Deploying OpenTalking + OmniRT + QuickTalk on Windows

This guide targets Windows laptop deployment scenarios. The goal is to go from git clone to running OpenTalking + OmniRT + QuickTalk in WSL2 Ubuntu, and completing an E2E Benchmark.

---

0. Final Deployment Structure

Windows Host
  └── WSL2 Ubuntu
        ├── OpenTalking: Web / API / Session orchestration
        ├── OmniRT: QuickTalk inference backend
        └── CUDA: RTX 3050 Laptop GPU

Recommended directory structure:

/root/test/
├── opentalking/
│   ├── .venv/
│   ├── apps/
│   ├── configs/
│   ├── scripts/
│   └── models/quicktalk/checkpoints/
├── omnirt/
│   ├── .venv/
│   └── models/quicktalk/
└── models/
    └── quicktalk -> /root/test/opentalking/models/quicktalk

Place code in WSL2's own Linux filesystem (e.g., /root/test or /home/<user>/test), not directly under /mnt/d/....

---

1. Windows-Side Prerequisites

1.1 NVIDIA Driver

Confirm the GPU and driver work on Windows:

nvidia-smi

Expected output:

NVIDIA GeForce RTX 3050 Laptop GPU
CUDA Version: 13.0

Note: The CUDA Version here indicates the highest CUDA version supported by the driver, not that CUDA Toolkit must be installed. PyTorch can be installed with CUDA wheels directly.

---

1.2 Install and Verify WSL2

In an administrator PowerShell:

wsl --version
wsl --status

Expected:

Default version: 2

If Ubuntu is not installed yet:

wsl --install -d Ubuntu-22.04

For a manually imported Ubuntu:

wsl --import Ubuntu-22.04 D:\wsl\Ubuntu-22.04 D:\wsl\downloads\ubuntu-22.04-wsl.rootfs.tar.gz --version 2

Enter WSL2:

wsl -d Ubuntu-22.04

Verify GPU in WSL2:

nvidia-smi

If WSL2 can see the RTX 3050, the CUDA inference prerequisites are met.

---

2. WSL2 Base Dependencies

The following commands run inside WSL2 Ubuntu. If running as root, sudo is not needed; otherwise prepend sudo to apt commands.

apt update
apt install -y \
  python3-pip python3-venv python3-dev \
  build-essential pkg-config \
  curl wget git git-lfs rsync unzip \
  ffmpeg nodejs npm jq \
  iproute2 procps psmisc \
  libgl1 libglib2.0-0

git lfs install

Check basic tools:

python3 --version
ffmpeg -version
node --version
npm --version
nvidia-smi

---

3. Clone from GitHub

Enter the working directory:

mkdir -p /root/test
cd /root/test

Clone both repositories:

git clone https://github.com/datascale-ai/opentalking.git
git clone https://github.com/datascale-ai/omnirt.git

The final structure should be:

/root/test/opentalking
/root/test/omnirt

Verify:

ls /root/test/opentalking
ls /root/test/omnirt

Path Notes

If the code already exists on Windows, copy it to WSL2:

rsync -a --info=progress2 /mnt/d/test_opentalking/opentalking/ /root/test/opentalking/
rsync -a --info=progress2 /mnt/d/test_opentalking/omnirt/ /root/test/omnirt/

If the code is already downloaded on a server, sync it to WSL2:

rsync -avP root@<your-server-ip>:/root/lyf/temp/opentalking/ /root/test/opentalking/
rsync -avP root@<your-server-ip>:/root/lyf/temp/omnirt/ /root/test/omnirt/

---

4. Install uv and Configure Mirrors

Install uv in WSL2:

python3 -m pip install -U uv

If uv is not in PATH:

export PATH="/usr/local/bin:$HOME/.local/bin:$PATH"
echo 'export PATH="/usr/local/bin:$HOME/.local/bin:$PATH"' >> ~/.bashrc

Verify:

uv --version

Write commonly used mirror environment variables:

cat >> ~/.bashrc <<'EOF'
export PIP_INDEX_URL=https://pypi.org/simple
export UV_DEFAULT_INDEX=https://pypi.org/simple
export UV_INDEX_URL=https://pypi.org/simple
export HF_ENDPOINT=https://huggingface.co
export npm_config_registry=https://registry.npmjs.org
export CUDA_VISIBLE_DEVICES=0
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb=128
EOF

source ~/.bashrc

---

5. Configure OpenTalking Environment

Enter OpenTalking:

cd /root/test/opentalking

Create an isolated virtual environment:

uv venv --python python3 .venv
source .venv/bin/activate

Confirm Python path:

which python

Expected:

/root/test/opentalking/.venv/bin/python

Install base packages:

uv pip install -U pip setuptools wheel

Install OpenTalking dependencies:

uv pip install -e ".[dev,models]"

Install CUDA PyTorch:

uv pip install \
  torch==2.9.1+cu128 \
  torchvision==0.24.1+cu128 \
  torchaudio==2.9.1+cu128 \
  --find-links https://download.pytorch.org/whl/cu128/

Verify CUDA:

python -c "import torch; print('torch=', torch.__version__); print('torch cuda=', torch.version.cuda); print('cuda=', torch.cuda.is_available()); print('gpu=', torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'NO CUDA')"

Expected output:

torch= 2.9.1+cu128
torch cuda= 12.8
cuda= True
gpu= NVIDIA GeForce RTX 3050 Laptop GPU

Notes

Installing CUDA PyTorch on Linux/WSL2 will pull nvidia-cudnn-cu12, nvidia-cublas-cu12, triton, and other dependencies — these are large, which is normal. The key is that cuda=True must be confirmed.

---

6. Prepare QuickTalk Weights

QuickTalk weights should be placed at:

/root/test/opentalking/models/quicktalk/checkpoints/

The complete structure should look like:

checkpoints/
├── quicktalk.pth
├── repair.npy
├── chinese-hubert-large/
│   ├── config.json
│   ├── preprocessor_config.json
│   └── pytorch_model.bin
└── auxiliary/models/buffalo_l/
    ├── det_10g.onnx
    ├── w600k_r50.onnx
    ├── 2d106det.onnx
    └── ...

Check key files:

cd /root/test/opentalking

ls -lh models/quicktalk/checkpoints/quicktalk.pth
ls -lh models/quicktalk/checkpoints/repair.npy
ls -lh models/quicktalk/checkpoints/chinese-hubert-large/pytorch_model.bin
ls -lh models/quicktalk/checkpoints/auxiliary/models/buffalo_l/det_10g.onnx
ls -lh models/quicktalk/checkpoints/auxiliary/models/buffalo_l/w600k_r50.onnx
ls -lh models/quicktalk/checkpoints/auxiliary/models/buffalo_l/2d106det.onnx

Path Notes

When starting QuickTalk in OmniRT, point directly to the checkpoints directory:

export OMNIRT_QUICKTALK_MODEL_ROOT=/root/test/opentalking/models/quicktalk/checkpoints

If the benchmark script expects /root/test/models/quicktalk, create a symlink:

mkdir -p /root/test/models
ln -sfn /root/test/opentalking/models/quicktalk /root/test/models/quicktalk

---

7. Configure OmniRT Environment

Enter OmniRT:

cd /root/test/omnirt

Create an isolated virtual environment:

uv venv --python python3 .venv
source .venv/bin/activate

Confirm Python path:

which python

Expected:

/root/test/omnirt/.venv/bin/python

Install dependencies:

uv pip install -U pip setuptools wheel

uv pip install -e ".[dev,server,quicktalk-cuda]" \
  --find-links https://download.pytorch.org/whl/cu128/

If needed, reinstall CUDA torch:

uv pip install \
  torch==2.9.1+cu128 \
  torchvision==0.24.1+cu128 \
  torchaudio==2.9.1+cu128 \
  --find-links https://download.pytorch.org/whl/cu128/

Check:

python -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"
which omnirt
omnirt --help

Expected:

/root/test/omnirt/.venv/bin/omnirt

Sync models to OmniRT:

mkdir -p /root/test/omnirt/models/quicktalk
rsync -a --info=progress2 \
  /root/test/opentalking/models/quicktalk/ \
  /root/test/omnirt/models/quicktalk/

---

8. Configure OpenTalking .env

Enter OpenTalking:

cd /root/test/opentalking
cp -n .env.example .env

Recommended key configurations:

cat >> .env <<'EOF'
OPENTALKING_TTS_DEFAULT_PROVIDER=edge
OPENTALKING_REDIS_MODE=memory
OMNIRT_ENDPOINT=http://127.0.0.1:9000
EOF

If LLM / STT is needed, configure the corresponding keys separately:

OPENTALKING_LLM_API_KEY=your_LLM_API_KEY
OPENTALKING_STT_DASHSCOPE_API_KEY=your_STT_API_KEY

Notes

LLM, TTS, and STT are independent providers. Edge TTS does not require a key and is best for initially verifying the pipeline. An expired or invalid DashScope key will affect LLM / STT but does not mean QuickTalk inference is unavailable.

---

9. Start OmniRT QuickTalk Backend

Open a new WSL2 terminal and enter OmniRT:

cd /root/test/omnirt
source .venv/bin/activate

Set environment variables:

export CUDA_VISIBLE_DEVICES=0
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb=128

export OMNIRT_QUICKTALK_RUNTIME=1
export OMNIRT_QUICKTALK_MODEL_ROOT=/root/test/opentalking/models/quicktalk/checkpoints
export OMNIRT_QUICKTALK_DEVICE=cuda:0
export OMNIRT_QUICKTALK_HUBERT_DEVICE=cuda:0

Start:

.venv/bin/python -m omnirt.cli.main serve-avatar-ws \
  --host 0.0.0.0 \
  --port 9000 \
  --compat flashtalk \
  --backend cuda \
  --avatar-runtime fake

Key parameters:

Parameter	Purpose
OMNIRT_QUICKTALK_RUNTIME=1	Enable QuickTalk runtime
OMNIRT_QUICKTALK_MODEL_ROOT=.../checkpoints	Point to QuickTalk weights directory
--compat flashtalk	Compatible with OpenTalking WebSocket protocol
--avatar-runtime fake	Let QuickTalk runtime take effect, avoiding the FlashTalk resident path

---

10. Start OpenTalking

Open another WSL2 terminal and enter OpenTalking:

cd /root/test/opentalking
source .venv/bin/activate

Start:

bash scripts/start_unified.sh \
  --backend omnirt \
  --model quicktalk \
  --omnirt http://127.0.0.1:9000 \
  --host 0.0.0.0

Frontend URL:

http://127.0.0.1:5173

Verify QuickTalk connection status:

curl -s http://127.0.0.1:8000/models \
  | python3 -c "import sys,json; [print(s) for s in json.load(sys.stdin)['statuses'] if s['id']=='quicktalk']"

Expected:

connected=true
reason=omnirt

Check GPU memory:

nvidia-smi --query-gpu=memory.used --format=csv,noheader

---

11. Common Path Issues

Location	Correct approach
Code directory	Place in /root/test/opentalking and /root/test/omnirt
OpenTalking venv	/root/test/opentalking/.venv
OmniRT venv	/root/test/omnirt/.venv
QuickTalk weights	/root/test/opentalking/models/quicktalk/checkpoints
OmniRT QuickTalk root	Point to .../checkpoints
Benchmark compatibility path	ln -sfn /root/test/opentalking/models/quicktalk /root/test/models/quicktalk
Low-VRAM config	resolution=160/128, batch=1, HuBERT=cpu

---

12. Final Checklist

Before running the benchmark, verify each item:

# WSL2 GPU
nvidia-smi

# OpenTalking environment
cd /root/test/opentalking
source .venv/bin/activate
python -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"

# QuickTalk weights
ls -lh models/quicktalk/checkpoints/quicktalk.pth
ls -lh models/quicktalk/checkpoints/chinese-hubert-large/pytorch_model.bin
ls -lh models/quicktalk/checkpoints/auxiliary/models/buffalo_l/det_10g.onnx

# OmniRT environment
cd /root/test/omnirt
source .venv/bin/activate
which omnirt
omnirt --help

# Frontend / basic tools
ffmpeg -version
node --version
npm --version

If all checks pass, run:

cd /root/test/opentalking
source .venv/bin/activate
bash scripts/run_opentalking_e2e_benchmark.sh \
  --tester xxx \
  --model quicktalk \
  --backend omnirt \
  --gpu-index 0 \
  --timeout 300

---

13. Summary

Deploying OpenTalking + OmniRT + QuickTalk on Windows is best done in WSL2:

Windows Host
  └── WSL2 Ubuntu
        ├── OpenTalking .venv
        ├── OmniRT .venv
        ├── QuickTalk checkpoints
        ├── OmniRT QuickTalk runtime
        └── E2E Benchmark

Advantages of this approach:

• More compatible with official bash scripts;
• Convenient use of Linux toolchains;
• CUDA can access the RTX 3050 normally through WSL2;
• OpenTalking and OmniRT environments are isolated, making troubleshooting clearer;
• The path from environment setup to full E2E benchmark is fully reproducible.