NVIDIA's recipe box for its Blackwell desktop

What it does

This repository is a curated stack of Jupyter notebooks that document how to install, configure, and run AI/ML workloads on NVIDIA’s DGX Spark, a Blackwell-based device. The playbooks span inference engines, fine-tuning frameworks, robotics simulators, and multi-device networking. It is largely setup documentation and integration glue — NVIDIA’s official cheat sheet for turning the hardware into a functional lab.

The interesting bit

The breadth is almost comically wide: you can chain three Sparks in a ring topology, quantize models with NVFP4, run Isaac Sim, or build a video-search agent, all from the same starting point. That variety suggests NVIDIA is treating the Spark less like a peripheral and more like a self-contained platform ecosystem.

Key highlights

• Covers inference across multiple engines: llama.cpp, vLLM, SGLang, TRT LLM, NIM, and Ollama.
• Includes multi-device recipes: ring topologies, switch networking, and NCCL for two Sparks.
• Ventures beyond LLMs into robotics (Isaac Sim, Isaac Lab), biology (single-cell RNA sequencing), and finance (portfolio optimization).
• Each notebook bundles prerequisites, troubleshooting, and example code.

Verdict

Grab this if you own a DGX Spark and want official, pre-vetted paths through the modern AI tooling maze. Skip it if you are looking for a reusable software library — this is a hardware-specific recipe box.