Overview
vLLM Ascend integrates the Ascend NPU backend with vLLM through a hardware-pluggable interface. The integration is designed to remain decoupled from vLLM core while supporting Transformer-like models, Mixture-of-Experts models, embedding models, and multimodal LLMs. The repository is active and not archived; its primary language is C++.
Key features
- Hardware plugin for running vLLM on supported Ascend NPUs
- Hardware-pluggable integration that decouples Ascend support from vLLM core
- Support categories include Transformer-like, Mixture-of-Experts, embedding, and multimodal models
- Versioned documentation, quick-start guides, and a model and feature support matrix
- Maintained main, release, and RFC collaboration branches
- Community support through an issue tracker, user forum, Slack channel, and weekly meeting
Requirements, installation and quick start
Usage
Model compatibility and use cases
The repository states support for popular open-source Transformer-like, Mixture-of-Experts, embedding, and multimodal LLMs, but does not provide a complete model list in the supplied excerpt. Consult https://docs.vllm.ai/projects/ascend/en/latest/user_guide/support_matrix/ for model and feature compatibility. Agent frameworks, MCP, and RAG-specific support are not stated in the repository metadata. Embedding-model support may be relevant to retrieval pipelines, but a complete RAG system is not described.
License and risk notes
Licensed under Apache License 2.0. Repository metadata identifies the SPDX license as Apache-2.0 and provides https://api.github.com/licenses/apache-2.0.
Release and maintenance
The supplied README lists v0.23.0rc1 as the latest release candidate, announced for July 2026, and v0.18.0 as the latest stable version, announced for May 2026. It also lists v0.13.0, v0.11.0, v0.9.1, and v0.7.3. The repository metadata reports creation on 2025-01-29 and last push on 2026-08-03. These values are reported as supplied and were not independently verified.