vllm-xpu-kernels
A vLLM component that provides custom SYCL kernels for Intel GPUs, covering normalization, attention, quantization and mixture-of-experts operations, to accelerate large-language-model inference on Intel hardware.
vllm-xpu-kernels supplies custom PyTorch operator implementations for the Intel XPU (GPU) backend used by vLLM. The kernels are written in SYCL and DPC++ and build on oneDNN's deep-learning primitives, following vLLM's existing custom-operator registration and dispatch pattern so that importing the package at startup registers the operators for vLLM to use automatically.
Supported operators span normalization, activation functions, flash and grouped-query attention variants, rotary positional encoding, mixture-of-experts routing and gathering, LoRA, and FP8 and MxFP4 quantization and matrix multiplication. Building it requires a matching PyTorch build for XPU, the Intel oneAPI base toolkit, CMake and the Ninja build system, and it is released under the Apache-2.0 license.