Skip to content

test(e2e): nightly real-hardware check for masked GPU index refusal #730

Open

Description

@rominf

Once #517 merges, the rocm serve scenarios that refuse a GPU index hidden by a visibility mask (serve-11, -13, -15, -16, -19, -20, -21) run on simulated machines on every PR. That keeps them on the required hosted lane and off the GPU queues.

The cost: no lane, nightly included, checks that refusal against a real amd-smi / rocminfo and a real HIP_VISIBLE_DEVICES / ROCR_VISIBLE_DEVICES mask. #517's two @nightly drift scenarios check the sysfs/KFD/lspci layout the simulated machine imitates. They do not check how the tools report devices once a mask is set.

Proposal: add one @nightly @requires-real-gpu scenario that sets a visibility mask hiding a real device, asks rocm serve for that index, and asserts both the refusal message and that no engine process started. That brings real-tool coverage back on the nightly lanes without slowing PR merges.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions