Turn cuda-python into a pure Python wheel that depends on the cuda-bindings wheel

.github/workflows/build-and-test.yml

@@ -123,7 +123,6 @@ jobs:
           name: ${{ env.CUDA_CORE_ARTIFACT_NAME }}
           path: ${{ env.CUDA_CORE_ARTIFACTS_DIR }}/*.whl
           if-no-files-found: error
-          overwrite: 'true'
 
       - name: Set up mini CTK
         uses: ./.github/actions/fetch_ctk
@@ -170,7 +169,35 @@ jobs:
           name: ${{ env.CUDA_BINDINGS_ARTIFACT_NAME }}
           path: ${{ env.CUDA_BINDINGS_ARTIFACTS_DIR }}/*.whl
           if-no-files-found: error
-          overwrite: 'true'
+
+      # upload-artifact's "overwrite: true" option has a race condition among parallel
+      # jobs, so we let job 0 do the work
+      - name: Build and check cuda-python wheel
+        if: ${{ strategy.job-index == 0 }}
+        run: |
+          pushd cuda_python
+          pip wheel -v --no-deps .
+          twine check *.whl
+          popd
+
+      - name: List the cuda-python artifacts directory
+        if: ${{ strategy.job-index == 0 }}
+        run: |
+          if [[ "${{ matrix.host-platform }}" == win* ]]; then
+            export CHOWN=chown
+          else
+            export CHOWN="sudo chown"
+          fi
+          $CHOWN -R $(whoami) cuda_python/*.whl
+          ls -lahR cuda_python
+
+      - name: Upload cuda-python build artifacts
+        if: ${{ strategy.job-index == 0 }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: cuda-python-wheel
+          path: cuda_python/*.whl
+          if-no-files-found: error
 
       - name: Pass environment variables to the next runner
         id: pass_env

.github/workflows/build-docs.yml

@@ -70,6 +70,17 @@ jobs:
           echo "CUDA_BINDINGS_ARTIFACT_NAME=${CUDA_BINDINGS_ARTIFACT_BASENAME}-${{ github.sha }}" >> $GITHUB_ENV
           echo "CUDA_BINDINGS_ARTIFACTS_DIR=$(realpath "$REPO_DIR/cuda_bindings/dist")" >> $GITHUB_ENV
 
+      - name: Download cuda-python build artifacts
+        uses: actions/download-artifact@v4
+        with:
+          name: cuda-python-wheel
+          path: .
+
+      - name: Display structure of downloaded cuda-python artifacts
+        run: |
+          pwd
+          ls -lahR .
+
       - name: Download cuda.bindings build artifacts
         uses: actions/download-artifact@v4
         with:
@@ -102,6 +113,8 @@ jobs:
           pip install *.whl
           popd
 
+          pip install cuda_python*.whl
+
       - name: Build all (latest) docs
         id: build
         run: |

.github/workflows/test-wheel.yml

@@ -86,14 +86,21 @@ jobs:
           dependencies: "build-essential jq wget"
           dependent_exes: "gcc jq wget"
 
+      - name: Download cuda-python build artifacts
+        if: ${{ env.SKIP_CUDA_BINDINGS_TEST == '0'}}
+        uses: actions/download-artifact@v4
+        with:
+          name: cuda-python-wheel
+          path: .
+
       - name: Download cuda.bindings build artifacts
         if: ${{ env.SKIP_CUDA_BINDINGS_TEST == '0'}}
         uses: actions/download-artifact@v4
         with:
           name: ${{ env.CUDA_BINDINGS_ARTIFACT_NAME }}
           path: ${{ env.CUDA_BINDINGS_ARTIFACTS_DIR }}
 
-      - name: Download cuda.bindings build artifacts from the prior branch
+      - name: Download cuda-python & cuda.bindings build artifacts from the prior branch
         if: ${{ env.SKIP_CUDA_BINDINGS_TEST == '1'}}
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
@@ -120,6 +127,14 @@ jobs:
           mkdir -p "${{ env.CUDA_BINDINGS_ARTIFACTS_DIR }}"
           mv $OLD_BASENAME/*.whl "${{ env.CUDA_BINDINGS_ARTIFACTS_DIR }}"/
 
+          gh run download $LATEST_PRIOR_RUN_ID -p cuda-python-wheel -R NVIDIA/cuda-python
+          mv cuda-python-wheel/*.whl .
+
+      - name: Display structure of downloaded cuda-python artifacts
+        run: |
+          pwd
+          ls -lahR .
+
       - name: Display structure of downloaded cuda.bindings artifacts
         run: |
           pwd
@@ -203,3 +218,11 @@ jobs:
           pip install -r "tests/requirements-cu${TEST_CUDA_MAJOR}.txt"
           pytest -rxXs tests/
           popd
+
+      - name: Ensure cuda-python installable
+        run: |
+          if [[ "${{ inputs.local-ctk }}" == 1 ]]; then
+            pip install cuda_python*.whl
+          else
+            pip install $(ls cuda_python*.whl)[all]
+          fi

cuda_bindings/cuda/bindings/_version.py

@@ -6,4 +6,4 @@
 # this software and related documentation outside the terms of the EULA
 # is strictly prohibited.
 
-__version__ = "12.6.2.post1"
+__version__ = "12.6.3"

cuda_bindings/pyproject.toml

@@ -11,7 +11,7 @@ requires = ["setuptools", "cython", "pyclibrary"]
 build-backend = "setuptools.build_meta"
 
 [project]
-name = "cuda-python"
+name = "cuda-bindings"
 description = "Python bindings for CUDA"
 authors = [{name = "NVIDIA Corporation", email = "[email protected]"},]
 license = {file = "LICENSE"}

cuda_bindings/setup.py

@@ -308,9 +308,9 @@ def build_extension(self, ext):
             # Allow extensions to discover libraries at runtime
             # relative their wheels installation.
             if ext.name == "cuda.bindings._bindings.cynvrtc":
-                ldflag = f"-Wl,--disable-new-dtags,-rpath,$ORIGIN/../../../nvidia/cuda_nvrtc/lib"
+                ldflag = "-Wl,--disable-new-dtags,-rpath,$ORIGIN/../../../nvidia/cuda_nvrtc/lib"
             elif ext.name == "cuda.bindings._internal.nvjitlink":
-                ldflag = f"-Wl,--disable-new-dtags,-rpath,$ORIGIN/../../../nvidia/nvjitlink/lib"
+                ldflag = "-Wl,--disable-new-dtags,-rpath,$ORIGIN/../../../nvidia/nvjitlink/lib"
             else:
                 ldflag = None
 
@@ -326,7 +326,7 @@ def build_extension(self, ext):
 cmdclass = {
     "bdist_wheel": WheelsBuildExtensions,
     "build_ext": ParallelBuildExtensions,
-    }
+}
 
 # ----------------------------------------------------------------------
 # Setup

cuda_bindings/tests/test_nvjitlink.py

@@ -55,9 +55,7 @@ def ptx_header(version, arch):
 def check_nvjitlink_usable():
     from cuda.bindings._internal import nvjitlink as inner_nvjitlink
 
-    if inner_nvjitlink._inspect_function_pointer("__nvJitLinkVersion") == 0:
-        return False
-    return True
+    return inner_nvjitlink._inspect_function_pointer("__nvJitLinkVersion") != 0
 
 
 pytestmark = pytest.mark.skipif(

cuda_core/cuda/core/experimental/_launcher.py

@@ -2,7 +2,6 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 
-import importlib.metadata
 from dataclasses import dataclass
 from typing import Optional, Union
 
@@ -11,7 +10,7 @@
 from cuda.core.experimental._kernel_arg_handler import ParamHolder
 from cuda.core.experimental._module import Kernel
 from cuda.core.experimental._stream import Stream
-from cuda.core.experimental._utils import CUDAError, check_or_create_options, handle_return
+from cuda.core.experimental._utils import CUDAError, check_or_create_options, get_binding_version, handle_return
 
 # TODO: revisit this treatment for py313t builds
 _inited = False
@@ -25,7 +24,7 @@ def _lazy_init():
 
     global _use_ex
     # binding availability depends on cuda-python version
-    _py_major_minor = tuple(int(v) for v in (importlib.metadata.version("cuda-python").split(".")[:2]))
+    _py_major_minor = get_binding_version()
     _driver_ver = handle_return(cuda.cuDriverGetVersion())
     _use_ex = (_driver_ver >= 11080) and (_py_major_minor >= (11, 8))
     _inited = True

cuda_core/cuda/core/experimental/_module.py

@@ -2,10 +2,9 @@
 #
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 
-import importlib.metadata
 
 from cuda import cuda
-from cuda.core.experimental._utils import handle_return, precondition
+from cuda.core.experimental._utils import get_binding_version, handle_return, precondition
 
 _backend = {
     "old": {
@@ -30,7 +29,7 @@ def _lazy_init():
 
     global _py_major_ver, _driver_ver, _kernel_ctypes
     # binding availability depends on cuda-python version
-    _py_major_ver = int(importlib.metadata.version("cuda-python").split(".")[0])
+    _py_major_ver, _ = get_binding_version()
     if _py_major_ver >= 12:
         _backend["new"] = {
             "file": cuda.cuLibraryLoadFromFile,

cuda_core/cuda/core/experimental/_utils.py

@@ -3,6 +3,7 @@
 # SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
 
 import functools
+import importlib.metadata
 from collections import namedtuple
 from typing import Callable, Dict
 
@@ -134,3 +135,11 @@ def get_device_from_ctx(ctx_handle) -> int:
         assert ctx_handle == handle_return(cuda.cuCtxPopCurrent())
         handle_return(cuda.cuCtxPushCurrent(prev_ctx))
     return device_id
+
+
+def get_binding_version():
+    try:
+        major_minor = importlib.metadata.version("cuda-bindings").split(".")[:2]
+    except importlib.metadata.PackageNotFoundError:
+        major_minor = importlib.metadata.version("cuda-python").split(".")[:2]
+    return tuple(int(v) for v in major_minor)

cuda_core/pyproject.toml

@@ -47,9 +47,8 @@ dependencies = [
 ]
 
 [project.optional-dependencies]
-# TODO: change this once cuda-bindings is packaged, see NVIDIA/cuda-python#105
-cu11 = ["cuda-python==11.8.*"]
-cu12 = ["cuda-python==12.*"]
+cu11 = ["cuda-bindings==11.8.*"]
+cu12 = ["cuda-bindings==12.*"]
 
 [project.urls]
 homepage = "https://nvidia.github.io/cuda-python/"

cuda_python/LICENSE

@@ -0,0 +1 @@
+../LICENSE

cuda_python/README.md

@@ -0,0 +1 @@
+../README.md

cuda_python/pyproject.toml

@@ -0,0 +1,41 @@
+# Copyright 2023-2025 NVIDIA Corporation.  All rights reserved.
+#
+# SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
+
+[build-system]
+requires = ["setuptools",]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "cuda-python"
+description = "CUDA Python: Performance meets Productivity"
+readme = {file = "README.md", content-type = "text/markdown"}
+authors = [{name = "NVIDIA Corporation", email = "[email protected]"},]
+license = {file = "LICENSE"}
+classifiers = [
+    "Operating System :: POSIX :: Linux",
+    "Operating System :: Microsoft :: Windows",
+    "Topic :: Software Development :: Libraries",
+    "Topic :: Education",
+    "Topic :: Scientific/Engineering",
+    "Intended Audience :: Developers",
+    "Intended Audience :: Science/Research",
+    "Intended Audience :: End Users/Desktop",
+    "License :: Other/Proprietary License",
+    "Programming Language :: Python :: 3 :: Only",
+    "Programming Language :: Python :: 3.9",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
+    "Programming Language :: Python :: Implementation :: CPython",
+    "Environment :: GPU :: NVIDIA CUDA",
+    "Environment :: GPU :: NVIDIA CUDA :: 12",
+]
+dynamic = ["version", "dependencies", "optional-dependencies"]
+
+[project.urls]
+homepage = "https://nvidia.github.io/cuda-python/"
+documentation = "https://nvidia.github.io/cuda-python/"
+repository = "https://github.com/NVIDIA/cuda-python/"
+issues = "https://github.com/NVIDIA/cuda-python/issues/"

cuda_python/setup.py

@@ -0,0 +1,23 @@
+# Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. ALL RIGHTS RESERVED.
+#
+# SPDX-License-Identifier: LicenseRef-NVIDIA-SOFTWARE-LICENSE
+
+from setuptools import setup
+
+# We want to keep the version in sync with cuda.bindings, but setuptools would not let
+# us to refer to any files outside of the project root, so we have to employ our own
+# run-time lookup using setup()...
+with open("../cuda_bindings/cuda/bindings/_version.py") as f:
+    exec(f.read())
+version = __version__  # noqa: F821
+del __version__  # noqa: F821
+
+setup(
+    version=version,
+    install_requires=[
+        f"cuda-bindings~={version}",
+    ],
+    extras_require={
+        "all": [f"cuda-bindings[all]~={version}"],
+    },
+)