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).
linux gpu triton quantization multi-gpu xe2 oneapi level-zero cachyos intel-xpu fp8 intel-arc vllm llm-inference bmg-g31 battlemage xe-driver tensor-parallel intel-arc-pro intel-arc-b70
-
Updated
Jun 13, 2026 - Shell