vllm-project/vllm-ascend

Community maintained hardware plugin for vLLM on Ascend

What it solves

It enables the vLLM inference engine to run seamlessly on Huawei Ascend NPUs, decoupling the hardware integration from the core vLLM codebase through a pluggable interface.

How it works

It acts as a hardware plugin that implements a specific interface (based on a hardware-pluggable RFC) to bridge vLLM's requirements with the Ascend NPU architecture. This allows various model types—including Transformer-based LLMs, Mixture-of-Experts (MoE), Embedding models, and Multi-modal LLMs—to be executed on Ascend hardware.

Who it’s for

Developers and AI engineers who use Huawei Ascend hardware (such as Atlas 800I or Atlas A2/A3 series) and want to deploy high-performance LLM inference using the vLLM framework.

Highlights

  • Support for a wide range of model architectures including MoE and multimodal LLMs.
  • Adheres to a hardware-pluggable architecture to keep the main vLLM project clean.
  • Compatible with CANN 9.1.0 and specific PyTorch/TorchNPU versions.
  • Community-maintained with dedicated support for various Atlas NPU series.

Related

  • Project
  • Project
  • Project
  • Project
  • Project