[FEA]: Introduce Python module with CCCL headers (NVIDIA#3201)

* Add cccl/python/cuda_cccl directory and use from cuda_parallel, cuda_cooperative

* Run `copy_cccl_headers_to_aude_include()` before `setup()`

* Create python/cuda_cccl/cuda/_include/__init__.py, then simply import cuda._include to find the include path.

* Add cuda.cccl._version exactly as for cuda.cooperative and cuda.parallel

* Bug fix: cuda/_include only exists after shutil.copytree() ran.

* Use `f"cuda-cccl @ file://{cccl_path}/python/cuda_cccl"` in setup.py

* Remove CustomBuildCommand, CustomWheelBuild in cuda_parallel/setup.py (they are equivalent to the default functions)

* Replace := operator (needs Python 3.8+)

* Fix oversights: remove `pip3 install ./cuda_cccl` lines from README.md

* Restore original README.md: `pip3 install -e` now works on first pass.

* cuda_cccl/README.md: FOR INTERNAL USE ONLY

* Remove `$pymajor.$pyminor.` prefix in cuda_cccl _version.py (as suggested under NVIDIA#3201 (comment))

Command used: ci/update_version.sh 2 8 0

* Modernize pyproject.toml, setup.py

Trigger for this change:

* NVIDIA#3201 (comment)

* NVIDIA#3201 (comment)

* Install CCCL headers under cuda.cccl.include

Trigger for this change:

* NVIDIA#3201 (comment)

Unexpected accidental discovery: cuda.cooperative unit tests pass without CCCL headers entirely.

* Factor out cuda_cccl/cuda/cccl/include_paths.py

* Reuse cuda_cccl/cuda/cccl/include_paths.py from cuda_cooperative

* Add missing Copyright notice.

* Add missing __init__.py (cuda.cccl)

* Add `"cuda.cccl"` to `autodoc.mock_imports`

* Move cuda.cccl.include_paths into function where it is used. (Attempt to resolve Build and Verify Docs failure.)

* Add # TODO: move this to a module-level import

* Modernize cuda_cooperative/pyproject.toml, setup.py

* Convert cuda_cooperative to use hatchling as build backend.

* Revert "Convert cuda_cooperative to use hatchling as build backend."

This reverts commit 61637d6.

* Move numpy from [build-system] requires -> [project] dependencies

* Move pyproject.toml [project] dependencies -> setup.py install_requires, to be able to use CCCL_PATH

* Remove copy_license() and use license_files=["../../LICENSE"] instead.

* Further modernize cuda_cccl/setup.py to use pathlib

* Trivial simplifications in cuda_cccl/pyproject.toml

* Further simplify cuda_cccl/pyproject.toml, setup.py: remove inconsequential code

* Make cuda_cooperative/pyproject.toml more similar to cuda_cccl/pyproject.toml

* Add taplo-pre-commit to .pre-commit-config.yaml

* taplo-pre-commit auto-fixes

* Use pathlib in cuda_cooperative/setup.py

* CCCL_PYTHON_PATH in cuda_cooperative/setup.py

* Modernize cuda_parallel/pyproject.toml, setup.py

* Use pathlib in cuda_parallel/setup.py

* Add `# TOML lint & format` comment.

* Replace MANIFEST.in with `[tool.setuptools.package-data]` section in pyproject.toml

* Use pathlib in cuda/cccl/include_paths.py

* pre-commit autoupdate (EXCEPT clang-format, which was manually restored)

* Fixes after git merge main

* Resolve warning: AttributeError: '_Reduce' object has no attribute 'build_result'

```
=========================================================================== warnings summary ===========================================================================
tests/test_reduce.py::test_reduce_non_contiguous
  /home/coder/cccl/python/devenv/lib/python3.12/site-packages/_pytest/unraisableexception.py:85: PytestUnraisableExceptionWarning: Exception ignored in: <function _Reduce.__del__ at 0x7bf123139080>

  Traceback (most recent call last):
    File "/home/coder/cccl/python/cuda_parallel/cuda/parallel/experimental/algorithms/reduce.py", line 132, in __del__
      bindings.cccl_device_reduce_cleanup(ctypes.byref(self.build_result))
                                                       ^^^^^^^^^^^^^^^^^
  AttributeError: '_Reduce' object has no attribute 'build_result'

    warnings.warn(pytest.PytestUnraisableExceptionWarning(msg))

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
============================================================= 1 passed, 93 deselected, 1 warning in 0.44s ==============================================================
```

* Move `copy_cccl_headers_to_cuda_cccl_include()` functionality to `class CustomBuildPy`

* Introduce cuda_cooperative/constraints.txt

* Also add cuda_parallel/constraints.txt

* Add `--constraint constraints.txt` in ci/test_python.sh

* Update Copyright dates

* Switch to https://github.com/ComPWA/taplo-pre-commit (the other repo has been archived by the owner on Jul 1, 2024)

For completeness: The other repo took a long time to install into the pre-commit cache; so long it lead to timeouts in the CCCL CI.

* Remove unused cuda_parallel jinja2 dependency (noticed by chance).

* Remove constraints.txt files, advertise running `pip install cuda-cccl` first instead.

* Make cuda_cooperative, cuda_parallel testing completely independent.

* Run only test_python.sh [skip-rapids][skip-matx][skip-docs][skip-vdc]

* Try using another runner (because V100 runners seem to be stuck) [skip-rapids][skip-matx][skip-docs][skip-vdc]

* Fix sign-compare warning (NVIDIA#3408) [skip-rapids][skip-matx][skip-docs][skip-vdc]

* Revert "Try using another runner (because V100 runners seem to be stuck) [skip-rapids][skip-matx][skip-docs][skip-vdc]"

This reverts commit ea33a21.

Error message: NVIDIA#3201 (comment)

* Try using A100 runner (because V100 runners still seem to be stuck) [skip-rapids][skip-matx][skip-docs][skip-vdc]

* Also show cuda-cooperative site-packages, cuda-parallel site-packages (after pip install) [skip-rapids][skip-matx][skip-docs][skip-vdc]

* Try using l4 runner (because V100 runners still seem to be stuck) [skip-rapids][skip-matx][skip-docs][skip-vdc]

* Restore original ci/matrix.yaml [skip-rapids]

* Use for loop in test_python.sh to avoid code duplication.

* Run only test_python.sh [skip-rapids][skip-matx][skip-docs][skip-vdc][skip pre-commit.ci]

* Comment out taplo-lint in pre-commit config [skip-rapids][skip-matx][skip-docs][skip-vdc]

* Revert "Run only test_python.sh [skip-rapids][skip-matx][skip-docs][skip-vdc][skip pre-commit.ci]"

This reverts commit ec206fd.

* Implement suggestion by @shwina (NVIDIA#3201 (review))

* Address feedback by @leofang

---------

Co-authored-by: Bernhard Manfred Gruber <bernhardmgruber@gmail.com>

Author: 2 people

.pre-commit-config.yaml

@@ -43,6 +43,17 @@ repos:
     hooks:
     - id: ruff  # linter
     - id: ruff-format  # formatter
+
+  # TOML lint & format
+  - repo: https://github.com/ComPWA/taplo-pre-commit
+    rev: v0.9.3
+    hooks:
+      # See https://github.com/NVIDIA/cccl/issues/3426
+      # - id: taplo-lint
+      #   exclude: "^docs/"
+      - id: taplo-format
+        exclude: "^docs/"
+
   - repo: https://github.com/codespell-project/codespell
     rev: v2.3.0
     hooks:

ci/test_python.sh

@@ -8,25 +8,28 @@ print_environment_details
 
 fail_if_no_gpu
 
-readonly prefix="${BUILD_DIR}/python/"
-export PYTHONPATH="${prefix}:${PYTHONPATH:-}"
+begin_group "⚙️ Existing site-packages"
+pip freeze
+end_group "⚙️ Existing site-packages"
 
-pushd ../python/cuda_cooperative >/dev/null
+for module in cuda_parallel cuda_cooperative; do
 
-run_command "⚙️  Pip install cuda_cooperative" pip install --force-reinstall --upgrade --target "${prefix}" .[test]
-run_command "🚀  Pytest cuda_cooperative" python -m pytest -v ./tests
+  pushd "../python/${module}" >/dev/null
 
-popd >/dev/null
+  TEMP_VENV_DIR="/tmp/${module}_venv"
+  rm -rf "${TEMP_VENV_DIR}"
+  python -m venv "${TEMP_VENV_DIR}"
+  . "${TEMP_VENV_DIR}/bin/activate"
+  echo 'cuda-cccl @ file:///home/coder/cccl/python/cuda_cccl' > /tmp/cuda-cccl_constraints.txt
+  run_command "⚙️  Pip install ${module}" pip install -c /tmp/cuda-cccl_constraints.txt .[test]
+  begin_group "⚙️ ${module} site-packages"
+  pip freeze
+  end_group "⚙️ ${module} site-packages"
+  run_command "🚀  Pytest ${module}" python -m pytest -v ./tests
+  deactivate
 
-pushd ../python/cuda_parallel >/dev/null
+  popd >/dev/null
 
-# Temporarily install the package twice to populate include directory as part of the first installation
-# and to let manifest discover these includes during the second installation. Do not forget to remove the
-# second installation after https://github.com/NVIDIA/cccl/issues/2281 is addressed.
-run_command "⚙️  Pip install cuda_parallel once" pip install --force-reinstall --upgrade --target "${prefix}" .[test]
-run_command "⚙️  Pip install cuda_parallel twice" pip install --force-reinstall --upgrade --target "${prefix}" .[test]
-run_command "🚀  Pytest cuda_parallel" python -m pytest -v ./tests
-
-popd >/dev/null
+done
 
 print_time_summary

ci/update_version.sh

@@ -37,6 +37,7 @@ CUB_CMAKE_VERSION_FILE="lib/cmake/cub/cub-config-version.cmake"
 LIBCUDACXX_CMAKE_VERSION_FILE="lib/cmake/libcudacxx/libcudacxx-config-version.cmake"
 THRUST_CMAKE_VERSION_FILE="lib/cmake/thrust/thrust-config-version.cmake"
 CUDAX_CMAKE_VERSION_FILE="lib/cmake/cudax/cudax-config-version.cmake"
+CUDA_CCCL_VERSION_FILE="python/cuda_cccl/cuda/cccl/_version.py"
 CUDA_COOPERATIVE_VERSION_FILE="python/cuda_cooperative/cuda/cooperative/_version.py"
 CUDA_PARALLEL_VERSION_FILE="python/cuda_parallel/cuda/parallel/_version.py"
 
@@ -110,6 +111,7 @@ update_file "$CUDAX_CMAKE_VERSION_FILE" "set(cudax_VERSION_MAJOR \([0-9]\+\))" "
 update_file "$CUDAX_CMAKE_VERSION_FILE" "set(cudax_VERSION_MINOR \([0-9]\+\))" "set(cudax_VERSION_MINOR $minor)"
 update_file "$CUDAX_CMAKE_VERSION_FILE" "set(cudax_VERSION_PATCH \([0-9]\+\))" "set(cudax_VERSION_PATCH $patch)"
 
+update_file "$CUDA_CCCL_VERSION_FILE" "^__version__ = \"\([0-9.]\+\)\"" "__version__ = \"$major.$minor.$patch\""
 update_file "$CUDA_COOPERATIVE_VERSION_FILE" "^__version__ = \"\([0-9.]\+\)\"" "__version__ = \"$pymajor.$pyminor.$major.$minor.$patch\""
 update_file "$CUDA_PARALLEL_VERSION_FILE" "^__version__ = \"\([0-9.]\+\)\"" "__version__ = \"$pymajor.$pyminor.$major.$minor.$patch\""
 

docs/repo.toml

@@ -347,6 +347,7 @@ autodoc.mock_imports = [
     "numba",
     "pynvjitlink",
     "cuda.bindings",
+    "cuda.cccl",
     "llvmlite",
     "numpy",
 ]

python/cuda_cccl/.gitignore

@@ -0,0 +1,2 @@
+cuda/cccl/include
+*egg-info

python/cuda_cccl/README.md

@@ -0,0 +1,3 @@
+## Note
+
+This package is currently FOR INTERNAL USE ONLY and not meant to be used/installed explicitly.

python/cuda_cccl/cuda/cccl/__init__.py

@@ -0,0 +1,8 @@
+# Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. ALL RIGHTS RESERVED.
+#
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+from cuda.cccl._version import __version__
+from cuda.cccl.include_paths import get_include_paths
+
+__all__ = ["__version__", "get_include_paths"]

python/cuda_cccl/cuda/cccl/_version.py

@@ -0,0 +1,7 @@
+# Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. ALL RIGHTS RESERVED.
+#
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+# This file is generated by ci/update_version.sh
+# Do not edit this file manually.
+__version__ = "2.8.0"

python/cuda_cccl/cuda/cccl/include_paths.py

@@ -0,0 +1,63 @@
+# Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. ALL RIGHTS RESERVED.
+#
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+import os
+import shutil
+from dataclasses import dataclass
+from functools import lru_cache
+from pathlib import Path
+from typing import Optional
+
+
+def _get_cuda_path() -> Optional[Path]:
+    cuda_path = os.environ.get("CUDA_PATH")
+    if cuda_path:
+        cuda_path = Path(cuda_path)
+        if cuda_path.exists():
+            return cuda_path
+
+    nvcc_path = shutil.which("nvcc")
+    if nvcc_path:
+        return Path(nvcc_path).parent.parent
+
+    default_path = Path("/usr/local/cuda")
+    if default_path.exists():
+        return default_path
+
+    return None
+
+
+@dataclass
+class IncludePaths:
+    cuda: Optional[Path]
+    libcudacxx: Optional[Path]
+    cub: Optional[Path]
+    thrust: Optional[Path]
+
+    def as_tuple(self):
+        # Note: higher-level ... lower-level order:
+        return (self.thrust, self.cub, self.libcudacxx, self.cuda)
+
+
+@lru_cache()
+def get_include_paths() -> IncludePaths:
+    # TODO: once docs env supports Python >= 3.9, we
+    # can move this to a module-level import.
+    from importlib.resources import as_file, files
+
+    cuda_incl = None
+    cuda_path = _get_cuda_path()
+    if cuda_path is not None:
+        cuda_incl = cuda_path / "include"
+
+    with as_file(files("cuda.cccl.include")) as f:
+        cccl_incl = Path(f)
+    assert cccl_incl.exists()
+
+    return IncludePaths(
+        cuda=cuda_incl,
+        libcudacxx=cccl_incl / "libcudacxx",
+        cub=cccl_incl,
+        thrust=cccl_incl,
+    )

python/cuda_cccl/pyproject.toml

@@ -0,0 +1,29 @@
+# Copyright (c) 2025, NVIDIA CORPORATION & AFFILIATES. ALL RIGHTS RESERVED.
+#
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+[build-system]
+requires = ["setuptools>=61.0.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "cuda-cccl"
+description = "Experimental Package with CCCL headers to support JIT compilation"
+authors = [{ name = "NVIDIA Corporation" }]
+classifiers = [
+  "Programming Language :: Python :: 3 :: Only",
+  "Environment :: GPU :: NVIDIA CUDA",
+  "License :: OSI Approved :: Apache Software License",
+]
+requires-python = ">=3.9"
+dynamic = ["version", "readme"]
+
+[project.urls]
+Homepage = "https://github.com/NVIDIA/cccl"
+
+[tool.setuptools.dynamic]
+version = { attr = "cuda.cccl._version.__version__" }
+readme = { file = ["README.md"], content-type = "text/markdown" }
+
+[tool.setuptools.package-data]
+cuda = ["cccl/include/**/*"]