Skip to content

cuda.core: VirtualMemoryResource host locations ignore device_id and report is_device_accessible=False with peers #3466

Open

Description

@Andy-Jost

Follow-up to #2910, which is now closed: its other items are fixed (#2917, #3149, #3343). Two gaps remain for host locations (location_type="host" or "host_numa"). Both were checked on main at ba7dcd0.

  1. device_id is validated and then ignored. VirtualMemoryResource.__init__ builds Device(device_id) and then sets self.device = None for host locations. The argument is required and must name a valid device, but it has no effect.
  2. A host resource with peers reports is_device_accessible == False. The resource maps its memory for each device in peers, but is_device_accessible is true only for location_type="device".

Both are small. Each needs a decision on the intended behavior before a fix.

Activity

  1. added this to the cuda.core next milestone on Nov 15, 2026
  2. added
    bugSomething isn't working
    P2Low priority - Nice to have
    on Nov 15, 2026
  3. self-assigned this
    on Nov 15, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

P2Low priority - Nice to havebugSomething isn't workingcuda.coreEverything related to the cuda.core module

Type

No type

Projects

No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions