pyproject.toml

[project]
name            = "gigl"
description     = "GIgantic Graph Learning Library"
readme          = "README.md"
version         = "0.2.0"
classifiers     = [
    "Programming Language :: Python",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.11",
]
requires-python = "==3.11.*" # Limited by tfx-bsl not having wheels available for newer python versions.

dependencies = [
  "argo-workflows",
  "chardet",
  # gigl-core hosts all C++ / CUDA / pybind11 extensions. Separate wheel per torch
  # variant (cpu/cu128). Version must match gigl exactly.
  "gigl-core==0.2.0",
  "google-cloud-aiplatform",
  "google-cloud-dataproc",
  "google-cloud-logging",
  "google-cloud-pipeline-components",
  "google-cloud-storage",
  # ipython and ipykernel are required for us to run jupyter notebooks as a part of our test suite.
  # See https://stackoverflow.com/questions/69759351/error-jupyter-client-kernelspec-nosuchkernel-no-such-kernel-named-python3-occu
  "ipykernel",
  "ipython",
  "kfp>=2.0.0",
  "matplotlib",
  "mmh3",
  "msgpack",
  "nbconvert",
  "nbformat",
  "numpy",
  "omegaconf>=2.3.0,<3.0.0",
  "pandas",
  "pip~=25.3", # Used by KFP Python components - need to install pip as a package
  "protobuf",
  "python-dotenv",
  "pyyaml",
  "scikit-learn",
  # When updating to a new tensorflow version, please ensure that the version changes align with
  # `transform` deps listed below. Package versions for `apache-beam`, `tensorflow-transform`, etc
  "tensorflow~=2.16",
]


[project.optional-dependencies]
transform = [
  # Apache Beam version has to match with what is specified in `containers/Dockerfile.dataflow.base`
  "apache-beam[gcp]==2.56.0",
  "pyarrow==10.0.1",
  # Tensorflow transform packages.
  # There is no official MAC support for these packages - thus we ignore them on darwin machines.
  # These packages are also required to be in sync.
  # See the following for the compatibility matrix: https://pypi.org/project/tensorflow-transform/
  "tensorflow-data-validation==1.16.1 ; sys_platform != 'darwin'",
  "tensorflow-metadata==1.16.1 ; sys_platform != 'darwin'",
  "tensorflow-transform==1.16.0 ; sys_platform != 'darwin'",
  # Oct 28, 2025 - Latest ver tfx-bsl==1.17.1 requires latest version of Linux distros using glibc 2.39+
  # i.e. Ubuntu 24.04. Currently, not compatible with our docker images.
  "tfx-bsl==1.16.1 ; sys_platform != 'darwin'",
]
pyg27-torch28-cpu = [
  "fbgemm-gpu~=1.3.0; sys_platform != 'darwin'", # Compatiblity table: https://docs.pytorch.org/FBGEMM/general/Releases.html#fbgemm-releases-compatibility
  "pyg_lib; sys_platform != 'darwin'",
  "torch==2.8",
  # Oct 24, 25 - 2.7 is latest version and torch ver support is inconsistant accross minor version changes
  "torch_cluster; sys_platform != 'darwin'",
  "torch_geometric==2.7",
  "torch_scatter; sys_platform != 'darwin'",
  "torch_sparse; sys_platform != 'darwin'",
  "torch_spline_conv; sys_platform != 'darwin'",
  "torchrec ; sys_platform != 'darwin'",
]
pyg27-torch28-cu128 = [
  "fbgemm-gpu~=1.3.0; sys_platform != 'darwin'",
  "pyg_lib; sys_platform != 'darwin'",
  "torch==2.8 ; sys_platform != 'darwin'",
  "torch_cluster; sys_platform != 'darwin'",
  "torch_geometric==2.7 ; sys_platform != 'darwin'",
  "torch_scatter; sys_platform != 'darwin'",
  "torch_sparse; sys_platform != 'darwin'",
  "torch_spline_conv; sys_platform != 'darwin'",
  "torchrec ; sys_platform != 'darwin'",

]
experimental = [
  "hydra-core==1.3.2",
]

# Declares what deps are conflicting so they are not resolved together.
# In the case below, we cannot have both pyg27-torch28-cpu and pyg27-torch28-cu128
# installed at the same time (one is cuda and one is cpu and pull from seperate indexes).
# Thus, they are treated as conflicting and `uv.lock` will be generated in a way that
# transitive dependencies between them are resolved separately.
[tool.uv]
conflicts = [
  [{ extra = "pyg27-torch28-cpu" },{ extra = "pyg27-torch28-cu128" }],
]
required-environments = [
    "sys_platform == 'linux' and platform_machine == 'x86_64'",
    "sys_platform == 'darwin' and platform_machine == 'arm64'",
]
# gigl-core must build against torch, which is not on PyPI and cannot be installed
# into an isolated build environment. Disabling isolation lets cmake find torch from
# the ambient venv (installed by make install_dev_deps / base Docker images).
no-build-isolation-package = ["gigl-core"]

[dependency-groups]
# scikit-build-core is gigl-core's PEP 517 build backend. With no-build-isolation-package
# set, uv does not install [build-system].requires automatically, so it must be present
# in the ambient environment before any gigl-core build (uv sync, uv build, Dockerfiles).
gigl-core-build-backend = ["scikit-build-core>=0.10"]
dev = [
  {include-group = "gigl-core-build-backend"},
  {include-group = "docs"},
  {include-group = "lint"},
  {include-group = "test"},
  {include-group = "typing-stubs"},
  "pre-commit==3.3.2",
]
docs = [
  "astroid==3.3.11", # Newer versions of astroid are not compatible with sphinx==7.4.7
  "mistune>=2.0.3",
  "myst-parser==2.0.0",
  "myst_nb==1.2.0",
  "nbconvert>=7.16.2",
  "nbsphinx==0.9.3",
  "sphinx==7.4.7",
  "sphinx-autoapi==3.6.0",
  "sphinx-autodoc-typehints==2.3.0",
  "sphinx-copybutton==0.5.2",
  "sphinx-design==0.6.0",
  "sphinx-hoverxref==1.3.0",
  "sphinx-rtd-theme==2.0.0",
  "pydata-sphinx-theme==0.16.1",
  "sphinx-tabs==3.4.5",

]
typing-stubs = [
  "pandas-stubs==2.2.2.240807",
  "types-psutil==7.0.0.20250401",
  "types-PyYAML~=6.0.12",
  "types-requests==2.31.0.6",
  "types-tqdm==4.67.0.20250513",
]
test = [
  "parameterized==0.9.0",
]
lint = [
  "ruff==0.15.10",
  "mdformat==0.7.22",
  "mdformat_tables==1.0.0",
  "mypy==1.8.0",
  "mypy-extensions",
  "mypy-protobuf==3.3.0",
]



# =============================== PyTorch Indexes =============================
[[tool.uv.index]]
name = "torch-cpu"
url = "https://download.pytorch.org/whl/cpu"
# Setting this to true prevents packages from being installed from this index
# unless explicitly pinned to it in [tool.uv.sources]
explicit = true

[[tool.uv.index]]
name = "torch-cu128"
url = "https://download.pytorch.org/whl/cu128"
explicit = true

# =============================== PyG Indexes ==================================
[[tool.uv.index]]
name = "pyg-torch28-cpu"
url = "https://data.pyg.org/whl/torch-2.8.0+cpu.html"
explicit = true
# By default, [[tool.uv.index]] entries are assumed to be PyPI-style registries
# that implement the PEP 503. However, pyg is hosted in a "flat" index, which is
# just an HTML page that contain flat lists of wheels and source distributions.
format = "flat"

[[tool.uv.index]]
name = "pyg-torch28-cu128"
url = "https://data.pyg.org/whl/torch-2.8.0+cu128.html"
explicit = true
format = "flat"

# =============================== Google Artifact Registry Index ========================
# Two separate indexes — one per torch variant — so pip always fetches the correct wheel.
# If both variants lived in the same index, pip would pick between them arbitrarily since
# both wheels share the same platform tag (linux_x86_64). Separate indexes remove the
# ambiguity and are essential once any extension contains CUDA kernels.
#
# The CPU registry reuses the original `gigl` repo for backwards compatibility.
# The CUDA registry is a new `gigl-cu128` repo that must be created in GCP project
# `external-snap-ci-github-gigl` before the release workflow will succeed.
[[tool.uv.index]]
name = "gcp-release-registry-cpu"
url = "https://us-central1-python.pkg.dev/external-snap-ci-github-gigl/gigl/simple/"
publish-url = "https://us-central1-python.pkg.dev/external-snap-ci-github-gigl/gigl"
explicit = true

[[tool.uv.index]]
name = "gcp-release-registry-cu128"
url = "https://us-central1-python.pkg.dev/external-snap-ci-github-gigl/gigl-cu128/simple/"
publish-url = "https://us-central1-python.pkg.dev/external-snap-ci-github-gigl/gigl-cu128"
explicit = true

# ========== Mapping individual packages to their appropriate index =============
[tool.uv.sources]
# ============= PyTorch hosted Package Index Mappings
torch = [
  { extra = "pyg27-torch28-cpu", index = "torch-cpu" },
  { extra = "pyg27-torch28-cu128", index = "torch-cu128" },
]
torchrec = [
  { extra = "pyg27-torch28-cpu", index = "torch-cpu" },
  { extra = "pyg27-torch28-cu128", index = "torch-cu128" },
]
fbgemm-gpu = [
  { extra = "pyg27-torch28-cpu", index = "torch-cpu" },
  { extra = "pyg27-torch28-cu128", index = "torch-cu128" },
]
# ============= PyG hosted Package Index Mappings
# PyG wheels are built for specific versions of torch @ a time. Thus the packages
# are mapped to the appropriate PyG index based on the torch version.
pyg_lib = [
  { extra = "pyg27-torch28-cpu", index = "pyg-torch28-cpu" },
  { extra = "pyg27-torch28-cu128", index = "pyg-torch28-cu128" },
]
torch_scatter = [
  { extra = "pyg27-torch28-cpu", index = "pyg-torch28-cpu" },
  { extra = "pyg27-torch28-cu128", index = "pyg-torch28-cu128" },
]
torch_sparse = [
  { extra = "pyg27-torch28-cpu", index = "pyg-torch28-cpu" },
  { extra = "pyg27-torch28-cu128", index = "pyg-torch28-cu128" },
]
torch_cluster = [
  { extra = "pyg27-torch28-cpu", index = "pyg-torch28-cpu" },
  { extra = "pyg27-torch28-cu128", index = "pyg-torch28-cu128" },
]
torch_spline_conv = [
  { extra = "pyg27-torch28-cpu", index = "pyg-torch28-cpu" },
  { extra = "pyg27-torch28-cu128", index = "pyg-torch28-cu128" },
]
# gigl-core is a local path dependency. In dev, uv resolves it from the adjacent
# gigl-core/ directory and builds the C++ extension via scikit-build-core.
# At release install time, published gigl and gigl-core wheels are resolved
# from the same GCP registry.
gigl-core = { path = "gigl-core" }


# ===================== Build/Project Configurations ===========================
# gigl is a pure-Python package. All C++/CUDA kernels live in the gigl-core
# wheel (see gigl-core/pyproject.toml). Release builds two variants per package:
# cpu and cu128. gigl's build stays fast because no CMake runs here.
[build-system]
requires      = ["setuptools>=68"]
build-backend = "setuptools.build_meta"

[tool.setuptools.packages.find]
include = ["gigl*", "snapchat*"]

[tool.setuptools.package-data]
# Include dep_vars.env from the root directory
"gigl" = ["dep_vars.env", "**/*.yaml"]
"gigl.scripts" = ["*.sh"]

[project.urls]
Homepage = "https://github.com/snapchat/gigl"

[project.scripts]
gigl-post-install = "gigl.scripts.post_install:main"

[tool.ruff]
# Skip generated proto files.
exclude = ["*_pb2.py", "*_pb2.pyi"]

[tool.ruff.lint]
# Rules reference: https://docs.astral.sh/ruff/rules/
#
# F401 - Unused imports (pyflakes)
#   Detects imported names that are never used in the module.
#   Replaces: autoflake --remove-all-unused-imports
#   Docs: https://docs.astral.sh/ruff/rules/unused-import/
#
# I - Import sorting (isort)
#   Enforces a consistent import order (stdlib → third-party → first-party).
#   Replaces: isort
#   Docs: https://docs.astral.sh/ruff/rules/#isort-i
select = ["F401", "I"]

[tool.ruff.lint.per-file-ignores]
# __init__.py files re-export symbols for the public API, so unused-import
# warnings (F401) are expected and should be suppressed.
"__init__.py" = ["F401"]

[tool.ruff.lint.isort]
# Docs: https://docs.astral.sh/ruff/settings/#lint_isort_known-first-party
known-first-party = ["gigl", "tests", "snapchat", "scripts"]