[Cherry-pick] Backport RHEL/UBI Dockerfile improvements to releases/v0.15.1

.cd/Dockerfile.rhel.ubi.vllm

@@ -4,9 +4,18 @@
 # Global build arguments - declared before first FROM to be available in all stages
 ARG ARTIFACTORY_URL="vault.habana.ai"
 ARG SYNAPSE_VERSION=1.23.0
+# Use an exact revision (e.g. 695) or "latest" to auto-detect the newest available revision.
+# If you do not want metadata to show "latest", pass an exact SYNAPSE_REVISION.
 ARG SYNAPSE_REVISION=695
 ARG BASE_NAME=rhel9.6
+ARG OS_VERSION=9.6
+ARG OS_STRING=rhel96
 ARG PT_VERSION=2.9.0
+ARG PT_MODULES_REPO_NAME=gaudi-pt-modules
+# for internal use - to support a different package name for RHEL 9.4 with Python 3.12
+ARG PT_PACKAGE_NAME_NON_DEFAULT_PYTHON_SUBSTRING=
+ARG PYPI_INDEX_URL="https://pypi.org/simple/"
+ARG HABANA_RPM_REPO_PATH="rhel/9/${OS_VERSION}"
 # can be upstream or fork
 ARG TORCH_TYPE=upstream
 ARG VLLM_GAUDI_COMMIT=v0.15.1
@@ -15,24 +24,29 @@ ARG VLLM_PROJECT_COMMIT=v0.15.1
 # ============================================================================
 # Stage 1: gaudi-base - Base system setup with Habana drivers
 # ============================================================================
-FROM registry.access.redhat.com/ubi9/ubi:9.6 AS gaudi-base
+FROM registry.access.redhat.com/ubi9/ubi:${OS_VERSION} AS gaudi-base
 
 # Re-declare global ARGs to use them in this stage
 ARG ARTIFACTORY_URL
 ARG SYNAPSE_VERSION
 ARG SYNAPSE_REVISION
-ARG BASE_NAME
+ARG OS_VERSION
+ARG OS_STRING
 ARG PT_VERSION
+ARG PYPI_INDEX_URL
+ARG HABANA_RPM_REPO_PATH
 ARG TORCH_TYPE
 
-# Labels for RHEL certification
+# Labels for RHEL certification.
+# Note: when SYNAPSE_REVISION=latest, this metadata intentionally records "latest"
+# (the requested input), not the detected numeric revision.
 LABEL vendor="Habanalabs Ltd." \
     release="${SYNAPSE_VERSION}-${SYNAPSE_REVISION}"
 
 # Environment variables - Habana-specific paths and configurations
 ENV TORCH_TYPE=${TORCH_TYPE} \
     PT_VERSION=${PT_VERSION} \
-    OS_STRING="rhel96" \
+    OS_STRING="${OS_STRING}" \
     VLLM_TARGET_DEVICE="hpu" \
     GC_KERNEL_PATH=/usr/lib/habanalabs/libtpc_kernels.so \
     HABANA_LOGS=/var/log/habana_logs/ \
@@ -43,10 +57,12 @@ ENV TORCH_TYPE=${TORCH_TYPE} \
 COPY LICENSE /licenses/
 
 # System setup - Remove FIPS provider and add EPEL
-RUN dnf install -y python3-dnf-plugin-versionlock && \
+RUN dnf install -y --allowerasing python3-dnf-plugin-versionlock && \
     dnf versionlock add redhat-release* && \
-    rpm -e --nodeps openssl-fips-provider-so && \
-    dnf install -y https://dl.fedoraproject.org/pub/epel/epel-release-latest-9.noarch.rpm && \
+    dnf -y update && \
+    # '|| true' is added to support RHEL 9.4 in which openssl-fips-provider-so is not installed
+    rpm -e --nodeps openssl-fips-provider-so || true && \
+    dnf install -y --allowerasing https://dl.fedoraproject.org/pub/epel/epel-release-latest-9.noarch.rpm && \
     dnf clean all
 
 # Add CentOS repositories for additional packages
@@ -55,10 +71,12 @@ RUN printf "[BaseOS]\\nname=CentOS Linux 9 - BaseOS\\nbaseurl=https://mirror.str
     printf "[CRB]\\nname=CentOS Linux 9 - CRB\\nbaseurl=https://mirror.stream.centos.org/9-stream/CRB/x86_64/os\\ngpgkey=https://www.centos.org/keys/RPM-GPG-KEY-CentOS-Official-SHA256\\ngpgcheck=1\\n" > /etc/yum.repos.d/CentOS-Linux-CRB.repo
 
 # Install system dependencies
-RUN dnf install -y \
-    wget git jq \
+RUN dnf install -y --allowerasing \
+    wget git jq libomp \
     # Image processing dependencies (needed for pillow-simd -> habana-media-loader)
-    zlib-devel libjpeg-devel \
+    zlib-devel libjpeg-turbo-devel \
+    # habanalabs-thunk dependency
+    libfdt-devel \
     # Python development
     python3-devel python3.12 python3.12-devel python3.12-pip \
     # NUMA support
@@ -72,11 +90,11 @@ RUN alternatives --install /usr/bin/python3 python3 /usr/bin/python3.12 2 && \
     ln -s /usr/bin/python3 /usr/bin/python
 
 # Install base Python packages
-RUN pip install setuptools==79.0.1 wheel setuptools_scm && \
-    pip install --upgrade Jinja2 protobuf urllib3 requests
+RUN pip install --no-cache-dir setuptools==79.0.1 wheel setuptools_scm && \
+    pip install --no-cache-dir --upgrade Jinja2 protobuf urllib3 requests
 
 # Setup Habana repository and install Habana packages
-RUN printf "[habanalabs]\\nname=Habana RH9 Linux repo\\nbaseurl=https://${ARTIFACTORY_URL}/artifactory/rhel/9/9.6\\ngpgkey=https://${ARTIFACTORY_URL}/artifactory/rhel/9/9.6/repodata/repomd.xml.key\\ngpgcheck=1\\n" > /etc/yum.repos.d/habanalabs.repo && \
+RUN printf "[habanalabs]\\nname=Habana RH9 Linux repo\\nbaseurl=https://${ARTIFACTORY_URL}/artifactory/${HABANA_RPM_REPO_PATH}\\ngpgkey=https://${ARTIFACTORY_URL}/artifactory/${HABANA_RPM_REPO_PATH}/repodata/repomd.xml.key\\ngpgcheck=1\\n" > /etc/yum.repos.d/habanalabs.repo && \
     echo "=== Content of habanalabs.repo ===" && \
     cat /etc/yum.repos.d/habanalabs.repo && \
     echo "=== End of habanalabs.repo ===" && \
@@ -85,17 +103,30 @@ RUN printf "[habanalabs]\\nname=Habana RH9 Linux repo\\nbaseurl=https://${ARTIFA
     rpm --import "${_GPG_TEMP}/habana_pubkey" && \
     rm -rf "${_GPG_TEMP}" && \
     dnf makecache && \
+    if [ "${SYNAPSE_REVISION}" = "latest" ]; then \
+    dnf install -y \
+    habanalabs-rdma-core-"$SYNAPSE_VERSION"* \
+    habanalabs-thunk-"$SYNAPSE_VERSION"* \
+    habanalabs-firmware-tools-"$SYNAPSE_VERSION"* \
+    habanalabs-graph-"$SYNAPSE_VERSION"*; \
+    else \
     dnf install -y \
     habanalabs-rdma-core-"$SYNAPSE_VERSION"-"$SYNAPSE_REVISION"* \
     habanalabs-thunk-"$SYNAPSE_VERSION"-"$SYNAPSE_REVISION"* \
     habanalabs-firmware-tools-"$SYNAPSE_VERSION"-"$SYNAPSE_REVISION"* \
-    habanalabs-graph-"$SYNAPSE_VERSION"-"$SYNAPSE_REVISION"* && \
+    habanalabs-graph-"$SYNAPSE_VERSION"-"$SYNAPSE_REVISION"*; \
+    fi && \
+    DETECTED_SYNAPSE_REVISION=$(rpm -q --qf '%{RELEASE}\n' habanalabs-rdma-core | head -n1 | sed 's/\..*$//') && \
+    mkdir -p /etc/habanalabs && \
+    echo "${DETECTED_SYNAPSE_REVISION}" > /etc/habanalabs/synapse_revision && \
+    echo "Detected Synapse revision: ${DETECTED_SYNAPSE_REVISION}" && \
     dnf clean all && \
     chmod +t /var/log/habana_logs && \
     rm -f /etc/yum.repos.d/habanalabs.repo
 
 # Install Habana media loader and configure Python path
-RUN pip install habana-media-loader=="${SYNAPSE_VERSION}"."${SYNAPSE_REVISION}" && \
+RUN SYNAPSE_REVISION_RUNTIME=$(cat /etc/habanalabs/synapse_revision) && \
+    pip install --no-cache-dir habana-media-loader=="${SYNAPSE_VERSION}"."${SYNAPSE_REVISION_RUNTIME}" --extra-index-url ${PYPI_INDEX_URL} && \
     echo "/usr/lib/habanalabs" > $(python3 -c "import sysconfig; print(sysconfig.get_path('platlib'))")/habanalabs-graph.pth
 
 # ============================================================================
@@ -105,36 +136,38 @@ FROM gaudi-base AS gaudi-pytorch
 
 # Re-declare global ARGs needed in this stage
 ARG ARTIFACTORY_URL
+ARG OS_STRING
 ARG PT_VERSION
+ARG PT_MODULES_REPO_NAME
+ARG PT_PACKAGE_NAME_NON_DEFAULT_PYTHON_SUBSTRING
+ARG PYPI_INDEX_URL
 ARG SYNAPSE_VERSION
 ARG SYNAPSE_REVISION
 ARG TORCH_TYPE
 
 # Environment variables inherited from base, OS_STRING needed for PyTorch install
-ENV OS_STRING="rhel96"
+ENV OS_STRING="${OS_STRING}"
 
 # Use installer script from Habana to install Pytorch
-RUN PT_PACKAGE_NAME="pytorch_modules-v${PT_VERSION}_${SYNAPSE_VERSION}_${SYNAPSE_REVISION}.tgz" && \
-    PT_ARTIFACT_PATH="https://${ARTIFACTORY_URL}/artifactory/gaudi-pt-modules/${SYNAPSE_VERSION}/${SYNAPSE_REVISION}/pytorch/${OS_STRING}" && \
+RUN SYNAPSE_REVISION_RUNTIME=$(cat /etc/habanalabs/synapse_revision) && \
+    PT_PACKAGE_NAME="pytorch_modules${PT_PACKAGE_NAME_NON_DEFAULT_PYTHON_SUBSTRING}-v${PT_VERSION}_${SYNAPSE_VERSION}_${SYNAPSE_REVISION_RUNTIME}.tgz" && \
+    PT_ARTIFACT_PATH="https://${ARTIFACTORY_URL}/artifactory/${PT_MODULES_REPO_NAME}/${SYNAPSE_VERSION}/${SYNAPSE_REVISION_RUNTIME}/pytorch/${OS_STRING}" && \
     TMP_PATH=$(mktemp --directory) && \
     wget --no-verbose "${PT_ARTIFACT_PATH}/${PT_PACKAGE_NAME}" && \
     tar -zxf "${PT_PACKAGE_NAME}" -C "${TMP_PATH}" && \
     cd "${TMP_PATH}" && \
     export SKIP_INSTALL_DEPENDENCIES=1 && \
-    ./install.sh $SYNAPSE_VERSION $SYNAPSE_REVISION $TORCH_TYPE && \
+    PYTHON_INDEX_URL="--extra-index-url ${PYPI_INDEX_URL}" ./install.sh $SYNAPSE_VERSION $SYNAPSE_REVISION_RUNTIME $TORCH_TYPE && \
     cd / && \
-    rm -rf "${TMP_PATH}" "${PT_PACKAGE_NAME}"
-
-# System update
-RUN dnf -y update --nobest --allowerasing --skip-broken && \
+    rm -rf "${TMP_PATH}" "${PT_PACKAGE_NAME}" && \
     dnf clean all
 
 WORKDIR /workspace
 
 # ============================================================================
 # Stage 3: vllm-final - Install vLLM and configure runtime
 # ============================================================================
-FROM gaudi-pytorch
+FROM gaudi-pytorch AS vllm-openai
 
 # Re-declare global ARGs needed in this stage
 ARG VLLM_GAUDI_COMMIT
@@ -148,10 +181,9 @@ ENV BASE_NAME=${BASE_NAME} \
     VLLM_PATH2=/workspace/vllm-gaudi
 
 # Install additional system dependencies
-RUN dnf update -y --nobest && \
-    dnf install -y gettext jq git --allowerasing && \
-    ln -sf /usr/bin/python3 /usr/bin/python && \
-    dnf clean all
+RUN dnf install -y gettext jq git --allowerasing && \
+    dnf clean all && \
+    ln -sf /usr/bin/python3 /usr/bin/python
 
 WORKDIR /root
 
@@ -173,11 +205,12 @@ RUN set -e && \
     git remote add upstream https://github.com/vllm-project/vllm.git && \
     git fetch upstream --tags && \
     git checkout ${VLLM_PROJECT_COMMIT} && \
-    python use_existing_torch.py && \
-    VLLM_TARGET_DEVICE=empty pip install --no-build-isolation . && \
+    pip install --no-cache-dir -r <(sed '/^torch/d' requirements/build.txt) && \
+    VLLM_TARGET_DEVICE=empty pip install --no-cache-dir --no-build-isolation . && \
     cd $VLLM_PATH2 && \
     git checkout ${VLLM_GAUDI_COMMIT} && \
-    VLLM_TARGET_DEVICE=hpu pip install -v . --no-build-isolation
+    VLLM_TARGET_DEVICE=hpu pip install --no-cache-dir -v . --no-build-isolation && \
+    pip check
 
 # The scripts below are used for benchmarks testing and autocalc:
 # RUN pip3 install -v -e $VLLM_PATH/tests/vllm_test_utils
@@ -186,11 +219,13 @@ RUN set -e && \
 #
 ## Copy utility scripts and configuration
 #RUN mkdir -p /root/scripts/
-#COPY templates /root/scripts/templates/
-#COPY entrypoints /root/scripts/entrypoints/
-#COPY server /root/scripts/server/
-#COPY benchmark /root/scripts/benchmark/
+#COPY .cd/templates /root/scripts/templates/
+#COPY .cd/entrypoints /root/scripts/entrypoints/
+#COPY .cd/server /root/scripts/server/
+#COPY .cd/benchmark /root/scripts/benchmark/
 #WORKDIR /root/scripts
+# Set testing entrypoint script
+#ENTRYPOINT ["python3", "-m", "entrypoints.entrypoint_main"]
 
 # Setup non-root user for OpenShift compatibility
 RUN umask 002 && \
@@ -203,3 +238,4 @@ COPY LICENSE /licenses/vllm.md
 
 USER 2000
 ENTRYPOINT ["python3", "-m", "vllm.entrypoints.openai.api_server"]
+