Merge branch 'main' into dependabot/github_actions/actions-monthly-67a8adbf42

Author: leofang

ci/tools/tests/test_fetch_ctk_redistrib.py

@@ -44,7 +44,6 @@ subpackage in the `cuda-python` monorepo.
   - run: `pytest tests/cython/`
 - **Examples**: example coverage is pytest-based under `examples/`.
 - **Benchmarks**: run with `pytest --benchmark-only benchmarks/` when needed.
-- **Orchestrated run**: from repo root, `scripts/run_tests.sh bindings`.
 
 ## Build and environment notes
 

cuda_bindings/AGENTS.md

@@ -2,7 +2,7 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 #
-# This code was automatically generated across versions from 12.9.1 to 13.3.0, generator version 0.3.1.dev1602+g3c8d84404. Do not modify it directly.
+# This code was automatically generated across versions from 12.9.1 to 13.3.0, generator version 0.3.1.dev1779+ga8cc71818.d20260626. Do not modify it directly.
 
 cimport cython  # NOQA
 
@@ -26815,13 +26815,13 @@ cpdef object unit_get_devices(intptr_t unit):
     """
     cdef unsigned int[1] deviceCount = [0]
     with nogil:
-        __status__ = nvmlUnitGetDevices(<nvmlUnit_t *>unit, <unsigned int*>deviceCount, NULL)
+        __status__ = nvmlUnitGetDevices(<nvmlUnit_t>unit, <unsigned int*>deviceCount, NULL)
     check_status_size(__status__)
     if deviceCount[0] == 0:
         return view.array(shape=(1,), itemsize=sizeof(intptr_t), format="P", mode="c")[:0]
     cdef view.array deviceArray = view.array(shape=(deviceCount[0],), itemsize=sizeof(intptr_t), format="P", mode="c")
     with nogil:
-        __status__ = nvmlUnitGetDevices(<nvmlUnit_t *>unit, <unsigned int*>deviceCount, <nvmlDevice_t *>deviceArray.data)
+        __status__ = nvmlUnitGetDevices(<nvmlUnit_t>unit, <unsigned int*>deviceCount, <nvmlDevice_t *>deviceArray.data)
     check_status(__status__)
     return deviceArray
 

cuda_bindings/cuda/bindings/nvml.pyx

@@ -3,8 +3,7 @@
 [build-system]
 requires = [
     "setuptools>=80.0.0",
-    "setuptools_scm[simple]>=8,<10.1",
-    "vcs-versioning<2.0",
+    "setuptools_scm[simple]>=8,!=10.1",
     "cython>=3.2,<3.3",
     "pyclibrary>=0.1.7",
     "cuda-pathfinder>=1.5",

cuda_bindings/pyproject.toml

@@ -0,0 +1,32 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
+
+import pytest
+
+from cuda.bindings import nvml
+
+
+@pytest.fixture
+def unit_handles(nvml_init):
+    """Yield all unit handles, or skip if no units are present."""
+    count = nvml.unit_get_count()
+    if count == 0:
+        pytest.skip("No NVML units present on this system")
+    return [nvml.unit_get_handle_by_index(i) for i in range(count)]
+
+
+@pytest.mark.agent_authored(model="claude-sonnet-4-6")
+def test_unit_get_devices_returns_array(unit_handles):
+    """Regression test: unit_get_devices must not crash due to double-pointer cast.
+
+    Previously nvmlUnitGetDevices was called with <nvmlUnit_t *>unit instead of
+    <nvmlUnit_t>unit, causing undefined behaviour because nvmlUnit_t is itself an
+    opaque pointer handle.
+    """
+    for unit in unit_handles:
+        devices = nvml.unit_get_devices(unit)
+        # Result is a memoryview / cython array of intptr_t device handles.
+        assert hasattr(devices, "__len__")
+        # Each element must be a non-zero pointer (valid device handle).
+        for dev in devices:
+            assert dev != 0

cuda_bindings/tests/nvml/test_unit.py

@@ -40,7 +40,6 @@ This file describes `cuda_core`, the high-level Pythonic CUDA subpackage in the
   - run: `pytest tests/cython/`
 - **Examples**: validate affected examples in `examples/` when changing user
   workflows or public APIs.
-- **Orchestrated run**: from repo root, `scripts/run_tests.sh core`.
 
 ## Runtime/build environment notes
 

cuda_core/AGENTS.md

@@ -24,12 +24,6 @@ To run these tests:
 * `python -m pytest tests/` with editable installations
 * `pytest tests/` with installed packages
 
-Alternatively, from the repository root you can use a simple script:
-
-* `./scripts/run_tests.sh core` to run only `cuda_core` tests
-* `./scripts/run_tests.sh` to run all package tests (pathfinder → bindings → core)
-* `./scripts/run_tests.sh smoke` to run meta-level smoke tests under `tests/integration`
-
 ### Cython Unit Tests
 
 Cython tests are located in `tests/cython` and need to be built. These builds have the same CUDA Toolkit header requirements as [those of cuda.bindings](https://nvidia.github.io/cuda-python/cuda-bindings/latest/install.html#requirements) where the major.minor version must match `cuda.bindings`. To build them:

cuda_core/README.md

@@ -110,8 +110,12 @@ class _PatchedProperty(metaclass=_PatchedPropMeta):
 from cuda.core._tensor_map import TensorMapDescriptor, TensorMapDescriptorOptions
 
 # isort: split
+# Texture/surface types live under the cuda.core.texture namespace (not the
+# flat cuda.core namespace); import the subpackage so it is available as
+# `cuda.core.texture` after `import cuda.core`.
 # Must come after the cuda.core._* extension imports above: loading graph
 # earlier interacts badly with the merged-wheel __path__ rewrite and leaves
 # Graph/GraphBuilder/GraphCompleteOptions/GraphDebugPrintOptions missing from
 # cuda.core.graph.
 import cuda.core.graph
+import cuda.core.texture