DGX Spark vLLM

Pre-built vLLM environment for NVIDIA DGX Spark (GB10, sm_121).

About

This package contains the vLLM environment extracted from NVIDIA's official nvcr.io/nvidia/vllm:25.09-py3 Docker image, configured for local installation on DGX Spark without Docker.

Why? The official vLLM doesn't support sm_121 (GB10) yet. NVIDIA's Docker image has patches for Blackwell support, but requires running everything in a container. This package lets you run vLLM natively on DGX Spark.

Requirements

• NVIDIA DGX Spark (GB10 chip, sm_121)
• Ubuntu with CUDA 13.0 installed
• Python 3.12
• aarch64 (ARM64) architecture

Installation

pip install dgx-spark-vllm
sudo dgx-spark-vllm-install

Then restart your shell or run:

source /etc/profile.d/nvidia-venv.sh

Included Versions

Usage

import torch
print(torch.__version__)  # 2.9.0a0+50eac811a6.nv25.09
print(torch.cuda.get_device_name(0))  # NVIDIA GB10

import vllm
print(vllm.__version__)  # 0.10.1.1+381074ae.nv25.09

Run vLLM server:

python3 -m vllm.entrypoints.openai.api_server --model <model_name>

What's Installed

• /home/srpost/shared/nvidia-venv/ - Python packages
• /opt/hpcx/ - HPC-X libraries (UCX, UCC, OpenMPI)
• /usr/lib/aarch64-linux-gnu/libcudnn* - cuDNN
• /usr/lib/aarch64-linux-gnu/libnccl* - NCCL
• /usr/local/lib/libnvpl* - NVPL (NVIDIA Performance Libraries)
• /etc/ld.so.conf.d/00-hpcx.conf - Library path config
• /etc/profile.d/nvidia-venv.sh - PYTHONPATH config

License

The installer is Apache 2.0. The installed packages retain their original licenses (PyTorch BSD, vLLM Apache 2.0, NVIDIA components under NVIDIA license).