Tests that pass without testing anything

[haanjack]

Collected while auditing what CI actually verifies. Each of these reports as a pass.

test_lora_async.py::test_chunked_vs_non_chunked_tp2 skips itself everywhere. tests/integration/lora/test_lora_async.py:95-98:

def _run_chunking_test(self, tp_size=1):
    if tp_size > 1:
        pytest.skip("TP > 1 requires torchrun")

The guard checks only the parameter, never os.environ / dist.is_initialized() / GPU count — so it skips even when correctly launched under torchrun --nproc_per_node=2, which is exactly how the NP2 list runs this file. Dead since 2026-04-12. Its siblings test_lora_checkpoint.py and test_lora_correctness.py guard correctly.

test_ep_plus_tp_combined returns instead of skipping. tests/multi_gpu/test_expert_parallel.py:429-436 guards with if world_size < 4: print(...); return. Nothing in scripts/ or .github/workflows/ ever launches --nproc_per_node=4, so it always returns before its assertions — and because it is a bare return rather than pytest.skip, it is counted as a pass, invisible in the summary. Zero EP+TP combined-sharding coverage since 2026-02-18.

assert True after a real distributed optimizer step. tests/multi_gpu/test_distributed_optimizer.py:114, test_step_with_ddp: runs a full DDP forward/backward/step, then asserts True with the comment "verify no errors occurred" — strictly weaker than no assertion at all, since pytest already fails on exceptions. Its sibling test_parameter_consistency_after_step does this properly.

test_holistic_summary. tests/integration/test_integration.py:486-524 — bare assert True after printing a summary. The docstring is honest that it is decorative; listing it so it is not mistaken for coverage.

Lower priority, but worth correcting while nearby:

• tests/multi_gpu/test_grad_norm.py:405-436 test_param_norm_ep2_moe only checks torch.isfinite, unlike its DP2 sibling which cross-rank-compares values. The code comment is honest; the name is not.
• tests/multi_gpu/test_grad_norm.py:13,325-328 carries a stale comment claiming EP2 tests are skipped for a known MoE backward bug. No skip mechanism exists and the bug was fixed in fix: resolve MoE backward pass contiguity error in EP dispatch #31. A future real failure could be waved off as "known".

[haanjack]

The single-GPU items are done. #97 fixed the always-true skip in test_lora_async, the bare return in test_ep_plus_tp_combined, and the assert True in test_step_with_ddp; #110 deleted test_holistic_summary, whose hardcoded manifest had also gone stale.

What is left needs two GPUs and lives in tests/multi_gpu/: test_param_norm_ep2_moe only checks torch.isfinite where its DP2 sibling cross-rank-compares values, and the stale comment in test_grad_norm.py claims EP2 tests are skipped for a MoE backward bug that was fixed in #31 — no skip mechanism exists, so a future real failure could be waved off as known.

Neither reproduces here: RCCL refuses two ranks on one device, and test_grad_norm.py is one of the five files that hang under the gloo workaround. Per #107 nothing currently serves the mp pool either.