Feature Overview
This feature enables support for NVIDIA time-sliced GPUs (vGPUs) by integrating
Cyborg with Nova’s mediated device (mdev) framework. The primary goal is to
transition vGPU management from Nova-native logic to a Cyborg-managed model.
This provides a consistent lifecycle for hardware accelerators and allows for
more granular resource scheduling via Placement. By supporting mdevs in Cyborg,
operators can leverage NVIDIA’s time-slicing capabilities for multi-tenant
workloads while maintaining a unified API for all accelerator types.

Goals

• Finalize Nova Generic MDEV Testing: Complete the validation of the generic
  mdev implementation within Nova to ensure a stable foundation for
  cross-component handoffs.
• Replicate MDEV Management in Cyborg: Port and adapt the mdev discovery,
  creation, and inventory management logic from Nova into Cyborg’s
  driver framework.
• Cross-Component Integration: Enable Nova to request and attach vGPUs
  specifically managed by Cyborg, ensuring proper handling of mdev
  types and UUIDs.
• Parity and Extension: Achieve functional parity with Nova's legacy vGPU
  support while adding Cyborg's advanced device profiling and
  metadata capabilities.

Requirements

Done - Acceptance Criteria

• A Virtual Machine successfully boots in Nova using a vGPU resource
  allocated and managed by Cyborg.
• The "NVIDIA time-sliced" mdev types are correctly discovered by Cyborg
  and visible in the Cyborg API.
• Nova’s libvirt driver correctly generates the domain XML for Cyborg-managed
  mdev devices, including the correct UUID and parent address.
• The integrated testing suite confirms that mdev resources are correctly
  cleaned up in both Cyborg and Nova upon instance termination.
• Successful replication of the Nova generic mdev test suite results
  within the Cyborg CI environment.

Use Cases - i.e. User Experience & Workflow:

• Time-Sliced GPU Allocation: A user requests a "small" GPU profile via a
  Cyborg device profile; the system selects a physical NVIDIA GPU and
  creates a specific time-sliced vGPU (mdev) for that instance.
• Shared Hardware Management: An administrator manages a pool of NVIDIA
  GPUs via Cyborg, adjusting the supported mdev types without
  reconfiguring Nova.

Out of Scope

• Support for non-mdev based GPU virtualization (e.g., direct PCI
  passthrough is handled by existing paths).
• Multi-Instance GPU (MIG) dynamic partitioning (this focus is strictly
  on time-slicing/mdev).

Documentation Considerations

• Reference Spec: https://review.opendev.org/c/openstack/nova-specs/+/967515
• Update the Cyborg Driver documentation to include NVIDIA-specific
  mdev configuration.
• Provide a guide for migrating vGPU workloads from Nova-managed to
  Cyborg-managed vGPUs.

Questions to Answer

• How will Cyborg handle the naming conflicts if both Nova and Cyborg
  attempt to manage mdevs on the same host during a transition period?
• Does the replication of Nova's mdev logic require new dependencies
  in Cyborg's requirements.txt?

Background and Strategic Fit
This feature aligns with the 2026.1 "cyborg-vgpu-support" approved spec.
It is a critical step in de-coupling hardware-specific logic from Nova,
moving toward the "Cyborg-as-the-source-of-truth" model for all
pluggable accelerators.

Customer Considerations
Customers utilizing NVIDIA vGPU software will need to ensure their host
drivers are compatible with the mdev bus before switching management to
Cyborg.

Team Sign Off