migrate to chatterbox
This commit is contained in:
parent
716e42a57d
commit
988c01c4f7
7 changed files with 480 additions and 1786 deletions
|
|
@ -1,41 +1,15 @@
|
|||
[project]
|
||||
name = "property-map-video-tts"
|
||||
version = "0.1.0"
|
||||
description = "Qwen3-TTS narration generator for the homepage demo video."
|
||||
requires-python = ">=3.12,<3.13"
|
||||
dependencies = [
|
||||
"qwen-tts>=0.1.1",
|
||||
# Host driver is CUDA 12.4 (see `nvidia-smi`). torch 2.7+ dropped cu124
|
||||
# wheels, so we cap below that and pull the cu124 build from PyTorch's
|
||||
# own index (configured below). torchaudio must match torch's CUDA build:
|
||||
# the PyPI default ships a CUDA 13 binary that fails to load
|
||||
# libcudart.so.13 on this host.
|
||||
"torch>=2.5,<2.7",
|
||||
"torchaudio>=2.5,<2.7",
|
||||
"soundfile>=0.12",
|
||||
"numpy>=1.26",
|
||||
]
|
||||
|
||||
[project.optional-dependencies]
|
||||
# Flash-attention prebuilt wheel matched to torch 2.6 + cu12 + cp312, old
|
||||
# CXX ABI (PyTorch's cu124 wheel reports compiled_with_cxx11_abi() == False
|
||||
# and only exports the old-ABI c10::Error constructor). Pinned to
|
||||
# 2.7.4.post1 because 2.8.x's torch2.6/abiFALSE wheels were mislabelled:
|
||||
# they ship new-ABI symbols and fail to import. Building from source needs
|
||||
# nvcc which isn't on the host. Enable via `uv sync --extra gpu`; render.sh
|
||||
# does this automatically when nvidia-smi reports a GPU.
|
||||
gpu = [
|
||||
"flash-attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1%2Bcu12torch2.6cxx11abiFALSE-cp312-cp312-linux_x86_64.whl ; sys_platform == 'linux' and platform_machine == 'x86_64'",
|
||||
]
|
||||
description = "Chatterbox narration generator + ffmpeg muxer for the homepage demo video."
|
||||
requires-python = ">=3.12"
|
||||
# Deliberately dependency-free. synth.py talks to the Chatterbox TTS server
|
||||
# over HTTP (urllib) and reads WAV headers with the stdlib `wave` module;
|
||||
# mux.py shells out to ffmpeg. The previous Qwen3-TTS implementation ran the
|
||||
# model in-process, which dragged in torch + torchaudio + a CUDA-matched
|
||||
# flash-attn wheel (~700MB of downloads, pinned to the host's driver version
|
||||
# and to cp312). Generating over HTTP moved all of that onto the server.
|
||||
dependencies = []
|
||||
|
||||
[tool.uv]
|
||||
environments = ["sys_platform == 'linux' and python_version < '3.13'"]
|
||||
|
||||
[tool.uv.sources]
|
||||
torch = [{ index = "pytorch-cu124" }]
|
||||
torchaudio = [{ index = "pytorch-cu124" }]
|
||||
|
||||
[[tool.uv.index]]
|
||||
name = "pytorch-cu124"
|
||||
url = "https://download.pytorch.org/whl/cu124"
|
||||
explicit = true
|
||||
environments = ["sys_platform == 'linux' and python_version < '3.14'"]
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue