pyproject.toml

[build-system]
requires = ["setuptools>=68", "wheel"]
build-backend = "setuptools.build_meta"

[project]
name = "agentflow-pro"
version = "0.1.0"
description = "Trainable multi-agent reasoning framework with DAPO + PRM"
requires-python = ">=3.11"
dependencies = [
    "httpx>=0.28.0",
    "openai>=1.75.0",
    "pydantic>=2.11.0",
    "rich>=13.9.0",
    "typer>=0.15.0",
    "python-dotenv>=1.1.0",
    "tavily-python>=0.5.0",
    "sympy>=1.13.0",
    "numpy>=2.0.0",
]

[project.optional-dependencies]
tools = [
    "fastmcp>=2.3.0",
]
eval = [
    "datasets>=3.5.0",
    "math-verify>=0.5.0",
]
memory = [
    "qdrant-client>=1.14.0",
    "sentence-transformers>=3.4.0",
]
rl = [
    # DAPO lives in TRL (loss_type="dapo" is the default; clip-higher via
    # epsilon_high, soft overlong punishment via trl.rewards). The policy is
    # trained as bf16 + a PEFT LoRA adapter — NO unsloth/bitsandbytes, which keeps
    # the dependency set small and reproducible (a 48GB GPU fits the 8B in bf16).
    # transformers is pinned <5: the 5.x line breaks trl/peft (validated 2026-05).
    "transformers>=4.55.0,<5.0",
    "trl>=1.5.0,<2.0",
    "peft>=0.17.0",
    "torch>=2.6.0",
    # torch's Linux CUDA wheels need cuSPARSELt at runtime; pin it explicitly so a
    # lock generated on a non-CUDA box (macOS) still installs it on the Linux GPU.
    "nvidia-cusparselt-cu12; sys_platform == 'linux'",
    "accelerate>=1.6.0",
    "datasets>=3.5.0",
]
dev = [
    "pytest>=8.3.0",
    "pytest-asyncio>=0.25.0",
]

[tool.setuptools.packages.find]
where = ["."]
include = ["core*", "rl*", "tools*", "eval*", "train*"]