Skip the accelerator-only CPU optimizer tests on the cpu accelerator

[delock]

TestCPUAdamGPUError and TestCPUAdagradGPUError assert that the CPU optimizer rejects a param living on a discrete accelerator:

device = get_accelerator().device_name(0)  # 'cuda:0' or 'xpu:0'
param = torch.nn.Parameter(torch.randn(model_size, device=device))
...
with pytest.raises(AssertionError):
    optimizer.step()

On the cpu accelerator the param is already on cpu, so the assert cannot fire — and the test never returns. In the first cpu-torch-multi-latest nightly (issue #8834, run https://github.com/deepspeedai/DeepSpeed/actions/runs/38849719748) two of the four xdist workers were stuck on exactly these two tests: they were the only items of the 851-test phase that started and never reported a result, and the run sat at 93-94% until the 120m phase budget killed it.

This marks both tests with an item-level skipif so they are skipped before tests/unit/common.py launches the rank pool, matching the guard TestCPULionGPUError already has:

@pytest.mark.skipif(get_accelerator().device_name() == 'cpu', reason="requires a discrete accelerator device")

Verified on a CPU-only machine (8 vCPU), with DS_MULTIRANK_ONLY=1 LOCAL_SIZE=4: both tests skip in ~2s without launching a rank pool, where the unmarked versions leave pytest wedged.

The shared test harness has related robustness gaps (an unbounded pool.starmap/join in DistributedTest._close_pool, one master port handed to every pool, and /dev/shm allreduce segments the op never unlinks). Those are deliberately left out of this PR; it only stops the nightly from selecting tests that cannot pass on cpu.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-16T09:05:52.591760Z	dfb8393	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.