vLLM Kunlun
vLLM Kunlun is a community-maintained vLLM hardware plugin that runs large language model inference on Baidu's Kunlun XPU accelerators, supporting more than 20 mainstream and multimodal models.
vLLM Kunlun is a hardware-pluggable interface, following vLLM's RFC for pluggable hardware, that lets vLLM run on Baidu's Kunlun XPU accelerators without modifying vLLM's own source code. It integrates through standard Python entry points and is presented as the recommended way to add Kunlun backend support to the vLLM ecosystem.
It supports more than 20 mainstream large language models, including Transformer-style, Mixture-of-Experts, embedding and multimodal models such as Qwen, Llama, DeepSeek, GLM, Gemma4 and Kimi-K2. Quantization support covers W8A8 (INT8), AWQ, GPTQ and compressed-tensors W4A16 for both MoE and dense models, LoRA and multi-LoRA fine-tuning is supported for Qwen-series models, and recent releases add multi-token prediction for DeepSeek-V3.2 with RoPE and decode-stage kernel optimizations.