Merge branch 'main' into rparolin/192_issue_fix

Author: rparolin

.gitattributes

@@ -6,6 +6,9 @@ cuda/_version.py export-subst
 # we do not own any headers checked in, don't touch them
 *.h binary
 *.hpp binary
+# Exception: headers we own (cuda_core C++ implementation)
+cuda_core/cuda/core/_cpp/*.h -binary text diff
+cuda_core/cuda/core/_cpp/*.hpp -binary text diff
 # git should not convert line endings in PNG files
 *.png binary
 *.svg binary

.gitignore

@@ -11,6 +11,7 @@ __pycache__/
 .pytest_cache/
 .benchmarks/
 *.cpp
+!*_impl.cpp
 !cuda_bindings/cuda/bindings/_lib/param_packer.cpp
 !cuda_bindings/cuda/bindings/_bindings/loader.cpp
 cache_driver

ci/tools/merge_cuda_core_wheels.py

@@ -150,15 +150,21 @@ def merge_wheels(wheels: List[Path], output_dir: Path, show_wheel_contents: bool
             "__init__.py",
             "_version.py",
             "_include",
+            "_cpp",  # Headers for Cython development
             "cu12",
             "cu13",
         )
+        # _resource_handles is shared (not CUDA-version-specific) and must stay
+        # at top level. It's imported early in __init__.py before versioned code.
+        items_to_keep_prefix = ("_resource_handles",)
         all_items = os.scandir(base_wheel / base_dir)
         removed_count = 0
         for f in all_items:
             f_abspath = f.path
             if f.name in items_to_keep:
                 continue
+            if any(f.name.startswith(prefix) for prefix in items_to_keep_prefix):
+                continue
             if f.is_dir():
                 print(f"  Removing directory: {f.name}", file=sys.stderr)
                 shutil.rmtree(f_abspath)

cuda_bindings/cuda/bindings/_lib/utils.pxd.in

@@ -50,6 +50,7 @@ cdef class _HelperCUpointer_attribute:
     # Return values
     cdef driver.CUcontext _ctx
     cdef unsigned int _uint
+    cdef int _int
     cdef driver.CUdeviceptr _devptr
     cdef void** _void
     cdef driver.CUDA_POINTER_ATTRIBUTE_P2P_TOKENS _token

cuda_bindings/cuda/bindings/_lib/utils.pxi.in

@@ -247,12 +247,14 @@ cdef class _HelperCUpointer_attribute:
             else:
                 self._cptr = <void*><void_ptr>init_value.getPtr()
         elif self._attr in ({{if 'CU_POINTER_ATTRIBUTE_MEMORY_TYPE'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_MEMORY_TYPE,{{endif}}
-                            {{if 'CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL,{{endif}}
                             {{if 'CU_POINTER_ATTRIBUTE_ALLOWED_HANDLE_TYPES'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_ALLOWED_HANDLE_TYPES,{{endif}}
                             {{if 'CU_POINTER_ATTRIBUTE_IS_GPU_DIRECT_RDMA_CAPABLE'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_IS_GPU_DIRECT_RDMA_CAPABLE,{{endif}}
                             {{if 'CU_POINTER_ATTRIBUTE_ACCESS_FLAGS'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_ACCESS_FLAGS,{{endif}}):
             self._uint = init_value
             self._cptr = <void*>&self._uint
+        elif self._attr in ({{if 'CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL,{{endif}}):
+            self._int = init_value
+            self._cptr = <void*>&self._int
         elif self._attr in ({{if 'CU_POINTER_ATTRIBUTE_DEVICE_POINTER'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_DEVICE_POINTER,{{endif}}
                             {{if 'CU_POINTER_ATTRIBUTE_RANGE_START_ADDR'}}cydriver.CUpointer_attribute_enum.CU_POINTER_ATTRIBUTE_RANGE_START_ADDR,{{endif}}):
             if self._is_getter:

cuda_bindings/cuda/bindings/_nvml.pyx

@@ -53,6 +53,12 @@
 # This pattern also affects html_static_path and html_extra_path.
 exclude_patterns = []
 
+# Include object entries (methods, attributes, etc.) in the table of contents
+# This enables the "On This Page" sidebar to show class methods and properties
+# Requires Sphinx 5.1+
+toc_object_entries = True
+toc_object_entries_show_parents = "domain"
+
 # -- Options for HTML output -------------------------------------------------
 
 # The theme to use for HTML and HTML Help pages.  See the documentation for
@@ -69,6 +75,10 @@
         "version-switcher",
         "navbar-nav",
     ],
+    # Use custom secondary sidebar that includes autodoc entries
+    "secondary_sidebar_items": ["page-toc"],
+    # Show more TOC levels by default
+    "show_toc_level": 3,
 }
 if os.environ.get("CI"):
     if int(os.environ.get("BUILD_PREVIEW", 0)):

cuda_bindings/docs/source/conf.py

@@ -0,0 +1,26 @@
+.. SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+.. SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
+
+.. module:: cuda.bindings
+
+``cuda-bindings`` 13.1.X Release notes
+======================================
+
+Highlights
+----------
+
+Experimental
+------------
+
+Bugfixes
+--------
+
+* Fixed an issue where the ``CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL`` attribute was
+  retrieved as an unsigned int, rather than a signed int.
+  (`PR #1336 <https://github.com/NVIDIA/cuda-python/pull/1336>`_)
+
+Known issues
+------------
+
+* Updating from older versions (v12.6.2.post1 and below) via ``pip install -U cuda-python`` might not work. Please do a clean re-installation by uninstalling ``pip uninstall -y cuda-python`` followed by installing ``pip install cuda-python``.
+* The graphics APIs in ``cuda.bindings.runtime`` are inadvertently disabled in 13.0.2.  Users needing these APIs should update to 13.0.3.

cuda_bindings/docs/source/release/13.1.X-notes.rst

@@ -11,14 +11,14 @@ def test_nvlink_get_link_count(all_devices):
     """
     for device in all_devices:
         fields = nvml.FieldValue(1)
-        fields[0].field_id = nvml.FI.DEV_NVLINK_LINK_COUNT
+        fields[0].field_id = nvml.FieldId.DEV_NVLINK_LINK_COUNT
         value = nvml.device_get_field_values(device, fields)[0]
         assert value.nvml_return == nvml.Return.SUCCESS or value.nvml_return == nvml.Return.ERROR_NOT_SUPPORTED, (
             f"Unexpected return {value.nvml_return} for link count field query"
         )
 
         # Use the alternative argument to device_get_field_values
-        value = nvml.device_get_field_values(device, [nvml.FI.DEV_NVLINK_LINK_COUNT])[0]
+        value = nvml.device_get_field_values(device, [nvml.FieldId.DEV_NVLINK_LINK_COUNT])[0]
         assert value.nvml_return == nvml.Return.SUCCESS or value.nvml_return == nvml.Return.ERROR_NOT_SUPPORTED, (
             f"Unexpected return {value.nvml_return} for link count field query"
         )

cuda_bindings/tests/nvml/test_nvlink.py

@@ -370,6 +370,22 @@ def test_cuda_pointer_attr():
     assert err == cuda.CUresult.CUDA_SUCCESS
 
 
+@pytest.mark.skipif(
+    driverVersionLessThan(11030) or not supportsManagedMemory(), reason="When new attributes were introduced"
+)
+def test_pointer_get_attributes_device_ordinal():
+    attributes = [
+        cuda.CUpointer_attribute.CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL,
+    ]
+
+    attrs = cuda.cuPointerGetAttributes(len(attributes), attributes, 0)
+
+    # device ordinals are always small numbers.  A large number would indicate
+    # an overflow error.
+
+    assert abs(attrs[1][0]) < 256
+
+
 @pytest.mark.skipif(not supportsManagedMemory(), reason="When new attributes were introduced")
 def test_cuda_mem_range_attr(device):
     size = 0x1000