ACE-Step 1.5 - Intel XPU Setup Guide

Quick start guide for running ACE-Step on Intel Arc GPUs and integrated graphics.

• Intel Arc GPU: A770, A750, A580, A380, or Intel integrated graphics
• Python 3.11: Download from python.org
• Latest Intel GPU drivers: Install from Intel's website
• Internet connection: For first-time setup
• Disk space: ~5-10 GB for dependencies

1. Run the setup script:

   setup_xpu.bat
   

2. Wait for installation (takes a few minutes on first run)

3. Launch the Gradio UI:

   start_gradio_ui_xpu.bat
   

4. Open your browser: http://127.0.0.1:7860

1. Create virtual environment:

   python -m venv venv_xpu
   

2. Activate it:

   call venv_xpu\Scripts\activate
   

3. Upgrade pip:

   python -m pip install --upgrade pip
   

4. Install XPU dependencies:

   pip install -r requirements-xpu.txt
   

5. Launch:

   start_gradio_ui_xpu.bat
   

Models should be placed in the checkpoints folder. If you already have models from a previous installation, they will be automatically detected.

• DiT Model: acestep-v15-turbo (fast generation)
• LM Model: acestep-5Hz-lm-4B (best quality, uses CPU offload)

1. Automatic (uses defaults):

   start_gradio_ui_xpu.bat
   

2. Manual (choose models interactively):

   start_gradio_ui_xpu_manual.bat
   

3. API Server (REST API access):

   start_api_server_xpu.bat
   

The bat files automatically set these performance optimizations:

set SYCL_CACHE_PERSISTENT=1
set SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
set PYTORCH_DEVICE=xpu
set TORCH_COMPILE_BACKEND=eager

These settings improve XPU performance and are based on verified working configurations.

Create a .env file in the root directory to customize settings:

# Gradio UI Settings
PORT=7860
SERVER_NAME=127.0.0.1
LANGUAGE=en

# Model Settings
ACESTEP_CONFIG_PATH=acestep-v15-turbo
ACESTEP_LM_MODEL_PATH=acestep-5Hz-lm-4B
ACESTEP_OFFLOAD_TO_CPU=true

# API Settings
ACESTEP_API_KEY=your-secret-key

Run setup_xpu.bat to create the virtual environment.

1. Check that your GPU drivers are up to date
2. Verify PyTorch XPU installation:

   call venv_xpu\Scripts\activate
   python -c "import torch; print(torch.xpu.is_available())"
   

Reinstall PyTorch XPU:

call venv_xpu\Scripts\activate
pip uninstall torch torchaudio torchvision
pip install --pre torch torchaudio torchvision --index-url https://download.pytorch.org/whl/nightly/xpu

1. Use a smaller LM model (0.6B or 1.7B instead of 4B)
2. Enable CPU offload in the UI
3. Close other GPU-intensive applications

• MP3/Opus/AAC files use torchaudio with ffmpeg backend (bundled)
• FLAC/WAV files use soundfile (fastest)
• If issues occur, try converting to WAV format

• ✅ WAV/FLAC: Native support via soundfile (fastest)
• ✅ MP3: Supported via torchaudio with ffmpeg backend
• ✅ Opus/AAC: Supported via torchaudio

No additional codec installation needed!

1. First launch takes longer as models are initialized
2. CPU offload is recommended for 4B LM on GPUs with <=16GB VRAM
3. torch.compile is disabled (not fully supported on XPU yet)
4. Python 3.11 is recommended for best compatibility

• Check the main documentation: README.md
• Verify XPU installation: Run verification commands above
• Update check: Bat files automatically check for updates on startup

Once setup is complete, simply run:

start_gradio_ui_xpu.bat

And start creating music with ACE-Step on your Intel GPU!