-certificates pkg-config \
    libopenblas-dev libcurl4-openssl-dev ccache \
 && rm -rf /var/lib/apt/lists/*

# --- build llama.cpp with BLAS + server
RUN git clone --depth=1 https://github.com/ggerganov/llama.cpp  /app && \
    cmake -S /app -B /app/build \
      -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS \
      -DGGML_NATIVE=OFF \
      -DLLAMA_BUILD_SERVER=ON \
      -DLLAMA_BUILD_EXAMPLES=OFF \
      -DLLAMA_BUILD_TESTS=OFF \
      -DCMAKE_C_COMPILER_LAUNCHER=ccache \
      -DCMAKE_CXX_COMPILER_LAUNCHER=ccache && \
    cmake --build /app/build --target llama-server -j2 && \
    ln -s /app/build/bin/llama-server /usr/local/bin/llama-server

# --- runtime env + user
RUN mkdir -p /models /workspace /data \
 && useradd -m -u 1000 user \
 && chown -R user:user /models /workspace /data

WORKDIR /workspace

# Defaults (override at deploy time as needed)
ENV MODEL_REPO="mradermacher/LFM2-VL-450M-GGUF" \
    MODEL_FILE="LFM2-VL-450M.Q3_K_L.gguf" \
    MMPROJ_FILE="LFM2-VL-450M.mmproj-Q8_0.gguf" \
    CTX_SIZE="4096" \
    THREADS="4" \
    PORT="7860"

# Use bash for safer heredoc
SHELL ["/bin/bash", "-lc"]

# --- startup script (downloads model if missing, then starts server)
RUN cat > /usr/local/bin/start.sh <<'EOF' \
 && chmod +x /usr/local/bin/start.sh
#!/usr/bin/env bash
set -euo pipefail

PORT="${PORT:-7860}"
CTX_SIZE="${CTX_SIZE:-4096}"
THREADS="${THREADS:-4}"
MODEL_REPO="${MODEL_REPO:-TheBloke/Llama-2-7B-GGUF}"
MODEL_FILE="${MODEL_FILE:-llama-2-7b.Q4_K_M.gguf}"
MMPROJ_FILE="${MMPROJ_FILE:-LFM2-VL-450M.mmproj-Q8_0.gguf}"

MODEL_PATH="/data/${MODEL_FILE}"
MODEL_PATH2="/data/${MMPROJ_FILE}"

HF_URL="https://huggingface.co/${MODEL_REPO}/resolve/main/${MODEL_FILE}"
HF_URL2="https://huggingface.co/${MODEL_REPO}/resolve/main/${MMPROJ_FILE}"

mkdir -p /data

if [[ ! -f "$MODEL_PATH" ]]; then
  echo "Downloading ${MODEL_REPO}/${MODEL_FILE} to /data..."
  if [[ -n "${HF_TOKEN:-}" ]]; then
    curl -fL --progress-bar -H "Authorization: Bearer ${HF_TOKEN}" -o "$MODEL_PATH" "$HF_URL"
  else
    curl -fL --progress-bar -o "$MODEL_PATH" "$HF_URL"
  fi
fi

if [[ ! -f "$MODEL_PATH2" ]]; then
  echo "Downloading ${MODEL_REPO}/${MMPROJ_FILE} to /data..."
  if [[ -n "${HF_TOKEN:-}" ]]; then
    curl -fL --progress-bar -H "Authorization: Bearer ${HF_TOKEN}" -o "$MODEL_PATH2" "$HF_URL2"
  else
    curl -fL --progress-bar -o "$MODEL_PATH2" "$HF_URL2"
  fi
fi

# Correct flags for llama-server:
exec llama-server \
  --model "$MODEL_PATH" \
  --mmproj "$MODEL_PATH2" \
  --host 0.0.0.0 \
  --port "$PORT" \
  --ctx-size "$CTX_SIZE" \
  --threads 2 \
  --jinja 
  
EOF

EXPOSE 7860
USER user
ENTRYPOINT ["/usr/local/bin/start.sh"]