Add fabric manager boot and support diagnostics

audit/internal/app/support_bundle.go

@@ -22,6 +22,8 @@ var supportBundleServices = []string{
 	"bee-selfheal.service",
 	"bee-selfheal.timer",
 	"bee-sshsetup.service",
+	"nvidia-dcgm.service",
+	"nvidia-fabricmanager.service",
 }
 
 var supportBundleCommands = []struct {
@@ -48,6 +50,43 @@ else
 fi
 `}},
 	{name: "system/nvidia-smi-q.txt", cmd: []string{"nvidia-smi", "-q"}},
+	{name: "system/nvidia-smi-topo.txt", cmd: []string{"sh", "-c", `
+if command -v nvidia-smi >/dev/null 2>&1; then
+  nvidia-smi topo -m 2>&1 || true
+else
+  echo "nvidia-smi not found"
+fi
+`}},
+	{name: "system/systemctl-nvidia-units.txt", cmd: []string{"sh", "-c", `
+if ! command -v systemctl >/dev/null 2>&1; then
+  echo "systemctl not found"
+  exit 0
+fi
+echo "=== unit files ==="
+systemctl list-unit-files --no-pager --all 'nvidia*' 'fabric*' 2>&1 || true
+echo
+echo "=== active units ==="
+systemctl list-units --no-pager --all 'nvidia*' 'fabric*' 2>&1 || true
+echo
+echo "=== failed units ==="
+systemctl --failed --no-pager 2>&1 | grep -iE 'nvidia|fabric' || echo "no failed nvidia/fabric units"
+`}},
+	{name: "system/fabric-manager-paths.txt", cmd: []string{"sh", "-c", `
+for candidate in \
+  /usr/bin/nvidia-fabricmanager \
+  /usr/bin/nv-fabricmanager \
+  /usr/bin/nvidia-fabricmanagerd \
+  /usr/bin/nvlsm; do
+  if [ -e "$candidate" ]; then
+    echo "=== $candidate ==="
+    ls -l "$candidate" 2>&1 || true
+    echo
+  fi
+done
+if ! ls /usr/bin/nvidia-fabricmanager /usr/bin/nv-fabricmanager /usr/bin/nvidia-fabricmanagerd /usr/bin/nvlsm >/dev/null 2>&1; then
+  echo "no fabric manager binaries found"
+fi
+`}},
 	{name: "system/lspci-nvidia-bridges-vv.txt", cmd: []string{"sh", "-c", `
 if ! command -v lspci >/dev/null 2>&1; then
   echo "lspci not found"
@@ -195,6 +234,10 @@ var supportBundleOptionalFiles = []struct {
 }{
 	{name: "system/kern.log", src: "/var/log/kern.log"},
 	{name: "system/syslog.txt", src: "/var/log/syslog"},
+	{name: "system/fabricmanager.log", src: "/var/log/fabricmanager.log"},
+	{name: "system/nvlsm.log", src: "/var/log/nvlsm.log"},
+	{name: "system/fabricmanager/fabricmanager.log", src: "/var/log/fabricmanager/fabricmanager.log"},
+	{name: "system/fabricmanager/nvlsm.log", src: "/var/log/fabricmanager/nvlsm.log"},
 }
 
 const supportBundleGlob = "????-??-?? (BEE-SP*)*.tar.gz"

audit/internal/platform/runtime.go

@@ -28,6 +28,8 @@ var runtimeTrackedServices = []string{
 	"bee-audit",
 	"bee-web",
 	"bee-sshsetup",
+	"nvidia-dcgm",
+	"nvidia-fabricmanager",
 }
 
 func (s *System) CollectRuntimeHealth(exportDir string) (schema.RuntimeHealth, error) {

iso/builder/VERSIONS

@@ -1,6 +1,7 @@
 DEBIAN_VERSION=12
 DEBIAN_KERNEL_ABI=auto
 NVIDIA_DRIVER_VERSION=590.48.01
+NVIDIA_FABRICMANAGER_VERSION=590.48.01-1
 NCCL_VERSION=2.28.9-1
 NCCL_CUDA_VERSION=13.0
 NCCL_SHA256=2e6faafd2c19cffc7738d9283976a3200ea9db9895907f337f0c7e5a25563186

iso/builder/build.sh

@@ -1262,6 +1262,7 @@ fi
 # --- substitute version placeholders in package list and archive ---
 if [ "$BEE_GPU_VENDOR" = "nvidia" ]; then
     sed -i \
+        -e "s/%%NVIDIA_FABRICMANAGER_VERSION%%/${NVIDIA_FABRICMANAGER_VERSION}/g" \
         -e "s/%%DCGM_VERSION%%/${DCGM_VERSION}/g" \
         "${BUILD_WORK_DIR}/config/package-lists/bee-gpu.list.chroot"
 elif [ "$BEE_GPU_VENDOR" = "amd" ]; then

iso/builder/config/hooks/normal/9000-bee-setup.hook.chroot

@@ -43,6 +43,7 @@ systemctl enable bee-journal-mirror@ttyS1.service 2>/dev/null || true
 # Enable GPU-vendor specific services
 if [ "$GPU_VENDOR" = "nvidia" ]; then
     systemctl enable nvidia-dcgm.service 2>/dev/null || true
+    systemctl enable nvidia-fabricmanager.service 2>/dev/null || true
     systemctl enable bee-nvidia.service
 elif [ "$GPU_VENDOR" = "amd" ]; then
     # ROCm symlinks (packages install to /opt/rocm-*/bin/)

iso/builder/config/package-lists/bee-nvidia.list.chroot

@@ -5,6 +5,7 @@
 # DCGM 4 is packaged per CUDA major. The image ships NVIDIA driver 590 with
 # CUDA 13 userspace, so install the CUDA 13 build plus proprietary components
 # explicitly.
+nvidia-fabricmanager=%%NVIDIA_FABRICMANAGER_VERSION%%
 datacenter-gpu-manager-4-cuda13=1:%%DCGM_VERSION%%
 datacenter-gpu-manager-4-proprietary=1:%%DCGM_VERSION%%
 datacenter-gpu-manager-4-proprietary-cuda13=1:%%DCGM_VERSION%%

iso/overlay/usr/local/bin/bee-nvidia-load

@@ -258,6 +258,22 @@ else
     log "WARN: nvidia-smi not found — cannot enable persistence mode"
 fi
 
+# Start or refresh Fabric Manager after the NVIDIA stack is ready. On NVSwitch
+# systems CUDA/DCGM can report "system not yet initialized" until fabric
+# training completes under nvidia-fabricmanager.
+if command -v systemctl >/dev/null 2>&1 && systemctl list-unit-files --no-legend 2>/dev/null | grep -q '^nvidia-fabricmanager\.service'; then
+    if systemctl restart nvidia-fabricmanager.service >/dev/null 2>&1; then
+        log "nvidia-fabricmanager restarted"
+    elif systemctl start nvidia-fabricmanager.service >/dev/null 2>&1; then
+        log "nvidia-fabricmanager started"
+    else
+        log "WARN: failed to start nvidia-fabricmanager.service"
+        systemctl status nvidia-fabricmanager.service --no-pager 2>&1 | sed 's/^/  fabricmanager: /' || true
+    fi
+else
+    log "WARN: nvidia-fabricmanager.service not installed"
+fi
+
 # Start DCGM host engine so dcgmi can discover GPUs.
 # nv-hostengine must run after the NVIDIA modules and device nodes are ready.
 # If it started too early (for example via systemd before bee-nvidia-load), it can