deploy/runtime/Dockerfile

# ROCm image of the vLLM Semantic Router built-in model runtime (`vllm-sr-runtime`)
# for the Decision 2.0 Studio hosts.
#
# Build context: `src/model-runtime` of a vllm-project/semantic-router commit on
# `main`. BASE_IMAGE is the Decision 2.0 packages' release ROCm image (PyTorch
# 2.12.0 on ROCm 7.2, Triton 3.7.1, FLA 0.5.2, causal-conv1d 1.7.0). The runtime's
# `exact` profile answers bit for bit like the released packages on that image;
# another PyTorch, Triton or FLA build can move answers by rounding.
#
#   docker build -f deploy/runtime/Dockerfile \
#     --build-arg BASE_IMAGE=<release-image> --build-arg VLLM_SR_COMMIT=<40-hex> \
#     -t decision-studio/vllm-sr-runtime-rocm:<12-hex> <checkout>/src/model-runtime
ARG BASE_IMAGE
FROM ${BASE_IMAGE}
ARG VLLM_SR_COMMIT
ENV PYTHONDONTWRITEBYTECODE=1 \
    PYTHONUNBUFFERED=1 \
    PIP_NO_CACHE_DIR=1 \
    PIP_DISABLE_PIP_VERSION_CHECK=1 \
    HF_HUB_DISABLE_TELEMETRY=1
WORKDIR /opt/vllm-sr-runtime
COPY pyproject.toml README.md ./
COPY vllm_sr_runtime ./vllm_sr_runtime
RUN test -n "$VLLM_SR_COMMIT" \
    && python3 -m pip install --no-deps --no-build-isolation . \
    && python3 -c "import causal_conv1d, fla, huggingface_hub, numpy, prometheus_client, safetensors, starlette, tokenizers, torch, triton, uvicorn, yaml, vllm_sr_runtime; print('torch', torch.__version__, 'triton', triton.__version__, 'fla', fla.__version__, 'causal_conv1d', causal_conv1d.__version__)"
LABEL org.opencontainers.image.source="https://github.com/vllm-project/semantic-router" \
      org.opencontainers.image.revision="${VLLM_SR_COMMIT}"
WORKDIR /
ENTRYPOINT ["vllm-sr-runtime"]
CMD ["--help"]