rahulunair/torch repository overview

⁠rahulunair/torch — PyTorch + Triton XPU for Intel Battlemage

PyTorch + pytorch-triton-xpu on top of rahulunair/xpu:latest-base⁠. Reusable foundation for any XPU LLM-serving project (vLLM, SGLang, …) — FROM rahulunair/torch:latest-xpu and skip rebuilding the full oneAPI + GPU userspace stack underneath.

⁠What's in it

Inherits everything from rahulunair/xpu:*-base, plus:

The pip wheel name pytorch-triton-xpu ships the Python module triton. Inside the image, import triton resolves to that wheel (verified at build time with from triton._C.libtriton import intel).

⁠ABI compatibility chain

Three things need to match for the XPU stack to work end-to-end. This image is built so they do:

[ vllm-xpu-kernels 0.1.7 ]   builds against
        ↓
[ torch 2.11.0+xpu ]   has DT_NEEDED libsycl.so.8 and uses urKernelSetArgValue@LIBUR_LOADER_0.12
        ↓
[ Intel oneAPI 2025.3.x ]   ships libsycl.so.8 and the matching libur_loader
        ↓
[ rahulunair/xpu:latest-base ]

If any link in this chain is bumped, the chain breaks. PyTorch nightly's XPU wheel (2.13.0.dev*+xpu) currently fails to load against oneAPI 2026.0 (libsycl.so.9, no urKernelSetArgValue). vllm-xpu-kernels source code references torch 2.11's torch::Library::_def symbol which torch 2.13 removed, so source-building from main against torch 2.13 also fails. This image therefore stays on the stable triple (kernels 0.1.7 ↔ torch 2.11 ↔ oneAPI 2025.3), and the pin advances when all three upstream components have releases for the same newer ABI.

⁠Run

docker run --rm --privileged --device /dev/dri \
  -e ZE_AFFINITY_MASK=0 \
  rahulunair/torch:latest-xpu \
  python3 -c "import torch; print(torch.__version__, torch.xpu.is_available(), torch.xpu.device_count())"

Or run the embedded smoke including @torch.compile codegen via Triton-XPU:

docker run --rm --privileged --device /dev/dri \
  rahulunair/torch:latest-xpu run-xpu-tests xpu-torch

That exercises base smokes + torch.xpu.is_available(), eager matmul on Battlemage, @torch.compile (which drives the Triton-XPU → ocloc codegen path).

⁠Use as a base for your own XPU project

FROM rahulunair/torch:latest-xpu
RUN pip install --break-system-packages --no-deps your-package
ENTRYPOINT ["your-server"]

--no-deps matters: many XPU-related packages still pin torch versions in their requirements that may differ from this image's installed version. If you need vllm-xpu-kernels alongside your code, install the prebuilt wheel matched to the torch in this image (vllm_xpu_kernels-0.1.7 for torch 2.11.0+xpu).

⁠Tags

• latest-xpu — most recent green build
• <torch-version>-xpu — derived from torch.__version__ (e.g. 2.11.0-xpu)
• last 7 versioned tags retained

⁠Source

github.com/rahulunair/vllm-ci⁠ — Dockerfile.bleeding, --target xpu-torch.