Personal / community image. Maintained by @rahulunair for personal experimentation and benchmarking on Intel Arc Pro B70. Not an official Intel image and not affiliated with, endorsed by, or supported by Intel Corporation, the vLLM project, or PyTorch. Provided as-is; for production use refer to Intel's official containers and PyTorch's official XPU documentation.
Foundation image for Intel XPU LLM-serving stacks on Intel Arc Pro B70 / Battlemage / Xe2-HPG. Provides a complete oneAPI envelope plus current GPU userspace, ready to be the FROM for rahulunair/torch, rahulunair/vllm, or any other XPU project.
| Layer | Source | Version |
|---|---|---|
| OS | docker.io/library/ubuntu | 24.04 |
| oneAPI compiler + libsycl + oneMKL/oneTBB/oneDNN/oneCCL/MPI | Intel oneAPI apt repo | {{INTEL_ONEAPI_COMPILER_DPCPP_CPP}} |
| Level Zero loader + headers | built from oneapi-src/level-zero | master HEAD |
| compute-runtime debs | intel/compute-runtime GitHub release | {{COMPUTE_RUNTIME}} |
| Intel Graphics Compiler debs | intel/intel-graphics-compiler GitHub release | {{IGC}} |
The Intel oneAPI version on this image determines which libsycl ABI is available, which in turn dictates the torch and vllm-xpu-kernels versions that can use it.
| oneAPI release | libsycl SONAME | Compatible torch+xpu | Compatible vllm-xpu-kernels |
|---|---|---|---|
| 2025.3.x (this image) | libsycl.so.8 | 2.11.0+xpu (stable index) | 0.1.7 (latest release) |
| 2026.0.x | libsycl.so.9 | not yet released for .so.9 | not yet released for .so.9 |
To see what's actually installed in this image:
docker run --rm rahulunair/xpu:latest-base cat /etc/xpu-versions
This image carries headers and compilers; it doesn't run a service. Use it as a FROM in your Dockerfile, or open a shell to verify GPU access:
docker run --rm -it --privileged --device /dev/dri \
rahulunair/xpu:latest-base bash
sycl-ls
zello_world
docker run --rm --privileged --device /dev/dri \
rahulunair/xpu:latest-base run-xpu-tests xpu-base
Runs zello_world (level-zero device discovery + kernel exec), sycl_hello (SYCL parallel_for on GPU), onednn_smoke (oneDNN GPU engine + stream), oneccl_smoke (oneCCL single-rank SYCL communicator). Verified on Intel Arc Pro B70 (Intel(R) Graphics [0xe223]) — Level Zero V2 adapter selected automatically; no V1 fallback needed.
latest-base — most recent green build<oneapi-version>-base — pinned to the Intel oneAPI compiler version (e.g. 2025.3.3-base)github.com/rahulunair/vllm-ci — Dockerfile.bleeding, --target xpu-base.
Content type
Image
Digest
sha256:799959ecb…
Size
2.6 GB
Last updated
6 months ago
docker pull rahulunair/xpu:latest-base