Repository navigation
feat: Intel oneAPI / SYCL backend support #27
Description
Activity
Status — not slated for V3.1.1. The V3.1.1 accelerator focus is AMD ROCm + Apple Metal; Intel oneAPI / SYCL is downstream of those. The umbrella tracker is #65. Work scope mirrors the ROCm path: inference Dockerfile variant, entrypoint script, bootstrap installer detection. Open to a PR from anyone with the hardware.
partial coverage update — Vulkan landed in #114 (commit f88a717) and the Mesa ANV driver is solid enough on Arc that intel users can run atlas today via
atlas init --backend vulkan+ thedocker-compose.vulkan.ymloverlay. so the "intel has no path at all" blocker is gone.still keeping this open though because vulkan is the slower path on intel. native sycl via oneAPI lets us:
- use the level-zero backend directly instead of going through the vulkan abstraction
- pick up the AMX + XMX matrix engines on the newer Arc / Lunar Lake / Battlemage parts
- get proper xpu device topology instead of the generic
/dev/driexposure that vulkan uses
scope when we pick this up would be a
Dockerfile.syclthat pullsintel/oneapi-basekit:2024.xas the build base, builds llama.cpp with-DGGML_SYCL=ON, and wiresentrypoint-v3.1-9b.shfor thesyclbackend case. probably also a_check_sycl_via_docker()inatlas/cli/commands/doctor.pymirroring the vulkan check.priority dropped from "blocking intel" to "perf optimization for intel". if anyone's running atlas on Arc via the vulkan path and has tokens/sec numbers please share so we can quantify the sycl gap before doing the work.
Status (2026-07-02): Intel Arc/Iris Xe are served today through the Vulkan fallback image (
inference/Dockerfile.vulkan, Mesa ANV) — documented in SETUP.md and SUPPORT_MATRIX.md. A native SYCL backend remains roadmap; per the support policy it stays out of the runtime until it's complete and validated on hardware.- added a parent issue
on Sep 28, 2026
Description
llama.cpp supports SYCL for Intel GPU acceleration. We need to add Intel Arc GPU support to ATLAS.
Requirements
Context
Similar scope to ROCm support (#26). llama.cpp's SYCL backend supports Intel Arc GPUs. The main work is containerization, testing, and documentation.
Hardware Needed
Contributors need access to an Intel Arc GPU with 16GB+ VRAM and oneAPI toolkit installed.