#Release Notes

#v26.3.3-1

First release of Alauda GPU Management.

#Highlights

• Unified delivery: one OLM operator and one ClusterPolicy manage the GPU driver, container toolkit, device plugin, GPU Feature Discovery, and DCGM-Exporter. This supersedes the standalone NVIDIA GPU Device Plugin (v0.18.x) and the standalone DCGM-Exporter.
• Precompiled driver for immutable OS: a validated driver tuple (R580, SUSE Linux Micro 5.5, kernel 5.14.21-150500.55.166) ships in the product package and keeps the host image unchanged. Validated on RTX 4090 and Tesla T4, including reboot survival.
• Preinstalled host driver mode (driver.enabled=false), validated on RHEL 9.6 with Tesla P100.
• OLM skipRange upgrade: any earlier version upgrades directly to the target version.

#Version note

The product version advances from the standalone device-plugin line (v0.18.x) to v26.3.x because it now tracks the upstream GPU Operator line. Migration from the standalone cluster-plugin model to the operator is not an in-place upgrade; uninstall the standalone plugins first. For the component mapping, see Versions and Components.

#Known limitations

• MIG is not validated on MIG hardware in this release and requires a user-supplied mig-manager image.
• The upstream mutable-OS driver mode is configurable but not validated in this release.
• A driver version change reloads kernel modules and interrupts GPU access on the node during the reload.

#Alauda Build of DCGM-Exporter

DCGM-Exporter is the NVIDIA GPU metrics component. In the unified delivery model it is an operand managed by Alauda GPU Management, and it can also be installed as a standalone plugin when a HAMi backend exposes the GPUs.

#v4.5.3-482.2

• Delivered as an Alauda GPU Management operand and as a standalone cluster plugin.
• Includes security fixes.

For installation and verification, see DCGM-Exporter. For enabling metrics as a scenario, see Enable GPU metrics.