pyproject.toml

[build-system]
requires = ["setuptools>=61.0", "wheel"]
build-backend = "setuptools.build_meta"

[project]
name = "complexity-framework"
version = "0.3.0"
description = "Modular Python framework for Transformer and Token-Routed language-model research"
readme = "README.md"
license = {text = "CC-BY-NC-4.0"}
authors = [
    {name = "Complexity-ML", email = "contact@complexity-ml.org"}
]
keywords = [
    "llm",
    "transformer",
    "deep-learning",
    "pytorch",
    "attention",
    "moe",
    "mixture-of-experts",
    "flash-attention",
    "token-routed",
    "tr-hash",
]
classifiers = [
    "Development Status :: 4 - Beta",
    "Intended Audience :: Developers",
    "Intended Audience :: Science/Research",
    "License :: Other/Proprietary License",
    "Programming Language :: Python :: 3",
    "Programming Language :: Python :: 3.10",
    "Programming Language :: Python :: 3.11",
    "Programming Language :: Python :: 3.12",
    "Topic :: Scientific/Engineering :: Artificial Intelligence",
]
requires-python = ">=3.10"
dependencies = [
    # NOTE: torch is intentionally NOT in required deps. PyPI's default `torch`
    # wheel is the NVIDIA CUDA build — installing it via `pip install -e .` on a
    # ROCm host silently replaces the AMD-built torch and yanks in 1.5 GB of
    # nvidia-* packages. Pick a backend extra ([cuda] / [rocm] / [cpu]) so
    # torch resolves against the right wheel index.
    "numpy>=1.23.0",
    "einops>=0.6.0",
    "transformers>=4.30.0",
    "tokenizers>=0.13.0",
    "tiktoken>=0.7.0",
    "datasets>=2.0.0",
    "tqdm>=4.0.0",
    "wandb>=0.15.0",
    "safetensors>=0.4.0",   # save_pretrained + checkpointing
    "filelock>=3.16.1",     # process-safe lazy token-shard cache
    "jinja2>=3.0.0",        # standalone HF/vLLM/MLX chat-template rendering
    "pyyaml>=6.0",          # imported at module load in config/model_config.py
    "typer>=0.9.0",         # CLI (complexity.cli)
    "tensorboard>=2.14.0",  # complexity.training.callbacks.TensorBoardCallback
    "tomli>=2.0.0; python_version < '3.11'",  # portable Supervisor job manifests
]

[project.optional-dependencies]
# IMPORTANT: torch is intentionally NOT listed in any backend extra.
#
# Reason: pip cannot attach a `--index-url` to an extra. If we list torch here
# pip will resolve it against whatever indexes are in scope at install time,
# which on a stock setup means PyPI's CUDA wheel — even when the user wanted
# ROCm. Pinning local-version tags (e.g. `torch>=2.9+rocm6.4`) is fragile
# because tags change per ROCm release.
#
# Instead, install torch in a SEPARATE step before the framework, against the
# backend-appropriate index. The `scripts/install_backend.sh` helper or the
# Makefile targets (`make install-rocm` / `install-cuda` / `install-cpu`) do
# this in one go. The extras below only carry backend-specific kernel libs
# that DO live on PyPI.
cuda = [
    "triton>=2.0.0",
    "liger-kernel>=0.5.0",  # fused linear+CE, RMSNorm, RoPE (Triton)
    # flash-attn removed: builds from source (~30min), SDPA covers the same path
]
rocm = [
    "pytorch-triton-rocm; platform_system == 'Linux'",
    "liger-kernel>=0.5.0",  # exact fused linear+CE and Triton kernels; supports ROCm through Triton
]
cpu = []
fp8 = [
    "torchao>=0.7.0",
]
image = [
    "Pillow>=10.0.0",
    "diffusers>=0.30.0",
    "accelerate>=0.30.0",
    "hf-xet>=1.1.0",
]
dev = [
    "pytest>=7.0.0",
    "pytest-cov>=4.0.0",
    "black>=23.0.0",
    "ruff>=0.1.0",
    "Pillow>=10.0.0",
]
tools = [
    "mcp>=1.0.0; python_version >= '3.10'",
]
pretrain-data = [
    "boto3>=1.34.0",
    "hf-xet>=1.1.0",
]
serve = [
    "fastapi>=0.115.0",
    "uvicorn>=0.30.0",
    "python-multipart>=0.0.9",
]
export = [
    "Pillow>=10.0.0",
    "onnx>=1.12.0; platform_system != 'Darwin'",
    "onnx>=1.12.0,<1.18.0; platform_system == 'Darwin' and python_version < '3.13'",
    "onnx>=1.20.0; platform_system == 'Darwin' and python_version >= '3.13'",
    "onnxslim>=0.1.82",
    "onnxruntime<1.20.0; python_version < '3.11'",
    "onnxruntime>=1.20.0; python_version >= '3.11'",
]
detection = [
    # Public computer-vision runtime stack aligned with Ultralytics. Torch is
    # still installed separately so CUDA, ROCm and CPU wheels cannot mix.
    "packaging>=23.0",
    "matplotlib>=3.3.0",
    "opencv-python>=4.7.0,!=4.13.0.90",
    "Pillow>=10.0.0",
    "requests>=2.23.0",
    "torchvision>=0.9.0",
    "psutil>=5.8.0",
    "polars>=0.20.0",
    "nvidia-ml-py>=12.0.0; platform_system == 'Linux'",
    "albumentations>=1.4.6",
    "faster-coco-eval>=1.6.7",
    "pycocotools>=2.0.7",
]
viz = [
    "matplotlib>=3.8.0",
    "plotly>=5.20.0",
    "scikit-learn>=1.4.0",
]

[project.scripts]
complexity = "complexity.cli.app:main"
cf-plan-run = "complexity.training.plan_run:main"
cf-plan-cluster = "complexity.training.cluster_plan:main"
cf-image-train = "complexity.generative.image.training:main"
cf-image-edit-train = "complexity.generative.image.edit_training:main"
cf-detector-train = "complexity.generative.detection.training:main"
cf-vision-pretrain = "complexity.generative.vision_language.pretraining:main"
cf-detector-serve = "complexity.generative.detection.service:main"
cf-sensor-fusion-train = "complexity.generative.sensor_fusion.training:main"
cf-sensor-fusion-submit = "complexity.generative.sensor_fusion.submission:main"

[project.urls]
Homepage = "https://github.com/Complexity-ML/complexity-framework"
Documentation = "https://github.com/Complexity-ML/complexity-framework/tree/main/docs"
Repository = "https://github.com/Complexity-ML/complexity-framework"
Issues = "https://github.com/Complexity-ML/complexity-framework/issues"

[tool.setuptools.packages.find]
where = ["."]
include = ["complexity*"]
exclude = ["docs*", "tests*"]

[tool.setuptools.package-data]
"*" = ["*.py"]

[tool.setuptools.exclude-package-data]
"*" = ["*.png", "*.jpg", "*.jpeg", "*.gif", "*.svg", "*.md"]

[tool.black]
line-length = 100
target-version = ["py310", "py311", "py312"]

[tool.ruff]
line-length = 100

[tool.ruff.lint]
select = ["E", "F", "W", "I"]
ignore = ["E501"]

[tool.pytest.ini_options]
markers = [
    "multigpu: requires more than one accelerator device",
]