Model-Optimizer/pyproject.toml

####################################################################################################
###############################  BUILD CONFIGURATION  ##############################################
####################################################################################################
[build-system]
requires = ["setuptools>=80", "setuptools-scm>=8,<10"]
build-backend = "setuptools.build_meta"

[tool.setuptools_scm]
fallback_version = "0.0.0"


####################################################################################################
###############################  PROJECT METADATA  #################################################
####################################################################################################
[project]
name = "nvidia-modelopt"
dynamic = ["version"]
description = "Nvidia Model Optimizer: A unified library of SOTA model optimization techniques like quantization, pruning, Neural Architecture Search (NAS), distillation, speculative decoding, etc. It compresses deep learning models for downstream deployment frameworks like TensorRT-LLM, TensorRT, vLLM, etc. to optimize inference speed."
readme = { text = "Checkout https://github.com/nvidia/Model-Optimizer for more information.", content-type = "text/markdown" }
license = "Apache-2.0"
license-files = ["LICENSE_HEADER"]
requires-python = ">=3.10,<3.15"
authors = [{ name = "NVIDIA Corporation" }]
classifiers = [
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Programming Language :: Python :: 3.13",
    "Intended Audience :: Developers",
    "Intended Audience :: Science/Research",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
]
dependencies = [
    # Common
    "ninja",            # for faster building of C++ / CUDA extensions
    "numpy",
    "nvidia-ml-py>=12",
    "packaging",
    "setuptools>=80",   # torch.utils.cpp_extension imports setuptools at load time
    "torch>=2.8",
    "tqdm",
    # modelopt.torch
    "PyYAML>=6.0",
    "omegaconf>=2.3.0",
    "pulp<4.0",         # breaking changes in upcoming 4.0 release
    "pydantic>=2.0",
    "regex",
    "rich",
    "safetensors",
    "scipy",
]

[project.optional-dependencies]
onnx = [
    "cppimport",
    "cupy-cuda12x; platform_machine != 'aarch64' and platform_machine != 'ARM64' and platform_system != 'Darwin'",
    "lief",
    "ml_dtypes",
    "onnx-graphsurgeon>=0.6.1",
    "onnx~=1.21.0",
    "onnxconverter-common~=1.16.0",
    # Python 3.10 support for ModelOpt ONNX is deprecated and will be removed in a future release.
    # Native Windows ARM64 does not support ModelOpt ONNX on Python 3.10.
    # ORT for Windows x64
    "onnxruntime-gpu==1.22.0; platform_system == 'Windows' and platform_machine == 'AMD64'",
    # ORT host and standalone TensorRT-RTX ABI EP for native Windows ARM64.
    "onnxruntime~=1.24.2; python_version > '3.10' and platform_system == 'Windows' and platform_machine == 'ARM64'",
    "onnxruntime-ep-nv-tensorrt-rtx-cu13==0.4.0; python_version > '3.10' and platform_system == 'Windows' and platform_machine == 'ARM64'",
    # ORT with Python <= 3.10 on supported platforms
    "onnxruntime~=1.22.0; python_version <= '3.10' and (platform_machine == 'aarch64' or platform_system == 'Darwin')",
    "onnxruntime-gpu~=1.22.0; python_version <= '3.10' and platform_machine != 'aarch64' and platform_system != 'Darwin' and platform_system != 'Windows'",
    # ORT with Python > 3.10
    "onnxruntime~=1.24.2; python_version > '3.10' and (platform_machine == 'aarch64' or platform_system == 'Darwin')",
    "onnxruntime-gpu~=1.24.2; python_version > '3.10' and platform_machine != 'aarch64' and platform_system != 'Darwin' and platform_system != 'Windows'",
    "onnxscript",
    "onnxslim>=0.1.76",
    "polygraphy>=0.53.4",
]
hf = [
    "accelerate>=1.0.0",
    "datasets>=3.0.0",
    "deepspeed>=0.9.6; platform_system != 'Darwin' and platform_system != 'Windows'",
    "diffusers>=0.32.2",
    "huggingface_hub>=0.24.0",
    "nltk",
    "peft>=0.17.0",
    "sentencepiece>=0.2.1",                                                           # Also implicitly used in test_unified_export_megatron, test_vllm_fakequant_megatron_export
    "tiktoken",
    "transformers>=4.57,<5.15",                                                       # Should match modelopt/torch/__init__.py and noxfile.py
    "wonderwords",
]

puzzletron = [ # Dependedencies for modelopt.torch.puzzletron subpackage
    "fire",
    "hydra-core~=1.3.4",
    "immutabledict",
    "lru-dict",
    "pandas",
    "typeguard",
]
dev-lint = [
    "bandit[toml]==1.7.9", # security/compliance checks
    "mypy==2.1.0",
    "pre-commit==4.6.0",
    "ruff==0.15.20",
]
# Docs are only built on Python >=3.12 (sphinx 9.x requires it); markers keep `uv lock` from resolving these for 3.10/3.11, which we don't care about for building docs.
dev-docs = [
    "autodoc_pydantic~=2.2.0; python_version >= '3.12'",
    "sphinx~=9.1.0; python_version >= '3.12'",
    "sphinx-argparse~=0.6.0; python_version >= '3.12'",
    "sphinx-autobuild==2025.8.25; python_version >= '3.12'",
    "sphinx-copybutton~=0.5.2; python_version >= '3.12'",
    "sphinx-inline-tabs==2025.12.21.14; python_version >= '3.12'",
    "shibuya~=2026.7.0; python_version >= '3.12'",
    "sphinx-togglebutton~=0.4.0; python_version >= '3.12'",
]
dev-test = [
    "coverage[toml]~=7.14.0", # a1_coverage.pth for subprocess tracking requires this
    "nox",
    "pytest~=9.1.0",
    "pytest-cov~=7.1.0",
    "pytest-instafail==0.5.0",
    "pytest-timeout~=2.4.0",
    "tomli; python_version < '3.11'",
    "uv",
    # test-specific dependencies
    "psutil", # tools/resource_monitor.py sidecar (tests/unit/tools)
    "timm",
    "torchprofile==0.0.4", # optional dependency for modelopt.torch
    "torchvision",
    "torch-geometric",
]
# Compound extras via self-references
mlflow = [ # Dependencies for modelopt.torch.utils.mlflow run tracking
    "mlflow-skinny>=2.9", # client only; the tracking server's own deps are not needed here
]

# mlflow is deliberately not in `all`: it is opt-in tracking for the examples, and the
# full client would land in the documented `nvidia-modelopt[all]` install path.
all = ["nvidia-modelopt[onnx,hf,puzzletron]"]
dev = ["nvidia-modelopt[all,mlflow,dev-docs,dev-lint,dev-test]"]

[project.urls]
Homepage = "https://github.com/NVIDIA/Model-Optimizer"

[tool.setuptools.packages.find]
include = ["modelopt*"]

[tool.setuptools.package-data]
modelopt = ["**/*.h", "**/*.cpp", "**/*.cu"]
modelopt_recipes = ["**/*.yml", "**/*.yaml"]

[tool.setuptools.exclude-package-data]
# huggingface/models is a backward-compat symlink to ../models. The recursive
# package-data glob above follows it, so drop the aliased copies here to avoid
# shipping every checkpoint recipe twice; setuptools' exclude glob is non-recursive,
# hence the explicit org/model/task depths. MANIFEST.in prunes the symlink dir entry
# itself (build_py can't copy a symlink-to-dir). Old huggingface/models/... --recipe
# paths keep working via the loader alias in modelopt/recipe/loader.py.
modelopt_recipes = [
    "huggingface/models",
    "huggingface/models/*/*/*/*.yaml", "huggingface/models/*/*/*/*.yml",
    "huggingface/models/*/*/*/*/*.yaml", "huggingface/models/*/*/*/*/*.yml",
]

[tool.uv]
managed = true
# override-dependencies = ["torch; sys_platform == 'never'"]


####################################################################################################
###############################  LINTING, FORMATTING CONFIGURATION  ################################
####################################################################################################
[tool.ruff]
target-version = "py310"
line-length = 100        # Line length limit for code
fix = true

[tool.ruff.format]
# Like Black, respect magic trailing commas.
skip-magic-trailing-comma = false
docstring-code-format = true
# Set the line length limit used when formatting code snippets in docstrings.
docstring-code-line-length = "dynamic"

[tool.ruff.lint]
# See available rules at https://docs.astral.sh/ruff/rules/
# Flake8 is equivalent to pycodestyle + pyflakes + mccabe.
select = [
    "C4",   # Flake8 comprehensions
    "D",    # pydocstyle
    "E",    # pycodestyle errors
    "F",    # pyflakes
    "FURB", # refurb
    "I",    # isort
    "ISC",  # flake8-implicit-str-concat
    "N",    # pep8 naming
    "PERF", # Perflint
    "PGH",  # pygrep-hooks
    "PIE",  # flake8-pie
    "PLE",  # pylint errors
    "PLR",  # pylint refactor
    "PT",   # flake8-pytest-style
    "RUF",  # ruff
    "SIM",  # flake8-simplify
    "TC",   # flake8-type-checking
    "UP",   # pyupgrade
    "W",    # pycodestyle warnings
]
extend-ignore = [
    "D105",
    "D417",
    "N812",
    "PLR0402",
    "PLR0912",
    "PLR0913",
    "PLR0915",
    "PLR2004",
    "PLR0911",
    "PT011",
    "PT018",
    "PT028",
    "PT030",
    "RUF002",
    "RUF012",
    "RUF059",
    "SIM102",
    "SIM108",
    "SIM115",
]


[tool.ruff.lint.per-file-ignores]
"__init__.py" = ["F401", "F403"]
"examples/*" = ["D"]
"noxfile.py" = ["D", "E501"]
"tests/*" = ["B017", "D", "E402", "PT012"]
"plugins/modelopt/skills/*/tests/test_*.py" = ["D", "E402"] # Skill test scripts: docstring (D) + sys.path import-order (E402) exemptions
"*/_[a-zA-Z]*" = ["D"] # Private packages (_abc/*.py) or modules (_xyz.py)
"*.ipynb" = ["D", "E501"] # Ignore missing docstrings or line length for Jupyter notebooks
"modelopt/torch/kernels/*" = ["N803", "N806", "E731"] # triton style
"modelopt/torch/puzzletron/*" = [
    "C4",
    "D",
    "E",
    "F",
    "N",
    "PERF",
    "RUF",
    "SIM",
    "UP",
] # TODO: Disabled for now, will enable once newer puzzletron code with ruff fixes is migrated

[tool.ruff.lint.pycodestyle]
max-line-length = 120 # Line length limit for comments and docstrings


[tool.ruff.lint.pydocstyle]
convention = "google"


[tool.ruff.lint.isort]
known-first-party = ["modelopt"]
known-third-party = ["vllm"]
split-on-trailing-comma = false


[tool.mypy]
files = "."
python_version = "3.10"
install_types = true
non_interactive = true
show_error_codes = true
disable_error_code = [
    "assignment",
    "operator",
    "has-type",
    "var-annotated",
    "override",
]
enable_error_code = [
    "ignore-without-code",  # require `# type: ignore[code]` instead of bare `# type: ignore`
    "deprecated",           # use of `@deprecated`-marked APIs
    # TODO: enable in a follow-up (each surfaces many errors needing case-by-case fixes):
    #   "possibly-undefined", "redundant-expr", "truthy-bool"
]
explicit_package_bases = true
namespace_packages = true
# strict checks
strict = true
disallow_subclassing_any = false
disallow_untyped_decorators = false
disallow_any_generics = false
disallow_untyped_calls = false
disallow_incomplete_defs = false
disallow_untyped_defs = false
warn_return_any = false


[[tool.mypy.overrides]]
module = ["tests.*"]
ignore_errors = true

[[tool.mypy.overrides]]
module = ["examples.*"]
disable_error_code = ["attr-defined"]

# Vendored verbatim from NVIDIA-NeMo/Automodel (Apache-2.0); kept faithful to upstream rather
# than annotated to modelopt's strict mypy (e.g. worker-global ``X | None`` initialized lazily).
[[tool.mypy.overrides]]
module = ["examples.diffusers.fastgen.preprocess.*"]
ignore_errors = true

[tool.bandit]
exclude_dirs = [".github/", "examples/", "noxfile.py", "tests/"]
# Do not change `skips`. It should be consistent with NVIDIA's Wheel-CI-CD bandit.yml config.
# Use of `# nosec BXXX` requires special approval
skips = [
    "B101", # assert_used
    "B110", # try_except_pass
    "B112", # try_except_continue
    "B303", # MD2, MD4, MD5, or SHA1
    "B311", # random
]


####################################################################################################
###############################  TESTING CONFIGURATION  ############################################
####################################################################################################
[tool.pytest.ini_options]
# Default additional options
# Show a short test summary info for all except passed tests with -ra flag
# print execution time for 50 slowest tests and generate coverage reports
addopts = "-v -ra --instafail --cov-report=term-missing --cov-report=html --cov-report=xml:coverage.xml --cov-config=pyproject.toml --durations=50 --strict-markers"
pythonpath = ["tests/"]
# Apply per-test timeouts (see tests/conftest.py) to the test call only, not fixture setup/teardown
timeout_func_only = true
markers = [
    "integration: Tests that require external services or other non-hermetic dependencies",
    "manual: Only run when --run-manual is given",
    "release: Regression tests that should be run before every release",
]


[tool.coverage.run]
branch = false
parallel = true
concurrency = ["multiprocessing", "thread"]
source_pkgs = ["modelopt", "modelopt_recipes"]


[tool.coverage.report]
skip_covered = true
ignore_errors = true
exclude_lines = [
    "pragma: no cover",
    # Don't complain about missing debug or verbose code
    "def __repr__",
    "if verbose",
    # Don't complain if tests don't hit defensive exception handling code
    "raise AssertionError",
    "raise NotImplementedError",
    "raise RuntimeError",
    "raise ValueError",
    "raise KeyError",
    "raise AttributeError",
    "except ImportError",
    # Don't complain if non-runnable code isn't run
    "if __name__ == \"__main__\":",
    "if TYPE_CHECKING:",
    # Don't complain about abstract methods, they aren't run
    "@(abc\\.)?abstractmethod",
]