vllm-omni/docker/Dockerfile.ci

# Use the official release image with its matching CUDA/PyTorch stack.
ARG VLLM_VERSION=0.31.0
ARG VLLM_BASE_IMAGE=vllm/vllm-openai
ARG VLLM_BASE_TAG=v${VLLM_VERSION}

# We use registry-based layer caching for the dependencies,
# where the tag is determined based on the hash of the contents
# of files like pyproject.toml, etc. The deps stage should have
# everything except the actual project dir, since that should not
# change much across branches.
FROM ${VLLM_BASE_IMAGE}:${VLLM_BASE_TAG} AS deps
ARG VLLM_VERSION
# Keep APP_DIR outside /workspace so k8s Buildkite workspace mounts cannot
# shadow the editable install target (#5364). Docker WORKDIR creates /opt if
# needed.
ARG APP_DIR=/opt/vllm-omni
WORKDIR ${APP_DIR}

RUN ln -sf /usr/bin/python3 /usr/bin/python

# Placeholder package tree so editable install can register APP_DIR.
# The final stage COPY --link fills this directory with the real sources;
# the editable .pth / egg-link keeps resolving imports there — no PYTHONPATH.
RUN mkdir -p vllm_omni && touch vllm_omni/__init__.py

# Install system dependencies
RUN apt-get update && \
    apt-get install -y espeak-ng git jq && \
    apt-get clean && \
    rm -rf /var/lib/apt/lists/*

# Copy dependency related things first and install deps/
# We set the version str to 0.0.0 since it doesn't matter
# for CI
COPY pyproject.toml setup.py ./
COPY requirements/ requirements/
# Scope the override to Omni; dependency source builds need their own versions.
ENV VLLM_OMNI_VERSION_OVERRIDE=0.0.0
# Editable install: deps + project metadata are cached in this layer, while
# the package itself is linked to APP_DIR (populated later by COPY --link).
# Do NOT use a non-editable install here — that copies an empty stub into
# site-packages and breaks ``import vllm_omni.entrypoints``.
RUN uv pip install --system -e ".[dev]"
# Fail the image build if dependency resolution replaced the selected vLLM.
# The CPU builder has no NVIDIA driver; CUDA imports are checked at the start
# of each GPU test job by .buildkite/cuda/scripts/check_vllm_runtime.py.
RUN python3 -c "import vllm; assert vllm.__version__ == '${VLLM_VERSION}', vllm.__version__"

# The final stage just copies the working dir on top of the
# dependencies image. This uses --link so that we can skip
# the image pull in Buildkite when we have a cache hit.
#
# IMPORTANT: Do not make any modifications that depend on the
# file system after the --link copy, otherwise we'd need to pull
# the heavy base image to build it. Editable install above already
# registered APP_DIR, so we must not re-pip-install after COPY.
FROM deps
ARG APP_DIR=/opt/vllm-omni
WORKDIR ${APP_DIR}

COPY --link . .

ENTRYPOINT []