Qwen-Image-2.1 Deployed on NVIDIA DGX Spark with GPUStack
A Juejin AI author reports deploying Qwen-Image-2.1 on an NVIDIA DGX Spark with the GB10 GPU, 121GB unified CPU/GPU memory, and ARM64 architecture. The GPUStack-managed service exposed an OpenAI-compatible API for text-to-image generation, image editing, native transparent backgrounds, and LoRA; reported results included 49 seconds for a 1024×1024, 40-step generation and 38GB resident memory. The report also documents ARM64 image building, dependency isolation, network troubleshooting, deployment, and gateway configuration.
The report presents a working ARM64 image-generation deployment on compact hardware, including an OpenAI-compatible API and documented deployment steps, though the claims come from a single source rather than independent corroboration.
一台 2025 年底上市的 ARM 小盒子。121GB 统一内存,一块 GB10,官方文档里写着这类模型跑不了它的架构。
现在它对外提供一个 OpenAI 兼容的图像接口,文生图、改图、原生透明底、LoRA 四项全通。硬件一点没动,驱动没升级,也没有第二台机器。
把镜像变成团队里人人能调的服务,中间那段是 GPUStack 接过去的。下面记录整个过程:先判断 ARM 到底能不能跑,再处理构建、网络、部署和网关问题。
01 · 文档说不支持,我先去验一句
机器是 DGX Spark,GB10,aarch64 架构。驱动 580.173.02,CUDA 13.0,121GB 内存由 CPU 和 GPU 共用。
| 项目 | 实测环境 |
|---|---|
| 设备 | NVIDIA DGX Spark |
| GPU | GB10 |
| 架构 | aarch64 / sm_121 |
| 驱动 | 580.173.02 |
| CUDA | 13.0 |
| 内存 | 121GB CPU/GPU 统一内存 |
#Qwen-Image-2.1#NVIDIA DGX Spark#NVIDIA GB10#GPUStack#vLLM-Omni