NVIDIA · isVoid · Aug 7, 2026
diff --git a/cuda_core/cuda/core/_device.pyx b/cuda_core/cuda/core/_device.pyx
@@ -1033,9 +1033,12 @@ class Device:
         tuple of Device
             A tuple containing instances of available devices.
         """
-        from cuda.core import system
-        total = system.get_num_devices()
-        return tuple(cls(device_id) for device_id in range(total))
+        return cls._get_all_devices_from_cuda_driver()
+
+    @classmethod
+    def _get_all_devices_from_cuda_driver(cls):
+        Device_ensure_cuda_initialized()
+        return tuple(Device_ensure_tls_devices(cls))
 
     def to_system_device(self) -> 'cuda.core.system.Device':
         """

diff --git a/cuda_core/docs/source/release/1.2.0-notes.rst b/cuda_core/docs/source/release/1.2.0-notes.rst
@@ -73,6 +73,11 @@ Fixes and enhancements
   Windows, both ``ctypes.CFUNCTYPE`` and ``ctypes.WINFUNCTYPE`` are accepted.
   (`#2439 <https://github.com/NVIDIA/cuda-python/issues/2439>`__)
 
+- CUDA device enumeration now queries the CUDA driver rather than using the
+  NVML system-device count. This prevents non-CUDA accelerators, such as an NPU,
+  from being treated as CUDA devices by :meth:`Device.get_all_devices`, examples,
+  and tests.
+
 Deprecation Notices
 -------------------
 

diff --git a/cuda_core/examples/show_device_properties.py b/cuda_core/examples/show_device_properties.py
@@ -15,7 +15,7 @@
 
 import sys
 
-from cuda.core import Device, system
+from cuda.core import Device
 
 
 # Convert boolean to YES or NO string
@@ -219,11 +219,12 @@ def print_device_properties(properties):
 
 # Print info about all CUDA devices in the system
 def show_device_properties():
-    ndev = system.get_num_devices()
+    devices = Device.get_all_devices()
+    ndev = len(devices)
     print(f"Number of GPUs: {ndev}")
 
-    for device_id in range(ndev):
-        device = Device(device_id)
+    for device in devices:
+        device_id = device.device_id
         print(f"DEVICE {device.name} (id={device_id})")
 
         device.set_current()

diff --git a/cuda_core/examples/simple_multi_gpu_example.py b/cuda_core/examples/simple_multi_gpu_example.py
@@ -17,7 +17,7 @@
 
 import cupy as cp
 
-from cuda.core import Device, LaunchConfig, Program, ProgramOptions, launch, system
+from cuda.core import Device, LaunchConfig, Program, ProgramOptions, launch
 
 dtype = cp.float32
 size = 50000
@@ -35,7 +35,7 @@ def __cuda_stream__(self):
 
 
 def main():
-    if system.get_num_devices() < 2:
+    if len(Device.get_all_devices()) < 2:
         print("this example requires at least 2 GPUs", file=sys.stderr)
         sys.exit(1)
 

diff --git a/cuda_core/tests/example_tests/test_basic_examples.py b/cuda_core/tests/example_tests/test_basic_examples.py
@@ -13,7 +13,7 @@
 import pytest
 from cuda_python_test_helpers.pep723 import has_package_requirements_or_skip
 
-from cuda.core import Device, ManagedMemoryResource, system
+from cuda.core import Device, ManagedMemoryResource
 from cuda.core._program import _can_load_generated_ptx
 
 
@@ -22,7 +22,7 @@ def has_compute_capability_9_or_higher() -> bool:
 
 
 def has_multiple_devices() -> bool:
-    return system.get_num_devices() >= 2
+    return len(Device.get_all_devices()) >= 2
 
 
 def has_display() -> bool:

diff --git a/cuda_core/tests/system/test_system_device.py b/cuda_core/tests/system/test_system_device.py
@@ -16,6 +16,7 @@
 import helpers
 import pytest
 
+from cuda.core import Device as CudaDevice
 from cuda.core import system
 from cuda.core.system import typing
 
@@ -132,7 +133,8 @@ def test_numa_node_id():
 
 
 def test_device_cuda_compute_capability():
-    for device in system.Device.get_all_devices():
+    for cuda_device in CudaDevice.get_all_devices():
+        device = cuda_device.to_system_device()
         cuda_compute_capability = device.cuda_compute_capability
         assert isinstance(cuda_compute_capability, tuple)
         assert len(cuda_compute_capability) == 2
@@ -310,7 +312,8 @@ def test_device_brand():
 
 
 def test_device_pci_bus_id():
-    for device in system.Device.get_all_devices():
+    for cuda_device in CudaDevice.get_all_devices():
+        device = cuda_device.to_system_device()
         pci_bus_id = device.pci_info.bus_id
         assert isinstance(pci_bus_id, str)
 
@@ -796,7 +799,8 @@ def test_pstates():
 
 
 def test_compute_running_processes():
-    for device in system.Device.get_all_devices():
+    for cuda_device in CudaDevice.get_all_devices():
+        device = cuda_device.to_system_device()
         with unsupported_before(device, "FERMI"):
             processes = device.compute_running_processes
         assert isinstance(processes, list)

diff --git a/cuda_core/tests/system/test_system_system.py b/cuda_core/tests/system/test_system_system.py
@@ -9,6 +9,7 @@
 from cuda_python_test_helpers.arch_check import skip_if_nvml_unsupported
 
 from cuda.bindings import driver
+from cuda.core import Device as CudaDevice
 from cuda.core import system
 from cuda.core._utils.cuda_utils import handle_return
 
@@ -57,8 +58,9 @@ def test_nvml_version():
 
 @skip_if_nvml_unsupported
 def test_get_process_name():
-    for device in system.Device.get_all_devices():
-        x = device.compute_running_processes
+    for cuda_device in CudaDevice.get_all_devices():
+        device = cuda_device.to_system_device()
+        _ = device.compute_running_processes
 
     try:
         process_name = system.get_process_name(os.getpid())

diff --git a/cuda_core/tests/test_green_context.py b/cuda_core/tests/test_green_context.py
@@ -291,12 +291,12 @@ def test_configure_scope_with_enum(self, wq_resource, scope):
         assert wq_resource.sharing_scope is scope
 
     def test_device_id_matches_source_multi_gpu(self):
-        from cuda.core import Device, system
+        from cuda.core import Device
 
-        if system.get_num_devices() < 2:
+        devices = Device.get_all_devices()
+        if len(devices) < 2:
             pytest.skip("requires 2+ GPUs")
-        dev0 = Device(0)
-        dev1 = Device(1)
+        dev0, dev1 = devices[:2]
         try:
             wq0 = dev0.resources.workqueue
             wq1 = dev1.resources.workqueue

diff --git a/cuda_core/tests/test_memory.py b/cuda_core/tests/test_memory.py
@@ -39,9 +39,6 @@
     VirtualMemoryResource,
     VirtualMemoryResourceOptions,
 )
-from cuda.core import (
-    system as ccx_system,
-)
 from cuda.core._dlpack import DLDeviceType
 from cuda.core._memory._ipc import IPCBufferDescriptor
 from cuda.core._utils.cuda_utils import CUDAError, handle_return
@@ -365,7 +362,7 @@ def test_buffer_external_host():
 
 @pytest.mark.parametrize("change_device", [True, False])
 def test_buffer_external_device(change_device):
-    n = ccx_system.get_num_devices()
+    n = len(Device.get_all_devices())
     if n < 1:
         pytest.skip("No devices found")
     dev_id = n - 1
@@ -389,7 +386,7 @@ def test_buffer_external_device(change_device):
 
 @pytest.mark.parametrize("change_device", [True, False])
 def test_buffer_external_pinned_alloc(change_device):
-    n = ccx_system.get_num_devices()
+    n = len(Device.get_all_devices())
     if n < 1:
         pytest.skip("No devices found")
     dev_id = n - 1
@@ -414,7 +411,7 @@ def test_buffer_external_pinned_alloc(change_device):
 
 @pytest.mark.parametrize("change_device", [True, False])
 def test_buffer_external_pinned_registered(change_device):
-    n = ccx_system.get_num_devices()
+    n = len(Device.get_all_devices())
     if n < 1:
         pytest.skip("No devices found")
     dev_id = n - 1
@@ -447,7 +444,7 @@ def test_buffer_external_pinned_registered(change_device):
 
 @pytest.mark.parametrize("change_device", [True, False])
 def test_buffer_external_managed(change_device):
-    n = ccx_system.get_num_devices()
+    n = len(Device.get_all_devices())
     if n < 1:
         pytest.skip("No devices found")
     dev_id = n - 1

diff --git a/cuda_core/tests/test_memory_peer_access.py b/cuda_core/tests/test_memory_peer_access.py
@@ -6,7 +6,7 @@
 from helpers.collection_interface_testers import assert_single_member_mutable_set_interface
 from helpers.constants import POOL_SIZE
 
-from cuda.core import Device, DeviceMemoryResource, DeviceMemoryResourceOptions, system
+from cuda.core import Device, DeviceMemoryResource, DeviceMemoryResourceOptions
 from cuda.core._memory import _peer_access_utils
 from cuda.core._memory._peer_access_utils import PeerAccessibleBySetProxy
 from cuda.core._utils.cuda_utils import CUDAError
@@ -231,7 +231,7 @@ def test_peer_accessible_by_silently_ignores_owner(isolated_dmr_x2):
 def test_peer_accessible_by_rejects_invalid_inputs(isolated_dmr_x2):
     """``add`` raises on out-of-range/unsupported inputs; lenient methods do not."""
     dmr, dev0, dev1 = isolated_dmr_x2
-    bad_id = system.get_num_devices()  # one past the last valid device ordinal
+    bad_id = len(Device.get_all_devices())  # one past the last valid CUDA device ordinal
 
     # add: validates strictly, propagates errors from Device(bad_id)
     with pytest.raises((ValueError, CUDAError)):

diff --git a/cuda_core/tests/test_object_protocols.py b/cuda_core/tests/test_object_protocols.py
@@ -27,7 +27,6 @@
     LaunchConfig,
     Program,
     Stream,
-    system,
 )
 from cuda.core._program import _can_load_generated_ptx
 from cuda.core.graph import GraphDefinition
@@ -170,7 +169,7 @@ def sample_program_nvvm(init_cuda):
 @pytest.fixture
 def sample_device_alt(init_cuda):
     """An alternate Device object (requires multi-GPU)."""
-    if system.get_num_devices() < 2:
+    if len(Device.get_all_devices()) < 2:
         pytest.skip("requires multi-GPU")
     device_alt = Device(1)
     device_alt.set_current()

diff --git a/cuda_core/tests/test_stream.py b/cuda_core/tests/test_stream.py
@@ -408,9 +408,7 @@ def test_default_stream_per_thread_when_env_set(monkeypatch):
 
 
 def _skip_unless_multi_gpu():
-    from cuda.core import system
-
-    if system.get_num_devices() < 2:
+    if len(Device.get_all_devices()) < 2:
         pytest.skip("requires 2+ GPUs")
 
 

diff --git a/cuda_core/tests/test_tensor_map.py b/cuda_core/tests/test_tensor_map.py
@@ -9,7 +9,6 @@
     Device,
     ManagedMemoryResourceOptions,
     TensorMapDescriptor,
-    system,
 )
 from cuda.core._dlpack import DLDeviceType
 from cuda.core._tensor_map import (
@@ -386,7 +385,7 @@ def test_replace_address_requires_device_accessible(self, dev, skip_if_no_tma):
             desc.replace_address(host_arr)
 
     def test_replace_address_rejects_tensor_from_other_device(self, dev, skip_if_no_tma):
-        if system.get_num_devices() < 2:
+        if len(Device.get_all_devices()) < 2:
             pytest.skip("requires multi-GPU")
 
         dev0 = dev
@@ -407,7 +406,7 @@ def test_replace_address_rejects_tensor_from_other_device(self, dev, skip_if_no_
             desc.replace_address(buf1)
 
     def test_replace_address_accepts_managed_buffer_on_nonzero_device(self, init_cuda):
-        if system.get_num_devices() < 2:
+        if len(Device.get_all_devices()) < 2:
             pytest.skip("requires multi-GPU")
 
         dev1 = Device(1)
@@ -431,7 +430,7 @@ class TestTensorMapMultiDeviceValidation:
     """Test multi-device validation for descriptor creation."""
 
     def test_from_tiled_rejects_tensor_from_other_device(self, init_cuda):
-        if system.get_num_devices() < 2:
+        if len(Device.get_all_devices()) < 2:
             pytest.skip("requires multi-GPU")
 
         dev0 = Device(0)
@@ -451,7 +450,7 @@ def test_from_tiled_rejects_tensor_from_other_device(self, init_cuda):
             )
 
     def test_from_tiled_accepts_managed_buffer_on_nonzero_device(self, init_cuda):
-        if system.get_num_devices() < 2:
+        if len(Device.get_all_devices()) < 2:
             pytest.skip("requires multi-GPU")
 
         dev1 = Device(1)