[ROCm][AMD][Installation] add LMCache kv-connector installation and runtime packages to docker image (#51208)

Signed-off-by: Hongxia Yang <hongxia.yang@amd.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
Hongxia Yang
2026-08-17 16:58:48 -05:00
committed by GitHub
co-authored by Andreas Karatzas
parent 8878ebd8fd
commit e68fb75b25
3 changed files with 88 additions and 2 deletions
+86 -1
View File
@@ -6,6 +6,7 @@ ARG COMMON_WORKDIR=/app
ARG BASE_IMAGE=rocm/vllm-dev:base
ARG CI_BASE_IMAGE=rocm/vllm-dev:ci_base
ARG ROCM_TRITON_KERNELS_COMMIT=0f380657dbf3ee86eb57558ff71df24f03b5d4e7
ARG INSTALL_LMCACHE=false
# NIC backend for MoRI RDMA support.
# By default (all), drivers and userspace libraries for all supported NIC types
# (ainic and bnxt) are installed; MoRI selects the appropriate one at runtime.
@@ -471,6 +472,71 @@ RUN cd /opt/nixl && \
/tmp/nixl_wheels/repaired/*.whl && \
cp /tmp/nixl_wheels/repaired/*.whl /app/install
# Fetch and verify pinned LMCache source.
FROM base AS lmcache_source
ARG LMCACHE_REPO="https://github.com/LMCache/LMCache.git"
ARG LMCACHE_REF="140819c9d57a975dbc5678a6459a218e544cb58b"
ARG LMCACHE_VERSION="0.5.3"
# Remove NVIDIA-only dependencies and accept the release image's NumPy version.
RUN printf '%s\n' "${LMCACHE_REF}" | grep -Eq '^[0-9a-f]{40}$' \
&& git clone --depth=1 --branch "v${LMCACHE_VERSION}" \
"${LMCACHE_REPO}" /app/lmcache \
&& test "$(git -C /app/lmcache rev-parse HEAD)" = "${LMCACHE_REF}" \
&& test "$(grep -Ec '^(cufile-python|nvtx)([<=>[:space:]]|$)' \
/app/lmcache/requirements/common.txt)" -eq 2 \
&& test "$(grep -Fxc 'numpy<=2.2.6' \
/app/lmcache/requirements/common.txt)" -eq 1 \
&& sed -i -E \
-e '/^(cufile-python|nvtx)([<=>[:space:]]|$)/d' \
-e 's/^numpy<=2\.2\.6$/numpy<=2.3.5/' \
/app/lmcache/requirements/common.txt \
&& rm -rf /app/lmcache/.git
FROM base AS build_lmcache
COPY --from=lmcache_source \
/app/lmcache/requirements/build.txt /tmp/lmcache-build-requirements.txt
RUN --mount=type=cache,id=vllm-rocm-uv,target=/root/.cache/uv \
uv pip install --system -r /tmp/lmcache-build-requirements.txt
ARG LMCACHE_VERSION="0.5.3"
ARG LMCACHE_ROCM_ARCH="gfx942;gfx950"
ARG USE_SCCACHE
COPY --from=lmcache_source /app/lmcache /app/lmcache
WORKDIR /app/lmcache
RUN --mount=type=cache,id=vllm-rocm-ccache,target=/root/.cache/ccache \
set -eu; \
printf '%s\n' "${LMCACHE_ROCM_ARCH}" \
| grep -Eq '^gfx[0-9a-z]+(;gfx[0-9a-z]+)*$'; \
offload_flags="$(printf '%s' "${LMCACHE_ROCM_ARCH}" \
| sed 's/;/ --offload-arch=/g; s/^/--offload-arch=/')"; \
if [ "${USE_SCCACHE:-0}" = "1" ]; then \
export HIP_CLANG_PATH=/opt/sccache-wrappers; \
fi; \
mkdir -p /app/install; \
SETUPTOOLS_SCM_PRETEND_VERSION="${LMCACHE_VERSION}" \
PYTORCH_ROCM_ARCH="${LMCACHE_ROCM_ARCH}" \
HIPCC_COMPILE_FLAGS_APPEND="${offload_flags}" \
BUILD_WITH_HIP=1 CXX=hipcc \
python3 setup.py bdist_wheel --dist-dir=/app/install; \
test "$(find /app/install -maxdepth 1 -name '*.whl' | wc -l)" -eq 1
RUN --mount=type=tmpfs,target=/tmp/lmcache-wheel \
set -eu; \
python3 -m zipfile -e /app/install/*.whl /tmp/lmcache-wheel; \
c_ops="$(find /tmp/lmcache-wheel -type f -name 'c_ops*.so')"; \
test "$(find /tmp/lmcache-wheel -type f -name 'c_ops*.so' | wc -l)" -eq 1; \
expected="$(printf '%s' "${LMCACHE_ROCM_ARCH}" | tr ';' '\n' | sort -u)"; \
actual="$(/opt/rocm/llvm/bin/llvm-objdump --offloading "${c_ops}" \
| grep -oE 'gfx[0-9a-z]+' | sort -u)"; \
if [ "${actual}" != "${expected}" ]; then \
printf 'LMCache code objects mismatch\nexpected:\n%s\nactual:\n%s\n' \
"${expected}" "${actual}" >&2; \
exit 1; \
fi
# ROCShmem build stage - split from DeepEP so changing DEEPEP_BRANCH does not
# invalidate the slow ROCShmem build.
FROM base AS build_rocshmem
@@ -797,6 +863,13 @@ RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system /tmp/vllm_test_utils \
&& rm -rf /tmp/vllm_test_utils
RUN --mount=type=bind,from=build_lmcache,src=/app/install,target=/lmcache_install \
--mount=type=bind,source=requirements/test/rocm.txt,target=/tmp/lmcache-constraints.txt \
--mount=type=cache,target=/root/.cache/uv \
uv pip install --system --constraint /tmp/lmcache-constraints.txt \
/lmcache_install/*.whl \
&& python3 -c "import lmcache, lmcache.c_ops; print('lmcache', lmcache.__version__)"
# -----------------------
# Test vLLM image (Tier 2) - vLLM-only layer on top of ci_base.
FROM ${CI_BASE_IMAGE} AS test
@@ -825,7 +898,7 @@ ENV VLLM_GPU_SYNC_CHECK=error
# -----------------------
# Final vLLM image
FROM mori_base AS final
FROM mori_base AS final_common
RUN python3 -m pip install --upgrade pip && rm -rf /var/lib/apt/lists/*
@@ -901,6 +974,18 @@ RUN echo "VLLM_BASE_IMAGE=${BASE_IMAGE}" >> ${COMMON_WORKDIR}/versions.txt \
CMD ["/bin/bash"]
FROM final_common AS final_lmcache_false
FROM final_common AS final_lmcache_true
RUN --mount=type=bind,from=build_lmcache,src=/app/install,target=/lmcache_install \
--mount=type=cache,target=/root/.cache/uv \
uv pip install --system /lmcache_install/*.whl \
&& test "$(python3 -c 'import numpy; print(numpy.__version__)')" = "2.3.5" \
&& uv pip check --system \
&& python3 -c "import lmcache, lmcache.c_ops; print('lmcache', lmcache.__version__)"
FROM final_lmcache_${INSTALL_LMCACHE} AS final
#Set entrypoint for vllm-openai official images
FROM final AS vllm-openai
ENTRYPOINT ["vllm", "serve"]