Intel Arc Pro B70 Inference Cookbook
A collection of pinned build recipes, benchmark records and failure notes for running LLM inference on Intel Arc Pro B60/B70 GPUs with vLLM and llama.cpp.
This cookbook documents specific, reproducible configurations for serving large language models on Intel's Arc Pro B60 and B70 (Battlemage, Xe2) GPUs, tracking what worked, what did not, and what to test next across vLLM's XPU backend and llama.cpp's SYCL backend.
It organizes recipes by goal, such as serving a single model on one card, running GGUF or vision models, handling mixture-of-experts or long-context models, or serving across two cards, each recipe naming its measured status and warning against mixing patch stacks between model families. A decision guide separates measured improvements from failed attempts and unpublished results, and a doctor script checks a given card and host before a first run.
The code is MIT licensed; measurement reports and prose are CC BY licensed.