[project]
name = "vibevoice-1.5b-onnx"
version = "0.1.0"
description = "VibeVoice-1.5B TTS — ONNX sub-parts (LLM decoder + acoustic/semantic tokenizers + diffusion head + connectors) with inference/eval drivers."
readme = "README.md"
requires-python = ">=3.10"

# Runtime deps for inference (onnxruntime + the diffusion scheduler/tokenizer from the VibeVoice source).
dependencies = [
    "numpy>=1.26",
    "onnxruntime>=1.20",          # swap for onnxruntime-gpu on a CUDA box to run the gpu_* builds
    "onnxruntime-genai>=0.14",
    "transformers>=5.4",
    "safetensors>=0.4",
    "soundfile>=0.12",
    "librosa>=0.10",
    # The VibeVoice source (DPMSolverMultistepScheduler + modular tokenizer/diffusion-head classes)
    # is VENDORED — it ships as the `vibevoice/` directory alongside these drivers, so there is NO
    # pip/git dependency on it. The drivers import individual leaf modules from that tree by path
    # (isolated-import shim), never `import vibevoice` (HANDOFF trap 1: its package __init__ collides
    # with transformers' native `vibevoice` registration and pulls diffusers + a renamed qwen2 tokenizer).
]

# Build the ONNX sub-parts (needs the source checkpoint + onnxruntime-genai ModelBuilder):
#   uv run optimize.py --device cpu --precision int4 1.5b     # or fp32; fp16/gpu need a CUDA box
#   uv run eval.py      --device cpu --precision int4 1.5b     # 9/9 component + pipeline parity
#   uv run inference.py --text "Hello world." onnx/1.5b/cpu_int4