ReleaseNVIDIANVIDIApublished Oct 23, 2026seen 21h

NVIDIA/nodewright-packages nvidia-tuned/0.11.0

NVIDIA/nodewright-packages

Open original ↗

Captured source

source ↗
published Oct 23, 2026seen 21hcaptured 21hhttp 200method plain

nvidia-tuned/0.11.0

Repository: NVIDIA/nodewright-packages

Tag: nvidia-tuned/0.11.0

Published: 2026-10-23T17:59:31Z

Prerelease: no

Release notes:

SUSE support

The configure-bootloader step, used by service=rke2, now supports SLES 15 SP7, SLES 16.0 and openSUSE Leap. Until now it failed on every distribution outside the Debian family, so an rke2 node on SUSE needed CONFIGURE_BOOTLOADER=false and got none of its profile's kernel arguments. Debian-family nodes are unchanged.

On SUSE the step writes no grub.d drop-in, because SUSE's grub2-mkconfig never reads /etc/default/grub.d/. It relies on tuned's own grub2 integration instead: the $tuned_params reference tuned adds to /etc/default/grub, and /etc/grub.d/00_tuned. The step requires LOADER_TYPE grub2 or grub2-efi and that reference, then runs update-bootloader --refresh. The regeneration is the part that matters: on SLES 15 SP7 tuned left a grub.cfg that set tuned_params but passed it on no linux line, so the node rebooted without the arguments. configure-bootloader-check now reads the generated grub.cfg on SUSE and fails on exactly that.

Upgrade note: a SUSE node whose /etc/sysconfig/bootloader sets LOADER_TYPE to grub2-bls or systemd-boot still fails the step; keep CONFIGURE_BOOTLOADER=false there.

On SUSE, uninstall clears tuned's bootloader cmdline itself: it empties TUNED_BOOT_CMDLINE and TUNED_BOOT_INITRD_ADD and regenerates grub.cfg, the same rollback tuned does when its daemon is stopped while running. uninstall-bootloader-check fails while grub.cfg still sets a non-empty tuned_params or tuned_initrd. Both act only when LOADER_TYPE is grub2 or grub2-efi, so a node on another loader, or with no /etc/sysconfig/bootloader, uninstalls cleanly. openSUSE Tumbleweed and Slowroll are not supported by the step.

Compatibility with OCI GB300 nodes on Oracle Cloud Agent 1.64 and oci-hpc-network-device-names 3.x.

Spectrum-X congestion control run by OCA

OCA 1.64 runs DOCA Spectrum-X congestion control itself on GB300, from persistent doca_spcx_cc@rdma_railN units. configure-spcx-cc treated those processes as a second owner and failed, leaving the node tainted. It now defers when the node has rails (mlx5_bond_* or rdma_railN) and each has exactly one doca_spcx_cc process from a persistent systemd unit that is not its own (the unit's main process, or one its command starts), and every instance of that unit's template is active. It then removes this package's units and udev rule, stopping any of its own processes still running, and the check verifies the other units report PCC host status Active in their current run. When deferring, the step does not check for DOCA or USER_PROGRAMMABLE_CC. Transient units, processes without a unit and strays in an unrelated unit are still refused, as before. Nodes where nothing else runs congestion control are unchanged.

The checks for this step now ignore not-found stubs of this package's units, as the uninstall check already did. They linger until reboot after the units are removed, and previously blocked the check after a handover or after switching spcx_cc off. Readiness is read from the unit's current run only, so a line left in a persistent journal by an earlier boot no longer counts.

NCCL topology file

The OCI GB300 NCCL topology file now names each NIC under both device naming schemes. NCCL matches the file's interfaces by name (_dma), and oci-hpc-network-device-names 3.x renames the VFs from mlx5_0..3 to rdma_vf_rail0..3. On those nodes the mlx5_N_dma entries matched nothing, so their 800 Gb/s speed was not applied; the VFs themselves report 200 Gb/s.

Each NIC now carries a second ` entry, identical apart from the rdma_vf_railN_dma name. NCCL 2.27.3 and later drop the entry that matches no device, so use 2.27.3 or later with this file. The file still lands at ${TOPO_PATH} (default /etc/nccl/topo.xml`), so workloads are unchanged on either image.

[0.11.0] - 2026-10-23

Bug Fixes

  • *(nvidia-tuned)* Support OCA 1.64 and rail device names on OCI GB300

New Features

  • *(nvidia-tuned)* Support SLES and openSUSE Leap in the bootloader step

Other Tasks

  • *(nvidia-tuned)* Release notes and changelog for 0.11.0