Adapt CUDA 12 NVML tests for current CI runners

Relocate the established hardware, driver, visibility and timing accommodations from the source-integration layer. Preserve their contents exactly; the FieldId correction moves with the test maintenance.

Signed-off-by: Ralf W. Grosse-Kunstleve <rgrossekunst@nvidia.com>

Author: rwgk

cuda_bindings_12/tests/nvml/test_cuda.py

@@ -1,6 +1,8 @@
 # SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
+import pytest
+
 import cuda.bindings.driver as cuda
 from cuda.bindings import nvml
 
@@ -15,10 +17,15 @@ def get_nvml_device_names():
         for idx in range(num_devices):
             handle = nvml.device_get_handle_by_index_v2(idx)
             name = nvml.device_get_name(handle)
-            info = nvml.device_get_pci_info_v3(handle)
-            assert isinstance(info.bus, int)
+            try:
+                info = nvml.device_get_pci_info_v3(handle)
+            except nvml.NotSupportedError:
+                bus_id = -1
+            else:
+                bus_id = info.bus
+            assert isinstance(bus_id, int)
             assert isinstance(name, str)
-            result.append({"name": name, "id": info.bus})
+            result.append({"name": name, "id": bus_id})
 
     return result
 
@@ -54,4 +61,15 @@ def test_cuda_device_order():
     cuda_devices = get_cuda_device_names()
     nvml_devices = get_nvml_device_names()
 
-    assert cuda_devices == nvml_devices, "CUDA and NVML device lists do not match"
+    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")
+
+    def compare(cuda_device, nvml_device):
+        return cuda_device["name"] == nvml_device["name"] and (
+            nvml_device["id"] == -1 or cuda_device["id"] == nvml_device["id"]
+        )
+
+    assert len(cuda_devices) <= len(nvml_devices)
+    for cuda_device in cuda_devices:
+        assert any(compare(cuda_device, nvml_device) for nvml_device in nvml_devices)

cuda_bindings_12/tests/nvml/test_init.py

@@ -6,7 +6,15 @@
 
 import pytest
 
-from cuda.bindings import nvml
+from cuda.bindings import driver, nvml
+
+
+def _driver_version_at_least_13_4():
+    (status,) = driver.cuInit(0)
+    assert status == driver.CUresult.CUDA_SUCCESS
+    status, version = driver.cuDriverGetVersion()
+    assert status == driver.CUresult.CUDA_SUCCESS
+    return version >= 13040
 
 
 def assert_nvml_is_initialized():
@@ -36,10 +44,14 @@ def test_devices_are_the_same_architecture(all_devices):
 
 
 @pytest.mark.skipif(sys.platform == "win32", reason="Test not supported on Windows")
+@pytest.mark.thread_unsafe(reason="nvml init affects other threads")
 def test_init_ref_count():
     """
     Verifies that we can call NVML shutdown and init(2) multiple times, and that ref counting works
     """
+    if _driver_version_at_least_13_4():
+        pytest.skip("Init behavior changed in CUDA 13.4")
+
     with pytest.raises(nvml.UninitializedError):
         nvml.shutdown()
 

cuda_bindings_12/tests/nvml/test_pynvml.py

@@ -4,8 +4,6 @@
 # A set of tests ported from https://github.com/gpuopenanalytics/pynvml/blob/11.5.3/pynvml/tests/test_nvml.py
 
 import os
-import time
-
 import pytest
 
 from cuda.bindings import nvml
@@ -147,22 +145,6 @@ def test_device_get_power_usage(ngpus, handles):
         assert power_mwatts >= 0.0
 
 
-def test_device_get_total_energy_consumption(ngpus, handles):
-    for i in range(ngpus):
-        with unsupported_before(handles[i], nvml.DeviceArch.VOLTA):
-            energy_mjoules1 = nvml.device_get_total_energy_consumption(handles[i])
-
-        for j in range(10):  # idle for 150 ms
-            time.sleep(0.015)  # and check for increase every 15 ms
-            with unsupported_before(handles[i], nvml.DeviceArch.VOLTA):
-                energy_mjoules2 = nvml.device_get_total_energy_consumption(handles[i])
-            assert energy_mjoules2 >= energy_mjoules1
-            if energy_mjoules2 > energy_mjoules1:
-                break
-        else:
-            raise AssertionError("energy did not increase across 150 ms interval")
-
-
 # [Skipping] pynvml.nvmlDeviceGetGpuOperationMode
 # [Skipping] pynvml.nvmlDeviceGetCurrentGpuOperationMode
 # [Skipping] pynvml.nvmlDeviceGetPendingGpuOperationMode

cuda_bindings_12/tests/nvml/util.py

@@ -43,5 +43,5 @@ def supports_ecc(device):
 
 def supports_nvlink(device):
     fields = nvml.FieldValue(1)
-    fields[0].field_id = nvml.FI.DEV_NVLINK_GET_STATE
+    fields[0].field_id = nvml.FieldId.DEV_NVLINK_GET_STATE
     return nvml.device_get_field_values(device, fields)[0].nvml_return == nvml.Return.SUCCESS