diff --git a/docker/build_cpu_image.sh b/docker/build_cpu_image.sh
new file mode 100755
index 000000000..53d72fdf1
--- /dev/null
+++ b/docker/build_cpu_image.sh
@@ -0,0 +1,25 @@
+# Current directory should be the root of the repository
+
+# You may need to pass 
+#   - VLLM_VERSION
+#   - MACA_VERSION
+# to build a specific version
+
+RHSM_SECRET_FILE=${RHSM_SECRET_FILE:-rhsm.env}
+if [ ! -f "$RHSM_SECRET_FILE" ] && [ -f docker/rhsm.env ]; then
+    RHSM_SECRET_FILE=docker/rhsm.env
+fi
+
+DOCKER_BUILDKIT=1 docker build \
+    --network host \
+    --secret id=rhsm,src=${RHSM_SECRET_FILE} \
+    -f docker/vllm-cpu.Dockerfile \
+    -t vllm_cpu:new_base_img \
+    --build-arg VLLM_VERSION=v0.18.0 \
+     .
+
+# debug dockerfile and run into shell with buildx:
+# ddocker () {
+#     BUILDX_EXPERIMENTAL=1 docker buildx debug --invoke /bin/bash --on=error $@
+# }
+
diff --git a/docker/rhsm.env b/docker/rhsm.env
new file mode 100644
index 000000000..318e2f633
--- /dev/null
+++ b/docker/rhsm.env
@@ -0,0 +1,2 @@
+RHSM_USER=''
+RHSM_PASS=''
diff --git a/docker/vllm-cpu.Dockerfile b/docker/vllm-cpu.Dockerfile
new file mode 100644
index 000000000..0dcd502b5
--- /dev/null
+++ b/docker/vllm-cpu.Dockerfile
@@ -0,0 +1,140 @@
+ARG BUILD_BASE_IMAGE=registry.redhat.io/rhai/base-image-cpu-rhel9:3.4.0-1777399554
+ARG PYTHON_VERSION=3.12
+# ARG UV_EXTRA_INDEX_URL=https://console.redhat.com/api/pypi/public-rhai/rhoai/3.4/cpu-ubi9/simple
+ARG UV_INDEX_URL=https://console.redhat.com/api/pypi/public-rhai/rhoai/3.4/cpu-ubi9/simple
+# may need passing a particular vllm version during build
+ARG VLLM_VERSION
+
+ARG RHSM_USER
+ARG RHSM_PASS
+
+#################### BASE IMAGE ####################
+FROM ${BUILD_BASE_IMAGE} AS base
+
+
+# # TODO: we most likely do not need all of the dnf installs below: they're already in the base image
+# RUN --mount=type=secret,id=rhsm \
+#     --mount=type=cache,target=/var/cache/dnf,sharing=locked \
+#     bash -euo pipefail -c '\
+#       source /run/secrets/rhsm; \
+#       cleanup() { \
+#         subscription-manager unregister >/dev/null 2>&1 || true; \
+#         subscription-manager clean >/dev/null 2>&1 || true; \
+#       }; \
+#       trap cleanup EXIT; \
+#       subscription-manager register \
+#         --username "$RHSM_USER" \
+#         --password "$RHSM_PASS" \
+#         --auto-attach; \
+#       subscription-manager repos \
+#         --enable codeready-builder-for-rhel-9-x86_64-rpms; \
+#       yum install -y https://dl.fedoraproject.org/pub/epel/epel-release-latest-9.noarch.rpm; \
+#       /usr/bin/crb enable; \
+#       yum makecache; \
+#       yum install -y --setopt=install_weak_deps=False \
+#         python3 \
+#         python3-devel \
+#         zeromq \
+#         bzip2 \
+#         cpio \
+#         elfutils-debuginfod-client \
+#         ffmpeg-free \
+#         fftw \
+#         file \
+#         freetype \
+#         gcc \
+#         gcc-c++ \
+#         gdal-libs \
+#         gdb \
+#         geos \
+#         git-core \
+#         glibc-langpack-en \
+#         glog \
+#         gmp \
+#         gzip \
+#         hdf5 \
+#         jemalloc \
+#         jq \
+#         krb5-libs \
+#         lcms2 \
+#         libaio \
+#         libev \
+#         libjpeg \
+#         libmpc \
+#         libomp \
+#         libpng \
+#         libpq \
+#         libqhull_r \
+#         libsndfile \
+#         libtiff \
+#         libunwind \
+#         libva \
+#         libwebp \
+#         libxml2 \
+#         libxslt \
+#         libzip \
+#         libzstd \
+#         loguru \
+#         lz4 \
+#         make \
+#         mariadb-connector-c \
+#         mpfr \
+#         netcdf \
+#         numactl \
+#         nvtop \
+#         openblas openblas-openmp openblas-openmp64 openblas-serial openblas-serial64 openblas-threads openblas-threads64 \
+#         openjpeg2 \
+#         openmpi \
+#         proj \
+#         protobuf \
+#         qpdf \
+#         re2 \
+#         snappy \
+#         spatialindex \
+#         tbb \
+#         tesseract \
+#         thrift \
+#         unixODBC \
+#         utf8proc \
+#         wget \
+#         xz \
+#         xz-libs \
+#         zlib \
+#         zstd; \
+#       yum clean all'
+
+WORKDIR /workspace
+
+RUN curl -LsSf https://astral.sh/uv/install.sh | sh
+
+ENV CC=/usr/bin/gcc CXX=/usr/bin/g++
+
+ENV PATH="/root/.local/bin:$PATH"
+ENV VIRTUAL_ENV="/opt/app-root/"
+ENV UV_PYTHON_INSTALL_DIR=/opt/app-root/python
+ARG PYTHON_VERSION
+# RUN uv venv  ${VIRTUAL_ENV} --python ${PYTHON_VERSION} --seed
+ENV PATH="$VIRTUAL_ENV/bin:$PATH"
+
+ENV UV_HTTP_TIMEOUT=500
+
+ENV LD_PRELOAD="/opt/app-root/lib/libiomp5.so"
+
+RUN echo 'ulimit -c 0' >> ~/.bashrc
+
+###################### BUILD IMAGE ####################
+FROM base AS vllm-cpu-build
+
+
+# Install Python dependencies
+ARG UV_EXTRA_INDEX_URL
+ARG UV_INDEX_URL
+ENV UV_EXTRA_INDEX_URL=${UV_EXTRA_INDEX_URL}
+ENV UV_INDEX_URL=${UV_INDEX_URL}
+ENV UV_INDEX_STRATEGY="unsafe-best-match"
+ENV UV_LINK_MODE="copy"
+
+ENV VLLM_LOGGING_LEVEL=DEBUG
+
+ARG VLLM_VERSION
+RUN uv pip install --torch-backend=cpu vllm==${VLLM_VERSION}
diff --git a/docker/vllm_ubi9_building_instructions.md b/docker/vllm_ubi9_building_instructions.md
deleted file mode 100644
index 670a07b26..000000000
--- a/docker/vllm_ubi9_building_instructions.md
+++ /dev/null
@@ -1,284 +0,0 @@
-Here we demonstrate how to build vLLM manually starting from a ubi9 docker image.
-
-# 0. start a base image container
-
-## start container
-```bash
-# you may modify privileged option and mount only specific GPU cards.
-# please refer to our docucments on https://developer.metax-tech.com
-docker run -d -it --net=host --uts=host --ipc=host --privileged=true --group-add video  \
-    --shm-size 100gb --ulimit memlock=-1 \
-    --security-opt seccomp=unconfined --security-opt apparmor=unconfined \
-    --device=/dev/dri --device=/dev/mxcd \
-    -v /usr/share/zoneinfo/Asia/Shanghai:/etc/localtime:ro \
-    --name base_image \
-    registry.access.redhat.com/ubi9/ubi:9.6 bash
-```
-Some packages needs subscription, in the following steps we ignore these packages.
-
-
-
-# 1. Installing Metax-Driver
-
-Reference: [MACA Repo](https://repos.metax-tech.com/gitea/repos/index/wiki/MACA.md#metax-driver)
-
-```bash
-# add repo source
-cat <<EOF> /etc/yum.repos.d/metax-driver-centos.repo 
-[metax-centos]
-name=Maca Driver Yum Repository
-baseurl=https://repos.metax-tech.com/r/metax-driver-centos-$(uname -m)/
-enabled=1
-gpgcheck=0
-EOF
-
-# would install the newest 3.1.0.x release
-# Metax-Driver mainly contains vbios and kmd file, which are not needed in a container.
-# Here we want to get the mx-smi management tool. 
-# kernel version mismatch errors are ignored
-yum makecache
-yum install -y metax-driver mxgvm
-
-# check
-rpm -qa | egrep "(metax|mxsmt|mxfw)"
-```
-
-
-# 2. Installing MACA SDK
-
-Reference: [MACA Repo](https://repos.metax-tech.com/gitea/repos/index/wiki/MACA.md#metax-driver)
-
-may need some time according to your network speed
-
-```bash
-cat <<EOF> /etc/yum.repos.d/maca-sdk-rpm.repo 
-[maca-sdk]
-name=Maca Sdk Yum Repository
-baseurl=https://repos.metax-tech.com/r/maca-sdk-rpm-$(uname -m)/
-enabled=1
-gpgcheck=0
-EOF
-
-yum makecache
-
-yum install -y maca_sdk
-
-rpm -qa | egrep "maca_sdk"
-
-# you may install specific MACA SDK version like this
-yum --showduplicates list |grep maca_sdk
-yum install maca_sdk-<version>-<release>
-```
-
-
-# 3. Install torch
-
-Reference: [MACA Repo](https://repos.metax-tech.com/gitea/repos/index/wiki/MACA.md#metax-driver)
-
-
-## 3.1 Setup a python environment using uv
-
-Our internal build pipeline uses conda. Here we use uv instead. 
-
-Attention: NOT fully tested!
-
-```bash
-curl -LsSf https://astral.sh/uv/install.sh | sh
-
-uv venv /opt/venv --python 3.10
-source /opt/venv/bin/activate
-```
-
-
-## 3.2 Install dependent yum packages
-
-The following packages need subscription, so we just SKIP them. It is about lapack and IB/RDMA utils, NOT fully tested, but should not cause problems.
-
-* `lapack-devel librdmacm-utils libibverbs-utils`
-
-```bash
-yum makecache && yum install -y \
-    openblas-devel \
-    gcc-c++ \
-    libibverbs librdmacm libibumad openssh-server \
-    && yum clean all
-```
-
-## 3.3 Install cu-bridge
-
-Cu-bridge is our cuda compatible package used to compile cuda code. Before installing torch, cu-bridge need to be installed.
-
-Please refer to [cu-bridge/02_User_Manual](https://gitee.com/metax-maca/cu-bridge/tree/master/docs/02_User_Manual)
-
-
-```bash
-# you may separate the building process in a single stage.
-
-yum install -y wget unzip cmake
-
-export MACA_PATH=/opt/maca
-
-wget https://gitee.com/metax-maca/cu-bridge/repository/archive/3.1.0.zip
-unzip 3.1.0.zip
-mv cu-bridge-3.1.0 cu-bridge
-chmod 755 cu-bridge -Rf
-cd cu-bridge
-mkdir build && cd ./build
-cmake -DCMAKE_INSTALL_PREFIX=/opt/maca/tools/cu-bridge ../
-make && make install
-```
-
-
-## 3.4 Some import environment settings
-
-You need to set the following envs to make sure maca-pytorch running properly. The cucc parts are manily used for compiling.
-
-```bash
-export MACA_PATH=/opt/maca
-
-export MACA_CLANG_PATH=${MACA_PATH}/mxgpu_llvm/bin
-
-export CUCC_PATH=${MACA_PATH}/tools/cu-bridge
-export CUDA_PATH=${CUCC_PATH}
-export PATH=$PATH:${CUCC_PATH}/tools:${CUCC_PATH}/bin
-
-export PATH=/opt/mxdriver/bin:${MACA_PATH}/bin:${MACA_CLANG_PATH}:${PATH}
-export LD_LIBRARY_PATH=/opt/mxdriver/lib:${MACA_PATH}/lib:${MACA_PATH}/mxgpu_llvm/lib:${MACA_PATH}/ompi/lib:${MACA_PATH}/ucx/lib:${LD_LIBRARY_PATH}
-```
-
-
-## 3.5 Install torch using uv 
-
-1. Currently in our internal CI/CD flow, we install from local wheel packages. Here we install packages from metax pypi source.
-2. Metax maca version is added to package's local version, e.g., "+metax3.1.0.4torch2.6", and packages's version should match with the installed maca SDK version.
-3. The only way to specify maca version is giving the full version name. It would be better to use separate channels for different maca versions, which is in our plan.
-
-
-```bash
-
-# `datasets` only has 3.1.0 in metax's pypi repo. first install from pulbic source.
-uv pip install datasets==4.1.1
-
-cat <<EOF>  ./requirements.txt
-apex==0.1+metax3.1.0.4
-causal_conv1d==1.5.0.post8+metax3.1.0.4torch2.6
-dropout_layer_norm==0.1+metax3.1.0.4torch2.6
-flash_attn==2.6.3+metax3.1.0.4torch2.6
-flash_linear_attention==0.1+metax3.1.0.4torch2.6
-flash_mla==1.0.1+metax3.1.0.4torch2.6
-flashinfer==0.2.2.post1+metax3.1.0.4torch2.6
-fused_dense_lib==2.6.3+metax3.1.0.4torch2.6
-mamba_ssm==2.2.4+metax3.1.0.4torch2.6
-mctlassEx==0.1.1+metax3.1.0.4torch2.6
-rotary_emb==0.1+metax3.1.0.4torch2.6
-sageattention==2.0.1+metax3.1.0.4torch2.6
-spconv==2.1.0+metax3.1.0.4torch2.6
-torch==2.6.0+metax3.1.0.4
-torchaudio==2.4.1+metax3.1.0.4
-torchvision==0.15.1+metax3.1.0.4
-triton==3.0.0+metax3.1.0.4
-xentropy_cuda_lib==0.1+metax3.1.0.4torch2.6
-xformers==0.0.22+metax3.1.0.4torch2.6
-EOF
-
-uv pip install -r ./requirements.txt -i https://repos.metax-tech.com/r/maca-pypi/simple --trusted-host repos.metax-tech.com
-
-# torch need to use numpy < 2.0
-uv pip install numpy==1.26.4
-```
-
-You may try the following commands:
-
-```bash
-# Search for available versions:
-pip index versions mcspconv -i https://repos.metax-tech.com/r/maca-pypi/simple --trusted-host repos.metax-tech.com
-
-# install a package
-uv pip install torch torchaudio torchvision -i https://repos.metax-tech.com/r/maca-pypi/simple  --trusted-host repos.metax-tech.com
-```
-
-Check torch installing result:
-
-```bash
-python -c "import torch; print(torch.cuda.is_available()); print(torch.cuda.device_count());"
-```
-
-
-
-# 4. Build and install vllm
-
-Reference: [vLLM-metax](https://github.com/MetaX-MACA/vLLM-metax)
-
-
-```bash
-yum install -y vim zip wget tar tzdata \
-    make cmake ninja-build gcc gcc-c++ procps-ng libxml2 openssh-server libXau \
-    openblas-devel \
-    libibverbs librdmacm libibumad \
-    && yum clean all
-
-uv pip install /opt/maca/share/mxsml/pymxsml-*.whl
-uv pip install tokenizers==0.20.3 orjson==3.10.6
-
-
-yum install -y git
-
-# build or install vllm
-uv pip install vllm==0.10.2 --no-deps
-
-
-git clone  --depth 1 --branch v0.10.2 https://github.com/MetaX-MACA/vLLM-metax.git
-cd vLLM-metax
-
-# build vllm on maca needs cuda 11.6
-wget https://developer.download.nvidia.com/compute/cuda/11.6.0/local_installers/cuda_11.6.0_510.39.01_linux.run && \
-    sh cuda_11.6.0_510.39.01_linux.run --silent --toolkit && \
-    rm cuda_11.6.0_510.39.01_linux.run
-
-
-# setup MACA path
-export MACA_PATH="/opt/maca"
-
-# setup CUDA && cu-bridge
-export CUDA_PATH="/usr/local/cuda"
-export CUCC_PATH="${MACA_PATH}/tools/cu-bridge"
-
-# update PATH
-export PATH=${CUDA_PATH}/bin:${MACA_PATH}/mxgpu_llvm/bin:${MACA_PATH}/bin:${CUCC_PATH}/tools:${CUCC_PATH}/bin:${PATH}
-export LD_LIBRARY_PATH=${MACA_PATH}/lib:${MACA_PATH}/ompi/lib:${MACA_PATH}/mxgpu_llvm/lib:${LD_LIBRARY_PATH}
-
-export VLLM_INSTALL_PUNICA_KERNELS=1
-
-
-# install requirements for building
-uv pip install -r requirements/build.txt
-# build wheels
-python setup.py bdist_wheel
-# install wheels
-uv pip install dist/*.whl
-
-```
-
-
-## 5. Install ray
-
-
-```bash
-yum install -y patch
-
-uv pip install click==8.2.1
-
-# the following packages cannot be installed. SKIP
-# uv pip install mcpy==2.1.9.4+b3.1.0.14 numbax==2.1.9.4+b3.1.0.14 -i https://repos.metax-tech.com/r/maca-pypi/simple  --trusted-host repos.metax-tech.com
-
-pip install ray==2.46.0
-
-unzip ray-patch.zip
-cp -rd ray-patch /workspace
-cd /workspace/ray-patch/ray_patch
-python apply_ray_patch.py mx_ray_2.46.batch
-
-if [ -f "/opt/conda/bin/ray" ]; then
-    ln -sf /opt/conda/bin/ray /bin/ray
-fi
-```