SOLVED: PyTorch 2.7.1+XPU Intel Arc Graphics Complete Setup Guide (Linux)"

Innomen · June 14, 2025, 4:58pm

Went down a bit of an AI tech support rabbit hole. 10 hours later I thought I’d share the bottom. This is a low effort AI post. Just hoping it saves people some time. This will likely be a drive by XD

SOLVED: PyTorch 2.7.1+XPU on Intel Arc Graphics - Complete Setup Guide

SUCCESS! PyTorch XPU Working on Intel Arc Graphics

After extensive troubleshooting, I’ve successfully got PyTorch XPU working with Intel Arc Graphics on Linux. Sharing the complete solution to help others facing similar issues.

Quick Verification:

import torch
print(f"PyTorch version: {torch.__version__}")
print(f"XPU compiled: {torch._C._xpu_getDeviceCount is not None}")
print(f"XPU available: {torch.xpu.is_available()}")
print(f"Device count: {torch.xpu.device_count()}")
print(f"Device name: {torch.xpu.get_device_name(0)}")

# Test tensor creation
x = torch.randn(3, 3, device='xpu')
print(f"Test tensor: {x.size()} {x.device}")

Output:

PyTorch version: 2.7.1+xpu
XPU compiled: True
XPU available: True
Device count: 1
Device name: Intel(R) Arc(TM) Graphics
Test tensor: torch.Size([3, 3]) xpu:0

Working Installation Command

The solution that worked after 8 full conversations of iterations of troubleshooting: Approx 10 hours. (Point and laugh if you must, I am 6 months into my 100% Linux daily driver journey XD)

# IMPORTANT: Complete clean uninstall first
pip uninstall -y torch torchvision torchaudio intel-cmplr-lib-rt intel-cmplr-lib-ur intel-cmplr-lic-rt intel-sycl-rt pytorch-triton-xpu tcmlib umf intel-pti

# Fresh install with all Intel runtime dependencies
pip install torch==2.7.1+xpu torchvision==0.22.1+xpu torchaudio==2.7.1+xpu intel-cmplr-lib-rt intel-cmplr-lib-ur intel-cmplr-lic-rt intel-sycl-rt pytorch-triton-xpu tcmlib umf intel-pti --index-url https://download.pytorch.org/whl/xpu --extra-index-url https://pypi.org/simple

Key Breakthrough Discovery

The Root Cause: Partial installations and version mixing between system packages and pip packages caused library path conflicts.

The Solution: Complete clean slate reinstallation of the entire PyTorch XPU + Intel runtime stack.

System Specifications

Hardware: Intel Meteor Lake-P with integrated Arc Graphics
OS: Arch Linux (kernel 6.x)
Python: 3.11.6 via pyenv
GPU: Intel Arc Graphics (128 compute units)
Driver: i915 with GuC/HuC firmware

Prerequisites

System Level Zero (install via package manager):

# Arch Linux
sudo pacman -S level-zero-loader level-zero-headers

# Ubuntu/Debian
sudo apt install level-zero level-zero-dev

Intel Compute Runtime (should be installed automatically with modern kernels)

Verify GPU Detection:

ls /dev/dri/  # Should show renderD128 or similar
clinfo        # Should list Intel GPU (if OpenCL tools installed)

Critical Success Factors

1. Use PyTorch 2.7.1+xpu (Not IPEX)

PyTorch 2.7.1+xpu has native Intel GPU support
No need for Intel Extension for PyTorch (IPEX)
XPU backend is built-in

2. Install ALL Intel Runtime Dependencies

The pip command above installs these essential Intel 2025.0.4 runtime components:

intel-cmplr-lib-rt-2025.0.4
intel-cmplr-lib-ur-2025.0.4
intel-cmplr-lic-rt-2025.0.4
intel-sycl-rt-2025.0.4
pytorch-triton-xpu-3.3.1

3. Clean Slate Installation

Don’t try to update existing installations
Uninstall everything PyTorch and Intel-related first
Fresh installation resolves library conflicts

Common Issues & Solutions

Issue: “XPU is not available”

Solution: Make sure you installed from the XPU index URL and included all Intel runtime dependencies.

Issue: Segmentation faults or crashes

Solution: Version conflicts between system and pip packages. Do complete uninstall/reinstall.

Issue: ImportError for Intel libraries

Solution: Install all Intel runtime dependencies listed in the command above.

Issue: Level Zero version warnings

Note: Version mismatch warnings between PyTorch expectations and system Level Zero are usually non-critical if basic tensor operations work.

Validation Steps

After installation, verify everything works:

import torch

# Basic checks
assert torch.xpu.is_available(), "XPU not available"
assert torch.xpu.device_count() > 0, "No XPU devices found"

# Create and operate on XPU tensor
x = torch.randn(1000, 1000, device='xpu')
y = torch.randn(1000, 1000, device='xpu')
z = torch.mm(x, y)  # Matrix multiplication on GPU

print(f"✅ Success! Tensor device: {z.device}")
print(f"✅ GPU: {torch.xpu.get_device_name(0)}")

What This Enables

With working PyTorch XPU, you can now:

Train neural networks on Intel GPU
Accelerate tensor computations
Use Intel Arc Graphics for AI workloads
Develop with native PyTorch GPU support

Complete Documentation

I’ve documented the entire troubleshooting journey with technical details, failed attempts, and system configuration info. The full technical report includes:

8 iteration attempts with detailed failure analysis
Dependency conflict resolution
Fish shell compatibility notes
Performance validation steps

Help Others

If this helped you, please:

Share your success in replies
Note any variations for your system
Report any issues you encounter