abh80 PyTorch XPU Setup & Usage
This skill covers end-to-end setup and usage of PyTorch on Intel GPUs (XPU) on Windows and Linux.
Important Context (as of 2026):
- Intel Extension for PyTorch (IPEX) reached end of life in March 2026. All XPU support is now native in upstream PyTorch.
- There are no conda packages for XPU. Only pip wheels exist from the XPU index URL.
- The recommended workflow: create a conda env → install via pip inside it.
Quick Reference — Installation
Prerequisites
Intel GPU Driver — required for all users.
- Windows: Download from Intel Arc & Iris Xe Graphics - Windows. Include LevelZeroSDK in the install for
torch.compilesupport. - Linux: Follow Client GPU installation instructions.
- Full prerequisites page: https://www.intel.com/content/www/us/en/developer/articles/tool/pytorch-prerequisites-for-intel-gpu/2-11.html
- Windows: Download from Intel Arc & Iris Xe Graphics - Windows. Include LevelZeroSDK in the install for
Intel Deep Learning Essentials — only needed if building from source. If installing from pip wheels, the runtime packages are bundled automatically.
Supported Hardware
| Platform | Supported GPUs |
|---|---|
| Windows 11 / Ubuntu 24.04+ | Intel Arc A-Series (Alchemist), Arc B-Series (Battlemage), Core Ultra with Arc Graphics (Meteor Lake-H, Arrow Lake-H, Lunar Lake) |
| Windows 11 / Ubuntu 25.10 | Panther Lake |
| Linux (RHEL 9.2, SUSE 15 SP5, Ubuntu 22.04+) | Intel Data Center GPU Max Series (Ponte Vecchio) |
Installation Steps
# 1. Create conda environment
conda create -n pytorch-xpu python=3.11 -y
conda activate pytorch-xpu
# 2. Install PyTorch with XPU support (stable release)
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu
# --- OR for nightly builds ---
pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu
# --- OR for a specific version (e.g. 2.10.0) ---
pip install torch==2.10.0 torchvision==0.25.0 torchaudio==2.10.0 --index-url https://download.pytorch.org/whl/xpu
# 3. Verify
python -c "import torch; print(torch.xpu.is_available()); print(torch.xpu.get_device_name(0))"
Python version note: Python 3.10–3.12 are well-tested. Python 3.13 has reported issues with some XPU nightly builds.
Troubleshooting Installation
| Symptom | Fix |
|---|---|
torch.xpu.is_available() returns False |
Update Intel GPU driver to latest version. Verify with sycl-ls (Linux) or Device Manager (Windows). |
AssertionError: Torch not compiled with XPU enabled |
You installed a non-XPU wheel. Reinstall with --index-url https://download.pytorch.org/whl/xpu. |
| Import errors about missing DLLs on Windows | Ensure LevelZeroSDK was included during driver install. |
torch.compile fails on Windows |
Requires PyTorch ≥ 2.7 + MSVC (Desktop Development with C++). |
Quick Reference — Code Usage
The torch.xpu API mirrors torch.cuda almost 1:1. Migrating from CUDA is a find-and-replace of "cuda" → "xpu".
For detailed code patterns (inference, training, AMP, torch.compile, memory management, multi-device, profiling), read the reference file:
→ references/code-patterns.md — Comprehensive code examples for all XPU workflows.
Minimal Example
import torch
device = torch.device("xpu" if torch.xpu.is_available() else "cpu")
model = MyModel().to(device)
data = torch.randn(32, 3, 224, 224).to(device)
with torch.no_grad():
output = model(data)
Key API Surface
| Function | Purpose |
|---|---|
torch.xpu.is_available() |
Check if XPU device is available |
torch.xpu.device_count() |
Number of Intel GPU devices |
torch.xpu.current_device() |
Index of selected device |
torch.xpu.get_device_name(i) |
Device name string |
torch.xpu.synchronize() |
Wait for all kernels to finish |
torch.xpu.empty_cache() |
Free unused cached memory |
torch.xpu.memory_allocated() |
Current memory usage (bytes) |
torch.xpu.max_memory_allocated() |
Peak memory usage |
torch.xpu.memory_stats() |
Dict of allocator statistics |
torch.xpu.Stream() |
Create a compute stream |
torch.xpu.Event() |
Create a timing event |
torch.autocast(device_type="xpu") |
Enable AMP |
torch.amp.GradScaler(device="xpu") |
Gradient scaling for AMP training |
Important Caveats
- GradScaler & FP64:
GradScalerrequires FP64 hardware support. Intel Arc A-Series does NOT support FP64 natively. On Arc A-series, useautocastwithoutGradScaler. - torch.compile on Windows: Supported from PyTorch 2.7+. Requires MSVC (Visual Studio 2022, Desktop Development with C++ workload).
- No IPEX needed: Do NOT install
intel-extension-for-pytorchfor new projects. It is EOL. Use nativetorch.xpudirectly. - No conda XPU packages: Always use pip with
--index-url https://download.pytorch.org/whl/xpuinside a conda env.
Official Documentation Links
| Resource | URL |
|---|---|
torch.xpu API Reference |
https://docs.pytorch.org/docs/stable/xpu.html |
| Getting Started on Intel GPU | https://docs.pytorch.org/docs/stable/notes/get_start_xpu.html |
| torch.compile on Windows XPU | https://docs.pytorch.org/tutorials/unstable/inductor_windows.html |
| Prerequisites for Intel GPUs | https://www.intel.com/content/www/us/en/developer/articles/tool/pytorch-prerequisites-for-intel-gpu/2-11.html |
| Intel GPU Driver (Windows) | https://www.intel.com/content/www/us/en/download/785597/intel-arc-iris-xe-graphics-windows.html |
| PyTorch Previous Versions | https://pytorch.org/get-started/previous-versions/ |