Skip to main content

AIsware v25.1.6

· 2 min read
Midokura Team
Midokura Team
Midokura Team

Version 25.1.6 of AIsware is now available.

Overview​

This release adds AMD ROCm GPU support, improves GPU scheduling and DNS reliability, and gives operators new tools to manage clusters.

What's new​

AMD ROCm GPU images​

AIsware now supports AMD GPU workloads. Provisioning uploads a new ubuntu-rocm image (ROCm 7.2) to Glance automatically. A gpu-optimized flavor row and a ROCm inventory example are added to the IaaS Console guide.

Ceph dashboard and metrics user​

Provisioning now configures the Ceph dashboard and creates a dedicated metrics user. This gives operators direct access to storage health and metrics.

GPU scheduling fix​

The scheduler failed with No valid host was found when a host was already in a Nova aggregate but missing from Placement. Provisioning now clears and re-adds all hosts to force a Placement re-sync. This prevents the Nova and Placement aggregate desync.

GPU availability diagnostic script​

A new gpu-status.sh script shows, for each GPU server, which physical GPU cards are free or in use. For each card in use, the script also shows the VM that holds it. Run the script inside the Ansible container shell.

DNS reliability improvements​

Two DNS problems are fixed:

  • The dnsmasq concurrent query limit is raised from 150 to 750. The default limit was saturated under load, which caused random DNS failures across the cluster.
  • Designate zone forwarding now points to the internal VIP instead of each controller host. This removes duplicate forwarding entries.

BGP dummy device support​

The openwrt_configure role now supports a dummy network device for the BGP interface. This replaces the previous veth workaround with a single always-up interface. The kmod-dummy module must be baked into the OpenWrt image at build time.

Tenant user provisioning​

A new script creates a per-tenant user and assigns the member role on the specified project. The script generates a password automatically and is idempotent.

Updated VM images​

The VPN, Kubernetes CAPI, and GPU-optimized images are updated to the 2026-07-14-1101016 build.

Operational tooling​

The following tools help operators manage clusters more safely:

  • Safe stop and cold start for a Phoenix OpenStack cluster.
  • Snapshot create and start scripts for a faster boot path.