Notebook Configuration

Follow this procedure to configure a workbench for Intel GPUs:

  1. Create a workbench in the RHOAI dashboard and select the Intel GPU hardware profile you created in Configuring GPU Profiles.

  2. Choose a notebook image that already contains the Intel GPU software stack, such as oneAPI, PyTorch with IPEX, or OpenVINO. RHOAI ships no Intel XPU workbench image, so add a custom image under Settings > Workbench images.

    Important

    The image must ship an Intel compute runtime new enough for your GPU. On Intel® Arc™ Pro B70 Graphics, intel/intel-extension-for-pytorch:2.8.10-xpu does not detect the device because its libze-intel-gpu is 25.18. Verify a candidate image before importing it, as described in Verifying GPU Access.

  3. Start the workbench and confirm that the pod received the GPU device.

    import torch
    
    print(f"torch: {torch.__version__}")
    print(f"XPU available: {torch.xpu.is_available()}")
    print(f"Device count: {torch.xpu.device_count()}")
    for i in range(torch.xpu.device_count()):
        print(f"  GPU {i}: {torch.xpu.get_device_name(i)}")
    

    PyTorch 2.5 and later support Intel XPU natively, so import intel_extension_for_pytorch is no longer required. Import it only if your image is built around an older PyTorch release.

  4. If the device count is 0, check the OpenCL platforms from a terminal in the workbench. When no Intel(R) OpenCL Graphics platform is listed, the image runtime is too old rather than the GPU being unavailable.

    clinfo -l
    ls -la /dev/dri/