feat: on arm64 (Grace-Blackwell GB200/GB300) skip fabric manager and install nvidia-imex

download.sh

@@ -48,11 +48,40 @@ install_fabric_manager () {
     mv /opt/gpu/fm_run_package_installer.sh /opt/gpu/fabricmanager-linux-${NVIDIA_FM_ARCH}-${DRIVER_VERSION}/sbin/fm_run_package_installer.sh
 }
 
-if [[ "${DRIVER_KIND}" == "cuda" ]]; then
-    # download fabricmanager for nvlink based systems, e.g. multi instance gpu vms.
+install_imex () {
+    # nvidia-imex is the cross-node NVLink (MNNVL / ComputeDomains) coordinator for
+    # Grace-Blackwell. It is NOT in the driver .run or the fabric-manager redist -- it
+    # ships only as a separate deb in the CUDA repo. Bundle the version-matched deb into
+    # the image; it is installed at node boot (see install.sh device_init). The exact deb
+    # revision suffix (e.g. -1ubuntu1) varies by version, so resolve the filename from the
+    # repo Packages index rather than hard-coding it.
+    local repo="https://developer.download.nvidia.com/compute/cuda/repos/ubuntu${VERSION_ID//./}/sbsa"
+    local deb
+    # NB: awk must read to EOF (no early `exit`) -- exiting mid-stream closes the pipe and
+    # SIGPIPEs curl, which `set -o pipefail` would turn into a build failure. Gate on a flag
+    # to keep only the first match instead.
+    deb="$(curl -fsSL "${repo}/Packages" \
+        | awk -v p="nvidia-imex_${DRIVER_VERSION}-" '$1=="Filename:" && index($2,p) && !f{print $2; f=1}')"
+    if [[ -z "${deb}" ]]; then
+        echo "nvidia-imex ${DRIVER_VERSION} not found in ${repo}"
+        exit 1
+    fi
+    curl -fsSLO "${repo}/${deb#./}"
+    mv "$(basename "${deb}")" /opt/gpu/
+}
+
+# download fabricmanager for nvlink based systems, but skip it on arm64:
+# arm64 = Grace-Blackwell (GB200/GB300), which uses IMEX, not a node-local FM.
+if [[ "${DRIVER_KIND}" == "cuda" && "${TARGETARCH}" != "arm64" ]]; then
    install_fabric_manager
 fi
 
+# download nvidia-imex for nvlink based arm64 systems (Grace-Blackwell GB200/GB300):
+# GB uses IMEX (not a node-local fabric manager) to coordinate cross-node NVLink.
+if [[ "${DRIVER_KIND}" == "cuda" && "${TARGETARCH}" == "arm64" ]]; then
+   install_imex
+fi
+
 
 # configure nvidia apt repo to cache packages
 curl -fsSLO https://nvidia.github.io/libnvidia-container/gpgkey

install.sh

@@ -166,14 +166,11 @@ initialize_nvidia_driver() {
     ldconfig
     nvidia-smi
 
-    # install fabricmanager for nvlink based systems
-    if [[ "${DRIVER_KIND}" == "cuda" ]]; then
-        NVIDIA_FM_ARCH=$ARCH
-        if [ "$NVIDIA_FM_ARCH" = "arm64" ]; then
-            # NVIDIA uses the name "SBSA" for ARM64 platforms for the fabric manager. See https://en.wikipedia.org/wiki/Server_Base_System_Architecture
-            NVIDIA_FM_ARCH="sbsa"
-        fi
-        bash /opt/gpu/fabricmanager-linux-${NVIDIA_FM_ARCH}-${DRIVER_VERSION}/sbin/fm_run_package_installer.sh
+    # install fabricmanager for nvlink based systems, but skip it on arm64:
+    # arm64 = Grace-Blackwell (GB200/GB300), which uses IMEX, not a node-local FM.
+    # (uname -m reports "aarch64" for arm64.)
+    if [[ "${DRIVER_KIND}" == "cuda" && "${ARCH}" != "aarch64" ]]; then
+        bash /opt/gpu/fabricmanager-linux-${ARCH}-${DRIVER_VERSION}/sbin/fm_run_package_installer.sh
     fi
 }
 
@@ -182,6 +179,18 @@ configure_nvidia_container_runtime() {
 
     install_nvidia_container_toolkit
 
+    # install nvidia-imex on arm64 (Grace-Blackwell): the cross-node NVLink (MNNVL)
+    # coordinator that GB uses in place of a node-local fabric manager. The binary must be
+    # present on the host so the NVIDIA DRA driver (ComputeDomains) can inject it into its
+    # per-workload IMEX daemon pods. --force-depends: the deb declares nvidia-modprobe,
+    # which is provided by the runfile driver (present on disk) rather than as a deb.
+    # The node-wide nvidia-imex.service stays OFF -- IMEX is orchestrated per ComputeDomain
+    # by DRA, not run cluster-wide from the host.
+    if [[ "${DRIVER_KIND}" == "cuda" && "${ARCH}" == "aarch64" ]]; then
+        dpkg -i --force-depends /opt/gpu/nvidia-imex_*_arm64.deb
+        systemctl disable --now nvidia-imex.service 2>/dev/null || true
+    fi
+
     mkdir -p /etc/containerd/config.d
     cp /opt/gpu/10-nvidia-runtime.toml /etc/containerd/config.d/10-nvidia-runtime.toml
 

test/install.bats

@@ -275,3 +275,82 @@ EOF
 
     [ "$status" -ne 0 ]
 }
+
+# --- arm64 (Grace-Blackwell GB200/GB300): fabric-manager skip + nvidia-imex ---
+# GB uses cross-node IMEX in place of a node-local fabric manager. initialize_nvidia_driver
+# must skip FM on aarch64 (installing the nonexistent aarch64 FM tarball previously bricked
+# node join), and configure_nvidia_container_runtime must instead stage the nvidia-imex deb
+# while leaving the node-wide service OFF (DRA orchestrates IMEX per ComputeDomain).
+
+@test "initialize_nvidia_driver installs the fabric manager on x86_64 cuda" {
+    DRIVER_KIND="cuda"; ARCH="x86_64"; DRIVER_VERSION="580.0.0"
+    nvidia-modprobe() { :; }
+    cp() { :; }
+    ldconfig() { :; }
+    nvidia-smi() { :; }
+    # the FM installer is invoked as `bash /opt/gpu/fabricmanager-.../fm_run_package_installer.sh`
+    bash() { echo "bash $*" >> "${TEST_TMP}/fm-calls"; }
+
+    run initialize_nvidia_driver
+
+    [ "$status" -eq 0 ]
+    run cat "${TEST_TMP}/fm-calls"
+    [ "$status" -eq 0 ]
+    [[ "$output" == *"/opt/gpu/fabricmanager-linux-x86_64-580.0.0/sbin/fm_run_package_installer.sh"* ]]
+}
+
+@test "initialize_nvidia_driver skips the fabric manager on arm64 (aarch64) cuda" {
+    DRIVER_KIND="cuda"; ARCH="aarch64"; DRIVER_VERSION="580.0.0"
+    nvidia-modprobe() { :; }
+    cp() { :; }
+    ldconfig() { :; }
+    nvidia-smi() { :; }
+    bash() { echo "bash $*" >> "${TEST_TMP}/fm-calls"; }
+
+    run initialize_nvidia_driver
+
+    [ "$status" -eq 0 ]
+    # FM installer must never run on aarch64 (GB has no node-local fabric manager)
+    [ ! -f "${TEST_TMP}/fm-calls" ]
+}
+
+@test "configure_nvidia_container_runtime installs nvidia-imex on arm64 (aarch64) and leaves the service off" {
+    DRIVER_KIND="cuda"; ARCH="aarch64"
+    install_nvidia_container_toolkit() { :; }
+    mkdir() { :; }
+    cp() { :; }
+    dirname() { command dirname "$@"; }
+    dpkg() { echo "dpkg $*" >> "${TEST_TMP}/imex-calls"; }
+    systemctl() { echo "systemctl $*" >> "${TEST_TMP}/imex-calls"; }
+    export AKSGPU_NVIDIA_CTK_BIN="${TEST_TMP}/bin/nvidia-ctk"
+    _stub_bin nvidia-ctk 0
+
+    run configure_nvidia_container_runtime
+
+    [ "$status" -eq 0 ]
+    run cat "${TEST_TMP}/imex-calls"
+    [ "$status" -eq 0 ]
+    [[ "$output" == *"dpkg -i --force-depends /opt/gpu/nvidia-imex_"*"_arm64.deb"* ]]
+    # node-wide IMEX service is disabled+stopped; DRA runs it per ComputeDomain
+    [[ "$output" == *"systemctl disable --now nvidia-imex.service"* ]]
+}
+
+@test "configure_nvidia_container_runtime does not install nvidia-imex on x86_64" {
+    DRIVER_KIND="cuda"; ARCH="x86_64"
+    install_nvidia_container_toolkit() { :; }
+    mkdir() { :; }
+    cp() { :; }
+    dirname() { command dirname "$@"; }
+    dpkg() { echo "dpkg $*" >> "${TEST_TMP}/imex-calls"; }
+    systemctl() { echo "systemctl $*" >> "${TEST_TMP}/imex-calls"; }
+    export AKSGPU_NVIDIA_CTK_BIN="${TEST_TMP}/bin/nvidia-ctk"
+    _stub_bin nvidia-ctk 0
+
+    run configure_nvidia_container_runtime
+
+    [ "$status" -eq 0 ]
+    run cat "${TEST_TMP}/imex-calls"
+    # imex must never be installed on x86_64 (fabric manager handles NVLink there)
+    [[ "$output" != *"dpkg"* ]]
+    [[ "$output" != *"nvidia-imex.service"* ]]
+}