Dockerfile.rocm

# default base image
ARG REMOTE_VLLM="0"
ARG COMMON_WORKDIR=/app
ARG BASE_IMAGE=rocm/vllm-dev:base

# Sccache configuration (only used in release pipeline)
ARG USE_SCCACHE
ARG SCCACHE_DOWNLOAD_URL
ARG SCCACHE_ENDPOINT
ARG SCCACHE_BUCKET_NAME=vllm-build-sccache
ARG SCCACHE_REGION_NAME=us-west-2
ARG SCCACHE_S3_NO_CREDENTIALS=0

FROM ${BASE_IMAGE} AS base

ARG ARG_PYTORCH_ROCM_ARCH
ENV PYTORCH_ROCM_ARCH=${ARG_PYTORCH_ROCM_ARCH:-${PYTORCH_ROCM_ARCH}}

# Install some basic utilities
RUN apt-get update -q -y && apt-get install -q -y \
    sqlite3 libsqlite3-dev libfmt-dev libmsgpack-dev libsuitesparse-dev \
    apt-transport-https ca-certificates wget curl
RUN python3 -m pip install --upgrade pip
# Remove sccache only if not using sccache (it exists in base image from Dockerfile.rocm_base)
ARG USE_SCCACHE
RUN if [ "$USE_SCCACHE" != "1" ]; then \
        apt-get purge -y sccache || true; \
        python3 -m pip uninstall -y sccache || true; \
        rm -f "$(which sccache)" || true; \
    fi

# Install UV — download first, then run, so a curl failure is not masked by the pipe
RUN curl -LsSf --retry 3 --retry-delay 5 https://astral.sh/uv/install.sh -o /tmp/uv-install.sh \
    && env UV_INSTALL_DIR="/usr/local/bin" sh /tmp/uv-install.sh \
    && rm -f /tmp/uv-install.sh \
    && uv --version

# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
# Reference: https://github.com/astral-sh/uv/pull/1694
ENV UV_HTTP_TIMEOUT=500
ENV UV_INDEX_STRATEGY="unsafe-best-match"
# Use copy mode to avoid hardlink failures with Docker cache mounts
ENV UV_LINK_MODE=copy

# Install sccache if USE_SCCACHE is enabled (for release builds)
ARG USE_SCCACHE
ARG SCCACHE_DOWNLOAD_URL
ARG SCCACHE_ENDPOINT
ARG SCCACHE_BUCKET_NAME
ARG SCCACHE_REGION_NAME
ARG SCCACHE_S3_NO_CREDENTIALS
RUN if [ "$USE_SCCACHE" = "1" ]; then \
        if command -v sccache >/dev/null 2>&1; then \
            echo "sccache already installed, skipping installation"; \
            sccache --version; \
        else \
            echo "Installing sccache..." \
            && SCCACHE_ARCH="x86_64" \
            && SCCACHE_VERSION="v0.8.1" \
            && SCCACHE_DL_URL="${SCCACHE_DOWNLOAD_URL:-https://github.com/mozilla/sccache/releases/download/${SCCACHE_VERSION}/sccache-${SCCACHE_VERSION}-${SCCACHE_ARCH}-unknown-linux-musl.tar.gz}" \
            && curl -L -o /tmp/sccache.tar.gz ${SCCACHE_DL_URL} \
            && tar -xzf /tmp/sccache.tar.gz -C /tmp \
            && mv /tmp/sccache-${SCCACHE_VERSION}-${SCCACHE_ARCH}-unknown-linux-musl/sccache /usr/bin/sccache \
            && chmod +x /usr/bin/sccache \
            && rm -rf /tmp/sccache.tar.gz /tmp/sccache-${SCCACHE_VERSION}-${SCCACHE_ARCH}-unknown-linux-musl \
            && sccache --version; \
        fi; \
    fi

# Set sccache environment variables only when USE_SCCACHE=1
# This prevents S3 config from leaking into images when sccache is not used
ARG USE_SCCACHE
ENV SCCACHE_BUCKET=${USE_SCCACHE:+${SCCACHE_BUCKET_NAME}}
ENV SCCACHE_REGION=${USE_SCCACHE:+${SCCACHE_REGION_NAME}}
ENV SCCACHE_S3_NO_CREDENTIALS=${USE_SCCACHE:+${SCCACHE_S3_NO_CREDENTIALS}}
ENV SCCACHE_IDLE_TIMEOUT=${USE_SCCACHE:+0}

ARG COMMON_WORKDIR
WORKDIR ${COMMON_WORKDIR}


# -----------------------
# vLLM fetch stages
FROM base AS fetch_vllm_0
ONBUILD COPY ./ vllm/
FROM base AS fetch_vllm_1
ARG VLLM_REPO="https://github.com/vllm-project/vllm.git"
ARG VLLM_BRANCH="main"
ENV VLLM_REPO=${VLLM_REPO}
ENV VLLM_BRANCH=${VLLM_BRANCH}
ONBUILD RUN git clone ${VLLM_REPO} \
	    && cd vllm \
	    && git fetch -v --prune -- origin ${VLLM_BRANCH} \
	    && git checkout FETCH_HEAD \
        && if [ ${VLLM_REPO} != "https://github.com/vllm-project/vllm.git" ] ; then \
               git remote add upstream "https://github.com/vllm-project/vllm.git" \
               && git fetch upstream ; fi
FROM fetch_vllm_${REMOTE_VLLM} AS fetch_vllm

# -----------------------
# vLLM build stages
FROM fetch_vllm AS build_vllm
# Build vLLM (setup.py auto-detects sccache in PATH)
RUN cd vllm \
    && python3 -m pip install -r requirements/rocm.txt \
    && python3 setup.py clean --all  \
    && python3 setup.py bdist_wheel --dist-dir=dist
FROM scratch AS export_vllm
ARG COMMON_WORKDIR
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/dist/*.whl /
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/requirements /requirements
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/benchmarks /benchmarks
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/tests /tests
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/examples /examples
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/docker/Dockerfile.rocm /docker/
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/.buildkite /.buildkite
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm/vllm/v1 /vllm_v1

# RIXL/UCX build stages
FROM base AS build_rixl
ARG RIXL_BRANCH="f33a5599"
ARG RIXL_REPO="https://github.com/ROCm/RIXL.git"
ARG UCX_BRANCH="da3fac2a"
ARG UCX_REPO="https://github.com/ROCm/ucx.git"
ENV ROCM_PATH=/opt/rocm
ENV UCX_HOME=/usr/local/ucx
ENV RIXL_HOME=/usr/local/rixl
ENV RIXL_BENCH_HOME=/usr/local/rixl_bench

# RIXL build system dependences and RDMA support
RUN apt-get -y update && apt-get -y install autoconf libtool pkg-config \
    libgrpc-dev \
    libgrpc++-dev \
    libprotobuf-dev \
    protobuf-compiler-grpc \
    libcpprest-dev \
    libaio-dev \
    librdmacm1 \
    librdmacm-dev \
    libibverbs1 \
    libibverbs-dev \
    ibverbs-utils \
    rdmacm-utils \
    ibverbs-providers \
    && rm -rf /var/lib/apt/lists/*

RUN uv pip install --system meson auditwheel patchelf tomlkit

RUN cd /usr/local/src && \
    git clone ${UCX_REPO} &&  \
    cd ucx  && \
    git checkout ${UCX_BRANCH} && \
    ./autogen.sh && \
    mkdir build && cd build && \
    ../configure \
        --prefix=/usr/local/ucx \
        --enable-shared \
        --disable-static \
        --disable-doxygen-doc \
        --enable-optimizations \
        --enable-devel-headers \
        --with-rocm=/opt/rocm \
        --with-verbs \
        --with-dm \
        --enable-mt && \
    make -j && \
    make install

ENV PATH=/usr/local/ucx/bin:$PATH
ENV LD_LIBRARY_PATH=${UCX_HOME}/lib:${LD_LIBRARY_PATH}

RUN git clone ${RIXL_REPO} /opt/rixl && \
    cd /opt/rixl && \
    git checkout ${RIXL_BRANCH} && \
    meson setup build --prefix=${RIXL_HOME} \
                     -Ducx_path=${UCX_HOME} \
                     -Drocm_path=${ROCM_PATH} && \
    cd build && \
    ninja && \
    ninja install

# Generate RIXL wheel
RUN cd /opt/rixl && mkdir -p /app/install && \
    ./contrib/build-wheel.sh \
        --output-dir /app/install \
        --rocm-dir ${ROCM_PATH} \
        --ucx-plugins-dir ${UCX_HOME}/lib/ucx \
        --nixl-plugins-dir ${RIXL_HOME}/lib/x86_64-linux-gnu/plugins

# DeepEP build stage
FROM base AS build_deep
ARG ROCSHMEM_BRANCH="ba0bf0f3"
ARG ROCSHMEM_REPO="https://github.com/ROCm/rocm-systems.git"
ARG DEEPEP_BRANCH="e84464ec"
ARG DEEPEP_REPO="https://github.com/ROCm/DeepEP.git"
ARG DEEPEP_NIC="cx7"
ENV ROCSHMEM_DIR=/opt/rocshmem

RUN git clone ${ROCSHMEM_REPO} \
 && cd rocm-systems \
 && git checkout ${ROCSHMEM_BRANCH} \
 && mkdir -p projects/rocshmem/build \
 && cd projects/rocshmem/build \
 && cmake .. \
    -DCMAKE_INSTALL_PREFIX="${ROCSHMEM_DIR}" \
    -DROCM_PATH=/opt/rocm \
    -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
    -DUSE_EXTERNAL_MPI=OFF \
 && make -j \
 && make install

# Build DeepEP wheel.
# DeepEP looks for rocshmem at ROCSHMEM_DIR.
RUN git clone ${DEEPEP_REPO} \
 && cd DeepEP \
 && git checkout ${DEEPEP_BRANCH} \
 && python3 setup.py --variant rocm --nic ${DEEPEP_NIC} bdist_wheel --dist-dir=/app/deep_install

# -----------------------
# vLLM wheel release build stage (for building distributable wheels)
# This stage pins dependencies to custom ROCm wheel versions and handles version detection
FROM fetch_vllm AS build_vllm_wheel_release

ARG COMMON_WORKDIR

# Create /install directory for custom wheels
RUN mkdir -p /install

# Copy custom ROCm wheels from docker/context if they exist
# COPY ensures Docker cache is invalidated when wheels change
# .keep file ensures directory always exists for COPY to work
COPY docker/context/base-wheels/ /tmp/base-wheels/
# This is how we know if we are building for a wheel release or not.
# If there are not wheels found there, we are not building for a wheel release. 
# So we exit with an error. To skip this stage.
RUN if [ -n "$(ls /tmp/base-wheels/*.whl 2>/dev/null)" ]; then \
        echo "Found custom wheels - copying to /install"; \
        cp /tmp/base-wheels/*.whl /install/ && \
        echo "Copied custom wheels:"; \
        ls -lh /install/; \
    else \
        echo "ERROR: No custom wheels found in docker/context/base-wheels/"; \
        echo "Wheel releases require pre-built ROCm wheels."; \
        exit 1; \
    fi

# GIT_REPO_CHECK: Verify repo is clean and tags are available (for release builds)
# This matches CUDA's Dockerfile behavior for proper version detection via setuptools_scm
ARG GIT_REPO_CHECK=0
RUN if [ "$GIT_REPO_CHECK" != "0" ]; then \
        echo "Running repository checks..."; \
        cd vllm && bash tools/check_repo.sh; \
    fi

# Extract version from git BEFORE any modifications (pin_rocm_dependencies.py modifies requirements/rocm.txt)
# This ensures setuptools_scm sees clean repo state for version detection
RUN --mount=type=bind,source=.git,target=vllm/.git \
    cd vllm \
    && pip install setuptools_scm regex \
    && VLLM_VERSION=$(python3 -c "import setuptools_scm; print(setuptools_scm.get_version())") \
    && echo "Detected vLLM version: ${VLLM_VERSION}" \
    && echo "${VLLM_VERSION}" > /tmp/vllm_version.txt

# Fail if git-based package dependencies are found in requirements files
# (uv doesn't handle git+ URLs well, and packages should be distributed on PyPI)
# Extra notes: pip install is able to handle git+ URLs, but uv doesn't.
RUN echo "Checking for git-based packages in requirements files..." \
    && echo "Checking common.txt for git-based packages:" \
    && if grep -q 'git+' ${COMMON_WORKDIR}/vllm/requirements/common.txt; then \
         echo "ERROR: Git-based packages found in common.txt:"; \
         grep 'git+' ${COMMON_WORKDIR}/vllm/requirements/common.txt; \
         echo "Please publish these packages to PyPI instead of using git dependencies."; \
         exit 1; \
       else \
         echo "  ✓ No git-based packages found in common.txt"; \
       fi \
    && echo "Checking rocm.txt for git-based packages:" \
    && if grep -q 'git+' ${COMMON_WORKDIR}/vllm/requirements/rocm.txt; then \
         echo "ERROR: Git-based packages found in rocm.txt:"; \
         grep 'git+' ${COMMON_WORKDIR}/vllm/requirements/rocm.txt; \
         echo "Please publish these packages to PyPI instead of using git dependencies."; \
         exit 1; \
       else \
         echo "  ✓ No git-based packages found in rocm.txt"; \
       fi \
    && echo "All requirements files are clean - no git-based packages found"

# Pin vLLM dependencies to exact versions of custom ROCm wheels
# This ensures 'pip install vllm' automatically installs correct torch/triton/torchvision/amdsmi
COPY tools/vllm-rocm/pin_rocm_dependencies.py /tmp/pin_rocm_dependencies.py
RUN echo "Pinning vLLM dependencies to custom wheel versions..." \
    && python3 /tmp/pin_rocm_dependencies.py /install ${COMMON_WORKDIR}/vllm/requirements/rocm.txt

# Install dependencies using custom wheels from /install
RUN cd vllm \
    && echo "Building vLLM with custom wheels from /install" \
    && python3 -m pip install --find-links /install -r requirements/rocm.txt \
    && python3 setup.py clean --all

# Build wheel using pre-extracted version to avoid dirty state from modified requirements/rocm.txt
# (setup.py auto-detects sccache in PATH)
RUN --mount=type=bind,source=.git,target=vllm/.git \
    cd vllm \
    && export SETUPTOOLS_SCM_PRETEND_VERSION=$(cat /tmp/vllm_version.txt) \
    && echo "Building wheel with version: ${SETUPTOOLS_SCM_PRETEND_VERSION}" \
    && python3 setup.py bdist_wheel --dist-dir=dist

FROM scratch AS export_vllm_wheel_release
ARG COMMON_WORKDIR
COPY --from=build_vllm_wheel_release ${COMMON_WORKDIR}/vllm/dist/*.whl /
COPY --from=build_vllm_wheel_release ${COMMON_WORKDIR}/vllm/requirements /requirements
COPY --from=build_vllm_wheel_release ${COMMON_WORKDIR}/vllm/benchmarks /benchmarks
COPY --from=build_vllm_wheel_release ${COMMON_WORKDIR}/vllm/tests /tests
COPY --from=build_vllm_wheel_release ${COMMON_WORKDIR}/vllm/examples /examples
COPY --from=build_vllm_wheel_release ${COMMON_WORKDIR}/vllm/docker/Dockerfile.rocm /docker/
COPY --from=build_vllm_wheel_release ${COMMON_WORKDIR}/vllm/.buildkite /.buildkite
COPY --from=build_vllm_wheel_release ${COMMON_WORKDIR}/vllm/vllm/v1 /vllm_v1

# -----------------------
# Test vLLM image
FROM base AS test

RUN python3 -m pip install --upgrade pip && rm -rf /var/lib/apt/lists/*

# Install vLLM using uv (inherited from base stage)
# Note: No -U flag to avoid upgrading PyTorch ROCm to CUDA version
RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
    --mount=type=cache,target=/root/.cache/uv \
    cd /install \
    && uv pip install --system -r requirements/rocm.txt \
    && uv pip install --system -r requirements/rocm-test.txt \
    && pip uninstall -y vllm \
    && uv pip install --system *.whl

# Verify that PyTorch is the ROCm build, not CUDA
RUN python3 -c "import torch; assert torch.version.hip is not None, \
    f'Expected ROCm PyTorch but got CUDA (torch.version.cuda={torch.version.cuda}, torch.version.hip={torch.version.hip})'; \
    print(f'Verified: PyTorch {torch.__version__} with ROCm (HIP {torch.version.hip})')"

# Install RIXL wheel
RUN --mount=type=bind,from=build_rixl,src=/app/install,target=/rixl_install \
    uv pip install --system /rixl_install/*.whl

# Install DeepEP wheel
RUN --mount=type=bind,from=build_deep,src=/app/deep_install,target=/deep_install \
    uv pip install --system /deep_install/*.whl
COPY --from=build_deep /opt/rocshmem /opt/rocshmem

# RIXL/MoRIIO runtime dependencies (RDMA userspace libraries)
RUN apt-get update -q -y && apt-get install -q -y \
    librdmacm1 \
    libibverbs1 \
    ibverbs-providers \
    ibverbs-utils \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /vllm-workspace
ARG COMMON_WORKDIR
COPY --from=build_vllm ${COMMON_WORKDIR}/vllm /vllm-workspace

# install development dependencies (for testing)
RUN cd /vllm-workspace \
    && python3 -m pip install -e tests/vllm_test_utils \
    && python3 -m pip install pytest-shard

# enable fast downloads from hf (for testing)
RUN --mount=type=cache,target=/root/.cache/uv \
    uv pip install --system hf_transfer
ENV HF_HUB_ENABLE_HF_TRANSFER=1

# install audio decode package `torchcodec` from source (required due to 
# ROCm and torch version mismatch) for tests with datasets package
COPY tools/install_torchcodec_rocm.sh /tmp/install_torchcodec.sh
RUN bash /tmp/install_torchcodec.sh \
    && rm /tmp/install_torchcodec.sh \
    && apt-get clean \
    && rm -rf /var/lib/apt/lists/*

# Copy in the v1 package (for python-only install test group)
COPY --from=export_vllm /vllm_v1 /usr/local/lib/python${PYTHON_VERSION}/dist-packages/vllm/v1

# Set MIOPEN ENVS to resolve performance regressions in MIOpen 3D convolution kernel
# See: https://github.com/pytorch/pytorch/issues/169857
ENV MIOPEN_DEBUG_CONV_DIRECT=0
ENV MIOPEN_DEBUG_CONV_GEMM=0

# Source code is used in the `python_only_compile.sh` test
# We hide it inside `src/` so that this source code
# will not be imported by other tests
RUN mkdir src && mv vllm src/vllm

# This is a workaround to ensure pytest exits with the correct status code in CI tests.
RUN echo "import os\n\ndef pytest_sessionfinish(session, exitstatus):\n    os._exit(int(exitstatus))" > /vllm-workspace/conftest.py

# -----------------------
# Final vLLM image
FROM base AS final

RUN python3 -m pip install --upgrade pip && rm -rf /var/lib/apt/lists/*

# Clean up sccache from release image (not needed at runtime)
# This removes the binary and wrappers that may have been installed during build
RUN rm -f /usr/bin/sccache || true \
    && rm -rf /opt/sccache-wrappers || true

# Unset sccache environment variables for the release image
# This prevents S3 bucket config from leaking into production images
ENV SCCACHE_BUCKET=
ENV SCCACHE_REGION=
ENV SCCACHE_S3_NO_CREDENTIALS=
ENV SCCACHE_IDLE_TIMEOUT=

# Error related to odd state for numpy 1.20.3 where there is no METADATA etc, but an extra LICENSES_bundled.txt.
# Manually remove it so that later steps of numpy upgrade can continue
RUN case "$(which python3)" in \
        *"/opt/conda/envs/py_3.9"*) \
            rm -rf /opt/conda/envs/py_3.9/lib/python3.9/site-packages/numpy-1.20.3.dist-info/;; \
        *) ;; esac

RUN --mount=type=cache,target=/root/.cache/uv \
    uv pip install --system --upgrade huggingface-hub[cli]

# Install vLLM using uv (inherited from base stage)
# Note: No -U flag to avoid upgrading PyTorch ROCm to CUDA version
RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
    --mount=type=cache,target=/root/.cache/uv \
    cd /install \
    && uv pip install --system -r requirements/rocm.txt \
    && pip uninstall -y vllm \
    && uv pip install --system *.whl

ARG COMMON_WORKDIR
ARG BASE_IMAGE

# Copy over the benchmark scripts as well
COPY --from=export_vllm /benchmarks ${COMMON_WORKDIR}/vllm/benchmarks
COPY --from=export_vllm /examples ${COMMON_WORKDIR}/vllm/examples
COPY --from=export_vllm /docker ${COMMON_WORKDIR}/vllm/docker

ENV TOKENIZERS_PARALLELISM=false

# ENV that can improve safe tensor loading, and end-to-end time
ENV SAFETENSORS_FAST_GPU=1

# Performance environment variable.
ENV HIP_FORCE_DEV_KERNARG=1

# Workaround for ROCm profiler limits
RUN echo "ROCTRACER_MAX_EVENTS=10000000" > ${COMMON_WORKDIR}/libkineto.conf
ENV KINETO_CONFIG="${COMMON_WORKDIR}/libkineto.conf"
RUN echo "VLLM_BASE_IMAGE=${BASE_IMAGE}" >> ${COMMON_WORKDIR}/versions.txt

CMD ["/bin/bash"]

#Set entrypoint for vllm-openai official images
FROM final AS vllm-openai
ENTRYPOINT ["vllm", "serve"]