Purpose

EngineArgs._resolve_device_ids composes --device-ids with CUDA_VISIBLE_DEVICES: when CVD is set, each --device-ids value is treated as an index into the visible set. The range check only guarded the upper bound:

for i in int_ids:
    if i >= len(cvd_ids):
        raise ValueError(f"--device-ids index {i} is out of range ...")
return [cvd_ids[i] for i in int_ids]

A negative index (for example --device-ids -1) makes i >= len(cvd_ids) evaluate to False, so no error is raised and cvd_ids[i] silently negative-indexes into the list, resolving to a valid-but-wrong physical device. With CUDA_VISIBLE_DEVICES=4,5 and --device-ids -1, the method returns physical device [5] instead of rejecting the input.

This completes the bounds check to not 0 <= i < len(cvd_ids) so an out-of-range negative index raises the same "out of range" ValueError as a too-large one, matching the intent of the existing test_device_ids_with_cvd_out_of_range test.

This changes only input validation (silent mis-selection becomes a clear error); it does not affect model output, accuracy, or serving numerics, so no model-evaluation results apply.

Not a duplicate: the only open PRs touching this area, #46132 and #41776, add MIG/UUID parsing for CUDA_VISIBLE_DEVICES; neither changes the _resolve_device_ids index bounds check. No open issue tracks this.

Test Plan

Added tests/engine/test_arg_utils.py::TestDeviceIds::test_device_ids_with_cvd_negative_index, mirroring the existing test_device_ids_with_cvd_out_of_range:

pytest tests/engine/test_arg_utils.py::TestDeviceIds -q

Test Result

_resolve_device_ids does not load a model, so I validated the exact logic on CPU by executing the method against a stubbed platform (a full CUDA build was not available on my machine); the added test runs in CI:

• Before: EngineArgs(model="m", device_ids=[-1]) with CUDA_VISIBLE_DEVICES=4,5 returned [5] (no error raised).
• After: it raises ValueError: --device-ids index -1 is out of range for CUDA_VISIBLE_DEVICES=4,5 (2 devices visible).
• Unchanged for valid input: device_ids=[0, 1] still resolves to [4, 5].

AI assistance disclosure: AI assistance was used to help identify and prepare this change, per the repository's contribution policy. The fix is a single bounds-check completion plus a regression test, and I have reviewed every changed line.