wehub-resource-sync
156 lines
5.8 KiB
Docker
156 lines
5.8 KiB
Docker
# The LMCache Dockerfile is used to build a LMCache image that is integrated
|
|
# to run with vLLM OpenAI server.
|
|
|
|
# Please update any changes made here to
|
|
# docs/source/developer_guide/docker_file.rst
|
|
# docs/source/getting_started/installation.rst
|
|
# docs/source/production/docker_deployment.rst
|
|
|
|
ARG CUDA_VERSION=13.0
|
|
ARG UBUNTU_VERSION=24.04
|
|
ARG NGC_VERSION=25.09
|
|
ARG BASE_IMAGE=nvcr.io/nvidia/cuda-dl-base:${NGC_VERSION}-cuda${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
|
|
|
|
#################### BASE BUILD IMAGE ####################
|
|
# Prepare basic build environment
|
|
|
|
FROM ${BASE_IMAGE} AS base
|
|
|
|
ARG CUDA_VERSION
|
|
ARG PYTHON_VERSION=3.12
|
|
ARG UBUNTU_VERSION
|
|
ENV DEBIAN_FRONTEND=noninteractive
|
|
ENV PATH="/opt/venv/bin:${PATH}"
|
|
|
|
# Install Python and other dependencies
|
|
RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
|
|
&& echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections \
|
|
&& apt-get update -y \
|
|
&& apt-get install -y --no-install-recommends \
|
|
ccache software-properties-common git curl sudo \
|
|
python3 python3-dev python3-venv python3-pip tzdata libxcb1-dev \
|
|
&& ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/ \
|
|
&& curl -LsSf https://astral.sh/uv/install.sh | sh \
|
|
&& mv ~/.local/bin/uv /usr/local/bin/ \
|
|
&& mv ~/.local/bin/uvx /usr/local/bin/ \
|
|
&& uv venv /opt/venv \
|
|
&& . /opt/venv/bin/activate \
|
|
&& python3 --version
|
|
|
|
WORKDIR /workspace
|
|
|
|
# CUDA arch list used by torch
|
|
ARG torch_cuda_arch_list='7.5 8.0 8.6 8.9 9.0 10.0 12.0+PTX'
|
|
ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
|
|
|
|
#################### vLLM IMAGE & LMCache (Build) ##########################
|
|
# Integrate vLLM nightly build and LMCache build, and expose vLLM OpenAI API
|
|
|
|
FROM base AS image-build
|
|
|
|
# install build dependencies
|
|
COPY ./requirements/build.txt build.txt
|
|
|
|
# Max jobs used by Ninja to build extensions
|
|
ARG max_jobs=2
|
|
ENV MAX_JOBS=${max_jobs}
|
|
|
|
# Number of threads used by nvcc
|
|
ARG nvcc_threads=8
|
|
ENV NVCC_THREADS=$nvcc_threads
|
|
|
|
ARG CUDA_VERSION
|
|
ARG VLLM_VERSION=nightly
|
|
|
|
RUN --mount=type=cache,target=/root/.cache/pip \
|
|
. /opt/venv/bin/activate && \
|
|
uv pip install -r build.txt
|
|
|
|
ARG LMCACHE_COMMIT_ID=1
|
|
|
|
COPY . /workspace/LMCache
|
|
WORKDIR /workspace/LMCache
|
|
|
|
|
|
RUN --mount=type=cache,target=/root/.cache/ccache,id=ccache \
|
|
--mount=type=cache,target=/root/.cache/uv,id=uv-cache,sharing=locked \
|
|
. /opt/venv/bin/activate && \
|
|
CUDA_TAG=cu$(echo ${CUDA_VERSION} | tr -d '.') && \
|
|
export LMCACHE_CUDA_MAJOR=$(echo ${CUDA_VERSION} | cut -d. -f1) && \
|
|
if [ "$VLLM_VERSION" = "nightly" ]; then \
|
|
VLLM_PRECOMPILED_WHEEL_VARIANT=${CUDA_TAG} uv pip install --prerelease=allow \
|
|
'vllm[runai,tensorizer,flashinfer]' \
|
|
--extra-index-url https://wheels.vllm.ai/nightly/${CUDA_TAG} \
|
|
--extra-index-url https://download.pytorch.org/whl/${CUDA_TAG} \
|
|
--index-strategy unsafe-first-match ; \
|
|
else \
|
|
VLLM_PRECOMPILED_WHEEL_VARIANT=${CUDA_TAG} uv pip install --prerelease=allow \
|
|
"vllm[runai,tensorizer,flashinfer]==${VLLM_VERSION}" ; \
|
|
fi && \
|
|
python3 -c 'import torch; print("TORCH=", torch.__version__)' && \
|
|
python3 setup.py bdist_wheel --dist-dir=dist_lmcache && \
|
|
uv pip install ./dist_lmcache/*.whl --verbose && \
|
|
|
|
uv pip install --reinstall-package "nixl-cu${LMCACHE_CUDA_MAJOR}" \
|
|
"nixl-cu${LMCACHE_CUDA_MAJOR}"
|
|
|
|
WORKDIR /workspace
|
|
ENTRYPOINT ["/opt/venv/bin/vllm", "serve"]
|
|
|
|
#################### vLLM IMAGE & LMCache (Release, cu13) #######################
|
|
# Integrate vLLM and LMCache stable releases, and expose vLLM OpenAI API.
|
|
# The default lmcache wheel on PyPI is built against cu13; the cu13 torch
|
|
# index is hinted explicitly so vLLM also resolves to its cu13 build.
|
|
|
|
FROM base AS image-release
|
|
|
|
ARG CUDA_VERSION
|
|
|
|
ARG LMCACHE_VERSION
|
|
|
|
RUN . /opt/venv/bin/activate && \
|
|
CUDA_TAG=cu$(echo ${CUDA_VERSION} | tr -d '.') && \
|
|
CUDA_MAJOR=$(echo ${CUDA_VERSION} | cut -d. -f1) && \
|
|
VER=$(echo "${LMCACHE_VERSION}" | sed 's/^v//') && \
|
|
VLLM_PRECOMPILED_WHEEL_VARIANT=${CUDA_TAG} uv pip install --prerelease=allow \
|
|
vllm[runai,tensorizer,flashinfer] \
|
|
--extra-index-url https://download.pytorch.org/whl/${CUDA_TAG} \
|
|
--index-strategy unsafe-best-match && \
|
|
uv pip install "lmcache${VER:+==${VER}}" \
|
|
--extra-index-url https://download.pytorch.org/whl/${CUDA_TAG} \
|
|
--index-strategy unsafe-best-match --verbose && \
|
|
|
|
uv pip install --reinstall-package "nixl-cu${CUDA_MAJOR}" "nixl-cu${CUDA_MAJOR}"
|
|
|
|
WORKDIR /workspace
|
|
ENTRYPOINT ["/opt/venv/bin/vllm", "serve"]
|
|
|
|
#################### vLLM IMAGE & LMCache (Release, cu129) ######################
|
|
# Installs the stable cu129 vLLM release wheel and
|
|
# lmcache from the v{tag}-cu129 GitHub Release.
|
|
# The cu129 torch index is hinted explicitly so vLLM resolves to its cu129 build.
|
|
|
|
FROM base AS image-release-cu129
|
|
|
|
ARG CUDA_VERSION=12.9
|
|
ARG LMCACHE_VERSION
|
|
|
|
RUN . /opt/venv/bin/activate && \
|
|
CUDA_TAG=cu$(echo ${CUDA_VERSION} | tr -d '.') && \
|
|
CUDA_MAJOR=$(echo ${CUDA_VERSION} | cut -d. -f1) && \
|
|
VER=$(echo ${LMCACHE_VERSION} | sed 's/^v//') && \
|
|
VLLM_PRECOMPILED_WHEEL_VARIANT=${CUDA_TAG} uv pip install --prerelease=allow \
|
|
vllm[runai,tensorizer,flashinfer] \
|
|
--extra-index-url https://download.pytorch.org/whl/${CUDA_TAG} \
|
|
--index-strategy unsafe-best-match && \
|
|
python3 -c 'import torch; print("TORCH=", torch.__version__)' && \
|
|
uv pip install lmcache==${VER} \
|
|
--extra-index-url https://download.pytorch.org/whl/${CUDA_TAG} \
|
|
--find-links https://github.com/LMCache/LMCache/releases/expanded_assets/v${VER}-cu129 \
|
|
--index-strategy unsafe-best-match --verbose && \
|
|
|
|
uv pip install --reinstall-package "nixl-cu${CUDA_MAJOR}" "nixl-cu${CUDA_MAJOR}"
|
|
|
|
WORKDIR /workspace
|
|
ENTRYPOINT ["/opt/venv/bin/vllm", "serve"]
|