pyproject.toml

[build-system]
requires = ["hatchling"]
build-backend = "hatchling.build"

[tool.hatch.metadata]
allow-direct-references = true

[project]
name = "musicvision"
version = "0.1.0"
description = "AI music video production pipeline: song → scenes → storyboard → video"
requires-python = ">=3.11"
license = "Apache-2.0"
license-files = ["LICENSE", "NOTICE", "THIRD-PARTY-NOTICES.md", "CLA.md"]
authors = [{ name = "Todd Green", email = "tsondo@gmail.com" }]
maintainers = [{ name = "Todd Green", email = "tsondo@gmail.com" }]

dependencies = [
    # --- Core ---
    "pydantic>=2.0",
    "pyyaml>=6.0",
    "rich>=13.0",           # CLI pretty-printing & progress bars
    # --- API ---
    "fastapi>=0.110",
    "uvicorn[standard]>=0.27",
    "python-multipart>=0.0.7",  # file uploads
    # --- Audio ---
    "librosa>=0.10",        # BPM detection, audio analysis
    "soundfile>=0.12",      # audio I/O
    # --- AI / ML (installed separately due to CUDA) ---
    # torch, torchvision, torchaudio  → install via pip with --index-url cu128
    # flash_attn                      → NOT required; vendored DiT uses SDPA fallback
    #                                   FA2 lacks sm_120 (Blackwell) kernels as of Feb 2026
    # transformers                     → for Whisper
    # diffusers                        → for FLUX
    # accelerate                       → for model offloading
    # safetensors                      → for weight loading
    # --- LLM ---
    "anthropic>=0.40",      # Claude API (cloud backend)
    "openai>=1.0",          # OpenAI-compatible client — used for local vLLM backend
    # --- Utilities ---
    "Pillow>=10.0",
    "numpy>=1.26",
    "torchcodec>=0.10.0",
]

[project.optional-dependencies]
ml = [
    "torch>=2.7",
    "torchvision>=0.22",
    "torchaudio>=2.7",
    "transformers>=4.40",
    "diffusers>=0.30",
    "accelerate>=0.30",
    "safetensors>=0.4",
    "sentencepiece>=0.2",   # T5 tokenizer for HuMo/Wan
    "ftfy>=6.0",             # text cleaning for Wan-AI tokenizer
    "regex>=2023.0",         # Wan-AI tokenizer uses regex (not re)
    "einops>=0.7",           # Wan-AI VAE uses einops.rearrange
    # --- Quantization (FLUX fp8/int8 on limited VRAM) ---
    "optimum-quanto>=0.2",  # fp8/int8 quantization for FLUX transformer
    # --- HuMo weight management ---
    "huggingface-hub>=0.21",  # hf_hub_download / snapshot_download for weight registry
    "gguf>=0.6",              # GGUF dequantization for GGUFLoader (Q6_K, Q4_K_M)
    # --- Demucs vocal separation ---
    # Pin to tested v4.0.1 from StemForge (matches demucs.apply/demucs.audio API used here).
    # demucs.pretrained.get_model auto-downloads weights (~300 MB) on first use.
    "demucs @ git+https://github.com/facebookresearch/demucs.git@v4.0.1",
    "julius==0.2.7",    # demucs dep: efficient audio resampling (julius.resample_frac)
    "lameenc==1.8.1",   # demucs dep: optional mp3 encode
    "openunmix==1.3.0", # demucs dep: UMX separation backend
    "dora-search==0.1.12", # demucs dep: experiment runner
    "submitit==1.5.4",  # demucs dep
    "diffq==0.2.4",     # required by mdx_extra_q only (quantised MDX model)
    # --- MelBandRoFormer vocal separation ---
    # audio-separator must be installed separately with GPU support:
    # pip install "audio-separator[gpu]"
    # (not listed here because it pulls onnxruntime-gpu which conflicts on some systems)
]
dev = [
    "pytest>=8.0",
    "pytest-asyncio>=0.23",
    "httpx>=0.27",          # async test client for FastAPI
    "ruff>=0.4",
]

[project.scripts]
musicvision = "musicvision.cli:main"

[tool.ruff]
target-version = "py311"
line-length = 120

[tool.ruff.lint]
select = ["E", "F", "I", "N", "W", "UP"]

[tool.pytest.ini_options]
asyncio_mode = "auto"
testpaths = ["tests"]