Release notes for NVIDIA Base Command™ Manager (BCM) 11.34.0#

Released: 11 September 2026

General#

New Features#

  • Added support for Slurm 26.05 (26.05.4)

  • Added an IPMI plugin to Slurm

  • Added the slurmXX.XX-builder package, which allows custom Slurm packages to be built on a BCM cluster to replace the Slurm packages shipped with BCM

  • Added CUDA 13.2 packages

  • Added arm64 packages for cm-nsight-systems-cli

  • Added Grafana dashboards and alerts for VRNVL72

  • Added instructions on configuring alerting in Grafana to the NVIDIA Mission Control 2.5 guide

  • Added support for RHEL add-on installations from the BCM Rocky ISO

  • Updated Slurm 25.05 and 25.11 to 25.05.9 and 25.11.8 respectively (CVE-2026-65107, CVE-2026-65108, CVE-2026-65109, CVE-2026-65138, CVE-2026-65139, CVE-2026-65140, CVE-2026-65165, CVE-2026-65168)

  • Updated Pyxis to 0.24.0 and Enroot to 4.2.0

  • Updated CUDA 12.8 to 12.8.2

  • Updated CUDA 12.9 to 12.9.2

  • Updated CUDA 13.0 to 13.0.3

  • Updated CUDA 13.1 to 13.1.2

  • Updated cm-docker to v29.4.3

  • Updated urllib3 to 2.7.0 in cm-python312

  • Updated Harbor to 2.15

  • Updated containerd to 2.3

  • Updated Helm to 4.20

  • Updated the Knative Kubernetes operator to 1.22

  • Updated MetalLB to 0.16

  • Updated the Tigera operator to 1.42

  • Added support for SLES15 SP7

  • Added support for Ubuntu 26.04

  • Added support for DGX OS 8.0.0/BaseOS 9 software images

  • Added support for Vera Rubin NVL72 bring up

  • Added Run:ai TUI setup wizard for RHEL9 based Linux distributions

CMDaemon#

New Features#

  • Added a Vital Product Data (VPD) info object, exposed as a vpd RPC in the REST API and as a vpd command in cmsh that lists PCI Vital Product Data

  • Added cluster-wide SSH multi-factor authentication (MFA) setup through cm-mfa-setup.py, with TOTP and math-test providers, a staged nullok rollout, and an --enforce option to require MFA

  • Added support for Topograph 4

  • EnablePrometheusExporterService now also accepts auto, which enables the Prometheus exporter when the license edition is NVIDIA Mission Control; auto is the new default

  • AWS: The architecture of the AMI is now validated against the machine type in use, to prevent booting from a wrong node-installer or netboot image

  • The target CPU architecture of a software image is now checked against the architecture of the cloud node instance, to prevent virtual machines from attempting to boot a wrong OS

  • Added an optional close.sh script hook that runs on the active head node when a device is closed

Fixed Issues#

  • Fixed the REST v2 response contract so that authentication failures, unsupported methods, and HEAD and OPTIONS requests all return the same JSON envelope; this includes 405 responses with an Allow header, and a Location header on 201 responses

  • Fixed Kubernetes drain status reporting in cmsh, and nodes remaining stuck in the “Draining” state

  • Fixed node deletion for worker nodes when the Kubelet role is unassigned; the bcm-kubelet-role-unassigned node role is now also removed when appropriate

  • Fixed CMDaemon device connectivity checks taking ICMP identifiers away from an operator-issued ping; Linux ping sockets (SOCK_DGRAM) are now used where available, so the kernel owns the identifier

  • Fixed duplicated mlxconfig firmware rows by skipping SR-IOV representors during adapter discovery

  • Fixed mlxconfig detail levels 1 and 2 showing almost no fields; advanced and expert fields are now selected by name, and the cmsh Field column is wider so that long keys are no longer truncated

  • Fixed the switch leak policy reporting a successful power-off while the NVLink switch was still reachable; a Redfish PowerState of Off is now required after an NVLink ForceOff, and a failure to restore PowerRestorePolicy is reported instead of leaving the BMC on LastState

  • Fixed disk layout validation rejecting RAID identifiers that contain a full device path such as /dev/md0, which made disks xml and HA cloning fail; the disks script now generates disklayout.xml with a dedicated RAID identifier type

  • Fixed compute nodes being unable to obtain their CMDaemon certificate, by binding privileged source ports when an edge director forwards JSON RPC to the active head node

  • Fixed a broken etcd role configuration that could be left behind by an earlier defect; cm-etcd-manage can now repair it, and a CMDaemon update instruction removes the erroneous option through MySQL

  • Fixed the cm-etcd-manage backup action failing when the first etcd member is down; it now falls back to another healthy etcd member

  • Fixed nondeterministic cloud node identification for nodes with mixed case hostnames

  • Fixed a crash in validate_gmc_node_changes when validating OCI GPU Memory Cluster (GMC) config changes after upgrading from an old CMDaemon version

  • Fixed a timeout during OCI GPU Memory Cluster (GMC) creation wrongly deleting the GMC

  • Fixed a possible race condition in the BCM CAPI provider that could occur when nodes previously used as control planes are repurposed as workers

  • Fixed a command injection vulnerability caused by a permissive regular expression in the cmburn SSH restrictor script

  • Fixed image updates deleting the CMDaemon-created Slurm pre-job prolog and the nvidia-dcgm enablement symlink

  • Added collection for all four PDB HSC current, power, temperature, and voltage sensors on GB200, GB300, and NVL72 Redfish metrics

  • Fixed generated /etc/genders lines exceeding the libgenders 64KiB buffer (GenderMaxLineLength) by wrapping them

  • Fixed software-image clone on XFS copying bind mounts, which could fill the disk

  • Fixed CMDaemon aborting during SNMP initialization when leftover Net-SNMP persistent backup files (snmpapp.0.conf) are present

Node Installer#

Fixed Issues#

  • Fixed CMDaemon failing to start on older distributions, by staging only the OS DCGM libraries into cuda-dcgm/current so that the distribution libstdc++ is no longer loaded

  • Fixed x86_64 ramdisk creation on aarch64 head nodes

  • Fixed node-installer fallback time synchronization on Ubuntu 26.04

  • Fixed a leftover PXE/ip= address surviving node-installer handoff causing issues with ifupdown

  • Updated the datacenter-gpu-manager packages in DGX OS images from the CUDA 12 to the CUDA 13 variant

Cluster Tools#

New Features#

  • Added the ability to clone all databases when setting up HA; cmha-setup has a new -a option and a new dialog for this. The previous behavior of cloning only the required databases remains the default

  • Non-latest Rocky images now use the Rocky vault repositories

Fixed Issues#

  • Fixed Ubuntu 26.04 image creation on Ubuntu 24.04

  • Fixed the first-stage ramdisk being unable to detect the control-plane BOOTIF, by adding the mlx5_core and ib_umad modules for all DGX systems in cm-create-image

  • cm-image create all now prints the location of the log file when ramdisk creation fails

  • Fixed cmha-setup GUI not honoring the skip flag for non-interactive mode, and CMDaemon adding a duplicate non-interactive flag

COD#

New Features#

  • Dropped left-over compatibility with BCM versions older than 9.2, which are no longer supported

  • COD-OCI: Changed the default image format to compressed qcow2

  • COD-Azure: Changed the default VM size to Standard_D2ls_v5

Fixed Issues#

  • Fixed cod-csp config dump for lists of strings that contain commas, such as inbound rules; such strings are now quoted so that the list is not split differently after a dump

  • COD-OCI: Added support for more than 24 head-node security rules; up to 119 rules are now supported, and the total rule limit is detected before the rules are created

  • COD-GCP: Fixed --on-error-undo not taking effect after an SSH or CMDaemon wait timeout

  • COD-Azure: Fixed cluster creation failing for longer cluster names, where the generated Azure resource names for the resource group and for the shared-storage private endpoint on HA clusters could exceed the Azure length limits. These resources are now named <cluster-name>-rg and <cluster-name>-shared-pe. Existing clusters are unaffected

  • COD-OCI: Fixed rollback not completing after a cluster creation timeout with --on-error-abort

  • COD-AWS: Improved node identification, to reduce orphaned cloud instances when instance creation times out but eventually completes. An instance is now identified by BCM as soon as it starts and enters netboot, which also speeds up bulk instance creation

cm-kubernetes-setup#

New Features#

  • Added the AgentGateway operator and its CRDs; the Helm chart is registered but disabled by default, and the operator can be selected manually

  • Minor Kubernetes versions are now pinned in non-air-gapped setups as well; air-gapped setups were already pinned

  • Pinned Knative Serving and Eventing to 1.21, the latest version compatible with Run:ai

Fixed Issues#

  • Fixed Run:ai pod crashes caused by outdated NVIDIA GPU Operator customizations

  • The NVIDIA GPU Operator now deploys the NVIDIA container toolkit itself when it is needed, so BCM no longer writes the containerd-nvidia-cri.toml drop-in for containerd. This fixes issues on nodes that use nvidia-container-runtime as a pass-through without a GPU

  • Fixed Knative Serving pods remaining in ImagePullBackOff during air-gapped deployment of multi-architecture image bundles

  • Improved the resilience of Helm-based operator installations when a release from a failed previous deployment is still present; the blocking release is now removed before the install is retried

  • Skipping or aborting a failed operator installation now reports which of the remaining operator installations are abandoned, and Gateway configuration now fails immediately with a clear message when a selected gateway operator was never installed, instead of retrying for 15 minutes

  • Fixed the dedicated section for configuring the Gateway TLS certificate, which reported success without installing the certificate

  • Fixed Kubelet failing to start with CDMM enabled on nodes with more than 8 NUMA nodes, the default maximum above which Kubelet’s topology manager would not bootstrap successfully

  • Fixed the cm-containerd containerd.service unit so that Restart=on-failure restarts containerd again after the GPU Operator sends it a SIGHUP

cm-mysql-setup#

New Features#

  • Added the ability to set up MySQL in HA mode for Slurm, with up to two MySQL nodes, including MySQL Router installation and configuration

cm-wlm-setup#

Deprecated or Removed Features#

  • PBS Professional, OpenPBS, and Altair Grid Engine setups are no longer possible on BCM 11

Fixed Issues#

  • Fixed Slurm setup failing on a cluster that has no compute nodes

cm-setup#

Fixed Issues#

  • Fixed cm-docker-setup failing against DGX images; nvidia-container-toolkit is now installed instead of cm-nvidia-container-toolkit on DGX images, and packages that conflict with cm-docker are removed

  • Fixed cm-setup silently not running post-install hooks for packages installed from a local .deb or .rpm path, a regression from BCM-34395

  • Fixed cm-container-registry-setup failing on ARM based systems

  • Fixed cm-container-registry-setup setting wrong permissions on the registry certificate files and directories; strict permissions are now applied when Docker certificates are distributed to other nodes

  • Fixed an issue with cm-ipmitool not finding the distribution freeipmi

  • Fixed cm-cloud-ha-setup sometimes hanging forever at the Setup License stage on Rocky head nodes

Base View#

New Features#

  • Mission Control license information now uses NVL naming for the firmware and rack power features, matching CMDaemon, and no longer lists the unused GPU Workload Power Profiles license attribute