FROM ghcr.io/avesed/vllm-ampere-optimized@sha256:6744eac95c9d0ee426c5c60ebaf72437da551e8a1786e8d2270e5164fb17dbfc

# Inherit the base's Torch 2.13.0+cu130 without duplicating CUDA wheels. Shadow
# calibration dependencies only in this venv; do not serve vLLM from it.
RUN python3 -m venv --system-site-packages /opt/mellum-quantize
COPY requirements.quantize.txt /opt/mellum/requirements.quantize.txt
RUN /opt/mellum-quantize/bin/python -m pip install --no-cache-dir -r /opt/mellum/requirements.quantize.txt \
    && /opt/mellum-quantize/bin/python -m pip freeze > /opt/mellum/quantizer-freeze.txt
COPY quantize.py audit_export.py check_quantizer.py recipe.yaml /opt/mellum/
ENV PATH="/opt/mellum-quantize/bin:${PATH}"
WORKDIR /opt/mellum
ENTRYPOINT ["python", "/opt/mellum/quantize.py"]