#Release Notes

#Product releases

The following Ascend NPU vendor base product is documented in this product line:

Use Versions and Components to check the delivery model and component differences between NPU Operator versions. Historical release notes remain below when they affect migration or support decisions.

#v26.6.0

Supports ACP v4.0-v4.3 on ARM Ascend 310P and 910B nodes.

The formal ARM64 package was validated in Ascend 910B3 Host Driver environments, Ascend 310P Host and Alauda OS environments, and Ascend 310P Host environments using HAMi full-device allocation, hard slicing, and soft slicing. The release scan found no Critical or High vulnerabilities in Alauda-maintained images.

#New and Optimized Features

• Adds operator-managed Host Driver installation for supported mutable operating systems, while retaining pre-installed Host Driver and managed precompiled Driver modes.
• Resolves precompiled Driver image tags automatically from the configured Driver version and node discovery labels.

#Fixed Issues

• An NPU workload that had entered Failed state while still terminating could be treated as released too early during a Driver upgrade. The Operator now waits for the Pod to disappear and stops the upgrade if the configured timeout is reached.
• A managed Host Driver that was temporarily unavailable after reboot could be treated as a fresh installation target. The installer now waits for the existing installation to become healthy and does not reinstall it once it is ready.
• Uninstalling the Operator could leave a managed policy and finalizer behind. The uninstall flow now removes those managed objects before the controller is removed.
• Pre-installed Host Driver deployments could report HCCN as unavailable because they do not use the precompiled-Driver readiness marker. HCCN Bootstrap observation now uses the Host utility path and remains fail-closed when the probe fails.

#Upgrade Notes

• For a pre-installed Host Driver, keep Driver management disabled and select Host mode.
• For an immutable operating system, import the matching ARM64 precompiled Driver image and configure the target workload cluster image mapping or ImageWhiteList before installation. The Driver image is delivered separately from the product package.
• v26.6.0 uses the ascend RuntimeClass backed by ascend-docker-runtime. The admission webhook injects this RuntimeClass into Pods that request Ascend resources, so workload manifests do not need to set it explicitly when the webhook is enabled.
• Do not use the CDI spec from v1.2.4 as a v26.6.0 readiness check. Verify the RuntimeClass, runtime integration DaemonSet, and admitted Pod instead.
• Do not run more than one device exposure owner for the same Ascend devices. When HAMi manages allocation or slicing, disable the NPU Operator-managed Ascend Device Plugin for those devices.

#Known Issues

• Automatic HCCN device IP initialization is not supported. HCCN Bootstrap is disabled by default; the validated profile is observe-only and does not rewrite preconfigured addresses when enabled.
• On a single-node control-plane deployment that uses host networking and fixed component ports, an upgrade can briefly wait for the previous Pod to release its port. Confirm that the old Pod has terminated before treating the replacement Pod as failed.

#Upgrade Path

• Upgrade from v1.2.4 to v26.6.0 through OperatorHub / OLM. Use a manual approval strategy for production clusters and verify the NPUOperatorCtl components and one NPU workload after the upgrade.
• To move from v1.1.3 or an earlier cluster-plugin release, uninstall the old cluster plugin first and then install v26.6.0 from OperatorHub. In-place conversion from the cluster-plugin delivery model is not supported.

#v1.2.4

Based on openFuyao npu-operator 1.2.0. This release changes the NPU Operator delivery model and adds immutable OS support.

#New and Optimized Features

• Supports Alauda OS and immutable OS nodes.
• Delivers the Ascend driver as a pre-built driver image keyed by HDK, chip, kernel, and OS.
• Supports pre-installed driver passthrough by disabling the Driver component.
• Uses CDI for device injection. New workloads can request the NPU resource directly without runtimeClassName: ascend.
• Adds operator-managed driver upgrade and node reboot workflow.
• Adds chip self-healing workflow controlled by recovery policy.
• Upgrades the MindCluster / Ascend component stack to the v7.3.0 train.
• Creates the NPU Exporter ServiceMonitor automatically.

#Breaking Changes

• Delivery changed from cluster plugin to OLM operator bundle.
• In-place upgrade from v1.1.3 or earlier cluster-plugin releases is not supported.
• The driver is delivered as a container image by default, not as a host .run package.
• Runtime integration is CDI-based. The legacy RuntimeClass path is kept only for compatibility.

#Deprecated and Removed Features

• The v1.1.x DKMS host-side driver installation flow is no longer used by the managed driver path.

#Fixed Issues

• Fixed NPU Exporter ServiceMonitor namespace and selector issues.
• Fixed upgrade validation that could rely only on driver pod readiness.
• Fixed runtime initialization fast-skip behavior on wedged chips.
• Includes community fixes for installation and detection logic on nodes without NPU cards.

#Known Issues

• Driver image availability depends on a tag matching the node kernel. Contact Alauda Customer Support if no tag exists.
• Driver upgrade and chip self-healing can require node reboot.

#Upgrade Notes

To move from v1.1.3 or earlier to v1.2.4, uninstall the old cluster plugin first, then install the v1.2.4 operator from OperatorHub.

#v1.1.3

Based on openFuyao npu-operator 1.1.1. This release uses the older cluster plugin delivery model and the MindCluster / Ascend v7.2.RC1 component stack.