diff --git a/v/vllm-rocm/vllm_v0.31.0_rocm_ubi_10.sh b/v/vllm-rocm/vllm_v0.31.0_rocm_ubi_10.sh
new file mode 100644
index 0000000000..4105747107
--- /dev/null
+++ b/v/vllm-rocm/vllm_v0.31.0_rocm_ubi_10.sh
@@ -0,0 +1,549 @@
+#!/bin/bash
+# -----------------------------------------------------------------------------
+#
+# Package       : vllm
+# Version       : v0.31.0
+# Source repo   : https://github.com/vllm-project/vllm
+# Tested on     : UBI 10 (ppc64le)
+# Language      : Python, C++, CUDA/HIP
+# Ci-Check      : True
+# Script License: Apache License, Version 2 or later
+# Maintainer    : Daniel Schenker <daniel.schenker@ibm.com>
+#
+# Disclaimer: This script has been tested in root mode on given
+# ==========  platform using the mentioned version of the package.
+#             It may not work as expected with newer versions of the
+#             package and/or distribution. In such case, please
+#             contact "Maintainer" of this script.
+#
+# ----------------------------------------------------------------------------
+#
+# ROCm installation mode (ROCM_INSTALL_MODE env var):
+#   rpms   (default) - Install ROCm RPMs from a provided repo URL
+#   path             - Assume ROCm is already present; use ROCM_PATH as-is
+#
+# PyTorch, torchvision, and torchaudio are installed from the IBM ROCm devpi
+# wheel index (DEVPI_ROCM_INDEX) rather than built from source.
+#
+# Runtime dependencies that are only available on the IBM devpi index
+# (xgrammar, sentencepiece, etc.) are installed from DEVPI_INDEX before the
+# wheel build so they are present in the environment at test time.
+#
+# Usage:
+#   ./vllm_v0.31.0_rocm_ubi_10.sh [v0.31.0]
+#
+# Environment variables honoured (can be set before running):
+#   PACKAGE_VERSION      - vLLM tag to build (default: v0.31.0)
+#   ROCM_INSTALL_MODE    - rpms (default) or path
+#   ROCM_PATH            - Path to ROCm installation (default: /opt/rocm)
+#   ROCM_REPO_URL        - RPM repo baseurl for ROCm
+#   MOONCAKE_WHEEL       - Path to a local mooncake-transfer-engine-rocm wheel.
+#                          Required until the wheel is published to DEVPI_ROCM_INDEX.
+#   DEVPI_INDEX          - IBM ppc64le devpi wheel index URL
+#                          (default: https://wheels.developerfirst.ibm.com/ppc64le/linux/+simple)
+#   DEVPI_ROCM_INDEX     - IBM ppc64le ROCm-specific devpi wheel index URL
+#                          (default: https://wheels.developerfirst.ibm.com/ppc64le/rocm/+simple)
+#
+# ---------------------------------------------------------------------------
+
+set -e
+
+PACKAGE_NAME=vllm
+PACKAGE_VERSION=${1:-v0.31.0}
+PACKAGE_URL=https://github.com/vllm-project/vllm.git
+CURRENT_DIR=$(pwd)
+OS_NAME=$(grep ^PRETTY_NAME /etc/os-release | cut -d= -f2)
+
+ROCM_INSTALL_MODE=${ROCM_INSTALL_MODE:-"rpms"}   # rpms | path
+ROCM_REPO_URL=${ROCM_REPO_URL:-"https://public.dhe.ibm.com/software/server/POWER/Linux/AMD/ROCm/RHEL/10/ppc64le"}
+ROCM_PATH=${ROCM_PATH:-/opt/rocm}
+
+DEVPI_INDEX=${DEVPI_INDEX:-"https://wheels.developerfirst.ibm.com/ppc64le/linux/+simple"}
+DEVPI_ROCM_INDEX=${DEVPI_ROCM_INDEX:-"https://wheels.developerfirst.ibm.com/ppc64le/rocm/+simple"}
+MOONCAKE_WHEEL=${MOONCAKE_WHEEL:-""}
+
+if [[ "$ROCM_INSTALL_MODE" != "rpms" && "$ROCM_INSTALL_MODE" != "path" ]]; then
+    echo "ERROR: ROCM_INSTALL_MODE must be one of: rpms, path"
+    exit 1
+fi
+
+echo "=== vLLM ROCm Build ==="
+echo "  PACKAGE_VERSION  : $PACKAGE_VERSION"
+echo "  ROCM_INSTALL_MODE: $ROCM_INSTALL_MODE"
+echo "  ROCM_PATH        : $ROCM_PATH"
+echo "  MOONCAKE_WHEEL   : ${MOONCAKE_WHEEL:-"(not set — will install from DEVPI_ROCM_INDEX)"}"
+echo "  DEVPI_INDEX      : $DEVPI_INDEX"
+echo "  DEVPI_ROCM_INDEX : $DEVPI_ROCM_INDEX"
+echo "======================="
+
+# ---------------------------------------------------------------------------
+# Install system build dependencies
+# ---------------------------------------------------------------------------
+
+# Python packages must appear first (wrapper script requirement).
+# Note: libdrm is not available in the UBI 10 repo — it is built from source below.
+yum install -y python3.12 python3.12-devel python3.12-pip \
+    gcc-toolset-15 gcc-toolset-15-gcc gcc-toolset-15-gcc-c++ \
+    git make wget patch cmake ninja-build \
+    zlib-devel curl \
+    meson pkgconf-pkg-config
+
+# Configure GCC Toolset 15
+if [[ -f /opt/rh/gcc-toolset-15/enable ]]; then
+    source /opt/rh/gcc-toolset-15/enable
+elif [[ -d /opt/rh/gcc-toolset-15/root/usr/bin ]]; then
+    export PATH="/opt/rh/gcc-toolset-15/root/usr/bin:$PATH"
+    export LD_LIBRARY_PATH="/opt/rh/gcc-toolset-15/root/usr/lib64:${LD_LIBRARY_PATH:-}"
+else
+    echo "ERROR: gcc-toolset-15 not found"
+    exit 1
+fi
+
+echo "Using gcc: $(gcc --version | head -1)"
+
+# Use Python 3.12 for the build so all produced wheels are cp312
+PYTHON=python3.12
+
+# ---------------------------------------------------------------------------
+# Build libdrm from source
+# libdrm is not available in the UBI 10 package repository.  We clone the
+# upstream freedesktop.org release, build it with meson, and install it into
+# /usr/local so every subsequent build step can find it via pkg-config.
+# ---------------------------------------------------------------------------
+LIBDRM_VERSION=${LIBDRM_VERSION:-"libdrm-2.4.124"}
+LIBDRM_URL="https://gitlab.freedesktop.org/mesa/drm.git"
+
+echo "Building libdrm ${LIBDRM_VERSION} from source"
+if [ -d "${CURRENT_DIR}/drm" ]; then
+    echo "drm directory already exists, reusing."
+    cd "${CURRENT_DIR}/drm"
+    git checkout "$LIBDRM_VERSION"
+else
+    if ! git clone --branch "$LIBDRM_VERSION" --depth 1 "$LIBDRM_URL" "${CURRENT_DIR}/drm"; then
+        echo "ERROR: Failed to clone libdrm ${LIBDRM_VERSION}"
+        exit 1
+    fi
+    cd "${CURRENT_DIR}/drm"
+fi
+
+meson setup build \
+    --prefix=/usr/local \
+    --buildtype=release \
+    -Damdgpu=enabled \
+    -Dradeon=enabled \
+    -Dintel=disabled \
+    -Dnouveau=disabled \
+    -Dvmwgfx=disabled \
+    -Dtests=false
+
+ninja -C build
+ninja -C build install
+
+# Make the freshly installed libdrm visible to pkg-config and the linker
+export PKG_CONFIG_PATH="/usr/local/lib64/pkgconfig:/usr/local/lib/pkgconfig:${PKG_CONFIG_PATH:-}"
+export LD_LIBRARY_PATH="/usr/local/lib64:/usr/local/lib:${LD_LIBRARY_PATH:-}"
+echo "libdrm installed: $(pkg-config --modversion libdrm)"
+
+cd "${CURRENT_DIR}"
+
+# ---------------------------------------------------------------------------
+# Install Rust (required by vLLM build)
+# ---------------------------------------------------------------------------
+if command -v cargo &>/dev/null; then
+    echo "Rust already installed: $(cargo --version)"
+else
+    echo "Installing Rust via rustup"
+    curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs > /tmp/rust-install.sh
+    sh /tmp/rust-install.sh --no-modify-path --default-toolchain stable --profile minimal -y
+    rm /tmp/rust-install.sh
+    export PATH="$PATH:$HOME/.cargo/bin"
+    echo "Rust installed: $(cargo --version)"
+fi
+
+# ---------------------------------------------------------------------------
+# MODE: rpms — install ROCm from a provided RPM repository
+# ---------------------------------------------------------------------------
+if [[ "$ROCM_INSTALL_MODE" == "rpms" ]]; then
+    if [[ ! "$ROCM_REPO_URL" =~ ^(https?|file):// ]]; then
+        echo "ERROR: ROCM_REPO_URL does not look like a valid URL (got: ${ROCM_REPO_URL})"
+        exit 1
+    fi
+    echo "Installing ROCm from ${ROCM_REPO_URL}"
+
+    ROCM_GPG_URL="https://public.dhe.ibm.com/software/server/POWER/Linux/AMD/RPM-GPG-KEY-PAMD"
+    ROCM_GPG_PATH="/etc/pki/rpm-gpg/RPM-GPG-KEY-PAMD"
+    echo "Importing ROCm GPG key from ${ROCM_GPG_URL}"
+    wget -q -O "${ROCM_GPG_PATH}" "${ROCM_GPG_URL}"
+    rpm --import "${ROCM_GPG_PATH}"
+
+    cat > /etc/yum.repos.d/rocm.repo <<EOF
+[ROCm]
+name=ROCm
+baseurl=${ROCM_REPO_URL}
+enabled=1
+gpgcheck=1
+gpgkey=file://${ROCM_GPG_PATH}
+EOF
+    dnf install -y rocm-complete
+    ROCM_PATH=/opt/rocm
+fi
+
+# Set ROCm path
+export ROCM_PATH
+export PATH=$ROCM_PATH/bin:$PATH
+export LD_LIBRARY_PATH="${ROCM_PATH}/lib:${ROCM_PATH}/lib64:${LD_LIBRARY_PATH:-}"
+
+# ROCm bundles its own copies of system libraries (lzma, drm, elfutils, …) under
+# lib/rocm_sysdeps/lib.  In containers (e.g. UBI) these OS packages are absent,
+# so we must make the sysdeps directory visible to both the runtime linker and the
+# link-time linker, and point pkg-config at the bundled .pc files.
+ROCM_SYSDEPS_LIB="${ROCM_PATH}/lib/rocm_sysdeps/lib"
+if [[ -d "$ROCM_SYSDEPS_LIB" ]]; then
+    export LD_LIBRARY_PATH="${ROCM_SYSDEPS_LIB}:${LD_LIBRARY_PATH}"
+    export LDFLAGS="-Wl,-rpath,${ROCM_SYSDEPS_LIB} ${LDFLAGS:-}"
+    export PKG_CONFIG_PATH="${ROCM_SYSDEPS_LIB}/pkgconfig:${PKG_CONFIG_PATH:-}"
+fi
+
+if ! command -v hipcc &>/dev/null; then
+    echo "ERROR: hipcc not found under ROCM_PATH=${ROCM_PATH}. Check your ROCm installation."
+    exit 1
+fi
+echo "ROCm hipcc: $(hipcc --version | head -1)"
+
+# ---------------------------------------------------------------------------
+# Install PyTorch, torchvision, and torchaudio from IBM ROCm devpi index
+# ---------------------------------------------------------------------------
+
+$PYTHON -m pip install --upgrade pip setuptools wheel
+
+echo "Installing torch, torchvision, and torchaudio from ROCm devpi index"
+$PYTHON -m pip install --prefer-binary \
+    --extra-index-url "${DEVPI_ROCM_INDEX}" \
+    --extra-index-url "${DEVPI_INDEX}" \
+    torch torchvision torchaudio
+
+# Verify torch is importable and ROCm is visible through it
+echo "Verifying torch install"
+export ROCPROFILER_LOG_LEVEL=0
+$PYTHON -c "import torch; print('torch version :', torch.__version__); print('ROCm available:', torch.cuda.is_available())"
+$PYTHON -c "import torchvision; print('torchvision version:', torchvision.__version__)"
+$PYTHON -c "import torchaudio; print('torchaudio version:', torchaudio.__version__)"
+
+# ---------------------------------------------------------------------------
+# Clone vLLM
+# ---------------------------------------------------------------------------
+echo "Cloning vLLM ${PACKAGE_VERSION}"
+if [ -d "${CURRENT_DIR}/vllm/.git" ]; then
+    echo "vllm directory already exists, reusing."
+    cd "${CURRENT_DIR}/vllm"
+    git fetch --tags origin
+    git checkout "$PACKAGE_VERSION"
+else
+    if ! git clone "$PACKAGE_URL" "${CURRENT_DIR}/vllm"; then
+        echo "------------------$PACKAGE_NAME:clone_fails---------------------------------------"
+        echo "$PACKAGE_URL $PACKAGE_NAME"
+        echo "$PACKAGE_NAME  |  $PACKAGE_URL | $PACKAGE_VERSION | GitHub | Fail |  Clone_Fails"
+        exit 1
+    fi
+    cd "${CURRENT_DIR}/vllm"
+    git checkout "$PACKAGE_VERSION"
+fi
+
+echo "Current vLLM commit:"
+git --no-pager log -1 --oneline
+
+# ---------------------------------------------------------------------------
+# Apply vLLM patches
+# ---------------------------------------------------------------------------
+
+# use_existing_torch.py — removes vLLM's bundled torch build in favour of
+# the torch wheel already installed above
+echo "Running vLLM use_existing_torch.py patch"
+[[ -f "tools/use_existing_torch.py" ]] || { echo "ERROR: use_existing_torch.py not found"; exit 1; }
+$PYTHON tools/use_existing_torch.py
+
+# Patch requirements/build/rocm.txt:
+#   1. Relax triton pin (==3.6.0 → >=3.6.0) for ppc64le wheel compatibility
+#   2. Remove amdsmi PyPI pin — installed from system ROCm below
+#   3. Pin apache-tvm-ffi to 0.1.10 — required by tilelang
+REQ_FILE="requirements/build/rocm.txt"
+[[ -f "$REQ_FILE" ]] || { echo "ERROR: $REQ_FILE not found"; exit 1; }
+
+if grep -q '^triton==3\.[0-9].*$' "$REQ_FILE"; then
+    sed -i 's/^triton==3\.[0-9].*$/triton>=3.6.0/' "$REQ_FILE"
+    echo "Patched $REQ_FILE: triton==3.x.y -> triton>=3.6.0"
+elif grep -q '^triton>=3\.6\.0$' "$REQ_FILE"; then
+    echo "$REQ_FILE already patched for triton>=3.6.0"
+else
+    echo "ERROR: Could not find expected triton==3.x line in $REQ_FILE"
+    exit 1
+fi
+
+if grep -q '^amdsmi' "$REQ_FILE"; then
+    sed -i '/^amdsmi/d' "$REQ_FILE"
+    echo "Patched $REQ_FILE: removed amdsmi line (installed from system ROCm instead)"
+fi
+
+# Pin apache-tvm-ffi to 0.1.10 — required by tilelang;
+# https://github.com/tile-ai/tilelang/issues/2367
+if grep -q '^apache-tvm-ffi' "$REQ_FILE"; then
+    sed -i 's|^apache-tvm-ffi.*$|apache-tvm-ffi==0.1.10|' "$REQ_FILE"
+    echo "Patched $REQ_FILE: apache-tvm-ffi pinned to 0.1.10"
+else
+    echo "apache-tvm-ffi==0.1.10" >> "$REQ_FILE"
+    echo "Patched $REQ_FILE: appended apache-tvm-ffi==0.1.10"
+fi
+
+# Inject ROCm bitcode path into CMAKE_HIP_FLAGS in setup.py
+ROCM_BITCODE_DIR="$ROCM_PATH/lib/llvm/amdgcn/bitcode"
+[[ -d "$ROCM_BITCODE_DIR" ]] || { echo "ERROR: ROCm bitcode directory not found: $ROCM_BITCODE_DIR"; exit 1; }
+
+echo "Patching setup.py CMAKE_HIP_FLAGS with ROCm bitcode path"
+$PYTHON - <<PY
+from pathlib import Path
+
+setup_file = Path("setup.py")
+text = setup_file.read_text()
+
+hip_flag = '            "-DCMAKE_HIP_FLAGS=--rocm-device-lib-path=${ROCM_BITCODE_DIR}/",\n'
+
+target = '''cmake_args = [
+            "-DCMAKE_BUILD_TYPE={}".format(cfg),
+            "-DVLLM_TARGET_DEVICE={}".format(VLLM_TARGET_DEVICE),
+        ]'''
+
+replacement = '''cmake_args = [
+            "-DCMAKE_BUILD_TYPE={}".format(cfg),
+            "-DVLLM_TARGET_DEVICE={}".format(VLLM_TARGET_DEVICE),
+''' + hip_flag + '''        ]'''
+
+if hip_flag.strip() in text:
+    print("setup.py already contains desired CMAKE_HIP_FLAGS patch.")
+elif target in text:
+    text = text.replace(target, replacement, 1)
+    setup_file.write_text(text)
+    print("setup.py patched successfully.")
+else:
+    raise SystemExit("Could not find the expected first cmake_args block in setup.py.")
+PY
+
+# Patch runai-model-streamer: replace [s3,gcs,azure] extras with [s3] only
+# (gcs and azure extras are not available for ppc64le)
+echo "Patching runai-model-streamer extras"
+$PYTHON - <<'PY'
+from pathlib import Path
+
+files_to_patch = [
+    "setup.py",
+    "requirements/rocm.txt",
+    "requirements/test/rocm.in",
+    "requirements/test/nightly-torch.txt",
+    "requirements/test/cuda.in",
+]
+
+old_extras = '"runai": ["runai-model-streamer[s3,gcs,azure] >= 0.15.7"],'
+new_extras = '"runai": ["runai-model-streamer[s3] >= 0.15.7"],'
+old_base   = "runai-model-streamer[s3,gcs,azure]=="
+new_base   = "runai-model-streamer[s3]>="
+
+patched = 0
+for fp in files_to_patch:
+    p = Path(fp)
+    if not p.exists():
+        print(f"  skip {fp} (not found)")
+        continue
+    text = p.read_text()
+    original = text
+    if fp == "setup.py":
+        if old_extras in text:
+            text = text.replace(old_extras, new_extras, 1)
+            print(f"  patched {fp}: extras declaration")
+            patched += 1
+        elif new_extras in text:
+            print(f"  {fp}: already patched")
+        else:
+            raise SystemExit(f"Could not find expected pattern in {fp}")
+    else:
+        if old_base in text:
+            text = text.replace(old_base, new_base)
+            print(f"  patched {fp}: base requirement")
+            patched += 1
+        elif new_base in text:
+            print(f"  {fp}: already patched")
+        else:
+            print(f"  skip {fp} (no runai-model-streamer found)")
+            continue
+    if text != original:
+        p.write_text(text)
+print(f"Patched {patched} file(s)")
+PY
+
+# ---------------------------------------------------------------------------
+# Install vLLM runtime dependencies from IBM devpi index
+#
+# Packages not available on PyPI for ppc64le are pulled from the IBM devpi
+# wheel index.  We install them before the wheel build so they are present
+# at import-test time and pip does not try to fetch them from PyPI when the
+# vllm wheel is installed later.
+# ---------------------------------------------------------------------------
+
+# Pre-built torch-family extras: triton, runai-model-streamer, z3-solver,
+# tilelang, opencv, grpcio — mirrors Daniel's script install_extra_dependencies()
+echo "Installing triton from devpi index"
+$PYTHON -m pip install --prefer-binary \
+    --extra-index-url "${DEVPI_INDEX}" \
+    "triton>=3.6.0"
+
+echo "Installing runai-model-streamer wheels from devpi index"
+$PYTHON -m pip install --prefer-binary \
+    --extra-index-url "${DEVPI_INDEX}" \
+    "runai-model-streamer" \
+    "runai-model-streamer[s3]"
+
+echo "Installing z3-solver from devpi index"
+$PYTHON -m pip install --prefer-binary \
+    --extra-index-url "${DEVPI_INDEX}" \
+    z3-solver==4.15.4.0
+# tilelang has moved to the ROCm-specific index; install with --no-build-isolation
+# so the already-installed torch is reused rather than a fresh copy downloaded.
+$PYTHON -m pip install cmake ninja \
+    "scikit-build-core[pyproject]>=0.10.0" \
+    "cython>=3.2.8" \
+    setuptools-scm
+echo "Installing tilelang from ROCm devpi index"
+$PYTHON -m pip install --prefer-binary --no-build-isolation \
+    --extra-index-url "${DEVPI_ROCM_INDEX}" \
+    tilelang
+
+echo "Installing opencv-python-headless and grpcio from devpi index"
+$PYTHON -m pip install --prefer-binary \
+    --extra-index-url "${DEVPI_INDEX}" \
+    "opencv-python-headless==4.13.0.92" \
+    grpcio
+
+echo "Installing xgrammar and sentencepiece from devpi index"
+$PYTHON -m pip install --prefer-binary \
+    --extra-index-url "${DEVPI_INDEX}" \
+    xgrammar \
+    sentencepiece \
+    tiktoken \
+    ijson \
+    llguidance \
+    msgspec \
+    cbor2 \
+    grpcio \
+    numba \
+    llvmlite \
+    onnx \
+    pandas \
+    scipy
+
+# Install pyarrow and datasets with explicit version caps:
+#   - datasets>=5.1.0 requires pyarrow>=24.0.0 which has no ppc64le binary on
+#     PyPI and is not yet on the IBM devpi index (highest available: 23.0.1).
+#     Pin datasets<5.1.0 so pip never needs pyarrow>=24.
+#   - Pre-install pyarrow from devpi so it is already present and resolved
+#     before the vLLM wheel install triggers datasets' dependency.
+echo "Installing pyarrow and datasets from devpi index"
+$PYTHON -m pip install --prefer-binary \
+    --extra-index-url "${DEVPI_INDEX}" \
+    "pyarrow>=21.0.0" \
+    "datasets<5.1.0"
+echo "Devpi runtime deps installed"
+
+# ---------------------------------------------------------------------------
+# Install mooncake-transfer-engine-rocm
+#
+# vLLM v0.31.0 requires mooncake-transfer-engine-rocm >= 0.3.13 in
+# requirements/rocm.txt.  The wheel is not yet on the IBM devpi index so it
+# is installed from a local path supplied via MOONCAKE_WHEEL.
+#
+# TODO: once the wheel is published to DEVPI_ROCM_INDEX, remove the
+# MOONCAKE_WHEEL local-path branch and the env var.
+# ---------------------------------------------------------------------------
+if [[ -n "$MOONCAKE_WHEEL" ]]; then
+    [[ -f "$MOONCAKE_WHEEL" ]] || { echo "ERROR: MOONCAKE_WHEEL file not found: ${MOONCAKE_WHEEL}"; exit 1; }
+    echo "Installing mooncake-transfer-engine-rocm from local wheel: ${MOONCAKE_WHEEL}"
+    $PYTHON -m pip install --prefer-binary \
+        --extra-index-url "${DEVPI_INDEX}" \
+        "${MOONCAKE_WHEEL}"
+else
+    echo "Installing mooncake-transfer-engine-rocm from ROCm devpi index"
+    $PYTHON -m pip install --prefer-binary \
+        --extra-index-url "${DEVPI_ROCM_INDEX}" \
+        "mooncake-transfer-engine-rocm>=0.3.13"
+fi
+
+# ---------------------------------------------------------------------------
+# Build vLLM wheel
+# ---------------------------------------------------------------------------
+cd "${CURRENT_DIR}/vllm"
+
+# Seed CMAKE_PREFIX_PATH with the installed torch cmake config so that vLLM's
+# CMake can find TorchConfig.cmake without a from-source build tree.
+TORCH_CMAKE_PREFIX=$($PYTHON -c 'import torch; print(torch.utils.cmake_prefix_path)')
+export CMAKE_PREFIX_PATH="${TORCH_CMAKE_PREFIX}:${ROCM_PATH}:${CMAKE_PREFIX_PATH:-}"
+
+# Must be exported (not just inline) so that vLLM's get_runtime_environment()
+# in setup.py sees VLLM_TARGET_DEVICE=rocm when it calls os.environ.get().
+# An inline-only assignment on the build command is not sufficient because
+# get_vllm_version() is invoked at module-scope before the env is propagated.
+export VLLM_TARGET_DEVICE="rocm"
+
+echo "Installing vLLM ROCm build requirements"
+[[ -f "requirements/build/rocm.txt" ]] || { echo "ERROR: requirements/build/rocm.txt not found"; exit 1; }
+$PYTHON -m pip install --prefer-binary -r requirements/build/rocm.txt --extra-index-url "${DEVPI_INDEX}"
+
+# Install matching AMDSMI Python bindings from system ROCm (not PyPI)
+AMDSMI_SRC="$ROCM_PATH/share/amd_smi"
+[[ -d "$AMDSMI_SRC" ]] || { echo "ERROR: AMDSMI source not found: $AMDSMI_SRC"; exit 1; }
+AMDSMI_TMP=$(mktemp -d)
+cp -a "$AMDSMI_SRC/." "$AMDSMI_TMP/"
+$PYTHON -m pip install --force-reinstall --no-deps "$AMDSMI_TMP"
+rm -rf "$AMDSMI_TMP"
+echo "Installed AMDSMI from system ROCm"
+
+echo "Building vLLM ROCm wheel (this will take a while)"
+# VLLM_VERSION_OVERRIDE bypasses get_vllm_version()'s own ROCm suffix logic
+# (which would append .rocm714 on top of +rocm7.14 and produce a double suffix).
+# The override is used as-is — producing e.g. vllm-0.28.0+rocm7.14-cp312-...
+unset BUILD_VERSION SETUPTOOLS_SCM_PRETEND_VERSION
+export VLLM_VERSION_OVERRIDE="${PACKAGE_VERSION#v}+rocm7.14"
+if ! $PYTHON setup.py bdist_wheel --universal --dist-dir "${CURRENT_DIR}"; then
+    echo "------------------$PACKAGE_NAME:install_fails---------------------------------------"
+    echo "$PACKAGE_URL $PACKAGE_NAME"
+    echo "$PACKAGE_NAME  |  $PACKAGE_URL | $PACKAGE_VERSION | $OS_NAME | GitHub | Fail | Install_Fails"
+    exit 1
+fi
+
+echo "Built vLLM wheel(s):"
+ls -lh "${CURRENT_DIR}"/vllm-*.whl
+
+$PYTHON -m pip install "${CURRENT_DIR}"/vllm-*.whl \
+    --prefer-binary \
+    --extra-index-url "${DEVPI_INDEX}" \
+    --extra-index-url "${DEVPI_ROCM_INDEX}"
+
+# ---------------------------------------------------------------------------
+# Import test
+# ---------------------------------------------------------------------------
+echo "Installing test-stage dependencies from devpi index"
+$PYTHON -m pip install --prefer-binary \
+    --extra-index-url "${DEVPI_INDEX}" \
+    "apache-tvm-ffi==0.1.10" \
+    fastsafetensors
+
+echo "Running import test"
+cd "${CURRENT_DIR}"
+
+if ! $PYTHON -c "import vllm; print('vllm version:', vllm.__version__)"; then
+    echo "------------------$PACKAGE_NAME:Install_success_but_test_fails---------------------"
+    echo "$PACKAGE_URL $PACKAGE_NAME"
+    echo "$PACKAGE_NAME  |  $PACKAGE_URL | $PACKAGE_VERSION | $OS_NAME | GitHub | Fail |  Install_success_but_test_Fails"
+    exit 2
+else
+    echo "------------------$PACKAGE_NAME:Install_&_test_both_success-------------------------"
+    echo "$PACKAGE_URL $PACKAGE_NAME"
+    echo "$PACKAGE_NAME  |  $PACKAGE_URL | $PACKAGE_VERSION | $OS_NAME | GitHub  | Pass |  Both_Install_and_Test_Success"
+    exit 0
+fi