Skip NVML tests if the hardware doesn't support NVML (#1585)

* Skip NVML tests if the hardware doesn't support NVML

* Also skip converting between Device types

* Fix test_to_system_device test

* Update cuda_bindings/cuda/bindings/_test_helpers/arch_check.py

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Fix test

* Simplify check

* Fix import when cuda_bindings is incompatible

* Fix logic

* Move over a single test from #1583

* Fix check

---------

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Author: mdboom

cuda_bindings/cuda/bindings/_test_helpers/arch_check.py

@@ -3,11 +3,30 @@
 
 
 from contextlib import contextmanager
+from functools import cache
 
 import pytest
 from cuda.bindings import nvml
 
 
+@cache
+def hardware_supports_nvml():
+    """
+    Tries to call the simplest NVML API possible to see if just the basics
+    works.  If not we are probably on one of the platforms where NVML is not
+    supported at all (e.g. Jetson Orin).
+    """
+    nvml.init_v2()
+    try:
+        nvml.system_get_driver_branch()
+    except (nvml.NotSupportedError, nvml.UnknownError):
+        return False
+    else:
+        return True
+    finally:
+        nvml.shutdown()
+
+
 @contextmanager
 def unsupported_before(device: int, expected_device_arch: nvml.DeviceArch | str | None):
     device_arch = nvml.device_get_architecture(device)

cuda_bindings/tests/nvml/__init__.py

@@ -1,2 +1,9 @@
-# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
+
+
+import pytest
+from cuda.bindings._test_helpers.arch_check import hardware_supports_nvml
+
+if not hardware_supports_nvml():
+    pytest.skip("NVML not supported on this platform", allow_module_level=True)

cuda_core/tests/system/conftest.py

@@ -1,13 +1,23 @@
-# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # SPDX-License-Identifier: Apache-2.0
 
 
 import pytest
 from cuda.core import system
 
+SHOULD_SKIP_NVML_TESTS = not system.CUDA_BINDINGS_NVML_IS_COMPATIBLE
+
+
+if system.CUDA_BINDINGS_NVML_IS_COMPATIBLE:
+    from cuda.bindings._test_helpers.arch_check import hardware_supports_nvml
+
+    SHOULD_SKIP_NVML_TESTS |= not hardware_supports_nvml()
+
+
 skip_if_nvml_unsupported = pytest.mark.skipif(
-    not system.CUDA_BINDINGS_NVML_IS_COMPATIBLE, reason="NVML support requires cuda.bindings version 12.9.6+ or 13.1.2+"
+    SHOULD_SKIP_NVML_TESTS,
+    reason="NVML support requires cuda.bindings version 12.9.6+ or 13.1.2+, and hardware that supports NVML",
 )
 
 

cuda_core/tests/test_device.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
 try:
@@ -35,6 +35,11 @@ def test_to_system_device(deinit_cuda):
             device.to_system_device()
         pytest.skip("NVML support requires cuda.bindings version 12.9.6+ or 13.1.2+")
 
+    from cuda.bindings._test_helpers.arch_check import hardware_supports_nvml
+
+    if not hardware_supports_nvml():
+        pytest.skip("NVML not supported on this platform")
+
     from cuda.core.system import Device as SystemDevice
 
     system_device = device.to_system_device()

cuda_core/tests/test_graph_mem.py

@@ -1,4 +1,4 @@
-# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 
@@ -13,6 +13,7 @@
     ProgramOptions,
     launch,
 )
+from cuda.core._utils.cuda_utils import CUDAError
 from helpers import IS_WINDOWS, IS_WSL
 from helpers.buffers import compare_buffer_to_constant, make_scratch_buffer, set_buffer
 
@@ -166,7 +167,12 @@ def test_graph_alloc_with_output(mempool_device, mode):
     out.copy_from(in_, stream=gb)
     launch(gb, LaunchConfig(grid=1, block=1), add_one, out, NBYTES)
     options = GraphCompleteOptions(auto_free_on_launch=True)
-    graph = gb.end_building().complete(options)
+    try:
+        graph = gb.end_building().complete(options)
+    except CUDAError as exc:
+        if "CUDA_ERROR_INVALID_VALUE" in str(exc):
+            pytest.skip("auto_free_on_launch not supported on this platform")
+        raise
 
     # Launch the graph. The output buffer is allocated and set to one.
     graph.upload(stream)