wehub-resource-sync
152 lines
5.5 KiB
Docker
152 lines
5.5 KiB
Docker
# The LMCache Standalone Dockerfile is used to build a LMCache image
|
|
# without vLLM integration.
|
|
|
|
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}
|
|
# Override to a stable version for releases.
|
|
ARG VLLM_VERSION=nightly
|
|
|
|
#################### 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 tzdata \
|
|
&& rm -rf /var/lib/apt/lists/* \
|
|
&& ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/ \
|
|
&& curl -LsSf https://astral.sh/uv/install.sh | sh \
|
|
&& mv ~/.local/bin/uv /usr/local/bin/ \
|
|
&& mv ~/.local/bin/uvx /usr/local/bin/ \
|
|
&& uv venv /opt/venv \
|
|
&& . /opt/venv/bin/activate \
|
|
&& python3 --version
|
|
|
|
WORKDIR /workspace
|
|
|
|
# Install small non-torch runtime dependencies; torch is installed in lmcache-build
|
|
# against the exact CUDA version so the compiled C extensions match at runtime.
|
|
# The CLI requirements (e.g. ``openai`` for the bench client) are also installed
|
|
# here so ``lmcache`` subcommand discovery in ``lmcache.cli.commands`` -- which
|
|
# eagerly imports every command module -- does not crash with
|
|
# ``ModuleNotFoundError: No module named 'openai'`` when the standalone image
|
|
# boots ``lmcache server``. See LMCache#3353.
|
|
RUN --mount=type=cache,target=/root/.cache/uv \
|
|
--mount=type=bind,source=requirements/cli.txt,target=/tmp/cli.txt \
|
|
. /opt/venv/bin/activate && \
|
|
uv pip install ray nvidia-ml-py -r /tmp/cli.txt
|
|
|
|
# CUDA arch list used by torch
|
|
ARG torch_cuda_arch_list='7.5 8.0 8.6 8.9 9.0 10.0 12.0+PTX'
|
|
ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
|
|
|
|
#################### TORCH RESOLVER ##########################
|
|
# Resolve the torch version vLLM uses for this CUDA tag into /torch.pin.
|
|
# vLLM is not installed; only the pin is consumed by downstream stages.
|
|
|
|
FROM base AS torch-resolver
|
|
|
|
ARG CUDA_VERSION
|
|
ARG VLLM_VERSION
|
|
|
|
RUN --mount=type=cache,target=/root/.cache/uv,id=uv-resolver \
|
|
. /opt/venv/bin/activate && \
|
|
CUDA_TAG=cu$(echo ${CUDA_VERSION} | tr -d '.') && \
|
|
if [ "$VLLM_VERSION" = "nightly" ]; then \
|
|
echo "vllm" > /tmp/req.in && \
|
|
uv pip compile /tmp/req.in --quiet --prerelease=allow \
|
|
--extra-index-url https://wheels.vllm.ai/nightly/${CUDA_TAG} \
|
|
--extra-index-url https://download.pytorch.org/whl/${CUDA_TAG} \
|
|
--index-strategy unsafe-best-match \
|
|
> /tmp/resolved.txt ; \
|
|
else \
|
|
echo "vllm==${VLLM_VERSION}" > /tmp/req.in && \
|
|
uv pip compile /tmp/req.in --quiet --prerelease=allow \
|
|
--extra-index-url https://download.pytorch.org/whl/${CUDA_TAG} \
|
|
> /tmp/resolved.txt ; \
|
|
fi && \
|
|
grep -E '^torch==' /tmp/resolved.txt > /torch.pin && \
|
|
echo "Resolved torch pin (matches vllm ${VLLM_VERSION} on ${CUDA_TAG}):" && \
|
|
cat /torch.pin
|
|
|
|
#################### LMCache Build ##########################
|
|
# Build LMCache wheel
|
|
|
|
FROM base AS lmcache-build
|
|
|
|
ARG CUDA_VERSION
|
|
|
|
# install build dependencies
|
|
COPY ./requirements/build.txt build.txt
|
|
COPY --from=torch-resolver /torch.pin /torch.pin
|
|
|
|
# Max jobs used by Ninja to build extensions
|
|
ARG max_jobs=2
|
|
ENV MAX_JOBS=${max_jobs}
|
|
|
|
# Number of threads used by nvcc
|
|
ARG nvcc_threads=8
|
|
ENV NVCC_THREADS=$nvcc_threads
|
|
|
|
# --index-url (singular) keeps PyPI out of torch resolution so uv can't
|
|
# pick PyPI's default cu13 wheel on a cu12.9 host.
|
|
RUN --mount=type=cache,target=/root/.cache/uv \
|
|
. /opt/venv/bin/activate && \
|
|
CUDA_TAG=cu$(echo ${CUDA_VERSION} | tr -d '.') && \
|
|
uv pip install -r /torch.pin \
|
|
--index-url https://download.pytorch.org/whl/${CUDA_TAG} && \
|
|
uv pip install -r build.txt
|
|
|
|
ARG LMCACHE_COMMIT_ID=1
|
|
|
|
COPY . /workspace/LMCache
|
|
WORKDIR /workspace/LMCache
|
|
|
|
# Build LMCache wheel (don't install yet)
|
|
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 && \
|
|
export LMCACHE_CUDA_MAJOR=$(echo ${CUDA_VERSION} | cut -d. -f1) && \
|
|
python3 -c 'import torch; print("TORCH=", torch.__version__)' && \
|
|
python3 setup.py bdist_wheel --dist-dir=dist_lmcache
|
|
|
|
#################### LMCache Final Image ##########################
|
|
# Clean install of LMCache from wheel
|
|
|
|
FROM base AS lmcache-final
|
|
|
|
ARG CUDA_VERSION
|
|
|
|
# Copy the built wheel from build stage
|
|
COPY --from=lmcache-build /workspace/LMCache/dist_lmcache/*.whl /tmp/
|
|
COPY --from=torch-resolver /torch.pin /torch.pin
|
|
|
|
# Install torch from the CUDA index first; the wheel's torch dep is then
|
|
# already satisfied, so uv won't re-resolve it against PyPI default cu13.
|
|
RUN --mount=type=cache,target=/root/.cache/uv \
|
|
. /opt/venv/bin/activate && \
|
|
CUDA_TAG=cu$(echo ${CUDA_VERSION} | tr -d '.') && \
|
|
uv pip install -r /torch.pin \
|
|
--index-url https://download.pytorch.org/whl/${CUDA_TAG} && \
|
|
uv pip install /tmp/*.whl --verbose && \
|
|
rm -rf /tmp/*.whl /torch.pin
|
|
|
|
WORKDIR /workspace
|
|
|
|
# Default shell entrypoint (no vLLM server)
|
|
CMD ["/bin/bash"]
|
|
|