# llama-server built from source with SM 120 (Blackwell / RTX 5070) CUDA kernels.
# Multi-stage: ~8GB devel toolkit stays in builder, runtime image is ~2GB.

# ── Builder ──────────────────────────────────────────────────────────────
FROM nvidia/cuda:12.8.1-devel-ubuntu24.04 AS builder

RUN apt-get update && apt-get install -y --no-install-recommends \
    cmake git build-essential curl ca-certificates \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /build

# Pin to a release tag for reproducibility
ARG LLAMA_CPP_VERSION=b5460
RUN git clone --depth 1 --branch ${LLAMA_CPP_VERSION} \
    https://github.com/ggerganov/llama.cpp.git

WORKDIR /build/llama.cpp

# CUDA driver symbols (cuMemCreate, etc.) are resolved at runtime by nvidia-container-runtime.
# --allow-shlib-undefined lets the linker accept unresolved refs in libggml-cuda.so.
RUN cmake -B build \
    -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_CUDA_ARCHITECTURES=120 \
    -DGGML_CUDA=ON \
    -DGGML_CUDA_FORCE_CUBLAS=ON \
    -DLLAMA_BUILD_SERVER=ON \
    -DLLAMA_CURL=OFF \
    -DCMAKE_EXE_LINKER_FLAGS="-Wl,--allow-shlib-undefined" \
    && cmake --build build --target llama-server -j$(nproc)

# ── Runtime ──────────────────────────────────────────────────────────────
FROM nvidia/cuda:12.8.1-runtime-ubuntu24.04

RUN apt-get update && apt-get install -y --no-install-recommends \
    curl ca-certificates libgomp1 \
    && rm -rf /var/lib/apt/lists/*

# Copy server binary and its shared libraries (libggml-*.so)
COPY --from=builder /build/llama.cpp/build/bin/llama-server /usr/local/bin/llama-server
COPY --from=builder /build/llama.cpp/build/bin/lib*.so /usr/local/lib/

RUN ldconfig

RUN useradd -u 1000 -m llama 2>/dev/null || true
USER 1000

ENTRYPOINT ["llama-server"]