Bootstrap: localimage
From: ./vllm-v0.24.0-cu129.sif

%labels
    org.opencontainers.image.title vllm-berag
    org.opencontainers.image.base.name vllm/vllm-openai:v0.24.0-cu129
    org.opencontainers.image.revision 2ca4f5b4fe3aa8018ee5f6bd014a6684570642d9
    org.opencontainers.image.source https://github.com/EriChen0615/vllm-berag

%files
    ./build/vllm-berag-src /opt/

%post
    set -eu

    VLLM_PACKAGE_DIR=/usr/local/lib/python3.12/dist-packages/vllm
    test -d "$VLLM_PACKAGE_DIR"
    test -d /opt/vllm-berag-src/vllm

    # Preserve the base image's compiled extensions while replacing and
    # adding the Python files from the BERAG source snapshot.
    cp -a /opt/vllm-berag-src/vllm/. "$VLLM_PACKAGE_DIR/"

%environment
    export HF_HOME=/cache/huggingface
    export PYTHONUNBUFFERED=1
    export VLLM_USAGE_SOURCE=production-sif-image

%test
    VLLM_PACKAGE_DIR=/usr/local/lib/python3.12/dist-packages/vllm
    test -f "$VLLM_PACKAGE_DIR/berag.py"
    grep -q 'def generate_berag' "$VLLM_PACKAGE_DIR/entrypoints/llm.py"
    find "$VLLM_PACKAGE_DIR" -maxdepth 1 \
        -name '_C_stable_libtorch*.so' -print -quit | grep -q .

%runscript
    exec vllm serve "$@"

%help
    vllm-berag based on vllm/vllm-openai:v0.24.0-cu129.

    The BERAG source snapshot is stored at /opt/vllm-berag-src. Its Python
    files are merged into the upstream package at build time so that the
    base image's compiled CUDA extensions are preserved.

    Run with NVIDIA GPU support, for example:

        apptainer exec --nv IMAGE.sif python3 -c \
            'import torch, vllm; print(vllm.__file__); print(torch.__version__)'

        apptainer run --nv IMAGE.sif MODEL_NAME --host 0.0.0.0 --port 8000