Fix nvbug6563848: Make tests properly skip on unsupported devices. (#2742)

* Fix nvbug6563848: Make tests properly skip on unsupported devices.

* Update cuda_bindings/tests/nvml/test_cuda.py

* test(nvml): skip unsupported Orin UUID lookups

---------

Co-authored-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>

Author: mdboom

cuda_bindings/tests/nvml/conftest.py

@@ -1,7 +1,6 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-from collections import namedtuple
 
 import pytest
 from cuda_python_test_helpers.arch_check import unsupported_before  # noqa: F401
@@ -26,56 +25,19 @@ def nvml_init():
         yield
 
 
-@pytest.fixture(scope="session", autouse=True)
-def device_info():
-    dev_count = None
-    bus_id_to_board_details = {}
-
-    with NVMLInitializer():
-        dev_count = nvml.device_get_count_v2()
-
-        # Store some details for each device now when we know NVML is in known state
-        for i in range(dev_count):
-            try:
-                dev = nvml.device_get_handle_by_index_v2(i)
-            except nvml.NoPermissionError:
-                continue
-            pci_info = nvml.device_get_pci_info_v3(dev)
-
-            name = nvml.device_get_name(dev)
-            # Get architecture name ex: Ampere, Kepler
-            arch_id = nvml.device_get_architecture(dev)
-
-            BoardCfg = namedtuple("BoardCfg", "name, ids_arr")
-            board = BoardCfg(name, ids_arr=[(pci_info.pci_device_id, pci_info.pci_sub_system_id)])
-
-            try:
-                serial = nvml.device_get_serial(dev)
-            except nvml.NvmlError:
-                serial = None
-
-            bus_id = pci_info.bus_id
-            device_id = pci_info.device_
-            uuid = nvml.device_get_uuid(dev)
-
-            BoardDetails = namedtuple("BoardDetails", "name, board, arch_id, bus_id, device_id, serial")
-            bus_id_to_board_details[uuid] = BoardDetails(name, board, arch_id, bus_id, device_id, serial)
-
-    return bus_id_to_board_details
-
-
-def get_devices(device_info):
-    for uuid in list(device_info.keys()):
+def get_devices():
+    dev_count = nvml.device_get_count_v2()
+    for i in range(dev_count):
         try:
-            yield nvml.device_get_handle_by_uuid(uuid)
+            yield nvml.device_get_handle_by_index_v2(i)
         except nvml.NoPermissionError:
             continue  # ignore devices that can't be accessed
 
 
 @pytest.fixture
-def all_devices(device_info):
+def all_devices():
     with NVMLInitializer():
-        yield sorted(set(get_devices(device_info)))
+        yield sorted(set(get_devices()))
 
 
 @pytest.fixture

cuda_bindings/tests/nvml/test_cuda.py

@@ -58,8 +58,9 @@ def test_cuda_device_order():
     cuda_devices = get_cuda_device_names()
     nvml_devices = get_nvml_device_names()
 
-    if any("Thor" in device["name"] for device in nvml_devices):
-        pytest.skip("Skipping test on Thor, which has non-standard device naming")
+    for kind in ("Orin", "Thor"):
+        if any(kind in device["name"] for device in nvml_devices):
+            pytest.skip(f"Skipping test on {kind}, which has non-standard device naming")
         return
 
     if "CUDA_VISIBLE_DEVICES" not in os.environ:

cuda_bindings/tests/nvml/test_device.py

@@ -63,11 +63,15 @@ def test_grid_licensable_features(all_devices):
             nvml.GridLicenseExpiry(feature.license_expiry)
 
 
-def test_get_handle_by_uuidv(all_devices):
+def test_get_handle_by_uuidv(all_devices, subtests):
     for device in all_devices:
-        uuid = nvml.device_get_uuid(device)
-        new_handle = nvml.device_get_handle_by_uuidv(nvml.UUIDType.ASCII, uuid.encode("ascii"))
-        assert new_handle == device
+        with subtests.test(device_index=nvml.device_get_index(device)):
+            uuid = nvml.device_get_uuid(device)
+            if "Orin" in nvml.device_get_name(device) and len(uuid) == 36:
+                pytest.skip("UUID lookup is unsupported on Orin, which reports a UUID without a GPU- prefix")
+            with unsupported_before(device, None):
+                new_handle = nvml.device_get_handle_by_uuidv(nvml.UUIDType.ASCII, uuid.encode("ascii"))
+            assert new_handle == device
 
 
 def test_get_nv_link_supported_bw_modes(all_devices, subtests):

cuda_bindings/tests/nvml/test_pynvml.py

@@ -52,9 +52,13 @@ def test_device_get_attributes(mig_handles):
         pytest.skip("No MIG devices found")
 
 
-def test_device_get_handle_by_uuid(ngpus, uuids):
-    handles = [nvml.device_get_handle_by_uuid(uuids[i]) for i in range(ngpus)]
-    assert len(handles) == ngpus
+def test_device_get_handle_by_uuid(ngpus, handles, uuids, subtests):
+    for i in range(ngpus):
+        with subtests.test(device_index=i):
+            uuid = uuids[i]
+            if "Orin" in nvml.device_get_name(handles[i]) and len(uuid) == 36:
+                pytest.skip("UUID lookup is unsupported on Orin, which reports a UUID without a GPU- prefix")
+            assert nvml.device_get_handle_by_uuid(uuid) == handles[i]
 
 
 def test_device_get_handle_by_pci_bus_id(ngpus, pci_info):
@@ -68,7 +72,7 @@ def test_device_get_memory_affinity(handles, scope, subtests):
     size = 1024
     for device_index, handle in enumerate(handles):
         with subtests.test(device_index=device_index):
-            with unsupported_before(handle, nvml.DeviceArch.KEPLER):
+            with unsupported_before(handle, None):
                 node_set = nvml.device_get_memory_affinity(handle, size, scope)
             assert node_set is not None
             assert len(node_set) == size
@@ -80,7 +84,7 @@ def test_device_get_cpu_affinity_within_scope(handles, scope, subtests):
     size = 1024
     for device_index, handle in enumerate(handles):
         with subtests.test(device_index=device_index):
-            with unsupported_before(handle, nvml.DeviceArch.KEPLER):
+            with unsupported_before(handle, None):
                 cpu_set = nvml.device_get_cpu_affinity_within_scope(handle, size, scope)
             assert cpu_set is not None
             assert len(cpu_set) == size