Intel now recommends using upstream PyTorch for its CPU and GPU platforms. The Intel Extension for PyTorch repository carries a retirement notice explaining that most of its features and optimisations have moved upstream and that quarterly releases ended after 2.8. The notice says the project will be maintained “as an open source project until the end of March 2026” and that Intel then plans to “mark the project End-of-Life unless there is a solid need to continue maintenance”. The repository is now archived, with a banner saying Intel will not provide or guarantee development of or support for it. Installing the old extension is no longer the general starting point for a new deployment.

What replaced it is a device string. In current PyTorch, Intel GPUs are reached with .to("xpu"), with torch.xpu.synchronize() and torch.autocast(device_type="xpu") alongside it.

SYCL and oneAPI

Khronos describes SYCL as an open, royalty-free, cross-platform abstraction layer that lets code for heterogeneous and offload processors be written in modern ISO C++. llama.cpp’s SYCL documentation adds the practical description: a single-source language for heterogeneous computing based on standard C++17, designed to improve productivity across CPUs, GPUs and FPGAs.

oneAPI is the toolchain around it. The same document describes it as an open, standards-based specification supporting multiple architectures, and names the key components: DPC++, the primary oneAPI SYCL implementation including the icpx and icx compilers; the oneAPI libraries such as oneMKL, oneMath and oneDNN; and Level Zero, a low-level interface for fine-grained control over Intel integrated and discrete GPUs.

Which Intel GPUs PyTorch covers

PyTorch’s Intel GPU page separates datacentre and client hardware. Its datacentre table names the Intel Data Center GPU Max Series and lists validated Linux distributions. Support is tied to that operating-system and driver combination, not just to the presence of an Intel GPU.

The client table includes Arc A-Series, Arc B-Series and supported Core Ultra integrated graphics. Windows, Linux and WSL support varies by processor family. Check the row for the actual device rather than applying another Arc or Core Ultra machine’s requirements to it.

The page notes that Intel GPU support arrived as a prototype in PyTorch 2.5, for client GPUs and the Data Center GPU Max Series on Linux and Windows, bringing Intel GPUs and the SYCL software stack into the official PyTorch stack. Installation is a wheel index: pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu. The GPU driver has to be installed first.

Serving runtimes

vLLM’s Intel GPU page says the project initially supports basic model inference and serving on the Intel GPU platform, listing Intel Data Center GPU and Intel Arc GPU as supported hardware. It depends on a separate package, vllm-xpu-kernels, and the documentation is emphatic that Python 3.12 is a must because the provided wheel is specific to it. Prebuilt XPU wheels come from wheels.vllm.ai, and each index also carries a triton==3.7.2+xpu shim, with PyTorch XPU packages served from the PyTorch XPU index, so both URLs are needed.

llama.cpp ships SYCL binaries in FP16 and FP32 variants for Ubuntu 24.04 x64, and a Windows SYCL build. Its doc states those Linux packages contain only the backend binaries and require the target machine to have matching Intel GPU drivers and oneAPI packages at the same version as the build, recommending Intel’s deep-learning-essentials Docker image. It also says the FP32 and FP16 packages differ in accuracy and performance on language models, so you should choose based on test results rather than assumption.

The SYCL changelog reports faster prompt processing after a oneAPI toolchain update on an Arc B570. That result compares builds of SYCL code. It is not a comparison against Vulkan and does not establish that SYCL wins on every Intel card or model.

Triton on Intel

Intel maintains an OpenAI Triton backend for its GPUs in a repository of its own, which the vLLM XPU shim reflects. PyTorch’s compiler guide says that for NVIDIA, AMD and Intel GPUs, torch.compile uses OpenAI Triton as the key building block, which explains the backend’s role in compiled PyTorch workloads.

FAQ

Do I still need the Intel Extension for PyTorch?

No. Its repository says active development is discontinued, official quarterly releases ended after 2.8 and Intel recommends using PyTorch directly for Intel CPU and GPU platforms. The repository is now archived.

Which device string do Intel GPUs use in PyTorch?

xpu. Intel GPUs are not mapped onto the cuda namespace the way ROCm is, so code written for CUDA needs the device changed.

Can vLLM serve on an Intel Arc card?

The documentation lists Intel Arc GPU as supported hardware for basic inference and serving, with Python 3.12 required and the vllm-xpu-kernels package as a dependency.

What is the difference between SYCL and oneAPI?

SYCL is the Khronos standard for single-source C++ heterogeneous programming. oneAPI is the specification and toolchain around it, including the DPC++ implementation of SYCL, the oneAPI libraries and Level Zero.