OVI Webui Manual Installation

Manual Install Method from my last Youtube video: https://youtu.be/2SvPfkXs3Nk


Manual Installation Guide for Ovi (Windows)

Step 1: Prerequisites
Install these first:
- Git – https://git-scm.com/download/win
- Python 3.10 (64-bit) – https://www.python.org/downloads/release/python-31011/
- NVIDIA CUDA Toolkit 12.8 (or newer driver) – https://developer.nvidia.com/cuda-downloads
  (You don’t need to add CUDA to PATH; we use PyTorch cu128 wheels.)

Step 2: Download & Enter Ovi
Open Command Prompt in the folder where you want it:
  git clone https://github.com/character-ai/Ovi.git
  cd Ovi

Step 3: Create Virtual Environment
  python -m venv venv
  CALL venv\Scripts\activate


Step 4: Install PyTorch (CUDA 12.8 build)
Known-good on Windows (Python 3.10):
  pip install --upgrade pip setuptools wheel
  pip install --index-url https://download.pytorch.org/whl/cu128 --extra-index-url https://pypi.org/simple ^
    torch==2.7.0+cu128 torchvision==0.22.0+cu128 torchaudio==2.7.0+cu128

Step 5: Install Ovi Dependencies
  pip install -r requirements.txt

Step 6: Faster Hugging Face Downloads (optional)
  pip install -U hf_transfer

Step 7: (Optional) FlashAttention for Speed
- SKIP this on RTX 50xx / Blackwell (not compatible).
- OK on 4000 series with this wheel:
  pip install triton-windows==3.4.0.post20
  pip install https://github.com/petermg/flash_attn_windows/releases/download/01/flash_attn-2.7.4.post1+cu128.torch270-cp310-cp310-win_amd64.whl
If it errors, skip, Ovi still works without it.

Step 8: Download Model Weights
  python download_weights.py

Step 9: Download the Launcher (for next time)
  curl -L -o LAUNCHER-OVI.bat "https://huggingface.co/Aitrepreneur/test/resolve/main/LAUNCHER-OVI.bat?download=true"

Step 10: Run the WebUI
If you’re tight on VRAM (<48 GB), add --cpu_offload.
python gradio_app.py --cpu_offload
Open local URL

Next time, double-click LAUNCHER-OVI.bat in the Ovi folder.


Troubleshooting (Windows)
- CUDA out of memory (OOM) → run with --cpu_offload WITH 5090 MINIMUM YOU HAVE BEEN WARNED!
- nvidia-smi not found → install/update NVIDIA driver (CUDA 12.8+).
- FlashAttention install fails → skip it.





Manual Installation Guide for RunPod

Step 1: Prerequisites
- Create account: https://bit.ly/runpodAi
- Deploy a pod with an NVIDIA GPU (≥32 GB VRAM recommended).
- Open a terminal in /workspace.

Step 2: System Packages
  apt-get update -y
  apt-get install -y --no-install-recommends git git-lfs ffmpeg build-essential python3-venv curl
  git lfs install

Step 3: Clone & Enter Ovi
  cd /workspace
  git clone https://github.com/character-ai/Ovi.git
  cd Ovi

Step 4: Virtual Environment
  python3 -m venv .venv
  source .venv/bin/activate
  python -m pip install -U pip setuptools wheel

Step 5: Install Dependencies
If your base image already has a compatible Torch/CUDA, you can skip installing Torch.
Otherwise, install Torch (cu128) first, then Ovi deps:
  pip install --index-url https://download.pytorch.org/whl/cu128 --extra-index-url https://pypi.org/simple \
    torch==2.8.0+cu128 torchvision==0.23.0+cu128 torchaudio==2.8.0+cu128 || true
  pip install -r requirements.txt

Step 6: Faster Hugging Face Downloads (optional)
  pip install -U hf_transfer || true
  export HF_HUB_ENABLE_HF_TRANSFER=1

Step 7: (Optional) FlashAttention
- Skip on RTX 50xx / Blackwell.
- You can try the wheel/build; if it fails, continue without it:
  set +e
  pip install flash_attn --no-build-isolation
  set -e || true

Step 8: Download Weights
  python3 download_weights.py

Step 9: Launch
  export GRADIO_SERVER_NAME=0.0.0.0
  export GRADIO_SERVER_PORT=3000
  # Add --cpu_offload if VRAM is tight
  python3 gradio_app.py --cpu_offload
Expose/forward port 3000 and open the public URL from RunPod.

Notes & Tips
- <48 GB VRAM? Use --cpu_offload.
- RTX 50xx / Blackwell? Don’t install FlashAttention.
- Slow weights download? Ensure HF_HUB_ENABLE_HF_TRANSFER=1 and hf_transfer is installed.