vllm-project/vllm-ascend

Implements vLLM as a pluggable hardware abstraction layer for Ascend NPUs, enabling support for diverse model architectures including Transformer, Mixture-of-Experts, embedding, and multi-modal LLMs without modifying core vLLM. Decouples Ascend backend integration through standardized hardware plugin interfaces while leveraging CANN runtime and PyTorch-NPU for optimized inference on Atlas hardware series (A2/A3 training and inference accelerators).

1,773 stars. Actively maintained with 344 commits in the last 30 days.