Installation with Nvidia GPU

In this section, we provide the installation guide for Nvidia GPU.

VeOmni also supports other hardware platform, please refer to Ascend.

Required Environment

CUDA 13.0 (the gpu extra targets +cu130 torch wheels and the nvcr.io/nvidia/pytorch:25.11-py3 base image).

Install with uv or pip

UV

Recommend to use uv for faster and easier installation.

git clone https://github.com/ByteDance-Seed/VeOmni.git
cd VeOmni

uv sync --locked --extra gpu
source .venv/bin/activate

gpu is a single full superset: cu130 torch, FA2 (cp311/cp312 prebuilt wheels) / FA3 (sm90 abi3 prebuilt wheel) / FA4 / FlashQLA, diffusion / audio / video / LoRA deps, and megatron-energon for the optional energon dataset format. See pyproject.toml for the full list.

Note: video/audio processing also needs ffmpeg installed at the OS level:

# Ubuntu/Debian
sudo apt-get install ffmpeg

# macOS
brew install ffmpeg

Pip

git clone https://github.com/ByteDance-Seed/VeOmni.git
cd VeOmni

pip3 install -e .[gpu]