forked from Karylab-cklius/vllm
d7045619c1
Signed-off-by: Zhewen Li <zhewenli@inferact.ai> Co-authored-by: Zhewen Li <zhewenli@inferact.ai>
68 lines
3.7 KiB
Docker
68 lines
3.7 KiB
Docker
ARG BASE_IMAGE=inferactinc/public:crusoe-kimi-pd-20260415
|
|
FROM ${BASE_IMAGE}
|
|
|
|
# Rebuild mooncake-transfer-engine from zhewenl/Mooncake-rdma (compat/ubuntu-22.04)
|
|
# on top of an already-built image whose mooncake wheel has the dmabuf
|
|
# sub-allocation-address bug (upstream kvcache-ai/Mooncake v0.3.10.post*).
|
|
#
|
|
# The fork is based on ivanium/Mooncake@yifan/dev, plus a LOG_EVERY_T→LOG shim
|
|
# for glog 0.4 (Ubuntu 22.04 ships glog 0.4; the fork requires 0.6). It fixes
|
|
# ibv_reg_dmabuf_mr to resolve the true CUDA allocation base via
|
|
# cuMemGetAddressRange, so vLLM KV caches (sub-views of a larger PyTorch
|
|
# allocation) register successfully instead of returning EFAULT.
|
|
#
|
|
# WITH_NVIDIA_PEERMEM=OFF enables the dmabuf code path (nvidia-peermem is not
|
|
# loadable on GB200, and dmabuf works on H200/H100 too).
|
|
#
|
|
# Usage:
|
|
# docker build -f docker/Dockerfile.mooncake-layer \
|
|
# --build-arg BASE_IMAGE=inferactinc/dev:svf-x86_64-cu130-b641ede \
|
|
# -t inferactinc/dev:svf-x86_64-cu130-b641ede-mcfix .
|
|
RUN set -e && \
|
|
PY=$(python3 -c "import sys; print(f'{sys.version_info.major}.{sys.version_info.minor}')") && \
|
|
apt-get update -y && \
|
|
apt-get install -y --no-install-recommends \
|
|
build-essential cmake ninja-build git wget unzip \
|
|
libibverbs-dev libgoogle-glog-dev libgtest-dev libjsoncpp-dev \
|
|
libunwind-dev libnuma-dev libpython3-dev libboost-all-dev libssl-dev \
|
|
libgrpc-dev libgrpc++-dev libprotobuf-dev libyaml-cpp-dev protobuf-compiler-grpc \
|
|
libcurl4-openssl-dev libhiredis-dev liburing-dev libjemalloc-dev \
|
|
libmsgpack-dev libzstd-dev libasio-dev libxxhash-dev \
|
|
pkg-config patchelf libc6-dev libc-bin && \
|
|
git clone --branch compat/ubuntu-22.04 --depth 1 --recurse-submodules \
|
|
https://github.com/zhewenl/Mooncake-rdma.git /tmp/Mooncake && \
|
|
cd /tmp/Mooncake/extern/yalantinglibs && mkdir -p build && cd build && \
|
|
cmake .. -DBUILD_EXAMPLES=OFF -DBUILD_BENCHMARK=OFF -DBUILD_UNIT_TESTS=OFF && \
|
|
cmake --build . -j$(nproc) && cmake --install . && \
|
|
cd /tmp/Mooncake && \
|
|
export LIBRARY_PATH=/usr/local/cuda/lib64/stubs:${LIBRARY_PATH} && \
|
|
mkdir -p build && cd build && \
|
|
cmake .. -DUSE_CUDA=ON -DWITH_NVIDIA_PEERMEM=OFF -DUSE_MNNVL=ON \
|
|
-DPython3_EXECUTABLE=$(which python3) \
|
|
-DPYTHON_EXECUTABLE=$(which python3) && \
|
|
make -j$(nproc) && make install && \
|
|
strings /usr/local/lib/python${PY}/dist-packages/mooncake/engine*.so \
|
|
| grep -q "cuMemGetAddressRange" || \
|
|
{ echo "ERROR: mooncake dmabuf fix missing in built binary"; exit 1; } && \
|
|
echo "VERIFIED: dmabuf sub-address fix compiled" && \
|
|
cd / && rm -rf /tmp/Mooncake && \
|
|
apt-get purge -y \
|
|
cmake ninja-build git wget unzip \
|
|
libgoogle-glog-dev libgtest-dev libjsoncpp-dev \
|
|
libunwind-dev libpython3-dev libboost-all-dev libssl-dev \
|
|
libgrpc-dev libgrpc++-dev libprotobuf-dev libyaml-cpp-dev protobuf-compiler-grpc \
|
|
libcurl4-openssl-dev libhiredis-dev liburing-dev libjemalloc-dev \
|
|
libmsgpack-dev libzstd-dev libasio-dev libxxhash-dev \
|
|
pkg-config patchelf && \
|
|
apt-get install -y --no-install-recommends \
|
|
libgoogle-glog0v5 libjsoncpp25 liburing2 libgflags2.2 \
|
|
libjemalloc2 libhiredis0.14 libcurl4 && \
|
|
apt-get autoremove -y && rm -rf /var/lib/apt/lists/* && \
|
|
if [ ! -f /usr/local/cuda/lib64/libcudart.so ] && [ -f /usr/local/cuda/lib64/libcudart.so.13 ]; then \
|
|
ln -s libcudart.so.13 /usr/local/cuda/lib64/libcudart.so; \
|
|
elif [ ! -f /usr/local/cuda/lib64/libcudart.so ] && [ -f /usr/local/cuda/lib64/libcudart.so.12 ]; then \
|
|
ln -s libcudart.so.12 /usr/local/cuda/lib64/libcudart.so; \
|
|
fi
|
|
|
|
ENTRYPOINT ["vllm", "serve"]
|