[PerryLink]

The defect

_ibv_get_device_list() calls ctypes.CDLL("libibverbs.so.1") unguarded (checkpoint_engine/device_utils.py:51). On any host without rdma-core installed — a slim container, a CI image, any machine with no RDMA NIC — the loader error escapes from this helper and reaches the caller instead of the error the code already raises for this case:

DeviceManager.rdma_device()            device_utils.py:267
  -> _get_my_rdma_device(..., _get_rdma_devices())   :271
       -> _ibv_get_device_list()       :82 -> :51
            -> ctypes.CDLL("libibverbs.so.1")

Measured on a Linux host without rdma-core:

$ python3 -c "import ctypes; ctypes.CDLL('libibverbs.so.1')"
OSError: libibverbs.so.1: cannot open shared object file: No such file or directory

So the user sees the dynamic loader's message, while _get_my_rdma_device() already has the actionable one ready at :89-90:

if not devices:
    raise RuntimeError("no rdma devices found")

That line is unreachable on such a host, because the exception is raised two frames earlier. A missing library is not an unexpected condition here — it means exactly what devices == [] means: there are no RDMA devices to enumerate.

The change

Two hunks in checkpoint_engine/device_utils.py, plus three tests:

1. _ibv_get_device_list() treats a missing libibverbs as an empty device list, so _get_my_rdma_device() reports no rdma devices found.
2. _get_rdma_devices() enumerates the HCA list once. When NCCL_IB_HCA parses to nothing, the or fallback called _ibv_get_device_list() a second time — a second CDLL and a second full enumeration for the same answer.

No behaviour change on a host where the library is present: _ibv_get_device_list() returns the same list as before.

Verification

Both new tests are written to fail against the unpatched source, so they detect the change rather than passing vacuously. Reverting only device_utils.py and keeping the tests:

                     with the patch        with the source reverted
new tests            3 passed              3 failed

and the third failure is the second hunk directly:

AssertionError: Expected '_ibv_get_device_list' to have been called once. Called 2 times.
Calls: [call(), call()].

The repository's own CPU gate is unchanged:

$ python -m pytest tests/ -m "not gpu" -q
2 failed, 108 passed, 1 skipped, 12 deselected

The two failures are pre-existing and unrelated to this change — they are Windows-only host artefacts I hit while working here (tests/test_store_barrier.py:34 relies on a libuv-enabled PyTorch, which the Windows wheel is not; tests/test_xpu_parity.py:84 asserts : as the path separator where Windows joins with ;). Both are inert on ubuntu-latest, and the passed count moved from 105 to 108 with these three tests added.

ruff check and ruff format --check are clean across checkpoint_engine/ and tests/ with the repository's own configuration.

What I could not verify

I have no RDMA hardware, no CUDA device, and no multi-node setup, so I did not exercise P2PStore or the P2P update path end to end — only the device-enumeration helper and its two callers, which is where the change is. The _ibv_get_device_list() path with the library present is covered by the existing test_detect_ibv_list, which skips without /sys/class/infiniband.