[project]
name = "vdn"
version = "0.1.0"
description = "MiniMax-H3 with hybrid window-softmax / linear attention: inference and the Stage A-D trainers"
readme = "README.md"
license = { file = "LICENSE" }   # Apache-2.0, the code only; the weights carry the MiniMax H3 agreement
requires-python = ">=3.12,<3.13"
dependencies = [
    # torch and diffusers are NOT listed: torch is installed first from the cu129 index
    # (flash-attn-4's dependency tree would otherwise pull the PyPI default, cu130) and
    # diffusers is the patched tree in ./diffusers, installed editable before this (README).
    "triton==3.7.1",
    "flash-attn-4==4.0.0b26",     # flash_attn.cute -- the FA4 CuTe kernels behind FlexAttention
    "transformers==5.15.0",
    "accelerate==1.14.0",
    "torchao==0.18.0",            # fp8 on the diffusers path (infer_diffusers.py --fp8)
    "peft==0.20.0",
    "safetensors==0.8.0",
    "huggingface_hub>=1.27.0",
    "omegaconf==2.3.1",
    "pyyaml==6.0.3",
    "einops==0.8.2",
    "numpy==2.5.2",
    "pillow==12.3.0",
    "torchvision==0.28.0",        # Qwen3-VL's processor needs it (src/inference/encode_prompt.py)
    "av==18.1.0",                 # mp4 muxing (diffusers.utils.export_utils.encode_video)
    "tqdm",
]

[tool.setuptools]
packages = { find = { include = ["src*"] } }