pyproject.toml

####################################################################################################
###############################  BUILD CONFIGURATION  ##############################################
####################################################################################################
[build-system]
requires = ["cmake>=3.27", "nanobind==2.10.2", "setuptools>=77"]
build-backend = "setuptools.build_meta"


[tool.setuptools]
include-package-data = true
package-dir = { "" = "." }

# version is supplied by setup.py so the CUDA local segment (e.g. 0.4.0+cu131)
# lands in the wheel METADATA — see setup.py. It stays in [project].dynamic below.
[tool.setuptools.packages.find]
where = ["."]
include = ["bionemo_ir*"]
# Every shipped Python package has an __init__.py. Regular package discovery
# prunes unrelated checkout directories before descending into them; implicit
# namespace discovery walks the entire workspace, including ignored build trees.
namespaces = false

[tool.setuptools.package-data]
"bionemo_ir" = [
  "_torch/custom_ops/*/configs/*.json",
  "_torch/attention_backend/*/configs/*.json",
]

[project]
name = "bionemo-ir"
description = "BioNeMo Inference Runtime: GPU accelerated framework for protein folding"
requires-python = ">=3.12,<4.0"
authors = [{ name = "NVIDIA-BioNeMo team" }]
dependencies = [
  "cuda-python>=13",
  "torch>=2.11",
  "cuequivariance==0.11.1",
  "cuequivariance-ops-cu13==0.11.1",
  "cuequivariance-ops-torch-cu13==0.11.1",
  "triton>=3.6.0,<3.7.1",
  "biopython>=1.87",
  "ray[serve,data]==2.56.1",  # Installs pandas and pyarrow.
  # uvloop 0.22 breaks Ray Serve: https://github.com/ray-project/ray/issues/57992
  "uvloop==0.21.0",
  "modelcif==1.6",  # Needed for CIFWriter.
  "rdkit==2025.9.6",  # Needed for the Boltz2 data pipeline.
  "nvidia-cutlass-dsl[cu13]==4.5.2",
  "nvidia-cudnn-frontend==1.23.0",
  "apache-tvm-ffi>=0.1.6,<0.2",
  "biotite==1.6.0",
  # https://github.com/huggingface/datasets/issues/7344
  "datasets==4.8.4",
  # hf_hub_download(), currently transitive through datasets.
  "huggingface_hub>=0.25,<2",
  "PyYAML>=6.0.3",
  "lru-dict==1.4.1",
  "kalign-python==3.6.0",
  "scipy==1.13.1",
  "gemmi==0.6.5",
  "einops>=0.8",  # rearrange() and einsum() are imported directly.
  "pydantic>=2,<3",
]
# readme is dynamic because setup.py rewrites it for a package index: an index
# renders it standalone, resolving no relative path and stripping no frontmatter.
dynamic = ["version", "readme"]
license = "Apache-2.0"
license-files = ["LICENSE", "THIRD_PARTY_NOTICES.md", "LICENSES/*.txt"]

# Rendered as the sidebar links on the PyPI project page, which is built from
# this metadata alone -- a published version's page cannot be edited afterwards.
[project.urls]
Homepage = "https://github.com/NVIDIA-BioNeMo/BioNeMo-Inference-Runtime"
Documentation = "https://docs.nvidia.com/bionemo/inference-runtime"
Repository = "https://github.com/NVIDIA-BioNeMo/BioNeMo-Inference-Runtime"
Issues = "https://github.com/NVIDIA-BioNeMo/BioNeMo-Inference-Runtime/issues"

[dependency-groups]
build = [
  # PEP 517 build environments are independent of project dependencies. Keep
  # the native build reproducible without narrowing the package's compatible
  # build-system floors below.
  "cmake==4.4.2",
  "nanobind==2.10.2",
  "setuptools==81.0.0",
]
coverage = [
  "coverage[toml]>=7.10.6",
]
lint = [
  "prek==0.4.12",  # Pinned for reproducible hook runs.
]
release = [
  { include-group = "build" },
  # Tools used after the build stay here rather than in the default developer
  # environment. Release CI syncs this complete group from uv.lock.
  "auditwheel==6.8.0",
  "build==1.5.0",
  # Match the digest-pinned NGC base: DALI 2.1.0 requires packaging<=26.0.
  "packaging==26.0",
  "pip==26.2.1",
  "wheel==0.47.0",
]
sbom = [
  # Isolated CI tool for inventorying the packages actually installed in NGC.
  "cyclonedx-bom==7.4.0",
]
test = [
  "pytest>=9.0.3",  # Older releases contain PYSEC-2026-1845.
  "pytest-cov>=7.1.0",
  "pytest-xdist>=3.8.0",
  # tests/contract runs setup.py.
  "setuptools>=77",
  "parameterized",
  "biotraj==1.2.2",
  "ml_collections==1.1.0",
]
dev = [
  { include-group = "lint" },
  { include-group = "test" },
]

[tool.uv]
required-version = ">=0.12.9,<0.13"
default-groups = ["dev"]
# CUTLASS's base and CUDA 13 wheels contain overlapping paths. Install in
# dependency order so the CUDA 13 wheel consistently supplies those files.
concurrent-installs = 1
# Keep uv's isolated editable builds on the same internal toolchain as the
# locked build group without narrowing the published PEP 517 compatibility
# floors in build-system.requires.
build-constraint-dependencies = [
  "cmake==4.4.2",
  "nanobind==2.10.2",
  "setuptools==81.0.0",
]
# BioIR only supports Linux. Keep both release architectures in every lock
# update so an x86_64 developer cannot silently break aarch64 resolution.
environments = [
  "sys_platform == 'linux' and platform_machine == 'x86_64' and python_version == '3.12'",
  "sys_platform == 'linux' and platform_machine == 'aarch64' and python_version == '3.12'",
]
required-environments = [
  "sys_platform == 'linux' and platform_machine == 'x86_64' and python_version == '3.12'",
  "sys_platform == 'linux' and platform_machine == 'aarch64' and python_version == '3.12'",
]
# uv otherwise notices only Python packaging metadata. These inputs and
# environment variables also determine the editable native extension.
cache-keys = [
  { file = "pyproject.toml" },
  { file = "setup.py" },
  { file = "build.env" },
  { file = "cpp/**/CMakeLists.txt" },
  { file = "cpp/cmake/**/*.cmake" },
  { file = "cpp/cmake/**/*.py" },
  { file = "cpp/kernels/**/*.cpp" },
  { file = "cpp/kernels/**/*.h" },
  { file = "cpp/kernels/cutedsl_*/cubins/index.json" },
  { file = "cpp/kernels/cutedsl_*/cubins/packs/*.tar.xz" },
  { env = "BIOIR_BUILD_CUTEDSL_KERNELS" },
  { env = "BIOIR_VERSION_LOCAL" },
  { env = "CUDA_TAG" },
  { env = "CC" },
  { env = "CXX" },
]

####################################################################################################
###############################  LINTING, FORMATTING AND TESTING CONFIGURATION  ####################
####################################################################################################
[tool.ruff]
line-length = 120
target-version = "py312"
# 3rdparty is vendored and never linted.
extend-exclude = ["3rdparty"]

[tool.ruff.lint]
# Enforced on changed files only, so the set stays low-noise: rules with little
# existing debt and mostly auto-fixable, to avoid bloating small MRs.
select = [
  # pycodestyle errors — PEP 8 violations (indentation, syntax-y style).
  # https://docs.astral.sh/ruff/rules/#pycodestyle-e-w
  "E",
  # Pyflakes — real bugs: unused imports/vars, undefined names, f-string issues.
  # https://docs.astral.sh/ruff/rules/#pyflakes-f
  "F",
  # pycodestyle warnings — whitespace, blank lines, deprecated constructs.
  # https://docs.astral.sh/ruff/rules/#pycodestyle-e-w
  "W",
  # isort — import ordering and grouping (auto-fixed).
  # https://docs.astral.sh/ruff/rules/#isort-i
  "I",
  # pyupgrade — modernize syntax for the target Python (e.g. `list` over `List`).
  # https://docs.astral.sh/ruff/rules/#pyupgrade-up
  "UP",
  # flake8-bugbear — likely bugs and footguns (mutable defaults, `raise` from).
  # https://docs.astral.sh/ruff/rules/#flake8-bugbear-b
  "B",
  # flake8-debugger — leftover `breakpoint()` / `pdb` calls.
  # https://docs.astral.sh/ruff/rules/#flake8-debugger-t10
  "T10",
  # flake8-comprehensions — simpler, faster list/dict/set comprehensions.
  # https://docs.astral.sh/ruff/rules/#flake8-comprehensions-c4
  "C4",
  # flake8-pie — remove redundant/dead constructs (dup dict keys, no-op code).
  # https://docs.astral.sh/ruff/rules/#flake8-pie-pie
  "PIE",
  # flynt — prefer f-strings over `str.join` / `%` / `.format`.
  # https://docs.astral.sh/ruff/rules/#flynt-fly
  "FLY",
  # pygrep-hooks — ban blanket `# noqa` / `# type: ignore` (enforces the
  # "avoid # type: ignore" rule in docs/coding.md).
  # https://docs.astral.sh/ruff/rules/#pygrep-hooks-pgh
  "PGH",
]
# The formatter owns line wrapping; long strings/URLs/comments shouldn't fail lint.
# E741: single-letter names like B/I/J/H/D and l are the domain's tensor
# shape-dim convention here; renaming would obscure the math and ripple across
# call sites for no readability gain.
# E731: assigning a lambda to a name is used deliberately for small local
# helpers; rewriting to def is low-value churn.
# UP042: rewriting `class X(str, Enum)` to `enum.StrEnum` is a behaviour change,
# not a style fix. New enums should still subclass `enum.StrEnum`.
# https://docs.astral.sh/ruff/rules/replace-str-enum/
ignore = ["E501", "E741", "E731", "UP042"]

[tool.ruff.lint.isort]
known-first-party = ["bionemo_ir"]

# Branch coverage over the shipped package. The test job runs two pytest phases
# (see scripts/run_tests.sh) and phase 1 fans out over xdist workers,
# so the data has to merge across processes and across runs.
[tool.coverage.run]
branch = true
# One data file per process (xdist worker, phase), combined before reporting.
parallel = true
# Store paths relative to the measurement root, so data files produced by
# different processes merge cleanly.
# https://coverage.readthedocs.io/en/latest/config.html#run-relative-files
relative_files = true
# The repo root, NOT ["bionemo_ir"]: coverage strips the configured source
# root from the Cobertura filenames, and CI only maps coverage onto the review
# diff when those filenames are repo-relative (bionemo_ir/foo.py). What
# actually gets measured is narrowed to the package by `--cov=bionemo_ir`
# in scripts/run_tests.sh; the omits below keep a bare
# `pytest --cov` honest too.
source = ["."]
omit = [
  # Vendored third-party trees (mirrors ruff's extend-exclude and prek.toml).
  "*/3rdparty/*",
  # The suite measures the shipped package, not itself.
  "tests/*",
  # Packaging script, not shipped package code.
  "setup.py",
]

[tool.coverage.report]
show_missing = true
precision = 2
exclude_also = [
  "if TYPE_CHECKING:",
  "raise NotImplementedError",
  "@(abc\\.)?abstractmethod",
  "if __name__ == .__main__.:",
]

[tool.commitizen]
# Conventional Commits. Used now only to lint commit/MR titles (`cz check`).
# Changelog generation (`cz changelog`) and version bumps (`cz bump`) are a
# follow-up — that's why the version config below is intentionally minimal.
name = "cz_conventional_commits"