NVIDIA/dgx-spark-playbooks

Collection of step-by-step playbooks for setting up AI/ML workloads on NVIDIA DGX Spark devices with Blackwell architecture.

NVIDIA DGX Spark Playbooks

A collection of step-by-step guides (playbooks) for setting up AI/ML workloads on NVIDIA DGX Spark devices (Blackwell architecture). Each playbook includes prerequisites, instructions, troubleshooting tips, and example code.

What's inside

  • Installing and configuring popular AI frameworks
  • Running inference with optimized models
  • Setting up development environments
  • Connecting and managing multiple DGX Spark devices

Playbook topics (broad categories)

  • Model serving: vLLM, SGLang, TensorRT-LLM, llama.cpp, LM Studio, Open WebUI + Ollama, NIM, Nemotron
  • Fine-tuning: Unsloth, LLaMA Factory, PyTorch, NeMo, FLUX.1, GR00T, Vision Language Models
  • Distributed computing: connecting 2 or 3 Sparks, ring topology, through a switch, NCCL, vLLM cluster
  • Development environments: VS Code, CLI coding agents, vibe coding, Tailscale remote access
  • Specialized workloads: cuPyNumeric, JAX, CUDA-X, cuOpt, healthcare agents, video processing, RAG, chatbots
  • System management: DGX Dashboard, MIG, multi-GPU setup, security (OpenShell)

Resources

License

Two license files are included in the repo: LICENSE (for the repository itself) and LICENSE-3rd-party (for third-party components).

Related

  • Project
    RaoFoundation/bittensor

    A decentralized platform and SDK for creating competitive markets for machine intelligence and other digital commodities, rewarded via a blockchain-based incentive system.

  • Project
    pytorch/gloo

    A collective communications library that provides primitives like allreduce and broadcast for efficient multi-machine machine learning training.

  • Project
    jax-md/jax-md

    A hardware-accelerated, differentiable molecular dynamics library built on JAX for simulating materials and small-scale interactions.

  • Project
    aeon-toolkit/aeon

    aeon is a scikit-learn compatible Python library that provides a unified, research-backed toolkit for a wide range of time series machine learning tasks, including classification, forecasting, and deep learning.