Reorganize graph tests and add device-side launch support

- Reorganize graph tests into tests/graph/ subdirectory:
  - test_basic.py: basic graph construction tests
  - test_conditional.py: conditional node tests (if, if-else, switch, while)
  - test_advanced.py: child graphs, update, stream lifetime
  - test_options.py: debug print, complete options, build mode
  - test_capture_alloc.py: graph memory resource tests (moved from test_graph_mem.py)
  - test_device_launch.py: new device-side graph launch tests

- Add Graph.handle property to expose CUgraphExec handle, consistent with
  Stream.handle and Event.handle patterns

- Add tests/helpers/graph_kernels.py with shared kernel compilation helpers

- Device-side graph launch tests require Hopper (sm_90+) architecture

Author: Andy-Jost

cuda_core/cuda/core/_graph.py

@@ -746,6 +746,18 @@ def close(self):
         """Destroy the graph."""
         self._mnff.close()
 
+    @property
+    def handle(self) -> driver.CUgraphExec:
+        """Return the underlying ``CUgraphExec`` object.
+
+        .. caution::
+
+            This handle is a Python object. To get the memory address of the underlying C
+            handle, call ``int()`` on the returned object.
+
+        """
+        return self._mnff.graph
+
     def update(self, builder: GraphBuilder):
         """Update the graph using new build configuration from the builder.
 

cuda_core/tests/graph/test_advanced.py

@@ -0,0 +1,186 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
+
+"""Advanced graph feature tests (child graphs, update, stream lifetime)."""
+
+import numpy as np
+import pytest
+from cuda.core import Device, LaunchConfig, LegacyPinnedMemoryResource, launch
+from helpers.graph_kernels import compile_common_kernels, compile_conditional_kernels
+
+
+@pytest.mark.skipif(tuple(int(i) for i in np.__version__.split(".")[:2]) < (2, 1), reason="need numpy 2.1.0+")
+def test_graph_child_graph(init_cuda):
+    mod = compile_common_kernels()
+    add_one = mod.get_kernel("add_one")
+
+    # Allocate memory
+    launch_stream = Device().create_stream()
+    mr = LegacyPinnedMemoryResource()
+    b = mr.allocate(8)
+    arr = np.from_dlpack(b).view(np.int32)
+    arr[0] = 0
+    arr[1] = 0
+
+    # Capture the child graph
+    gb_child = Device().create_graph_builder().begin_building()
+    launch(gb_child, LaunchConfig(grid=1, block=1), add_one, arr[1:].ctypes.data)
+    launch(gb_child, LaunchConfig(grid=1, block=1), add_one, arr[1:].ctypes.data)
+    launch(gb_child, LaunchConfig(grid=1, block=1), add_one, arr[1:].ctypes.data)
+    gb_child.end_building()
+
+    # Capture the parent graph
+    gb_parent = Device().create_graph_builder().begin_building()
+    launch(gb_parent, LaunchConfig(grid=1, block=1), add_one, arr.ctypes.data)
+
+    ## Add child
+    try:
+        gb_parent.add_child(gb_child)
+    except NotImplementedError as e:
+        with pytest.raises(
+            NotImplementedError,
+            match="^Launching child graphs is not implemented for versions older than CUDA 12",
+        ):
+            raise e
+        gb_parent.end_building()
+        b.close()
+        pytest.skip("Launching child graphs is not implemented for versions older than CUDA 12")
+
+    launch(gb_parent, LaunchConfig(grid=1, block=1), add_one, arr.ctypes.data)
+    graph = gb_parent.end_building().complete()
+
+    # Parent updates first value, child updates second value
+    assert arr[0] == 0
+    assert arr[1] == 0
+    graph.launch(launch_stream)
+    launch_stream.sync()
+    assert arr[0] == 2
+    assert arr[1] == 3
+
+    # Close the memory resource now because the garbage collected might
+    # de-allocate it during the next graph builder process
+    b.close()
+
+
+@pytest.mark.skipif(tuple(int(i) for i in np.__version__.split(".")[:2]) < (2, 1), reason="need numpy 2.1.0+")
+def test_graph_update(init_cuda):
+    mod = compile_conditional_kernels(int)
+    add_one = mod.get_kernel("add_one")
+
+    # Allocate memory
+    launch_stream = Device().create_stream()
+    mr = LegacyPinnedMemoryResource()
+    b = mr.allocate(12)
+    arr = np.from_dlpack(b).view(np.int32)
+    arr[0] = 0
+    arr[1] = 0
+    arr[2] = 0
+
+    def build_graph(condition_value):
+        # Begin capture
+        gb = Device().create_graph_builder().begin_building()
+
+        # Add Node A (sets condition)
+        handle = gb.create_conditional_handle(default_value=condition_value)
+
+        # Add Node B (while condition)
+        try:
+            gb_case = list(gb.switch(handle, 3))
+        except Exception as e:
+            with pytest.raises(RuntimeError, match="^(Driver|Binding) version"):
+                raise e
+            gb.end_building()
+            raise e
+
+        ## Case 0
+        gb_case[0].begin_building()
+        launch(gb_case[0], LaunchConfig(grid=1, block=1), add_one, arr.ctypes.data)
+        launch(gb_case[0], LaunchConfig(grid=1, block=1), add_one, arr.ctypes.data)
+        launch(gb_case[0], LaunchConfig(grid=1, block=1), add_one, arr.ctypes.data)
+        gb_case[0].end_building()
+
+        ## Case 1
+        gb_case[1].begin_building()
+        launch(gb_case[1], LaunchConfig(grid=1, block=1), add_one, arr[1:].ctypes.data)
+        launch(gb_case[1], LaunchConfig(grid=1, block=1), add_one, arr[1:].ctypes.data)
+        launch(gb_case[1], LaunchConfig(grid=1, block=1), add_one, arr[1:].ctypes.data)
+        gb_case[1].end_building()
+
+        ## Case 2
+        gb_case[2].begin_building()
+        launch(gb_case[2], LaunchConfig(grid=1, block=1), add_one, arr[2:].ctypes.data)
+        launch(gb_case[2], LaunchConfig(grid=1, block=1), add_one, arr[2:].ctypes.data)
+        launch(gb_case[2], LaunchConfig(grid=1, block=1), add_one, arr[2:].ctypes.data)
+        gb_case[2].end_building()
+
+        return gb.end_building()
+
+    try:
+        graph_variants = [build_graph(0), build_graph(1), build_graph(2)]
+    except Exception as e:
+        with pytest.raises(RuntimeError, match="^(Driver|Binding) version"):
+            raise e
+        b.close()
+        pytest.skip("Driver does not support conditional switch")
+
+    # Launch the first graph
+    assert arr[0] == 0
+    assert arr[1] == 0
+    assert arr[2] == 0
+    graph = graph_variants[0].complete()
+    graph.launch(launch_stream)
+    launch_stream.sync()
+    assert arr[0] == 3
+    assert arr[1] == 0
+    assert arr[2] == 0
+
+    # Update with second variant and launch again
+    graph.update(graph_variants[1])
+    graph.launch(launch_stream)
+    launch_stream.sync()
+    assert arr[0] == 3
+    assert arr[1] == 3
+    assert arr[2] == 0
+
+    # Update with third variant and launch again
+    graph.update(graph_variants[2])
+    graph.launch(launch_stream)
+    launch_stream.sync()
+    assert arr[0] == 3
+    assert arr[1] == 3
+    assert arr[2] == 3
+
+    # Close the memory resource now because the garbage collected might
+    # de-allocate it during the next graph builder process
+    b.close()
+
+
+def test_graph_stream_lifetime(init_cuda):
+    mod = compile_common_kernels()
+    empty_kernel = mod.get_kernel("empty_kernel")
+
+    # Create simple graph from device
+    gb = Device().create_graph_builder().begin_building()
+    launch(gb, LaunchConfig(grid=1, block=1), empty_kernel)
+    graph = gb.end_building().complete()
+
+    # Destroy simple graph and builder
+    gb.close()
+    graph.close()
+
+    # Create simple graph from stream
+    stream = Device().create_stream()
+    gb = stream.create_graph_builder().begin_building()
+    launch(gb, LaunchConfig(grid=1, block=1), empty_kernel)
+    graph = gb.end_building().complete()
+
+    # Destroy simple graph and builder
+    gb.close()
+    graph.close()
+
+    # Verify the stream can still launch work
+    launch(stream, LaunchConfig(grid=1, block=1), empty_kernel)
+    stream.sync()
+
+    # Destroy the stream
+    stream.close()

cuda_core/tests/graph/test_basic.py

@@ -0,0 +1,164 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
+
+"""Basic graph construction and topology tests."""
+
+import numpy as np
+import pytest
+from cuda.core import Device, GraphBuilder, LaunchConfig, LegacyPinnedMemoryResource, launch
+from helpers.graph_kernels import compile_common_kernels
+
+
+def test_graph_is_building(init_cuda):
+    gb = Device().create_graph_builder()
+    assert gb.is_building is False
+    gb.begin_building()
+    assert gb.is_building is True
+    gb.end_building()
+    assert gb.is_building is False
+
+
+def test_graph_straight(init_cuda):
+    mod = compile_common_kernels()
+    empty_kernel = mod.get_kernel("empty_kernel")
+    launch_stream = Device().create_stream()
+
+    # Simple linear topology
+    gb = Device().create_graph_builder().begin_building()
+    launch(gb, LaunchConfig(grid=1, block=1), empty_kernel)
+    launch(gb, LaunchConfig(grid=1, block=1), empty_kernel)
+    launch(gb, LaunchConfig(grid=1, block=1), empty_kernel)
+    graph = gb.end_building().complete()
+
+    # Sanity upload and launch
+    graph.upload(launch_stream)
+    graph.launch(launch_stream)
+    launch_stream.sync()
+
+
+def test_graph_fork_join(init_cuda):
+    mod = compile_common_kernels()
+    empty_kernel = mod.get_kernel("empty_kernel")
+    launch_stream = Device().create_stream()
+
+    # Simple diamond topology
+    gb = Device().create_graph_builder().begin_building()
+    launch(gb, LaunchConfig(grid=1, block=1), empty_kernel)
+
+    with pytest.raises(ValueError, match="^Invalid split count: expecting >= 2, got 1"):
+        gb.split(1)
+
+    left, right = gb.split(2)
+    launch(left, LaunchConfig(grid=1, block=1), empty_kernel)
+    launch(left, LaunchConfig(grid=1, block=1), empty_kernel)
+    launch(right, LaunchConfig(grid=1, block=1), empty_kernel)
+    launch(right, LaunchConfig(grid=1, block=1), empty_kernel)
+
+    with pytest.raises(ValueError, match="^Must join with at least two graph builders"):
+        GraphBuilder.join(left)
+
+    gb = GraphBuilder.join(left, right)
+
+    launch(gb, LaunchConfig(grid=1, block=1), empty_kernel)
+    graph = gb.end_building().complete()
+
+    # Sanity upload and launch
+    graph.upload(launch_stream)
+    graph.launch(launch_stream)
+    launch_stream.sync()
+
+
+def test_graph_is_join_required(init_cuda):
+    mod = compile_common_kernels()
+    empty_kernel = mod.get_kernel("empty_kernel")
+
+    # Starting builder is always primary
+    gb = Device().create_graph_builder()
+    assert gb.is_join_required is False
+    gb.begin_building()
+
+    # Create root node
+    launch(gb, LaunchConfig(grid=1, block=1), empty_kernel)
+
+    # First returned builder is always the original
+    first_split_builders = gb.split(3)
+    assert first_split_builders[0] is gb
+
+    # Only the original builder need not join
+    assert first_split_builders[0].is_join_required is False
+    for builder in first_split_builders[1:]:
+        assert builder.is_join_required is True
+
+    # Launch kernel on each split
+    for builder in first_split_builders:
+        launch(builder, LaunchConfig(grid=1, block=1), empty_kernel)
+
+    # Splitting on new builder will all require joining
+    second_split_builders = first_split_builders[-1]
+    first_split_builders = first_split_builders[0:-1]
+    second_split_builders = second_split_builders.split(3)
+    for builder in second_split_builders:
+        assert builder.is_join_required is True
+
+    # Launch kernel on each second split
+    for builder in second_split_builders:
+        launch(builder, LaunchConfig(grid=1, block=1), empty_kernel)
+
+    # Joined builder requires joining if all builder need to join
+    gb = GraphBuilder.join(*second_split_builders)
+    assert gb.is_join_required is True
+    gb = GraphBuilder.join(gb, *first_split_builders)
+    assert gb.is_join_required is False
+
+    # Create final node
+    launch(gb, LaunchConfig(grid=1, block=1), empty_kernel)
+    gb.end_building().complete()
+
+
+@pytest.mark.skipif(tuple(int(i) for i in np.__version__.split(".")[:2]) < (2, 1), reason="need numpy 2.1.0+")
+def test_graph_repeat_capture(init_cuda):
+    mod = compile_common_kernels()
+    add_one = mod.get_kernel("add_one")
+
+    # Allocate memory
+    launch_stream = Device().create_stream()
+    mr = LegacyPinnedMemoryResource()
+    b = mr.allocate(4)
+    arr = np.from_dlpack(b).view(np.int32)
+    arr[0] = 0
+
+    # Launch the graph once
+    gb = launch_stream.create_graph_builder().begin_building()
+    launch(gb, LaunchConfig(grid=1, block=1), add_one, arr.ctypes.data)
+    graph = gb.end_building().complete()
+
+    # Run the graph once
+    graph.launch(launch_stream)
+    launch_stream.sync()
+    assert arr[0] == 1
+
+    # Continue capturing to extend the graph
+    with pytest.raises(RuntimeError, match="^Cannot resume building after building has ended."):
+        gb.begin_building()
+
+    # Graph can be re-launched
+    graph.launch(launch_stream)
+    graph.launch(launch_stream)
+    graph.launch(launch_stream)
+    launch_stream.sync()
+    assert arr[0] == 4
+
+    # Close the memory resource now because the garbage collected might
+    # de-allocate it during the next graph builder process
+    b.close()
+
+
+def test_graph_capture_errors(init_cuda):
+    gb = Device().create_graph_builder()
+    with pytest.raises(RuntimeError, match="^Graph has not finished building."):
+        gb.complete()
+
+    gb.begin_building()
+    with pytest.raises(RuntimeError, match="^Graph has not finished building."):
+        gb.complete()
+    gb.end_building().complete()

cuda_core/tests/test_graph_mem.py …a_core/tests/graph/test_capture_alloc.pycuda_core/tests/test_graph_mem.py renamed to cuda_core/tests/graph/test_capture_alloc.py cuda_core/tests/test_graph_mem.py renamed to cuda_core/tests/graph/test_capture_alloc.py

@@ -2,6 +2,8 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 
+"""Graph memory resource tests."""
+
 import pytest
 from cuda.core import (
     Device,