Symptoms

When you configure compute for GPU passthrough and create a flavor: 

# openstack --insecure flavor create --ram 16384 --vcpus 8 --property "pci_passthrough:alias"="gpu:1" --public gpu-flavor

And create a VM without GPU. Then shut down the VM make offline resize to flavor with GPU alias and check status VM. The status VM became to 'Error' state, an error appears 'PCI alias gpu is not defined'
Error text in nova-scheduler log

"Exhausted all hosts available for retrying build failures for instance "

Documentation: https://docs.virtuozzo.com/virtuozzo_hybrid_infrastructure_4_7_admins_cmd_guide/index.html#creating-virtual-machines-with-physical-gpu.html 

Cause

The behavior is a consequence of VSTOR-47788. Alias is missing from the nova-compute config of the compute node (for which alias is defined in the backend), but present in nova-compute and nova-api config on the controller node.

Compute node

# grep alias /etc/kolla/nova-compute/nova.conf 
#

Backend(controller) nodes

# grep alias /etc/kolla/nova-api/nova.conf 
alias = {"vendor_id": "10de", "product_id": "1eb8", "device_type": "ANY", "name": "gpu"}
# grep alias /etc/kolla/nova-compute/nova.conf 
alias = {"vendor_id": "10de", "product_id": "1eb8", "device_type": "ANY", "name": "gpu"}

To resize guest with PCI device the PCI alias on the compute node should be configured as well.

Resolution

After the cluster upgrade to 4.7.1 manually call if you use PCI passthrough with generic devices:

# vinfra service compute set --pci-passthrough-config your_config.yaml