[project]
name = "nccl4py"
dynamic = ["version"]
requires-python = ">=3.10"
description = "nccl4py: Pythonic bindings for RCCL on ROCm (AMD fork of NVIDIA nccl4py)"
readme = {file = "DESCRIPTION.rst", content-type = "text/x-rst"}
license = "Apache-2.0"
license-files = ["LICENSE.txt"]
authors = [
    { name = "Advanced Micro Devices, Inc." }
]
classifiers = [
    "Intended Audience :: Developers",
    "Intended Audience :: Science/Research",
    "Intended Audience :: End Users/Desktop",
    "Natural Language :: English",
    "Operating System :: POSIX :: Linux",
    "Topic :: Education",
    "Topic :: Scientific/Engineering",
    "Topic :: Software Development :: Libraries",
    "Programming Language :: Cython",
    "Programming Language :: Python :: 3 :: Only",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Programming Language :: Python :: 3.13",
    "Programming Language :: Python :: 3.14",
    "Programming Language :: Python :: Implementation :: CPython",
]

dependencies = [
    "packaging",
    "numpy",
    # cuda.core / cuda-bindings / nvidia-* are replaced by the local
    # nccl._hip_compat.cuda_core_shim package backed by hip-python.
    "hip-python",
]

[dependency-groups]
test = [
    "pytest",
    "pytest-cov",
    "pytest-mpi",
    "pytest-run-parallel",
    "mpi4py",
    "ml-dtypes",
]

[project.urls]
homepage = "https://github.com/ROCm/rccl"
documentation = "https://rocm.docs.amd.com/projects/rccl/en/latest/"
repository = "https://github.com/ROCm/rocm-systems"
issues = "https://github.com/ROCm/rocm-systems/issues"

[build-system]
requires = [
    "setuptools>=77.0.3",
    "Cython>=3.1"
]
build-backend = "setuptools.build_meta"

[tool.setuptools]
include-package-data = false

[tool.setuptools.packages.find]
# Only the `nccl` package is shipped on disk. The HIP-backed
# `cuda.core` surface is registered in `sys.modules` lazily on
# `import nccl` (see `nccl/__init__.py` -> `_register_as_cuda_core`)
# so we do NOT install a top-level `cuda` package here. A regular
# (non PEP 420) `cuda/__init__.py` would shadow co-installed
# distributions that legitimately contribute to the `cuda.*`
# namespace (`cuda-bindings` pulled by PyTorch / RAPIDS / Triton /
# nsight tooling, etc.); whichever package wins the `__init__.py`
# race in `site-packages` silently hides the loser's `cuda.core`.
# `from cuda.core import ...` therefore resolves only after
# `import nccl` has run; users who reach `cuda.core` first get
# whatever package owns the namespace on disk, which is the
# correct behavior.
include = ["nccl*"]

[tool.setuptools.package-data]
"nccl.bindings" = ["cynccl.pxd"]

[tool.setuptools.dynamic]
version = { attr = "nccl._version.__version__" }

[tool.ruff]
line-length = 100

[tool.pyright]
include = ["nccl", "tests"]
exclude = ["build", "dist", "*.egg-info"]
pythonVersion = "3.10"
typeCheckingMode = "basic"

[tool.pytest.ini_options]
minversion = "6.0"
testpaths = ["tests"]
python_files = ["test_*.py"]
# The collectives + interop suites need >=2 MPI ranks, so they live
# under the `mpi` marker and are launched via
# `mpirun -np N pytest -m mpi tests/`. Declared here for `--strict-markers`.
markers = [
    "mpi: requires mpirun and >=2 ranks (uses pytest-mpi)",
]

[tool.cibuildwheel]
build = "cp310-* cp311-* cp312-* cp313-* cp314-* cp314t-*"
skip = "*-musllinux_*"
build-verbosity = 1

[tool.cibuildwheel.linux]
archs = ["native"]  # Fast local builds; override in CI for multi-arch
manylinux-x86_64-image = "manylinux_2_28"
manylinux-aarch64-image = "manylinux_2_28"