level-zero

A pure C inference engine designed for Intel Xe-LP iGPUs and Gemma 4 26B-A4B QAT.

Intel LevelZero JNI library for TornadoVM

Model Dispatch and Memory Allocator: an LLM inference engine for the hardware you own. Custom CPU, ROCm, Vulkan, CUDA, Metal and Level Zero kernels, exact memory admission, mTLS gRPC and private federation.

Pure-Go GPU/NPU compute wrapper for Intel NEO driver Level Zero APIs.

Field-tested guide: multi-GPU vLLM tensor-parallel (TP=2/TP=4) on Intel Arc Pro B70 (Battlemage BMG-G31, Xe2) on Linux. Driver setup (xe force_probe=e223), bare-metal vLLM + oneAPI 2025.3, the compute-runtime multi-root USM + triton-xpu init_devices fixes, FP8/int4-AutoRound quant, root-cause error reports. AI-agent readable (AGENTS.md).

Unofficial community hub for Intel Arc Pro B70/B-series, Intel XPU, oneAPI, OpenVINO, PyTorch XPU, vLLM, llama.cpp, setup guides, benchmarks, and patches.

Per-precision XMX (matrix engine) profiling for Intel Arc GPUs via Level Zero metric streamers — observes any workload without wrapping it

Pure-Rust reimplementation of VkFFT with typed planning, kernel IR, and multi-backend GPU runtimes

A simple utility to count the number of compute and copy engines on Intel GPUs.

Run LLMs on Intel GPUs at full speed — Ollama with SYCL backend, 2x faster than Vulkan (Arc A770, Core Ultra, Flex/Max)

Pioneer documentation: OpenFOAM CFD GPU offloading on Intel Arc Pro B70 Pro (Battlemage Xe2). Hardware excellent (FP64 96%, kernel latency CUDA-par), software stack not ready (GPU 66% idle, no working strong preconditioner in Ginkgo 1.10 SYCL). 14 bugs documented + direct VRAM/profiling data.

PoC for the acceleration of the core Math functions in Llama2.c to run on GPUs with Shared Memory

Native Intel Xe2 LLM inference: SYCL/DPAS kernels, paged KV, prefix caching, continuous batching, and XMX packed prefill

Experimental PyTorch backend and custom SHAVE/DPU kernels for Intel Lunar Lake NPU4000

Fork of comfy-aimdo adding Intel Arc (B580/DG2) XPU backend via oneAPI/Level-Zero (torch 2.14.0+xpu): cold-start OOM root-cause fixes (real-free-VRAM admission gate + USM-fragmentation recovery), 3-tier VRAM/RAM/disk offload, and a pinned runtime contract (provider.json).

vLLM running natively on Windows on two Intel Arc Pro B60s - two cards faster than one. Port notes, probes, build scripts and measurements.

Xe2 dual-B70 kernel and 2x2 parallelism lab (TP=2, PP=2)