vLLM Ascend Plugin

| About Ascend | Developer Slack (#sig-ascend) |

English | 中文

---

Latest News 🔥

• [2024/12] We are working with the vLLM community to support [RFC]: Hardware pluggable.

---

Overview

vLLM Ascend plugin (vllm-ascend) is a backend plugin for running vLLM on the Ascend NPU.

This plugin is the recommended approach for supporting the Ascend backend within the vLLM community. It adheres to the principles outlined in the [RFC]: Hardware pluggable, providing a hardware-pluggable interface that decouples the integration of the Ascend NPU with vLLM.

By using vLLM Ascend plugin, popular open-source models, including Transformer-like, Mixture-of-Expert, Embedding, Multi-modal LLMs can run seamlessly on the Ascend NPU.

Prerequisites

Support Devices

• Atlas A2 Training series (Atlas 800T A2, Atlas 900 A2 PoD, Atlas 200T A2 Box16, Atlas 300T A2)
• Atlas 800I A2 Inference series (Atlas 800I A2)

Dependencies

Requirement	Supported version	Recommended version	Note
vLLM	main	main	Required for vllm-ascend
Python	>= 3.9	3.10	Required for vllm
CANN	>= 8.0.RC2	8.0.RC3	Required for vllm-ascend and torch-npu
torch-npu	>= 2.4.0	2.5.1rc1	Required for vllm-ascend
torch	>= 2.4.0	2.5.1	Required for torch-npu and vllm

Find more about how to setup your environment in here.

Getting Started

Note

Currently, we are actively collaborating with the vLLM community to support the Ascend backend plugin, once supported you can use one line command pip install vllm vllm-ascend to compelete installation.

Installation from source code:

# Install vllm main branch according:
# https://docs.vllm.ai/en/latest/getting_started/installation/cpu/index.html#build-wheel-from-source
git clone --depth 1 https://github.com/vllm-project/vllm.git
cd vllm
pip install -r requirements-build.txt
VLLM_TARGET_DEVICE=empty pip install .

# Install vllm-ascend main branch
git clone https://github.com/vllm-project/vllm-ascend.git
cd vllm-ascend
pip install -e .

Run the following command to start the vLLM server with the Qwen/Qwen2.5-0.5B-Instruct model:

# export VLLM_USE_MODELSCOPE=true to speed up download
vllm serve Qwen/Qwen2.5-0.5B-Instruct
curl http://localhost:8000/v1/models

Please refer to vLLM Quickstart for more details.

Building

Build Python package from source

git clone https://github.com/vllm-project/vllm-ascend.git
cd vllm-ascend
pip install -e .

Build container image from source

git clone https://github.com/vllm-project/vllm-ascend.git
cd vllm-ascend
docker build -t vllm-ascend-dev-image -f ./Dockerfile .

See Building and Testing for more details, which is a step-by-step guide to help you set up development environment, build and test.

Feature Support Matrix

Feature	Supported	Note
Chunked Prefill	✗	Plan in 2025 Q1
Automatic Prefix Caching	✅	Imporve performance in 2025 Q1
LoRA	✗	Plan in 2025 Q1
Prompt adapter	✅
Speculative decoding	✅	Impore accuracy in 2025 Q1
Pooling	✗	Plan in 2025 Q1
Enc-dec	✗	Plan in 2025 Q1
Multi Modality	✅ (LLaVA/Qwen2-vl/Qwen2-audio/internVL)	Add more model support in 2025 Q1
LogProbs	✅
Prompt logProbs	✅
Async output	✅
Multi step scheduler	✅
Best of	✅
Beam search	✅
Guided Decoding	✗	Plan in 2025 Q1

Model Support Matrix

The list here is a subset of the supported models. See supported_models for more details:

Model	Supported	Note
Qwen 2.5	✅
Mistral		Need test
DeepSeek v2.5		Need test
LLama3.1/3.2	✅
Gemma-2		Need test
baichuan		Need test
minicpm		Need test
internlm	✅
ChatGLM	✅
InternVL 2.5	✅
Qwen2-VL	✅
GLM-4v		Need test
Molomo	✅
LLaVA 1.5	✅
Mllama		Need test
LLaVA-Next		Need test
LLaVA-Next-Video		Need test
Phi-3-Vison/Phi-3.5-Vison		Need test
Ultravox		Need test
Qwen2-Audio	✅

Contributing

We welcome and value any contributions and collaborations:

• Please feel free comments here about your usage of vLLM Ascend Plugin.
• Please let us know if you encounter a bug by filing an issue.
• Please see the guidance on how to contribute in CONTRIBUTING.md.

License

Apache License 2.0, as found in the LICENSE file.