vllm-xpu-kernels

It is a vLLM component that provides custom kernels—specialized computational routines—for Intel GPUs (XPU) to accelerate large language model (LLM) inference. ^[1] The kernels are written in SYCL/DPC++ (programming models for heterogeneous computing) and use the oneDNN library for deep learning primitives; importing the library registers PyTorch custom operations so vLLM dispatches them automatically on Intel GPU hardware. ^[2a][2b]