KBaba7/llama.cpp

1ARG UBUNTU_VERSION=22.042# This needs to generally match the container host's environment.3ARG CUDA_VERSION=12.6.04# Target the CUDA build image5ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}6 7ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}8 9FROM ${BASE_CUDA_DEV_CONTAINER} AS build10 11# CUDA architecture to build for (defaults to all supported archs)12ARG CUDA_DOCKER_ARCH=default13 14RUN apt-get update && \15    apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp116 17WORKDIR /app18 19COPY . .20 21RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \22    export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \23    fi && \24    cmake -B build -DGGML_NATIVE=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \25    cmake --build build --config Release -j$(nproc)26 27RUN mkdir -p /app/lib && \28    find build -name "*.so" -exec cp {} /app/lib \;29 30RUN mkdir -p /app/full \31    && cp build/bin/* /app/full \32    && cp *.py /app/full \33    && cp -r gguf-py /app/full \34    && cp -r requirements /app/full \35    && cp requirements.txt /app/full \36    && cp .devops/tools.sh /app/full/tools.sh37 38## Base image39FROM ${BASE_CUDA_RUN_CONTAINER} AS base40 41RUN apt-get update \42    && apt-get install -y libgomp1 curl\43    && apt autoremove -y \44    && apt clean -y \45    && rm -rf /tmp/* /var/tmp/* \46    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \47    && find /var/cache -type f -delete48 49COPY --from=build /app/lib/ /app50 51### Full52FROM base AS full53 54COPY --from=build /app/full /app55 56WORKDIR /app57 58RUN apt-get update \59    && apt-get install -y \60    git \61    python3 \62    python3-pip \63    && pip install --upgrade pip setuptools wheel \64    && pip install -r requirements.txt \65    && apt autoremove -y \66    && apt clean -y \67    && rm -rf /tmp/* /var/tmp/* \68    && find /var/cache/apt/archives /var/lib/apt/lists -not -name lock -type f -delete \69    && find /var/cache -type f -delete70 71 72ENTRYPOINT ["/app/tools.sh"]73 74### Light, CLI only75FROM base AS light76 77COPY --from=build /app/full/llama-cli /app78 79WORKDIR /app80 81ENTRYPOINT [ "/app/llama-cli" ]82 83### Server, Server only84FROM base AS server85 86ENV LLAMA_ARG_HOST=0.0.0.087 88COPY --from=build /app/full/llama-server /app89 90WORKDIR /app91 92HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]93 94ENTRYPOINT [ "/app/llama-server" ]95